From 121717594b0d1ac5db0f58da46ca2abe8b89c946 Mon Sep 17 00:00:00 2001 From: Seebs Date: Tue, 31 Mar 2020 15:10:01 -0500 Subject: [PATCH 1/8] improve inspect output, switch roaring over to using new unmarshal At some point the code changeover to use roaring iterators for unmarshal got dropped, but the old unmarshal code is way harder to make work for inspect, so this change is back. This exports some of the names from the things returned by Info, but also adds a roaring function to use the unmarshalling logic on arbitrary data, allowing us to get more insight into a file -- in particular, letting us distinguish between the bitmaps specified by the roaring data and the bitmaps resulting from applying the ops log. --- cmd/inspect.go | 3 + ctl/inspect.go | 152 +++++++++++++--- ctl/inspect_test.go | 4 +- roaring/roaring.go | 92 ++++++---- roaring/unmarshal_binary.go | 341 +++++++++++++++--------------------- 5 files changed, 334 insertions(+), 258 deletions(-) diff --git a/cmd/inspect.go b/cmd/inspect.go index 4a2e852f6..47a40cae8 100644 --- a/cmd/inspect.go +++ b/cmd/inspect.go @@ -45,5 +45,8 @@ Inspects a data file and provides stats. return inspector.Run(context.Background()) }, } + flags := inspectCmd.Flags() + flags.BoolVarP(&inspector.Quiet, "quiet", "q", false, "don't list details of containers") + flags.IntVarP(&inspector.Max, "max", "n", 0, "list at most max items (0 = unlimited)") return inspectCmd } diff --git a/ctl/inspect.go b/ctl/inspect.go index 48924b89d..11471caa9 100644 --- a/ctl/inspect.go +++ b/ctl/inspect.go @@ -33,6 +33,10 @@ import ( type InspectCommand struct { // Path to data file Path string + // don't list details of objects + Quiet bool + // list only this many objects + Max int // Standard input/output *pilosa.CmdIO @@ -45,6 +49,113 @@ func NewInspectCommand(stdin io.Reader, stdout, stderr io.Writer) *InspectComman } } +type pointerContext struct { + from, to uintptr +} + +func (p *pointerContext) pretty(c roaring.ContainerInfo) string { + var pointer string + if c.Mapped { + if c.Pointer >= p.from && c.Pointer < p.to { + pointer = fmt.Sprintf("@+0x%x", c.Pointer-p.from) + } else { + pointer = fmt.Sprintf("!0x%x!", c.Pointer) + } + } else { + pointer = fmt.Sprintf("0x%x", c.Pointer) + } + return fmt.Sprintf("%s \t%d \t%d \t%s ", c.Type, c.N, c.Alloc, pointer) +} + +func (cmd *InspectCommand) PrintOps(info roaring.BitmapInfo) { + fmt.Fprintln(cmd.Stdout, "== Ops ==") + tw := tabwriter.NewWriter(cmd.Stdout, 0, 8, 0, '\t', 0) + fmt.Fprintf(tw, "%s\t%s\t%s\t\n", "TYPE", "OpN", "SIZE") + printed := 0 + for _, op := range info.OpDetails { + fmt.Fprintf(tw, "%s\t%d\t%d\t\n", op.Type, op.OpN, op.Size) + printed++ + if cmd.Max != 0 && printed >= cmd.Max { + break + } + } + tw.Flush() +} + +func (cmd *InspectCommand) PrintContainers(info roaring.BitmapInfo, pC pointerContext) { + fmt.Fprintln(cmd.Stdout, "== Containers ==") + tw := tabwriter.NewWriter(cmd.Stdout, 0, 8, 0, '\t', 0) + fmt.Fprintf(tw, "\tRoaring\t\t\t\tOps\t\t\t\t\n") + fmt.Fprintf(tw, "%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t\n", "KEY", "TYPE", "N", "ALLOC", "OFFSET", "TYPE", "N", "ALLOC", "OFFSET") + c1s := info.Containers + c2s := info.OpContainers + l1 := len(c1s) + l2 := len(c2s) + i1 := 0 + i2 := 0 + var c1, c2 roaring.ContainerInfo + c1.Key = ^uint64(0) + c2.Key = ^uint64(0) + c1e := false + c2e := false + if i1 < l1 { + c1 = c1s[i1] + i1++ + c1e = true + } + if i2 < l2 { + c2 = c2s[i2] + i2++ + c2e = true + } + printed := 0 + for c1e || c2e { + c1used := false + c2used := false + var key uint64 + c1fmt := "-\t\t\t" + c2fmt := "-\t\t\t" + if !c2e || (c1e && c1.Key < c2.Key) { + c1fmt = pC.pretty(c1) + key = c1.Key + c1used = true + } else if !c1e || (c2e && c2.Key < c1.Key) { + c2fmt = pC.pretty(c2) + key = c2.Key + c2used = true + } else { + // c1e and c2e both set, and neither key is < the other. + c1fmt = pC.pretty(c1) + c2fmt = pC.pretty(c2) + key = c1.Key + c1used = true + c2used = true + } + if c1used { + if i1 < l1 { + c1 = c1s[i1] + i1++ + } else { + c1e = false + } + } + if c2used { + if i2 < l2 { + c2 = c2s[i2] + i2++ + } else { + c2e = false + } + } + fmt.Fprintf(tw, "%d\t%s\t%s\t\n", key, c1fmt, c2fmt) + printed++ + if cmd.Max > 0 && printed >= cmd.Max { + break + } + } + tw.Flush() +} + // Run executes the inspect command. func (cmd *InspectCommand) Run(_ context.Context) error { // Open file handle. @@ -72,39 +183,34 @@ func (cmd *InspectCommand) Run(_ context.Context) error { }() // Attach the mmap file to the bitmap. t := time.Now() - fmt.Fprintf(cmd.Stderr, "unmarshalling bitmap...") - bm := roaring.NewBitmap() - if err := bm.UnmarshalBinary(data); err != nil { - return errors.Wrap(err, "unmarshalling") + fmt.Fprintf(cmd.Stderr, "inspecting bitmap...") + info, err := roaring.InspectBinary(data) + if err != nil { + return errors.Wrap(err, "inspecting") } fmt.Fprintf(cmd.Stderr, " (%s)\n", time.Since(t)) - // Retrieve stats. - t = time.Now() - fmt.Fprintf(cmd.Stderr, "calculating stats...") - info := bm.Info() - fmt.Fprintf(cmd.Stderr, " (%s)\n", time.Since(t)) + pC := pointerContext{ + from: uintptr(unsafe.Pointer(&data[0])), + } + pC.to = pC.from + uintptr(len(data)) // Print top-level info. fmt.Fprintf(cmd.Stdout, "== Bitmap Info ==\n") - fmt.Fprintf(cmd.Stdout, "Containers: %d\n", len(info.Containers)) - fmt.Fprintf(cmd.Stdout, "Operations: %d\n", info.OpN) + fmt.Fprintf(cmd.Stdout, "Bits: %d\n", info.BitCount) + fmt.Fprintf(cmd.Stdout, "Containers: %d (%d roaring)\n", info.ContainerCount, len(info.Containers)) + fmt.Fprintf(cmd.Stdout, "Operations: %d (%d bits)\n", info.Ops, info.OpN) fmt.Fprintln(cmd.Stdout, "") // Print info for each container. - fmt.Fprintln(cmd.Stdout, "== Containers ==") - tw := tabwriter.NewWriter(cmd.Stdout, 0, 8, 0, '\t', 0) - fmt.Fprintf(tw, "%s\t%s\t% 8s \t% 8s\t%s\n", "KEY", "TYPE", "N", "ALLOC", "OFFSET") - for _, ci := range info.Containers { - fmt.Fprintf(tw, "%d\t%s\t% 8d \t% 8d \t0x%08x\n", - ci.Key, - ci.Type, - ci.N, - ci.Alloc, - uintptr(ci.Pointer)-uintptr(unsafe.Pointer(&data[0])), - ) + if !cmd.Quiet { + if info.ContainerCount > 0 { + cmd.PrintContainers(info, pC) + } + if info.Ops > 0 { + cmd.PrintOps(info) + } } - tw.Flush() return nil } diff --git a/ctl/inspect_test.go b/ctl/inspect_test.go index 1ff852712..2698dcc03 100644 --- a/ctl/inspect_test.go +++ b/ctl/inspect_test.go @@ -41,7 +41,7 @@ func TestInspectCommand_Run(t *testing.T) { file.Close() cm.Path = file.Name() err = cm.Run(context.Background()) - expectedError := "unmarshalling: " + expectedError := "inspecting: " if !strings.Contains(err.Error(), expectedError) { t.Fatalf("expected error '%s', got '%v'", expectedError, err) } @@ -52,7 +52,7 @@ func TestInspectCommand_Run(t *testing.T) { if err != nil { t.Fatalf("copying data: %v", err) } - if !strings.Contains(buf.String(), "unmarshalling bitmap...") { + if !strings.Contains(buf.String(), "inspecting bitmap...") { t.Fatalf("Inspect doesn't work: %s", err) } diff --git a/roaring/roaring.go b/roaring/roaring.go index cbb7dc583..3eb5ecef2 100644 --- a/roaring/roaring.go +++ b/roaring/roaring.go @@ -1724,8 +1724,9 @@ type roaringIterator interface { // Next yields the information about the next container Next() (key uint64, cType byte, n int, length int, pointer *uint16, err error) // Remaining yields the bytes left over past the end of the roaring data, - // which is typically an ops log in our case. - Remaining() []byte + // which is typically an ops log in our case, and also its offset in case + // we need to talk about truncation. + Remaining() ([]byte, int64) } // baseRoaringIterator holds values used by both Pilosa and official Roaring @@ -1884,11 +1885,11 @@ func (r *baseRoaringIterator) Done(err error) { r.currentDataOffset = 0 } -func (r *baseRoaringIterator) Remaining() []byte { +func (r *baseRoaringIterator) Remaining() ([]byte, int64) { if r.lastDataOffset == 0 { - return nil + return nil, 0 } - return r.data[r.lastDataOffset:] + return r.data[r.lastDataOffset:], r.lastDataOffset } func (r *pilosaRoaringIterator) Next() (key uint64, cType byte, n int, length int, pointer *uint16, err error) { @@ -2435,18 +2436,19 @@ func (b *Bitmap) roaringSize() (int64, int64) { } // Info returns stats for the bitmap. -func (b *Bitmap) Info() bitmapInfo { - info := bitmapInfo{ +func (b *Bitmap) Info() BitmapInfo { + info := BitmapInfo{ OpN: b.opN, Ops: b.ops, - Containers: make([]containerInfo, 0, b.Containers.Size()), + Containers: make([]ContainerInfo, 0, b.Containers.Size()), } - + info.ContainerCount = cap(info.Containers) citer, _ := b.Containers.Iterator(0) for citer.Next() { k, c := citer.Value() ci := c.info() ci.Key = k + info.BitCount += uint64(c.N()) info.Containers = append(info.Containers, ci) } return info @@ -2504,11 +2506,15 @@ func (b *Bitmap) Flip(start, end uint64) *Bitmap { return result } -// bitmapInfo represents a point-in-time snapshot of bitmap stats. -type bitmapInfo struct { - OpN int - Ops int - Containers []containerInfo +// BitmapInfo represents a point-in-time snapshot of bitmap stats. +type BitmapInfo struct { + OpN int + Ops int + OpDetails []OpInfo + BitCount uint64 + ContainerCount int + Containers []ContainerInfo + OpContainers []ContainerInfo } // Iterator represents an iterator over a Bitmap. @@ -3705,8 +3711,8 @@ func (c *Container) size() int { } // info returns the current stats about the container. -func (c *Container) info() containerInfo { - info := containerInfo{N: c.N()} +func (c *Container) info() ContainerInfo { + info := ContainerInfo{N: c.N(), Mapped: c.Mapped()} if c == nil { info.Type = "nil" info.Alloc = 0 @@ -3723,17 +3729,7 @@ func (c *Container) info() containerInfo { info.Type = "bitmap" info.Alloc = len(c.bitmap()) * 8 // sizeof(uint64) } - - if c.Mapped() { - if c.isArray() { - info.Pointer = unsafe.Pointer(&c.array()[0]) - } else if c.isRun() { - info.Pointer = unsafe.Pointer(&c.runs()[0]) - } else { - info.Pointer = unsafe.Pointer(&c.bitmap()[0]) - } - } - + info.Pointer = uintptr(unsafe.Pointer(c.pointer)) return info } @@ -3799,13 +3795,14 @@ func (c *Container) bitmapRepair() { c.setN(n) } -// containerInfo represents a point-in-time snapshot of container stats. -type containerInfo struct { - Key uint64 // container key - Type string // container type (array, bitmap, or run) - N int32 // number of bits - Alloc int // memory used - Pointer unsafe.Pointer // offset within the mmap +// ContainerInfo represents a point-in-time snapshot of container stats. +type ContainerInfo struct { + Key uint64 // container key + Type string // container type (array, bitmap, or run) + N int32 // number of bits + Alloc int // memory used + Pointer uintptr // address + Mapped bool // whether this container thinks it is mmapped } // flip returns a new container containing the inverse of all @@ -5267,6 +5264,15 @@ const ( opTypeRemoveRoaring = opType(5) ) +var opTypes = []string{ + "add", + "remove", + "addN", + "removeN", + "addRoaring", + "removeRoaring", +} + // op represents an operation on the bitmap. type op struct { typ opType @@ -5276,6 +5282,24 @@ type op struct { roaring []byte } +// OpInfo is a description of an op. +type OpInfo struct { + Type string + OpN int + Size int +} + +func (op *op) info() (info OpInfo) { + if int(op.typ) < len(opTypes) { + info.Type = opTypes[op.typ] + } else { + info.Type = fmt.Sprintf("unknown-type-%d", op.typ) + } + info.OpN = op.opN + info.Size = op.size() + return info +} + // apply executes the operation against a bitmap. func (op *op) apply(b *Bitmap) (changed bool) { switch op.typ { diff --git a/roaring/unmarshal_binary.go b/roaring/unmarshal_binary.go index c9071b1ca..59f65a117 100644 --- a/roaring/unmarshal_binary.go +++ b/roaring/unmarshal_binary.go @@ -1,4 +1,4 @@ -// Copyright 2017 Pilosa Corp. +// Copyright 2019 Pilosa Corp. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -15,232 +15,175 @@ package roaring import ( - "encoding/binary" - "fmt" - "unsafe" - - "github.com/pkg/errors" + "errors" + "io" ) -// UnmarshalBinary decodes b from a binary-encoded byte slice. data can be in -// either official roaring format or Pilosa's roaring format. -func (b *Bitmap) UnmarshalBinary(data []byte) error { +// UnmarshalBinary reads Pilosa's format, or upstream roaring (mostly; +// it may not handle some edge cases), and decodes them into the given +// bitmap, replacing the existing contents. +func (b *Bitmap) UnmarshalBinary(data []byte) (err error) { if data == nil { - // Nothing to unmarshal - return nil - } - statsHit("Bitmap/UnmarshalBinary") - // reset ops/opN since we're reading new data. - b.ops = 0 - b.opN = 0 - fileMagic := uint32(binary.LittleEndian.Uint16(data[0:2])) - if fileMagic == MagicNumber { // if pilosa roaring - return errors.Wrap(b.unmarshalPilosaRoaring(data), "unmarshaling as pilosa roaring") + return errors.New("no roaring bitmap provided") } + var itr roaringIterator + var itrKey uint64 + var itrCType byte + var itrN int + var itrLen int + var itrPointer *uint16 + var itrErr error - keyN, containerTyper, header, pos, haveRuns, err := readOfficialHeader(data) + itr, err = newRoaringIterator(data) if err != nil { - return errors.Wrap(err, "reading roaring header") + return err } - // Only the Pilosa roaring format has flags. The official Roaring format - // hasn't got space in its header for flags. - b.Flags = 0 - - b.Containers.ResetN(int(keyN)) - // Descriptive header section: Read container keys and cardinalities. - for i, buf := uint(0), data[header:]; i < uint(keyN); i, buf = i+1, buf[4:] { - card := int(binary.LittleEndian.Uint16(buf[2:4])) + 1 - b.Containers.PutContainerValues( - uint64(binary.LittleEndian.Uint16(buf[0:2])), - containerTyper(i, card), /// container type voodo with isRunBitmap - card, - true) + if itr == nil { + return errors.New("failed to create roaring iterator, but don't know why") } - // Read container offsets and attach data. - if haveRuns { - err := readWithRuns(b, data, pos, keyN) - if err != nil { - return errors.Wrap(err, "reading offsets from official roaring format") + b.Containers.Reset() + + itrKey, itrCType, itrN, itrLen, itrPointer, itrErr = itr.Next() + for itrErr == nil { + newC := &Container{ + typeID: itrCType, + n: int32(itrN), + len: int32(itrLen), + cap: int32(itrLen), + pointer: itrPointer, + flags: flagMapped, } - } else { - err := readOffsets(b, data, pos, keyN) - if err != nil { - return errors.Wrap(err, "reading official roaring format") + if !b.preferMapping { + newC.unmapOrClone() } + b.Containers.Put(itrKey, newC) + itrKey, itrCType, itrN, itrLen, itrPointer, itrErr = itr.Next() } - return nil -} - -func readOffsets(b *Bitmap, data []byte, pos int, keyN uint32) error { - - citer, _ := b.Containers.Iterator(0) - for i, buf := 0, data[pos:]; i < int(keyN); i, buf = i+1, buf[4:] { - // Verify the offset is fully formed - if len(buf) < 4 { - return fmt.Errorf("insufficient data for offsets: len=%d", len(buf)) - } - offset := binary.LittleEndian.Uint32(buf[0:4]) - // Verify the offset is within the bounds of the input data. - if int(offset) >= len(data) { - return fmt.Errorf("offset out of bounds: off=%d, len=%d", offset, len(data)) - } - - // Map byte slice directly to the container data. - citer.Next() - k, c := citer.Value() - if !c.Mapped() { - fmt.Printf("inexplicable: container %d (%d/%d) doesn't think it's mapped. fixing that.\n", - k, i, keyN) - c.setMapped(true) - } - switch c.typ() { - case containerArray: - c.setArray((*[0xFFFFFFF]uint16)(unsafe.Pointer(&data[offset]))[:c.N():c.N()]) - case containerBitmap: - c.setBitmap((*[0xFFFFFFF]uint64)(unsafe.Pointer(&data[offset]))[:bitmapN:bitmapN]) - default: - return fmt.Errorf("unsupported container type %d", c.typ()) - } - } - return nil -} - -func readWithRuns(b *Bitmap, data []byte, pos int, keyN uint32) error { - if len(data) < pos+runCountHeaderSize { - return fmt.Errorf("insufficient data for offsets(run): len=%d", len(data)) - } - citer, _ := b.Containers.Iterator(0) - for i := 0; i < int(keyN); i++ { - citer.Next() - k, c := citer.Value() - if !c.Mapped() { - fmt.Printf("inexplicable: container %d (%d/%d) doesn't think it's mapped. fixing that.\n", - k, i, keyN) - c.setMapped(true) - } - switch c.typ() { - case containerRun: - runCount := binary.LittleEndian.Uint16(data[pos : pos+runCountHeaderSize]) - c.setRuns((*[0xFFFFFFF]interval16)(unsafe.Pointer(&data[pos+runCountHeaderSize]))[:runCount:runCount]) - runs := c.runs() - - for o := range runs { // must convert from start:length to start:end :( - runs[o].last = runs[o].start + runs[o].last - } - pos += int((runCount * interval16Size) + runCountHeaderSize) - case containerArray: - c.setArray((*[0xFFFFFFF]uint16)(unsafe.Pointer(&data[pos]))[:c.N():c.N()]) - pos += int(c.N() * 2) - case containerBitmap: - c.setBitmap((*[0xFFFFFFF]uint64)(unsafe.Pointer(&data[pos]))[:bitmapN:bitmapN]) - pos += bitmapN * 8 - } - } - return nil -} - -func (b *Bitmap) unmarshalPilosaRoaring(data []byte) error { - if len(data) < headerBaseSize { - return errors.New("data too small") - } - - // Verify the first two bytes are a valid MagicNumber, and second two bytes match current storageVersion. - fileMagic := uint32(binary.LittleEndian.Uint16(data[0:2])) - fileVersion := uint32(data[2]) - b.Flags = data[3] - if fileMagic != MagicNumber { - return fmt.Errorf("invalid roaring file, magic number %v is incorrect", fileMagic) - } - - if fileVersion != storageVersion { - return fmt.Errorf("wrong roaring version, file is v%d, server requires v%d", fileVersion, storageVersion) - } - - // Read key count in bytes sizeof(cookie)+sizeof(flag):(sizeof(cookie)+sizeof(uint32)). - keyN := binary.LittleEndian.Uint32(data[3+1 : 8]) - if int64(len(data)) < headerBaseSize+int64(keyN)*12 { - return fmt.Errorf("insufficient data for header + offsets: key-cardinality not provided for %d containers", keyN) - } - - headerSize := headerBaseSize - b.Containers.ResetN(int(keyN)) - // Descriptive header section: Read container keys and cardinalities. - for i, buf := 0, data[headerSize:]; i < int(keyN); i, buf = i+1, buf[12:] { - b.Containers.PutContainerValues( - binary.LittleEndian.Uint64(buf[0:8]), - byte(binary.LittleEndian.Uint16(buf[8:10])), - int(binary.LittleEndian.Uint16(buf[10:12]))+1, - true) - } - opsOffset := int64(headerSize) + int64(keyN)*12 - - // Read container offsets and attach data. - citer, _ := b.Containers.Iterator(0) - // if you have enough containers that the *headers alone* exceed 4GB, we - // need to start with a higher cycle offset. - cycleOffset := opsOffset &^ ((1 << 32) - 1) - prevOffset32 := uint32(opsOffset) - for i, buf := 0, data[opsOffset:]; i < int(keyN); i, buf = i+1, buf[4:] { - offset32 := binary.LittleEndian.Uint32(buf[0:4]) - if offset32 < prevOffset32 { - cycleOffset += (1 << 32) - } - prevOffset32 = offset32 - offset := int64(offset32) + cycleOffset - // Verify the offset is within the bounds of the input data. - if offset >= int64(len(data)) { - return fmt.Errorf("offset out of bounds: off=%d, len=%d", offset, len(data)) - } - - // Map byte slice directly to the container data. - citer.Next() - k, c := citer.Value() - - // this shouldn't happen, since we don't normally store nils. - if c == nil { - continue - } - if !c.Mapped() { - fmt.Printf("inexplicable: container %d (%d/%d) doesn't think it's mapped. fixing that.\n", - k, i, keyN) - c.setMapped(true) - } - switch c.typ() { - case containerRun: - runCount := binary.LittleEndian.Uint16(data[offset : offset+runCountHeaderSize]) - c.setRuns((*[0xFFFFFFF]interval16)(unsafe.Pointer(&data[offset+runCountHeaderSize]))[:runCount:runCount]) - opsOffset = offset + runCountHeaderSize + int64(len(c.runs()))*interval16Size - case containerArray: - c.setArray((*[0xFFFFFFF]uint16)(unsafe.Pointer(&data[offset]))[:c.N():c.N()]) - opsOffset = offset + int64(len(c.array()))*2 // sizeof(uint32) - case containerBitmap: - c.setBitmap((*[0xFFFFFFF]uint64)(unsafe.Pointer(&data[offset]))[:bitmapN:bitmapN]) - opsOffset = offset + int64(len(c.bitmap()))*8 // sizeof(uint64) - } + // note: if we get a non-EOF err, it's possible that we made SOME + // changes but didn't log them. I don't have a good solution to this. + if itrErr != io.EOF { + return itrErr } // Read ops log until the end of the file. - buf := data[opsOffset:] - + b.ops = 0 + b.opN = 0 + buf, lastValidOffset := itr.Remaining() for { // Exit when there are no more ops to parse. if len(buf) == 0 { break } + // Unmarshal the op and apply it. var opr op if err := opr.UnmarshalBinary(buf); err != nil { - return newFileShouldBeTruncatedError(err, int64(opsOffset)) + return newFileShouldBeTruncatedError(err, int64(lastValidOffset)) } + opr.apply(b) + // Increase the op count. b.ops++ b.opN += opr.count() - opsOffset += int64(opr.size()) - // Move the buffer forward. - buf = data[opsOffset:] - } + // Move the buffer forward. + opSize := opr.size() + buf = buf[opSize:] + lastValidOffset += int64(opSize) + } return nil } + +// unmarshalPilosaRoaring treats data as being encoded in Pilosa's 64 bit +// roaring format and decodes it into b. +func InspectBinary(data []byte) (info BitmapInfo, err error) { + if data == nil { + return info, errors.New("no roaring bitmap provided") + } + var itr roaringIterator + var itrKey uint64 + var itrCType byte + var itrN int + var itrLen int + var itrPointer *uint16 + var itrErr error + + itr, err = newRoaringIterator(data) + if err != nil { + return info, err + } + if itr == nil { + return info, errors.New("failed to create roaring iterator, but don't know why") + } + + b := NewFileBitmap() + + itrKey, itrCType, itrN, itrLen, itrPointer, itrErr = itr.Next() + for itrErr == nil { + newC := &Container{ + typeID: itrCType, + n: int32(itrN), + len: int32(itrLen), + cap: int32(itrLen), + pointer: itrPointer, + flags: flagMapped, + } + b.Containers.Put(itrKey, newC) + itrKey, itrCType, itrN, itrLen, itrPointer, itrErr = itr.Next() + } + // note: if we get a non-EOF err, it's possible that we made SOME + // changes but didn't log them. I don't have a good solution to this. + if itrErr != io.EOF { + return info, itrErr + } + + // gather initial info + info = b.Info() + + // Read ops log until the end of the file. + b.ops = 0 + b.opN = 0 + buf, lastValidOffset := itr.Remaining() + for { + // Exit when there are no more ops to parse. + if len(buf) == 0 { + break + } + + // Unmarshal the op and apply it. + var opr op + if err := opr.UnmarshalBinary(buf); err != nil { + return info, err + } + opr.apply(b) + + // Increase the op count. + info.Ops++ + info.OpN += opr.count() + info.OpDetails = append(info.OpDetails, opr.info()) + + // Move the buffer forward. + opSize := opr.size() + buf = buf[opSize:] + lastValidOffset += int64(opSize) + } + // now we want to compute the actual container and bit counts after + // ops, and create a report of just the containers which got changed. + info.ContainerCount = 0 + info.BitCount = 0 + citer, _ := b.Containers.Iterator(0) + for citer.Next() { + k, c := citer.Value() + info.ContainerCount++ + info.BitCount += uint64(c.N()) + if c.Mapped() { + continue + } + ci := c.info() + ci.Key = k + info.OpContainers = append(info.OpContainers, ci) + } + return info, nil +} From ceaf5c15d1e3a6c685e4ffd5463728805faca7f9 Mon Sep 17 00:00:00 2001 From: Seebs Date: Tue, 7 Apr 2020 18:51:04 -0500 Subject: [PATCH 2/8] thread the holder through things, and improve snapshot queue logic This is logically two separate things, but the individual changes are thoroughly intertwined in the code. The first change is a logical change to the design of the snapshot queue, which is that it now adjusts the maxOpN the background scan targets, allowing it to lower that value over time when things are quiet. We do this because it turns out that on large data sets, this can make a factor-of-four difference in memory usage! So, in general, on a quiet system, each pass through the holder aims for about 1/4 of the existing fragments to get snapshotted. When there's more load, we adjust those values up. We also make the snapshot queue a bit less chatty, to make testing less annoying -- we only print stats if the queue enqueues at least two snapshots, or skips any. The second change is threading the holder through things. We've always threaded the logger through, and then added the snapshot queue, and some of the Inspect-related work led to wanting to have a way to thread options through, so what if we just threaded the holder itself through, and removed the direct copying around of the logger, snapshot queue, and so on. Similarly, everything can now use holder.PartitionN instead of having to get its own copy of PartitionN handed out to each index. This does imply ensuring that test cases always get a reasonable default holder. This is a precursor to adding additional information to the holder, such as whether it's in a special read-only mode, which would imply not modifying on-disk files. This is already semi-supported for the specific case of the background snapshot queue and cache flushing, which are attached to the (created in a previous commit) new holder Activate method, instead of being automatic on holder Open. The change to a snapshot queue can also cause races in tests, because the fragment.Clean method's "sanity check" accesses a fragment without a lock. Fix that. Since there's a couple of t.Fatalf(), but we need to release the lock before closing, we use an anonymous function with a defer to handle that. Whee! --- cluster_internal_test.go | 2 +- field.go | 41 +++++++--------- field_internal_test.go | 6 +-- field_test.go | 6 +-- fragment.go | 50 +++++++++---------- fragment_internal_test.go | 64 ++++++++++++------------- generation_test.go | 4 +- holder.go | 38 +++++++-------- index.go | 40 ++++++---------- index_internal_test.go | 2 +- index_test.go | 2 +- roaring/roaring.go | 4 +- roaring/unmarshal_binary.go | 5 +- server.go | 18 +++++-- snapshotqueue.go | 96 ++++++++++++++++++++++++++++++------- test/field.go | 4 +- test/index.go | 4 +- view.go | 42 ++++++++-------- view_internal_test.go | 2 +- 19 files changed, 235 insertions(+), 195 deletions(-) diff --git a/cluster_internal_test.go b/cluster_internal_test.go index b9e714326..059053423 100644 --- a/cluster_internal_test.go +++ b/cluster_internal_test.go @@ -96,7 +96,7 @@ func newIndexWithTempPath(name string) *Index { if err != nil { panic(err) } - index, err := NewIndex(path, name, DefaultPartitionN) + index, err := NewIndex(NewHolder(DefaultPartitionN), path, name) if err != nil { panic(err) } diff --git a/field.go b/field.go index b746e1800..f82f03543 100644 --- a/field.go +++ b/field.go @@ -31,7 +31,6 @@ import ( "github.com/gogo/protobuf/proto" "github.com/pilosa/pilosa/v2/internal" - "github.com/pilosa/pilosa/v2/logger" "github.com/pilosa/pilosa/v2/pql" "github.com/pilosa/pilosa/v2/roaring" "github.com/pilosa/pilosa/v2/stats" @@ -117,10 +116,6 @@ type Field struct { // Shards with data on any node in the cluster, according to this node. remoteAvailableShards *roaring.Bitmap - logger logger.Logger - - snapshotQueue snapshotQueue - translateStore TranslateStore // Instantiates new translation stores @@ -338,17 +333,17 @@ func OptFieldTypeBool() FieldOption { // that it's of the type `OptFieldType*`). This means // this function couldn't be used to set, for example, // `FieldOptions.Keys`. -func NewField(path, index, name string, opts FieldOption) (*Field, error) { +func NewField(holder *Holder, path, index, name string, opts FieldOption) (*Field, error) { err := validateName(name) if err != nil { return nil, errors.Wrap(err, "validating name") } - return newField(path, index, name, opts) + return newField(holder, path, index, name, opts) } // newField returns a new instance of field (without name validation). -func newField(path, index, name string, opts FieldOption) (*Field, error) { +func newField(holder *Holder, path, index, name string, opts FieldOption) (*Field, error) { // Apply functional option. fo := FieldOptions{} err := opts(&fo) @@ -372,7 +367,7 @@ func newField(path, index, name string, opts FieldOption) (*Field, error) { remoteAvailableShards: roaring.NewBitmap(), - logger: logger.NopLogger, + holder: holder, OpenTranslateStore: OpenInMemTranslateStore, } @@ -448,7 +443,7 @@ func (f *Field) loadAvailableShards() error { } // some other problem: if err != nil { - f.logger.Printf("available shards file present but unreadable, discarding: %v", err) + f.holder.Logger.Printf("available shards file present but unreadable, discarding: %v", err) err = os.Remove(path) if err != nil { return errors.Wrap(err, "deleting corrupt available shards list") @@ -457,7 +452,7 @@ func (f *Field) loadAvailableShards() error { } bm := roaring.NewBitmap() if err = bm.UnmarshalBinary(buf); err != nil { - f.logger.Printf("available shards file corrupt, discarding: %v", err) + f.holder.Logger.Printf("available shards file corrupt, discarding: %v", err) err = os.Remove(path) if err != nil { return errors.Wrap(err, "deleting corrupt available shards list") @@ -548,17 +543,17 @@ func (f *Field) Options() FieldOptions { func (f *Field) Open() error { if err := func() (err error) { // Ensure the field's path exists. - f.logger.Debugf("ensure field path exists: %s", f.path) + f.holder.Logger.Debugf("ensure field path exists: %s", f.path) if err := os.MkdirAll(f.path, 0777); err != nil { return errors.Wrap(err, "creating field dir") } - f.logger.Debugf("load meta file for index/field: %s/%s", f.index, f.name) + f.holder.Logger.Debugf("load meta file for index/field: %s/%s", f.index, f.name) if err := f.loadMeta(); err != nil { return errors.Wrap(err, "loading meta") } - f.logger.Debugf("load available shards for index/field: %s/%s", f.index, f.name) + f.holder.Logger.Debugf("load available shards for index/field: %s/%s", f.index, f.name) if err := f.loadAvailableShards(); err != nil { return errors.Wrap(err, "loading available shards") } @@ -570,17 +565,17 @@ func (f *Field) Open() error { } // Apply the field options loaded from meta (or set via setOptions()). - f.logger.Debugf("apply options for index/field: %s/%s", f.index, f.name) + f.holder.Logger.Debugf("apply options for index/field: %s/%s", f.index, f.name) if err := f.applyOptions(f.options); err != nil { return errors.Wrap(err, "applying options") } - f.logger.Debugf("open views for index/field: %s/%s", f.index, f.name) + f.holder.Logger.Debugf("open views for index/field: %s/%s", f.index, f.name) if err := f.openViews(); err != nil { return errors.Wrap(err, "opening views") } - f.logger.Debugf("open row attribute store for index/field: %s/%s", f.index, f.name) + f.holder.Logger.Debugf("open row attribute store for index/field: %s/%s", f.index, f.name) if err := f.rowAttrStore.Open(); err != nil { return errors.Wrap(err, "opening attrstore") } @@ -607,7 +602,7 @@ func (f *Field) Open() error { return err } - f.logger.Debugf("successfully opened field index/field: %s/%s", f.index, f.name) + f.holder.Logger.Debugf("successfully opened field index/field: %s/%s", f.index, f.name) return nil } func blockingWriteAvailableShards(fieldPath string, availableShardBytes []byte) { @@ -748,7 +743,7 @@ fileLoop: <-fieldQueue }() name := filepath.Base(fi.Name()) - f.logger.Debugf("open index/field/view: %s/%s/%s", f.index, f.name, fi.Name()) + f.holder.Logger.Debugf("open index/field/view: %s/%s/%s", f.index, f.name, fi.Name()) view := f.newView(f.viewPath(name), name) if err := view.open(); err != nil { return fmt.Errorf("opening view: view=%s, err=%s", view.name, err) @@ -770,7 +765,7 @@ fileLoop: } view.rowAttrStore = f.rowAttrStore - f.logger.Debugf("add index/field/view to field.viewMap: %s/%s/%s", f.index, f.name, view.name) + f.holder.Logger.Debugf("add index/field/view to field.viewMap: %s/%s/%s", f.index, f.name, view.name) mu.Lock() f.viewMap[view.name] = view mu.Unlock() @@ -1183,14 +1178,10 @@ func (f *Field) createViewIfNotExistsBase(name string) (*view, bool, error) { } func (f *Field) newView(path, name string) *view { - view := newView(path, f.index, f.name, name, f.options) - view.logger = f.logger + view := newView(f.holder, path, f.index, f.name, name, f.options) view.rowAttrStore = f.rowAttrStore view.stats = f.Stats view.broadcaster = f.broadcaster - if f.snapshotQueue != nil { - view.snapshotQueue = f.snapshotQueue - } return view } diff --git a/field_internal_test.go b/field_internal_test.go index db4a3b5e7..296e54be0 100644 --- a/field_internal_test.go +++ b/field_internal_test.go @@ -205,7 +205,7 @@ func NewTestField(t *testing.T, opts FieldOption) *TestField { if err != nil { t.Fatal(err) } - field, err := NewField(path, "i", "f", opts) + field, err := NewField(NewHolder(DefaultPartitionN), path, "i", "f", opts) if err != nil { t.Fatal(err) } @@ -235,7 +235,7 @@ func (f *TestField) Reopen() error { } path, index, name := f.Path(), f.Index(), f.Name() - f.Field, err = NewField(path, index, name, OptFieldTypeDefault()) + f.Field, err = NewField(NewHolder(DefaultPartitionN), path, index, name, OptFieldTypeDefault()) if err != nil { return err } @@ -730,7 +730,7 @@ func TestDecimalField_MinMaxBoundaries(t *testing.T) { }, } { t.Run("minmax"+strconv.Itoa(i), func(t *testing.T) { - _, err := NewField("no-path", "i", "f", OptFieldTypeDecimal(test.scale, test.min, test.max)) + _, err := NewField(NewHolder(DefaultPartitionN), "no-path", "i", "f", OptFieldTypeDecimal(test.scale, test.min, test.max)) if err != nil && test.expErr { if !strings.Contains(err.Error(), "is not supported") { t.Fatal(err) diff --git a/field_test.go b/field_test.go index 40b64728a..ab24b466f 100644 --- a/field_test.go +++ b/field_test.go @@ -144,7 +144,7 @@ func TestField_NameRestriction(t *testing.T) { if err != nil { panic(err) } - field, err := pilosa.NewField(path, "i", ".meta", pilosa.OptFieldTypeDefault()) + field, err := pilosa.NewField(pilosa.NewHolder(pilosa.DefaultPartitionN), path, "i", ".meta", pilosa.OptFieldTypeDefault()) if field != nil { t.Fatalf("unexpected field name %s", err) } @@ -177,13 +177,13 @@ func TestField_NameValidation(t *testing.T) { panic(err) } for _, name := range validFieldNames { - _, err := pilosa.NewField(path, "i", name, pilosa.OptFieldTypeDefault()) + _, err := pilosa.NewField(pilosa.NewHolder(pilosa.DefaultPartitionN), path, "i", name, pilosa.OptFieldTypeDefault()) if err != nil { t.Fatalf("unexpected field name: %s %s", name, err) } } for _, name := range invalidFieldNames { - _, err := pilosa.NewField(path, "i", name, pilosa.OptFieldTypeDefault()) + _, err := pilosa.NewField(pilosa.NewHolder(pilosa.DefaultPartitionN), path, "i", name, pilosa.OptFieldTypeDefault()) if err == nil { t.Fatalf("expected error on field name: %s", name) } diff --git a/fragment.go b/fragment.go index 3258d6f00..f70e9e71c 100644 --- a/fragment.go +++ b/fragment.go @@ -106,6 +106,10 @@ type fragment struct { field string view string shard uint64 + + // parent holder, used to find snapshot queue, etc. + holder *Holder + // debugging tool: addresses of current and previous maps prevdata, currdata struct{ from, to uintptr } @@ -152,12 +156,10 @@ type fragment struct { mutexVector vector stats stats.StatsClient - - snapshotQueue snapshotQueue } // newFragment returns a new instance of Fragment. -func newFragment(path, index, field, view string, shard uint64, flags byte) *fragment { +func newFragment(holder *Holder, path, index, field, view string, shard uint64, flags byte) *fragment { f := &fragment{ path: path, index: index, @@ -168,11 +170,10 @@ func newFragment(path, index, field, view string, shard uint64, flags byte) *fra CacheType: DefaultCacheType, CacheSize: DefaultCacheSize, - Logger: logger.NopLogger, + holder: holder, MaxOpN: defaultFragmentMaxOpN, - stats: stats.NopStatsClient, - snapshotQueue: defaultSnapshotQueue, + stats: stats.NopStatsClient, } f.snapshotCond = sync.Cond{L: &f.mu} return f @@ -188,13 +189,13 @@ func (f *fragment) Open() error { if err := func() error { // Initialize storage in a function so we can close if anything goes wrong. - f.Logger.Debugf("open storage for index/field/view/fragment: %s/%s/%s/%d", f.index, f.field, f.view, f.shard) + f.holder.Logger.Debugf("open storage for index/field/view/fragment: %s/%s/%s/%d", f.index, f.field, f.view, f.shard) if err := f.openStorage(true); err != nil { return errors.Wrap(err, "opening storage") } // Fill cache with rows persisted to disk. - f.Logger.Debugf("open cache for index/field/view/fragment: %s/%s/%s/%d", f.index, f.field, f.view, f.shard) + f.holder.Logger.Debugf("open cache for index/field/view/fragment: %s/%s/%s/%d", f.index, f.field, f.view, f.shard) if err := f.openCache(); err != nil { e2 := f.closeStorage() if e2 != nil { @@ -214,7 +215,7 @@ func (f *fragment) Open() error { return err } - f.Logger.Debugf("successfully opened index/field/view/fragment: %s/%s/%s/%d", f.index, f.field, f.view, f.shard) + f.holder.Logger.Debugf("successfully opened index/field/view/fragment: %s/%s/%s/%d", f.index, f.field, f.view, f.shard) return nil } @@ -273,7 +274,7 @@ func (f *fragment) importStorage(data []byte, file *os.File, newGen generation, } return false, fmt.Errorf("unmarshal storage: file=%s, err=%s", file.Name(), err) } - f.Logger.Printf("warning: unmarshal storage, file=%s, err=%v", file.Name(), err) + f.holder.Logger.Printf("warning: unmarshal storage, file=%s, err=%v", file.Name(), err) trunc, ok := cause.(roaring.FileShouldBeTruncatedError) if ok { // generation code looks for a FileShouldBeTruncatedError @@ -299,7 +300,7 @@ func (f *fragment) applyStorage(data []byte, file *os.File, newGen generation, m if file != nil { fi, err := file.Stat() if err != nil { - f.Logger.Printf("trying to apply new storage to existing bitmap, stat failed: %v", err) + f.holder.Logger.Printf("trying to apply new storage to existing bitmap, stat failed: %v", err) } if err == nil && fi != nil && fi.Size() == 0 { return f.emptyStorage(file) @@ -359,7 +360,7 @@ func (f *fragment) openStorage(unmarshalData bool) error { storageOp = f.applyStorage } var err error - f.gen, err = newGeneration(f.gen, f.path, unmarshalData, storageOp, f.Logger) + f.gen, err = newGeneration(f.gen, f.path, unmarshalData, storageOp, f.holder.Logger) if f.gen != nil { scratchData := f.gen.Bytes() f.prevdata = f.currdata @@ -407,7 +408,7 @@ func (f *fragment) openCache() error { // Unmarshal cache data. var pb internal.Cache if err := proto.Unmarshal(buf, &pb); err != nil { - f.Logger.Printf("error unmarshaling cache data, skipping: path=%s, err=%s", path, err) + f.holder.Logger.Printf("error unmarshaling cache data, skipping: path=%s, err=%s", path, err) return nil } @@ -435,13 +436,13 @@ func (f *fragment) Close() error { func (f *fragment) close() error { // Flush cache if closing gracefully. if err := f.flushCache(); err != nil { - f.Logger.Printf("fragment: error flushing cache on close: err=%s, path=%s", err, f.path) + f.holder.Logger.Printf("fragment: error flushing cache on close: err=%s, path=%s", err, f.path) return errors.Wrap(err, "flushing cache") } // Close underlying storage. if err := f.closeStorage(); err != nil { - f.Logger.Printf("fragment: error closing storage: err=%s, path=%s", err, f.path) + f.holder.Logger.Printf("fragment: error closing storage: err=%s, path=%s", err, f.path) return errors.Wrap(err, "closing storage") } @@ -688,7 +689,8 @@ func (f *fragment) unprotectedSetRow(row *Row, rowID uint64) (changed bool, err f.rowCache.Add(rowID, nil) // Snapshot storage. - f.snapshotQueue.Enqueue(f) + f.holder.SnapshotQueue.Enqueue(f) + f.stats.Count("setRow", 1, 1.0) return changed, nil } @@ -728,7 +730,7 @@ func (f *fragment) unprotectedClearRow(rowID uint64) (changed bool, err error) { f.rowCache.Add(rowID, nil) // Snapshot storage. - f.snapshotQueue.Enqueue(f) + f.holder.SnapshotQueue.Enqueue(f) return changed, nil } @@ -2006,11 +2008,11 @@ func (f *fragment) importPositions(set, clear []uint64, rowSet map[uint64]struct // we got an error. it's possible that the error indicates that something went wrong. mappedIn, mappedOut, unmappedIn, errs, e2 := f.storage.SanityCheckMapping(f.currdata.from, f.currdata.to) if errs != 0 { - f.Logger.Printf("transaction failed on %s. storage has %d mapped in range, %d mapped out of range, %d unmapped in range, %d errors total, last %v", + f.holder.Logger.Printf("transaction failed on %s. storage has %d mapped in range, %d mapped out of range, %d unmapped in range, %d errors total, last %v", f.path, mappedIn, mappedOut, unmappedIn, errs, e2) if f.prevdata.from != f.currdata.from { mappedIn, mappedOut, unmappedIn, errs, e2 = f.storage.SanityCheckMapping(f.prevdata.from, f.prevdata.to) - f.Logger.Printf("with previous map, storage would have %d mapped in range, %d mapped out of range, %d unmapped in range, %d errors total, last %v", + f.holder.Logger.Printf("with previous map, storage would have %d mapped in range, %d mapped out of range, %d unmapped in range, %d errors total, last %v", mappedIn, mappedOut, unmappedIn, errs, e2) } } @@ -2164,7 +2166,7 @@ func (f *fragment) importValue(columnIDs []uint64, values []int64, bitDepth uint // in theory, this should probably have been queued anyway, but if enough // of the bits matched existing bits, we'll be under our opN estimate, and // we want to ensure that the snapshot happens. - return f.snapshotQueue.Immediate(f) + return f.holder.SnapshotQueue.Immediate(f) } // importRoaring imports from the official roaring data format defined at @@ -2252,7 +2254,7 @@ func (f *fragment) incrementOpN(changed int) { f.opN += changed f.ops++ if f.opN > f.MaxOpN { - f.snapshotQueue.Enqueue(f) + f.holder.SnapshotQueue.Enqueue(f) } } @@ -2286,7 +2288,7 @@ func (f *fragment) snapshot() (err error) { // we can't see the actual values that were used to generate this, probably. if e2.Error() == "runtime error: invalid memory address or nil pointer dereference" { mappedIn, mappedOut, unmappedIn, errs, _ := f.storage.SanityCheckMapping(f.currdata.from, f.currdata.to) - f.Logger.Printf("transaction failed on %s. storage has %d mapped in range, %d mapped out of range, %d unmapped in range, %d errors total", + f.holder.Logger.Printf("transaction failed on %s. storage has %d mapped in range, %d mapped out of range, %d unmapped in range, %d errors total", f.path, mappedIn, mappedOut, unmappedIn, errs) } } else { @@ -2306,7 +2308,7 @@ func (f *fragment) snapshot() (err error) { func unprotectedWriteToFragment(f *fragment, bm *roaring.Bitmap) (n int64, err error) { // nolint: interfacer completeMessage := fmt.Sprintf("fragment: snapshot complete %s/%s/%s/%d", f.index, f.field, f.view, f.shard) start := time.Now() - defer track(start, completeMessage, f.stats, f.Logger) + defer track(start, completeMessage, f.stats, f.holder.Logger) // Create a temporary file to snapshot to. snapshotPath := f.path + snapshotExt @@ -3096,7 +3098,7 @@ func (s *fragmentSyncer) syncBlockFromPrimary(id int) error { // the primary node. nodes := s.Cluster.shardNodes(f.index, f.shard) if s.Node.ID != nodes[0].ID { - f.Logger.Debugf("non-primary replica expecting sync from primary: %s, index=%s, field=%s, shard=%d", nodes[0].ID, f.index, f.field, f.shard) + f.holder.Logger.Debugf("non-primary replica expecting sync from primary: %s, index=%s, field=%s, shard=%d", nodes[0].ID, f.index, f.field, f.shard) return nil } diff --git a/fragment_internal_test.go b/fragment_internal_test.go index 5912eb132..2d8c27b73 100644 --- a/fragment_internal_test.go +++ b/fragment_internal_test.go @@ -1477,7 +1477,7 @@ func BenchmarkFragment_Blocks(b *testing.B) { } // Open the fragment specified by the path. - f := newFragment(*FragmentPath, "i", "f", viewStandard, 0, 0) + f := newFragment(NewHolder(DefaultPartitionN), *FragmentPath, "i", "f", viewStandard, 0, 0) if err := f.Open(); err != nil { b.Fatal(err) } @@ -2006,7 +2006,7 @@ func BenchmarkFragment_Snapshot(b *testing.B) { b.ReportAllocs() // Open the fragment specified by the path. - f := newFragment(*FragmentPath, "i", "f", viewStandard, 0, 0) + f := newFragment(NewHolder(DefaultPartitionN), *FragmentPath, "i", "f", viewStandard, 0, 0) if err := f.Open(); err != nil { b.Fatal(err) } @@ -2121,7 +2121,7 @@ func BenchmarkImportRoaring(b *testing.B) { // care whether this succeeds, // but if it's happening we want // it to be done. - _ = f.snapshotQueue.Await(f) + _ = defaultSnapshotQueue.Await(f) f.Clean(b) b.Fatalf("import error: %v", err) } @@ -2162,7 +2162,7 @@ func BenchmarkImportRoaringConcurrent(b *testing.B) { err := frags[j].importRoaringT(data[j], false) // error unimportant if it happened, but we want // any snapshots to have finished. - _ = frags[j].snapshotQueue.Await(frags[j]) + _ = defaultSnapshotQueue.Await(frags[j]) return err }) } @@ -2203,7 +2203,7 @@ func BenchmarkImportRoaringUpdateConcurrent(b *testing.B) { if err != nil { b.Fatalf("importing roaring: %v", err) } - err = frags[j].snapshotQueue.Immediate(frags[j]) + err = defaultSnapshotQueue.Immediate(frags[j]) if err != nil { b.Fatalf("snapshot after import: %v", err) } @@ -2214,7 +2214,7 @@ func BenchmarkImportRoaringUpdateConcurrent(b *testing.B) { j := j eg.Go(func() error { err := frags[j].importRoaringT(updata, false) - err2 := frags[j].snapshotQueue.Await(frags[j]) + err2 := defaultSnapshotQueue.Await(frags[j]) if err == nil { err = err2 } @@ -2282,7 +2282,7 @@ func BenchmarkImportRoaringUpdate(b *testing.B) { if err != nil { b.Errorf("import error: %v", err) } - err = f.snapshotQueue.Immediate(f) + err = defaultSnapshotQueue.Immediate(f) if err != nil { b.Errorf("snapshot after import error: %v", err) } @@ -2292,7 +2292,7 @@ func BenchmarkImportRoaringUpdate(b *testing.B) { f.Clean(b) b.Errorf("import error: %v", err) } - err = f.snapshotQueue.Await(f) + err = defaultSnapshotQueue.Await(f) if err != nil { b.Errorf("snapshot after import error: %v", err) } @@ -2391,7 +2391,7 @@ func BenchmarkImportIntoLargeFragment(b *testing.B) { } origF.Close() fi.Close() - nf := newFragment(fi.Name(), "i", "f", viewStandard, 0, 0) + nf := newFragment(NewHolder(DefaultPartitionN), fi.Name(), "i", "f", viewStandard, 0, 0) err = nf.Open() if err != nil { b.Fatalf("opening fragment: %v", err) @@ -2428,7 +2428,7 @@ func BenchmarkImportRoaringIntoLargeFragment(b *testing.B) { } origF.Close() fi.Close() - nf := newFragment(fi.Name(), "i", "f", viewStandard, 0, 0) + nf := newFragment(NewHolder(DefaultPartitionN), fi.Name(), "i", "f", viewStandard, 0, 0) err = nf.Open() if err != nil { b.Fatalf("opening fragment: %v", err) @@ -2607,17 +2607,23 @@ func (f *fragment) sanityCheck(t testing.TB) { func (f *fragment) Clean(t testing.TB) { f.mu.Lock() - err := f.snapshotQueue.Await(f) - f.mu.Unlock() - if err != nil { - t.Fatalf("snapshot failed before sanity check: %v", err) - } - f.sanityCheck(t) - if f.storage != nil && f.storage.Source != nil { - if f.storage.Source.Dead() { - t.Fatalf("cleaning up fragment %s, source %s, source already dead", f.path, f.storage.Source.ID()) + // we need to ensure that we unlock the mutex before terminating + // the clean operation, but we need it held during the sanity + // check or else, in some cases, the background snapshot queue + // can decide to pick it up. + func() { + defer f.mu.Unlock() + err := defaultSnapshotQueue.Await(f) + if err != nil { + t.Fatalf("snapshot failed before sanity check: %v", err) } - } + f.sanityCheck(t) + if f.storage != nil && f.storage.Source != nil { + if f.storage.Source.Dead() { + t.Fatalf("cleaning up fragment %s, source %s, source already dead", f.path, f.storage.Source.ID()) + } + } + }() errc := f.Close() // prevent double-closes of generation during testing. f.gen = nil @@ -2626,10 +2632,6 @@ func (f *fragment) Clean(t testing.TB) { if errc != nil || errf != nil { t.Fatal("cleaning up fragment: ", errc, errf, errp) } - if f.snapshotQueue != nil { - f.snapshotQueue.Stop() - f.snapshotQueue = nil - } // not all fragments have cache files if errp != nil && !os.IsNotExist(errp) { t.Fatalf("cleaning up fragment cache: %v", errp) @@ -2649,10 +2651,6 @@ func (f *fragment) CleanKeep(t testing.TB) { if errc != nil { t.Fatal("closing fragment: ", errc, errp) } - if f.snapshotQueue != nil { - f.snapshotQueue.Stop() - f.snapshotQueue = nil - } // not all fragments have cache files if errp != nil && !os.IsNotExist(errp) { t.Fatalf("cleaning up fragment cache: %v", errp) @@ -2680,12 +2678,11 @@ func mustOpenFragmentFlags(index, field, view string, shard uint64, cacheType st cacheType = DefaultCacheType } - f := newFragment(file.Name(), index, field, view, shard, flags) + f := newFragment(NewHolder(DefaultPartitionN), file.Name(), index, field, view, shard, flags) f.CacheType = cacheType f.RowAttrStore = &memAttrStore{ store: make(map[uint64]map[string]interface{}), } - f.snapshotQueue = newSnapshotQueue(1, 1, nil) if err := f.Open(); err != nil { panic(err) @@ -3179,7 +3176,7 @@ func TestUnionInPlaceMapped(t *testing.T) { // it's used only in computation of things that usually don't go to // disk, which is why we handle this specially in testing and not // generically. - err = f.snapshotQueue.Immediate(f) + err = defaultSnapshotQueue.Immediate(f) if err != nil { t.Fatalf("snapshot after union-in-place: %v", err) } @@ -3332,7 +3329,6 @@ func TestImportClearRestart(t *testing.T) { cols: []uint64{1, 1, 1, 1, 1, 1}, }, } - for i, test := range tests { for _, maxOpN := range []int{0, 10000} { t.Run(fmt.Sprintf("%dMaxOpN%d", i, maxOpN), func(t *testing.T) { @@ -3385,7 +3381,7 @@ func TestImportClearRestart(t *testing.T) { check(t, f, exp) - f2 := newFragment(f.path, "i", "f", viewStandard, 0, 0) + f2 := newFragment(NewHolder(DefaultPartitionN), f.path, "i", "f", viewStandard, 0, 0) f2.MaxOpN = maxOpN f2.CacheType = f.CacheType @@ -3419,7 +3415,7 @@ func TestImportClearRestart(t *testing.T) { check(t, f2, exp) - f3 := newFragment(f2.path, "i", "f", viewStandard, 0, 0) + f3 := newFragment(NewHolder(DefaultPartitionN), f2.path, "i", "f", viewStandard, 0, 0) f3.MaxOpN = maxOpN f3.CacheType = f.CacheType diff --git a/generation_test.go b/generation_test.go index 75444fcf7..7df71fc98 100644 --- a/generation_test.go +++ b/generation_test.go @@ -39,14 +39,14 @@ func TestGenerationPanic(t *testing.T) { } prevData = f.gen.(*mmapGeneration).data f.mu.Lock() - _ = f.snapshotQueue.Immediate(f) + _ = defaultSnapshotQueue.Immediate(f) f.mu.Unlock() runtime.GC() for i := 0; i < (f.MaxOpN / 2); i++ { _, _ = f.setBit(0, uint64(i*32)+23) } f.mu.Lock() - f.snapshotQueue.Await(f) + defaultSnapshotQueue.Await(f) f.mu.Unlock() runtime.GC() newData := f.gen.(*mmapGeneration).data diff --git a/holder.go b/holder.go index de3cd0701..d06920cd3 100644 --- a/holder.go +++ b/holder.go @@ -77,9 +77,8 @@ type Holder struct { // The interval at which the cached row ids are persisted to disk. cacheFlushInterval time.Duration - Logger logger.Logger - - snapshotQueue snapshotQueue + Logger logger.Logger + SnapshotQueue SnapshotQueue // Instantiates new translation stores OpenTranslateStore OpenTranslateStoreFunc @@ -169,6 +168,8 @@ func NewHolder(partitionN int) *Holder { translationSyncer: NopTranslationSyncer, Logger: logger.NopLogger, + + SnapshotQueue: defaultSnapshotQueue, } } @@ -213,11 +214,6 @@ func (h *Holder) Open() error { return errors.Wrap(err, "reading directory") } - // Run snapshots asynchronously. The snapshotQueue will have a background - // task associated with it which flushes it and waits until this channel - // is closed, so we should always close this channel when done. - h.snapshotQueue = newSnapshotQueue(10, 2, h.Logger) - for _, fi := range fis { // Skip files or hidden directories. if !fi.IsDir() || strings.HasPrefix(fi.Name(), ".") { @@ -261,18 +257,25 @@ func (h *Holder) Open() error { h.Logger.Printf("open holder: complete") - // Periodically flush cache. - h.wg.Add(1) - go func() { defer h.wg.Done(); h.monitorCacheFlush() }() - h.Stats.Open() - h.snapshotQueue.ScanHolder(h) h.opened.Close() return nil } +// Activate runs the background tasks relevant to keeping a holder in a stable +// state, such as scanning it for needed snapshots, or flushing caches. This +// is separate from opening because, while a server would nearly always want +// to do this, other use cases (like consistency checks of a data directory) +// need to avoid it even getting started. +func (h *Holder) Activate() { + // Periodically flush cache. + h.wg.Add(2) + go func() { defer h.wg.Done(); h.monitorCacheFlush() }() + go func() { defer h.wg.Done(); h.SnapshotQueue.ScanHolder(h, h.closing) }() +} + // checkForeignIndex is a check before applying a foreign // index to a field; if the index is not yet available, // (because holder is still opening and may not have opened @@ -313,10 +316,6 @@ func (h *Holder) Close() error { return errors.Wrap(err, "closing index") } } - if h.snapshotQueue != nil { - h.snapshotQueue.Stop() - h.snapshotQueue = nil - } // Reset opened in case Holder needs to be reopened. h.opened.mu.Lock() @@ -587,19 +586,16 @@ func (h *Holder) createIndex(name string, opt IndexOptions) (*Index, error) { } func (h *Holder) newIndex(path, name string) (*Index, error) { - index, err := NewIndex(path, name, h.partitionN) + index, err := NewIndex(h, path, name) if err != nil { return nil, err } - index.logger = h.Logger index.Stats = h.Stats.WithTags(fmt.Sprintf("index:%s", index.Name())) index.broadcaster = h.broadcaster index.newAttrStore = h.NewAttrStore index.columnAttrs = h.NewAttrStore(filepath.Join(index.path, ".data")) - index.snapshotQueue = h.snapshotQueue index.OpenTranslateStore = h.OpenTranslateStore index.translationSyncer = h.translationSyncer - index.holder = h return index, nil } diff --git a/index.go b/index.go index 481d22cf4..4a7ee3f5f 100644 --- a/index.go +++ b/index.go @@ -27,7 +27,6 @@ import ( "github.com/gogo/protobuf/proto" "github.com/pilosa/pilosa/v2/internal" - "github.com/pilosa/pilosa/v2/logger" "github.com/pilosa/pilosa/v2/roaring" "github.com/pilosa/pilosa/v2/stats" "github.com/pkg/errors" @@ -46,9 +45,6 @@ type Index struct { trackExistence bool existenceFld *Field - // Partitions used by translation. - partitionN int - // Fields by name. fields map[string]*Field @@ -60,9 +56,6 @@ type Index struct { broadcaster broadcaster Stats stats.StatsClient - logger logger.Logger - snapshotQueue snapshotQueue - // Passed to field for foreign-index lookup. holder *Holder @@ -76,24 +69,23 @@ type Index struct { } // NewIndex returns a new instance of Index. -func NewIndex(path, name string, partitionN int) (*Index, error) { +func NewIndex(holder *Holder, path, name string) (*Index, error) { err := validateName(name) if err != nil { return nil, errors.Wrap(err, "validating name") } return &Index{ - path: path, - name: name, - partitionN: partitionN, - fields: make(map[string]*Field), + path: path, + name: name, + fields: make(map[string]*Field), newAttrStore: newNopAttrStore, columnAttrs: nopStore, broadcaster: NopBroadcaster, Stats: stats.NopStatsClient, - logger: logger.NopLogger, + holder: holder, trackExistence: true, translateStores: make(map[int]TranslateStore), @@ -155,18 +147,18 @@ func (i *Index) OpenWithTimestamp() error { return i.open(true) } func (i *Index) open(withTimestamp bool) (err error) { // Ensure the path exists. - i.logger.Debugf("ensure index path exists: %s", i.path) + i.holder.Logger.Debugf("ensure index path exists: %s", i.path) if err := os.MkdirAll(i.path, 0777); err != nil { return errors.Wrap(err, "creating directory") } // Read meta file. - i.logger.Debugf("load meta file for index: %s", i.name) + i.holder.Logger.Debugf("load meta file for index: %s", i.name) if err := i.loadMeta(); err != nil { return errors.Wrap(err, "loading meta file") } - i.logger.Debugf("open fields for index: %s", i.name) + i.holder.Logger.Debugf("open fields for index: %s", i.name) if err := i.openFields(withTimestamp); err != nil { return errors.Wrap(err, "opening fields") } @@ -181,15 +173,15 @@ func (i *Index) open(withTimestamp bool) (err error) { return errors.Wrap(err, "opening attrstore") } - i.logger.Debugf("open translate store for index: %s", i.name) + i.holder.Logger.Debugf("open translate store for index: %s", i.name) var g errgroup.Group var mu sync.Mutex - for partitionID := 0; partitionID < i.partitionN; partitionID++ { + for partitionID := 0; partitionID < i.holder.partitionN; partitionID++ { partitionID := partitionID g.Go(func() error { - store, err := i.OpenTranslateStore(i.TranslateStorePath(partitionID), i.name, "", partitionID, i.partitionN) + store, err := i.OpenTranslateStore(i.TranslateStorePath(partitionID), i.name, "", partitionID, i.holder.partitionN) if err != nil { return errors.Wrapf(err, "opening index translate store: partition=%d", partitionID) } @@ -239,7 +231,7 @@ fileLoop: defer func() { <-indexQueue }() - i.logger.Debugf("open field: %s", fi.Name()) + i.holder.Logger.Debugf("open field: %s", fi.Name()) mu.Lock() fld, err := i.newField(i.fieldPath(filepath.Base(fi.Name())), filepath.Base(fi.Name())) if withTimestamp { @@ -257,7 +249,7 @@ fileLoop: if err := fld.Open(); err != nil { return fmt.Errorf("open field: name=%s, err=%s", fld.Name(), err) } - i.logger.Debugf("add field to index.fields: %s", fi.Name()) + i.holder.Logger.Debugf("add field to index.fields: %s", fi.Name()) mu.Lock() i.fields[fld.Name()] = fld mu.Unlock() @@ -512,17 +504,13 @@ func (i *Index) createField(name string, opt *FieldOptions) (*Field, error) { } func (i *Index) newField(path, name string) (*Field, error) { - f, err := newField(path, i.name, name, OptFieldTypeDefault()) + f, err := newField(i.holder, path, i.name, name, OptFieldTypeDefault()) if err != nil { return nil, err } - f.logger = i.logger f.Stats = i.Stats f.broadcaster = i.broadcaster f.rowAttrStore = i.newAttrStore(filepath.Join(f.path, ".data")) - if i.snapshotQueue != nil { - f.snapshotQueue = i.snapshotQueue - } f.OpenTranslateStore = i.OpenTranslateStore return f, nil } diff --git a/index_internal_test.go b/index_internal_test.go index 026607da4..5b5a5b5a3 100644 --- a/index_internal_test.go +++ b/index_internal_test.go @@ -25,7 +25,7 @@ func mustOpenIndex(opt IndexOptions) *Index { if err != nil { panic(err) } - index, err := NewIndex(path, "i", DefaultPartitionN) + index, err := NewIndex(NewHolder(1), path, "i") if err != nil { panic(err) } diff --git a/index_test.go b/index_test.go index 63e9b6179..bfebee7f7 100644 --- a/index_test.go +++ b/index_test.go @@ -242,7 +242,7 @@ func TestIndex_InvalidName(t *testing.T) { if err != nil { panic(err) } - index, err := pilosa.NewIndex(path, "ABC", pilosa.DefaultPartitionN) + index, err := pilosa.NewIndex(pilosa.NewHolder(pilosa.DefaultPartitionN), path, "ABC") if err == nil { t.Fatalf("should have gotten an error on index name with caps") } diff --git a/roaring/roaring.go b/roaring/roaring.go index 3eb5ecef2..f79b20c5b 100644 --- a/roaring/roaring.go +++ b/roaring/roaring.go @@ -2513,8 +2513,8 @@ type BitmapInfo struct { OpDetails []OpInfo BitCount uint64 ContainerCount int - Containers []ContainerInfo - OpContainers []ContainerInfo + Containers []ContainerInfo // The containers found in the bitmap originally + OpContainers []ContainerInfo // The containers resulting from ops log changes. } // Iterator represents an iterator over a Bitmap. diff --git a/roaring/unmarshal_binary.go b/roaring/unmarshal_binary.go index 59f65a117..ee566236a 100644 --- a/roaring/unmarshal_binary.go +++ b/roaring/unmarshal_binary.go @@ -96,8 +96,9 @@ func (b *Bitmap) UnmarshalBinary(data []byte) (err error) { return nil } -// unmarshalPilosaRoaring treats data as being encoded in Pilosa's 64 bit -// roaring format and decodes it into b. +// InspectBinary reads a roaring bitmap, plus a possible ops log, +// and reports back on the contents, including distinguishing between +// the original ops log and the post-ops-log contents. func InspectBinary(data []byte) (info BitmapInfo, err error) { if data == nil { return info, errors.New("no roaring bitmap provided") diff --git a/server.go b/server.go index 131a6971c..75f5b2024 100644 --- a/server.go +++ b/server.go @@ -66,9 +66,10 @@ type Server struct { // nolint: maligned extensions []*ext.ExtensionInfo // External - systemInfo SystemInfo - gcNotifier GCNotifier - logger logger.Logger + systemInfo SystemInfo + gcNotifier GCNotifier + logger logger.Logger + snapshotQueue SnapshotQueue nodeID string uri URI @@ -533,6 +534,9 @@ func (s *Server) UpAndDown() error { func (s *Server) Open() error { s.logger.Printf("open server") + // Start background monitoring. + s.snapshotQueue = newSnapshotQueue(10, 2, s.logger) + // Log startup err := s.holder.logStartup() if err != nil { @@ -560,6 +564,8 @@ func (s *Server) Open() error { if err := s.holder.Open(); err != nil { return errors.Wrap(err, "opening Holder") } + // bring up the background tasks for the holder. + s.holder.Activate() if err := s.cluster.setNodeState(nodeStateReady); err != nil { return errors.Wrap(err, "setting nodeState") } @@ -571,7 +577,6 @@ func (s *Server) Open() error { // buffered channel. s.cluster.listenForJoins() - // Start background monitoring. s.wg.Add(3) go func() { defer s.wg.Done(); s.monitorAntiEntropy() }() go func() { defer s.wg.Done(); s.monitorRuntime() }() @@ -596,6 +601,11 @@ func (s *Server) Close() error { if s.holder != nil { errh = s.holder.Close() } + if s.snapshotQueue != nil { + s.holder.SnapshotQueue = nil + s.snapshotQueue.Stop() + s.snapshotQueue = nil + } // prefer to return holder error over cluster // error. This order is somewhat arbitrary. It would be better if we had // some way to combine all the errors, but probably not important enough to diff --git a/snapshotqueue.go b/snapshotqueue.go index 373e5c538..5dcc5dbf8 100644 --- a/snapshotqueue.go +++ b/snapshotqueue.go @@ -16,6 +16,7 @@ package pilosa import ( "fmt" + "math/bits" "os" "sync" "sync/atomic" @@ -39,12 +40,22 @@ import ( // Await, Enqueue, and Immediate should be called only with the fragment lock // held. // -// ScanHolder spawns a new goroutine. You don't need to use `go` on it. -type snapshotQueue interface { +// If you create a queue, it should get stopped at some point. The +// atomicSnapshotQueue implementation used as defaultSnapshotQueue has +// a Start function which will tell you whether it actually started a +// queue. This logic exists because in a normal server case, you probably +// want the queue to be shut down as part of server shutdown, but if you're +// running cluster tests, you probably want to start and shop the queue as +// part of the test, not stop it when any server terminates. +// +// It's less likely to be desireable to start/stop individual queues, +// because fragments use the defaultSnapshotQueue anyway. This design +// needs revisiting. +type SnapshotQueue interface { Immediate(*fragment) error Enqueue(*fragment) Await(*fragment) error - ScanHolder(*Holder) + ScanHolder(*Holder, chan struct{}) Stop() } @@ -64,20 +75,18 @@ func (q *queuelessSnapshotQueue) Immediate(f *fragment) error { return f.snapshot() } -func (q *queuelessSnapshotQueue) ScanHolder(h *Holder) { +func (q *queuelessSnapshotQueue) ScanHolder(h *Holder, done chan struct{}) { } func (q *queuelessSnapshotQueue) Stop() { } -// defaultSnapshotQueue is the fallback to use if none is available, -// and currently uses queueless -- it runs all snapshots immediately. -var defaultSnapshotQueue *queuelessSnapshotQueue +var defaultSnapshotQueue = &queuelessSnapshotQueue{} // newSnapshotQueue makes a new snapshot queue, of depth N, with // w worker threads. -func newSnapshotQueue(n int, w int, l logger.Logger) snapshotQueue { - sq := prioritySnapshotQueue{normal: make(chan snapshotRequest, n), urgent: make(chan snapshotRequest), background: make(chan snapshotRequest), done: make(chan struct{}), logger: l} +func newSnapshotQueue(n int, w int, l logger.Logger) SnapshotQueue { + sq := prioritySnapshotQueue{normal: make(chan snapshotRequest, n), urgent: make(chan snapshotRequest), background: make(chan snapshotRequest), done: make(chan struct{}), maxOpN: 10000, logger: l} if sq.logger == nil { sq.logger = logger.NewStandardLogger(os.Stderr) } @@ -108,9 +117,11 @@ type prioritySnapshotQueue struct { done chan struct{} mu sync.RWMutex scanWG, workerWG sync.WaitGroup + maxOpN int + observedOpN [16]int stats struct { - enqueued uint64 - skipped uint64 + enqueued uint32 + skipped uint32 } } @@ -192,7 +203,9 @@ func (sq *prioritySnapshotQueue) Stop() { sq.urgent = nil close(sq.background) sq.background = nil - if sq.stats.skipped > 0 || sq.stats.enqueued > 1 { + enqueued := atomic.LoadUint32(&sq.stats.enqueued) + skipped := atomic.LoadUint32(&sq.stats.skipped) + if skipped > 0 || enqueued > 1 { sq.logger.Printf("snapshot queue: enqueued %d, skipped %d\n", sq.stats.enqueued, sq.stats.skipped) } } @@ -217,10 +230,10 @@ func (sq *prioritySnapshotQueue) Enqueue(f *fragment) { // try to enqueue snapshot select { case sq.normal <- snapshotRequest{frag: f, when: time.Now()}: - atomic.AddUint64(&sq.stats.enqueued, 1) + atomic.AddUint32(&sq.stats.enqueued, 1) return default: - atomic.AddUint64(&sq.stats.skipped, 1) + atomic.AddUint32(&sq.stats.skipped, 1) f.snapshotPending = false return } @@ -230,6 +243,11 @@ func (sq *prioritySnapshotQueue) Enqueue(f *fragment) { // held. Await waits on a condition variable inside f, associated with the // fragment's lock, so this does not conflict with the lock being used for // snapshots. +// +// Note that workers don't stop just because the queue's been stopped; only +// the background scanner is stopped. So an Await shouldn't block forever +// even if the queue gets shut down. If you're reading this, possibly that +// analysis is incorrect. func (sq *prioritySnapshotQueue) Await(f *fragment) (err error) { for f.snapshotPending { f.snapshotCond.Wait() @@ -281,9 +299,19 @@ func (sq *prioritySnapshotQueue) needsSnapshot(f *fragment) bool { if f.snapshotPending { return false } - if f.opN > f.MaxOpN { + if f.opN > sq.maxOpN { return true } + // aka "log2(n) + 1", or 0 for n==0 + pow2 := 32 - bits.LeadingZeros32(uint32(f.opN)) + // 15 == 16384. we assume that since 16384 is higher than our + // normal maxOpN, it's always a reasonable value. + if pow2 > 15 { + pow2 = 15 + } + // store in inverse order so the lowest slot in the array is the + // highest cardinality + sq.observedOpN[15-pow2]++ return false } @@ -291,10 +319,10 @@ func (sq *prioritySnapshotQueue) needsSnapshot(f *fragment) bool { // indexes/fields/views/fragments, looking for fragments which have OpN // high enough to justify a snapshot but don't seem to have one pending. // It then dumps these in the low priority background queue. -func (sq *prioritySnapshotQueue) ScanHolder(h *Holder) { +func (sq *prioritySnapshotQueue) ScanHolder(h *Holder, done chan struct{}) { sq.mu.Lock() sq.scanWG.Add(1) - go sq.scanHolderWorker(h, sq.background, sq.done) + go sq.scanHolderWorker(h, sq.background, done) sq.mu.Unlock() } @@ -302,6 +330,9 @@ func (sq *prioritySnapshotQueue) ScanHolder(h *Holder) { // fragments which need snapshots taken. It's the cleanup task for snapshots // that would have been requested by Enqueue, but the queue was full. func (sq *prioritySnapshotQueue) scanHolderWorker(h *Holder, background chan snapshotRequest, done chan struct{}) { + // queueDone is global to this snapshotQueue, done is specific to this holder + // scanner. + queueDone := sq.done defer sq.scanWG.Done() var indexNames, fieldNames, viewNames []string var fragNums []uint64 @@ -379,9 +410,11 @@ func (sq *prioritySnapshotQueue) scanHolderWorker(h *Holder, background chan sna // background queue. If there's nothing that needs snapshots, // we pause frequently for a second or so at a time. counter++ - if counter == 100 { + if counter == 1000 { select { case <-time.After(1 * time.Second): + case <-queueDone: + return case <-done: return } @@ -400,9 +433,36 @@ func (sq *prioritySnapshotQueue) scanHolderWorker(h *Holder, background chan sna // No reason to be active if we're not finding anything. select { case <-time.After(60 * time.Second): + case <-queueDone: + return case <-done: return } } + sq.logger.Debugf("observedOpN by power of 2: %d\n", sq.observedOpN[:]) + total := 0 + for _, v := range sq.observedOpN { + total += v + } + target := total / 4 + subTotal := 0 + for i, v := range sq.observedOpN { + subTotal += v + if subTotal >= target { + prevMaxOpN := sq.maxOpN + sq.maxOpN = (1 << (15 - uint(i))) / 2 + if sq.maxOpN > 0 { + sq.maxOpN-- + } + if prevMaxOpN != sq.maxOpN { + sq.logger.Printf("background scan: %d/%d fragments considered have opN %d or higher\n", + subTotal, total, sq.maxOpN) + } + break + } + } + for i := range sq.observedOpN { + sq.observedOpN[i] = 0 + } } } diff --git a/test/field.go b/test/field.go index b95e4f88b..f56672834 100644 --- a/test/field.go +++ b/test/field.go @@ -33,7 +33,7 @@ func newField(opts pilosa.FieldOption) *Field { if err != nil { panic(err) } - field, err := pilosa.NewField(path, "i", "f", opts) + field, err := pilosa.NewField(pilosa.NewHolder(pilosa.DefaultPartitionN), path, "i", "f", opts) if err != nil { panic(err) } @@ -63,7 +63,7 @@ func (f *Field) reopen() error { } path, index, name := f.Path(), f.Index(), f.Name() - f.Field, err = pilosa.NewField(path, index, name, pilosa.OptFieldTypeDefault()) + f.Field, err = pilosa.NewField(pilosa.NewHolder(pilosa.DefaultPartitionN), path, index, name, pilosa.OptFieldTypeDefault()) if err != nil { return err } diff --git a/test/index.go b/test/index.go index bf65c6ae2..b90702ae2 100644 --- a/test/index.go +++ b/test/index.go @@ -32,7 +32,7 @@ func newIndex() *Index { if err != nil { panic(err) } - index, err := pilosa.NewIndex(path, "i", pilosa.DefaultPartitionN) + index, err := pilosa.NewIndex(pilosa.NewHolder(pilosa.DefaultPartitionN), path, "i") if err != nil { panic(err) } @@ -62,7 +62,7 @@ func (i *Index) Reopen() error { } path, name := i.Path(), i.Name() - i.Index, err = pilosa.NewIndex(path, name, pilosa.DefaultPartitionN) + i.Index, err = pilosa.NewIndex(pilosa.NewHolder(pilosa.DefaultPartitionN), path, name) if err != nil { return err } diff --git a/view.go b/view.go index 478725e1e..af73b7fe2 100644 --- a/view.go +++ b/view.go @@ -26,7 +26,6 @@ import ( "sync/atomic" "time" - "github.com/pilosa/pilosa/v2/logger" "github.com/pilosa/pilosa/v2/pql" "github.com/pilosa/pilosa/v2/roaring" "github.com/pilosa/pilosa/v2/stats" @@ -49,6 +48,8 @@ type view struct { field string name string + holder *Holder + fieldType string cacheType string cacheSize uint32 @@ -56,24 +57,24 @@ type view struct { // Fragments by shard. fragments map[uint64]*fragment - broadcaster broadcaster - stats stats.StatsClient - rowAttrStore AttrStore - logger logger.Logger - snapshotQueue snapshotQueue + broadcaster broadcaster + stats stats.StatsClient + rowAttrStore AttrStore knownShards *roaring.Bitmap knownShardsCopied uint32 } // newView returns a new instance of View. -func newView(path, index, field, name string, fieldOptions FieldOptions) *view { +func newView(holder *Holder, path, index, field, name string, fieldOptions FieldOptions) *view { return &view{ path: path, index: index, field: field, name: name, + holder: holder, + fieldType: fieldOptions.Type, cacheType: fieldOptions.CacheType, cacheSize: fieldOptions.CacheSize, @@ -82,7 +83,6 @@ func newView(path, index, field, name string, fieldOptions FieldOptions) *view { broadcaster: NopBroadcaster, stats: stats.NopStatsClient, - logger: logger.NopLogger, knownShards: roaring.NewSliceBitmap(), } } @@ -133,14 +133,14 @@ func (v *view) open() error { if err := func() error { // Ensure the view's path exists. - v.logger.Debugf("ensure view path exists: %s", v.path) + v.holder.Logger.Debugf("ensure view path exists: %s", v.path) if err := os.MkdirAll(v.path, 0777); err != nil { return errors.Wrap(err, "creating view directory") } else if err := os.MkdirAll(filepath.Join(v.path, "fragments"), 0777); err != nil { return errors.Wrap(err, "creating fragments directory") } - v.logger.Debugf("open fragments for index/field/view: %s/%s/%s", v.index, v.field, v.name) + v.holder.Logger.Debugf("open fragments for index/field/view: %s/%s/%s", v.index, v.field, v.name) if err := v.openFragments(); err != nil { return errors.Wrap(err, "opening fragments") } @@ -151,7 +151,7 @@ func (v *view) open() error { return err } - v.logger.Debugf("successfully opened index/field/view: %s/%s/%s", v.index, v.field, v.name) + v.holder.Logger.Debugf("successfully opened index/field/view: %s/%s/%s", v.index, v.field, v.name) return nil } @@ -190,12 +190,12 @@ fileLoop: // Parse filename into integer. shard, err := strconv.ParseUint(filepath.Base(fi.Name()), 10, 64) if err != nil { - v.logger.Debugf("WARNING: couldn't use non-integer file as shard in index/field/view %s/%s/%s: %s", v.index, v.field, v.name, fi.Name()) + v.holder.Logger.Debugf("WARNING: couldn't use non-integer file as shard in index/field/view %s/%s/%s: %s", v.index, v.field, v.name, fi.Name()) continue } workQueue <- struct{}{} - v.logger.Debugf("open index/field/view/fragment: %s/%s/%s/%d", v.index, v.field, v.name, shard) + v.holder.Logger.Debugf("open index/field/view/fragment: %s/%s/%s/%d", v.index, v.field, v.name, shard) eg.Go(func() error { defer func() { <-workQueue @@ -205,7 +205,7 @@ fileLoop: return fmt.Errorf("open fragment: shard=%d, err=%s", frag.shard, err) } frag.RowAttrStore = v.rowAttrStore - v.logger.Debugf("add index/field/view/fragment to view.fragments: %s/%s/%s/%d", v.index, v.field, v.name, shard) + v.holder.Logger.Debugf("add index/field/view/fragment to view.fragments: %s/%s/%s/%d", v.index, v.field, v.name, shard) mu.Lock() v.fragments[frag.shard] = frag v.addKnownShard(frag.shard) @@ -342,7 +342,7 @@ func (v *view) notifyIfNewShard(shard uint64) { // Broadcast a message that a new max shard was just created. err := v.broadcaster.SendSync(msg) if err != nil { - v.logger.Printf("broadcasting create shard: %v", err) + v.holder.Logger.Printf("broadcasting create shard: %v", err) } close(broadcastChan) }() @@ -352,19 +352,15 @@ func (v *view) notifyIfNewShard(shard uint64) { select { case <-broadcastChan: case <-time.After(50 * time.Millisecond): - v.logger.Debugf("broadcasting create shard took >50ms") + v.holder.Logger.Debugf("broadcasting create shard took >50ms") } } func (v *view) newFragment(path string, shard uint64) *fragment { - frag := newFragment(path, v.index, v.field, v.name, shard, v.flags()) + frag := newFragment(v.holder, path, v.index, v.field, v.name, shard, v.flags()) frag.CacheType = v.cacheType frag.CacheSize = v.cacheSize - frag.Logger = v.logger frag.stats = v.stats - if v.snapshotQueue != nil { - frag.snapshotQueue = v.snapshotQueue - } if v.fieldType == FieldTypeMutex { frag.mutexVector = newRowsVector(frag) } else if v.fieldType == FieldTypeBool { @@ -382,7 +378,7 @@ func (v *view) deleteFragment(shard uint64) error { return ErrFragmentNotFound } - v.logger.Printf("delete fragment: (%s/%s/%s) %d", v.index, v.field, v.name, shard) + v.holder.Logger.Printf("delete fragment: (%s/%s/%s) %d", v.index, v.field, v.name, shard) // Close data files before deletion. if err := fragment.Close(); err != nil { @@ -396,7 +392,7 @@ func (v *view) deleteFragment(shard uint64) error { // Delete fragment cache file. if err := os.Remove(fragment.cachePath()); err != nil { - v.logger.Printf("no cache file to delete for shard %d", shard) + v.holder.Logger.Printf("no cache file to delete for shard %d", shard) } delete(v.fragments, shard) diff --git a/view_internal_test.go b/view_internal_test.go index bb50003d2..bbb1a20f4 100644 --- a/view_internal_test.go +++ b/view_internal_test.go @@ -34,7 +34,7 @@ func mustOpenView(index, field, name string) *view { CacheSize: DefaultCacheSize, } - v := newView(path, index, field, name, fo) + v := newView(NewHolder(DefaultPartitionN), path, index, field, name, fo) if err := v.open(); err != nil { panic(err) } From 7e7051d38742c639f9dd0c0aa4e12fa9ab301101 Mon Sep 17 00:00:00 2001 From: Seebs Date: Tue, 12 May 2020 15:40:38 -0500 Subject: [PATCH 3/8] don't try to truncate files when invoked read-only --- fragment.go | 8 +++++++- holder.go | 5 +++++ 2 files changed, 12 insertions(+), 1 deletion(-) diff --git a/fragment.go b/fragment.go index f70e9e71c..9aeceb79a 100644 --- a/fragment.go +++ b/fragment.go @@ -223,6 +223,9 @@ func (f *fragment) Open() error { // get no data. It tries to write the current storage to the provided file, // which is assumed to be the file they didn't get any data from. func (f *fragment) emptyStorage(file *os.File) (bool, error) { + if f.holder.ReadOnly { + return false, errors.New("can't flush/create storage for read-only holder") + } // No data. We'll mark this for no mapping, clear any existing // mapped containers, and set the Source to nil. We also have no // ops. @@ -276,7 +279,10 @@ func (f *fragment) importStorage(data []byte, file *os.File, newGen generation, } f.holder.Logger.Printf("warning: unmarshal storage, file=%s, err=%v", file.Name(), err) trunc, ok := cause.(roaring.FileShouldBeTruncatedError) - if ok { + if ok && !f.holder.ReadOnly { + // if the holder is ReadOnly, we silently ignore the "advisory" + // error. This may be a bad idea. + // generation code looks for a FileShouldBeTruncatedError return false, trunc } diff --git a/holder.go b/holder.go index d06920cd3..90df16860 100644 --- a/holder.go +++ b/holder.go @@ -101,6 +101,11 @@ type Holder struct { // needs to be queued and completed after all indexes // have opened. opening bool + + // ReadOnly indicates that this holder's contents should not produce + // disk writes under any circumstances. It must be set before Open + // is called, and changing it is not supported. + ReadOnly bool } func (h *Holder) StartTransaction(ctx context.Context, id string, timeout time.Duration, exclusive bool) (*Transaction, error) { From 3e7f0b32e9f4a17cabb16432415196b91cf858c1 Mon Sep 17 00:00:00 2001 From: Seebs Date: Wed, 8 Apr 2020 13:43:46 -0500 Subject: [PATCH 4/8] drop old Call data while processing a list of calls We don't need the Calls anymore, and especially Precomputed calls (like Distinct) could be a significant memory load that's increased as we process additional calls, so we drop the Precomputed references. We can't drop the calls entirely -- translation can require lookups of call arguments. --- executor.go | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-) diff --git a/executor.go b/executor.go index 6891072ee..57bc2abca 100644 --- a/executor.go +++ b/executor.go @@ -381,6 +381,16 @@ func (e *executor) handlePreCalls(ctx context.Context, index string, c *pql.Call return nil } +// dumpPrecomputedCalls throws away precomputed call data. this is used so we +// can drop any large data associated with a call once we've processed +// the call. +func (e *executor) dumpPrecomputedCalls(ctx context.Context, c *pql.Call) { + for _, call := range c.Children { + e.dumpPrecomputedCalls(ctx, call) + } + c.Precomputed = nil +} + // handlePreCallChildren handles any pre-calls in the children of a given call. func (e *executor) handlePreCallChildren(ctx context.Context, index string, c *pql.Call, shards []uint64, opt *execOptions) error { for i := range c.Children { @@ -433,7 +443,7 @@ func (e *executor) execute(ctx context.Context, index string, q *pql.Query, shar // Execute each call serially. results := make([]interface{}, 0, len(q.Calls)) - for _, call := range q.Calls { + for i, call := range q.Calls { if err := validateQueryContext(ctx); err != nil { return nil, err } @@ -463,6 +473,11 @@ func (e *executor) execute(ctx context.Context, index string, q *pql.Query, shar return nil, err } results = append(results, v) + // Some Calls can have significant data associated with them + // that gets generated during processing, such as Precomputed + // values. Dumping the precomputed data, if any, lets the GC + // free the memory before we get there. + e.dumpPrecomputedCalls(ctx, q.Calls[i]) } return results, nil } From 176d49e4b514d6f4e5abe4e80a1c0bdd75bd68b1 Mon Sep 17 00:00:00 2001 From: Seebs Date: Fri, 22 May 2020 13:28:57 -0500 Subject: [PATCH 5/8] use syswrap to close file after opening it with syswrap --- generation.go | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/generation.go b/generation.go index 7c1361b1d..c22b73867 100644 --- a/generation.go +++ b/generation.go @@ -263,7 +263,7 @@ func (m *mmapGeneration) openFile() (shouldClose bool, err error) { } // do we actually want this in every openFile? I don't know. if err := syscall.Flock(int(m.file.Fd()), syscall.LOCK_EX|syscall.LOCK_NB); err != nil { - m.file.Close() + _ = syswrap.CloseFile(m.file) m.file = nil return false, fmt.Errorf("flock: %s", err) } From 2826ecc0b7c5db0705d268369455df514deca532 Mon Sep 17 00:00:00 2001 From: Seebs Date: Tue, 12 May 2020 15:38:11 -0500 Subject: [PATCH 6/8] track retries correctly in truncation case --- generation.go | 1 + 1 file changed, 1 insertion(+) diff --git a/generation.go b/generation.go index c22b73867..f3cc6ae88 100644 --- a/generation.go +++ b/generation.go @@ -333,6 +333,7 @@ func newGeneration(existing generation, path string, readData bool, setup func([ m := mmapGeneration{path: path, logger: logger} if existing != nil { m.generation = existing.Generation() + 1 + m.retries = existing.(*mmapGeneration).retries // we might keep a previous generation around just for its generation count. if !existing.Dead() { defer existing.Done() From 4d494f6699ab3202e8eee5022953cca5d7ab792d Mon Sep 17 00:00:00 2001 From: Seebs Date: Tue, 26 May 2020 16:39:04 -0500 Subject: [PATCH 7/8] shared/generic functionality for iterating holders This is sort of large, but it's annoyingly difficult to separate out. The basic idea is to allow us to have a single holder-iterating block of code, which is associated with the holder, that can be used for various things, like the snapshot queue background scan, or for inspect operations. We invent the concept of a HolderFilter, which is a thing that can decide what things in a holder it cares about, and a HolderOperator, which can also process those things selectively. In the process, we fix up a couple of subtle bugs in the inspect logic; specifically, the assumption that the mapped flag could tell you whether a container was modified by the ops log doesn't work with mmap, so we have a shiny new flag which is used to track that, internal to the roaring/container code. All of this leads to the actual *point* of this exercise, which is making it easier to create an /inspect endpoint which produces almost the same data we'd have gotten from `pilosa inspect` on a data directory; the distinction is that it doesn't try to identify the distinction between data from disk and data from operations since the file was loaded. Possibly it should, but it doesn't yet. The snapshot queue is now implemented using the HolderOperator design, which requires some subtle changes to how it works, but overall makes it easier to follow the snapshot queue logic, and also shares that logic with the way Inspect works. The holder's snapshot queue is now provided by the server, in a default environment. The queueless snapshot queue no longer triggers snapshots on enqueue -- it turns out that breaks badly, because a key point about enqueueing a snapshot is that it's safe to do it *during* a transaction on that fragment, and triggering a snapshot during a transaction actually causes horrible errors as the ops log ends up being the old file, which we close. Related to this, we also need to prevent closed fragments from trying to snapshot, so we track fragment openness when opening or closing, and bail on trying to snapshot a fragment which is closed. We also stop using the queueless snapshot queue during tests, because that's a horrible idea. We copy a little bit of the partition logic from the cluster code so we don't have to expose it all, this lets us check whether the node we're looking at is the one which should be primary for a given shard, and if not, identify which node would be. This works only when pointed at a data directory, for now. The test cases for the holder have to be internal, because pilosa doesn't export view/fragment, just Index/Field. This means that the holder test cases can't just use the test/* package, so they duplicate some of its logic, approximately. --- api.go | 4 + cmd/inspect.go | 4 + ctl/inspect.go | 221 ++++++++++++++++-- fragment.go | 50 ++++- fragment_internal_test.go | 9 +- handler.go | 28 +++ holder.go | 413 ++++++++++++++++++++++++++++++++++ holder_internal_test.go | 430 ++++++++++++------------------------ http/handler.go | 34 +++ roaring/container_stash.go | 44 ++-- roaring/roaring.go | 32 ++- roaring/roaring_test.go | 2 +- roaring/unmarshal_binary.go | 93 ++++++-- server.go | 1 + snapshotqueue.go | 366 ++++++++++++++++-------------- 15 files changed, 1206 insertions(+), 525 deletions(-) diff --git a/api.go b/api.go index 8b6f3249d..23c36cb0b 100644 --- a/api.go +++ b/api.go @@ -1515,6 +1515,10 @@ func (api *API) Info() serverInfo { } } +func (api *API) Inspect(ctx context.Context, req *InspectRequest) (*HolderInfo, error) { + return api.holder.Inspect(ctx, req) +} + // GetTranslateEntryReader provides an entry reader for key translation logs starting at offset. func (api *API) GetTranslateEntryReader(ctx context.Context, offsets TranslateOffsetMap) (_ TranslateEntryReader, err error) { span, ctx := tracing.StartSpanFromContext(ctx, "API.GetTranslateEntryReader") diff --git a/cmd/inspect.go b/cmd/inspect.go index 47a40cae8..86452905d 100644 --- a/cmd/inspect.go +++ b/cmd/inspect.go @@ -48,5 +48,9 @@ Inspects a data file and provides stats. flags := inspectCmd.Flags() flags.BoolVarP(&inspector.Quiet, "quiet", "q", false, "don't list details of containers") flags.IntVarP(&inspector.Max, "max", "n", 0, "list at most max items (0 = unlimited)") + flags.StringVarP(&inspector.InspectOpts.Indexes, "index", "i", "", "filter indexes") + flags.StringVarP(&inspector.InspectOpts.Views, "view", "v", "", "filter views") + flags.StringVarP(&inspector.InspectOpts.Fields, "field", "f", "", "filter fields") + flags.StringVarP(&inspector.InspectOpts.Shards, "shard", "s", "", "filter shards") return inspectCmd } diff --git a/ctl/inspect.go b/ctl/inspect.go index 11471caa9..eb24a0597 100644 --- a/ctl/inspect.go +++ b/ctl/inspect.go @@ -16,15 +16,23 @@ package ctl import ( "context" + "encoding/binary" "fmt" + "hash/fnv" "io" + "io/ioutil" "os" + "path/filepath" + "sort" + "strconv" + "strings" "syscall" "text/tabwriter" "time" - "unsafe" + "github.com/gogo/protobuf/proto" "github.com/pilosa/pilosa/v2" + "github.com/pilosa/pilosa/v2/internal" "github.com/pilosa/pilosa/v2/roaring" "github.com/pkg/errors" ) @@ -37,6 +45,8 @@ type InspectCommand struct { Quiet bool // list only this many objects Max int + // Filters: + InspectOpts pilosa.InspectRequest // Standard input/output *pilosa.CmdIO @@ -68,12 +78,12 @@ func (p *pointerContext) pretty(c roaring.ContainerInfo) string { } func (cmd *InspectCommand) PrintOps(info roaring.BitmapInfo) { - fmt.Fprintln(cmd.Stdout, "== Ops ==") + fmt.Fprintln(cmd.Stdout, " Ops:") tw := tabwriter.NewWriter(cmd.Stdout, 0, 8, 0, '\t', 0) - fmt.Fprintf(tw, "%s\t%s\t%s\t\n", "TYPE", "OpN", "SIZE") + fmt.Fprintf(tw, " \t%s\t%s\t%s\t\n", "TYPE", "OpN", "SIZE") printed := 0 for _, op := range info.OpDetails { - fmt.Fprintf(tw, "%s\t%d\t%d\t\n", op.Type, op.OpN, op.Size) + fmt.Fprintf(tw, "\t%s\t%d\t%d\t\n", op.Type, op.OpN, op.Size) printed++ if cmd.Max != 0 && printed >= cmd.Max { break @@ -83,10 +93,10 @@ func (cmd *InspectCommand) PrintOps(info roaring.BitmapInfo) { } func (cmd *InspectCommand) PrintContainers(info roaring.BitmapInfo, pC pointerContext) { - fmt.Fprintln(cmd.Stdout, "== Containers ==") + fmt.Fprintln(cmd.Stdout, " Containers:") tw := tabwriter.NewWriter(cmd.Stdout, 0, 8, 0, '\t', 0) - fmt.Fprintf(tw, "\tRoaring\t\t\t\tOps\t\t\t\t\n") - fmt.Fprintf(tw, "%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t\n", "KEY", "TYPE", "N", "ALLOC", "OFFSET", "TYPE", "N", "ALLOC", "OFFSET") + fmt.Fprintf(tw, " \t\tRoaring\t\t\t\tOps\t\t\t\tFlags\t\n") + fmt.Fprintf(tw, "\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t%s\t\n", "KEY", "TYPE", "N", "ALLOC", "OFFSET", "TYPE", "N", "ALLOC", "OFFSET", "FLAGS") c1s := info.Containers c2s := info.OpContainers l1 := len(c1s) @@ -115,10 +125,14 @@ func (cmd *InspectCommand) PrintContainers(info roaring.BitmapInfo, pC pointerCo var key uint64 c1fmt := "-\t\t\t" c2fmt := "-\t\t\t" + // If c2 exists, we'll always prefer its flags, + // if it doesn't, this gets overwritten. + flags := c2.Flags if !c2e || (c1e && c1.Key < c2.Key) { c1fmt = pC.pretty(c1) key = c1.Key c1used = true + flags = c1.Flags } else if !c1e || (c2e && c2.Key < c1.Key) { c2fmt = pC.pretty(c2) key = c2.Key @@ -147,7 +161,7 @@ func (cmd *InspectCommand) PrintContainers(info roaring.BitmapInfo, pC pointerCo c2e = false } } - fmt.Fprintf(tw, "%d\t%s\t%s\t\n", key, c1fmt, c2fmt) + fmt.Fprintf(tw, "\t%d\t%s\t%s\t%s\t\n", key, c1fmt, c2fmt, flags) printed++ if cmd.Max > 0 && printed >= cmd.Max { break @@ -157,7 +171,7 @@ func (cmd *InspectCommand) PrintContainers(info roaring.BitmapInfo, pC pointerCo } // Run executes the inspect command. -func (cmd *InspectCommand) Run(_ context.Context) error { +func (cmd *InspectCommand) Run(ctx context.Context) error { // Open file handle. f, err := os.Open(cmd.Path) if err != nil { @@ -169,7 +183,173 @@ func (cmd *InspectCommand) Run(_ context.Context) error { if err != nil { return errors.Wrap(err, "statting file") } + if fi.IsDir() { + total := 0 + infos, err := f.Readdir(0) + if err != nil { + return err + } + if len(infos) == 0 { + return errors.New("directory contains no files") + } + names := make([]string, len(infos)) + nameToInfo := make(map[string]os.FileInfo, len(infos)) + // find numeric-only names; we'll operate on + // either those, or the whole holder if we find + // a .topology file. + n := 0 + for _, fi := range infos { + name := fi.Name() + if name == ".topology" { + return cmd.InspectHolder(ctx, cmd.Path) + } + if _, err := strconv.Atoi(name); err == nil { + names[n] = name + nameToInfo[name] = fi + n++ + } + } + if n == 0 { + return fmt.Errorf("directory contains no fragments (looking for numeric names)") + } + names = names[:n] + fmt.Fprintf(cmd.Stdout, "%s contains %d fragments:\n", cmd.Path, n) + for _, name := range names { + f2, err := os.Open(filepath.Join(cmd.Path, name)) + if err != nil { + return fmt.Errorf("opening %q: %v", name, err) + } + fmt.Fprintf(cmd.Stdout, "%s/%s:\n", cmd.Path, name) + err = cmd.InspectFile(f2, nameToInfo[name]) + total++ + f2.Close() + if err != nil { + return fmt.Errorf("inspecting %q: %v", name, err) + } + } + return nil + } + return cmd.InspectFile(f, fi) +} + +// loadTopology is copied almost exactly from pilosa/cluster.go. +func loadTopology(path string) (topology internal.Topology, myID string, err error) { + buf, err := ioutil.ReadFile(filepath.Join(path, ".topology")) + if os.IsNotExist(err) { + return topology, myID, err + } else if err != nil { + return topology, myID, errors.Wrap(err, "reading file") + } + if err := proto.Unmarshal(buf, &topology); err != nil { + return topology, myID, errors.Wrap(err, "unmarshalling") + } + sort.Slice(topology.NodeIDs, + func(i, j int) bool { + return topology.NodeIDs[i] < topology.NodeIDs[j] + }) + buf, err = ioutil.ReadFile(filepath.Join(path, ".id")) + if os.IsNotExist(err) { + return topology, myID, err + } else if err != nil { + return topology, myID, nil + } + myID = strings.TrimSpace(string(buf)) + return topology, myID, nil +} + +var partitions = make(map[string]map[uint64]int) + +func findPartition(index string, shard uint64, partitionN int) (partition int) { + var shardMap map[uint64]int + var ok bool + if shardMap, ok = partitions[index]; !ok { + shardMap = make(map[uint64]int) + partitions[index] = shardMap + } + if partition, ok = shardMap[shard]; !ok { + var buf [8]byte + binary.BigEndian.PutUint64(buf[:], shard) + + // Hash the bytes and mod by partition count. + h := fnv.New64a() + _, _ = h.Write([]byte(index)) + _, _ = h.Write(buf[:]) + partition = int(h.Sum64() % uint64(partitionN)) + shardMap[shard] = partition + } + return partition +} + +func findPartitionPath(path string, partitionN int) (int, error) { + parts := strings.Split(path, "/") + shard, err := strconv.ParseUint(parts[len(parts)-1], 10, 64) + if err != nil { + return 0, err + } + return findPartition(parts[0], shard, partitionN), nil +} + +func (cmd *InspectCommand) InspectHolder(ctx context.Context, path string) error { + holder := pilosa.NewHolder(pilosa.DefaultPartitionN) + holder.Path = path + holder.Opts.Inspect = true + holder.Opts.ReadOnly = true + err := holder.Open() + if err != nil { + return fmt.Errorf("%s: holder open: %v", path, err) + } + holderInfo, err := holder.Inspect(ctx, &cmd.InspectOpts) + if err != nil { + return fmt.Errorf("%s: inspect: %v", path, err) + } + myPartition := 0 + topology, myID, err := loadTopology(path) + if err == nil { + fmt.Fprintf(cmd.Stdout, "Cluster ID: %q\n", topology.ClusterID) + if len(topology.NodeIDs) > 1 { + fmt.Fprintf(cmd.Stdout, "Cluster of %d nodes, this node %q\n", len(topology.NodeIDs), myID) + } else { + fmt.Fprintf(cmd.Stdout, "Cluster has only one node: %q\n", myID) + } + found := false + for i := range topology.NodeIDs { + if topology.NodeIDs[i] == myID { + found = true + myPartition = i + break + } + } + if !found { + fmt.Fprintf(cmd.Stdout, "Warning: node ID %q not found in topology (%q)\n", myID, topology.NodeIDs) + } + } else { + fmt.Fprintf(cmd.Stdout, "warning: reading topology failed: %v\n", err) + } + for _, name := range holderInfo.FragmentNames { + partition, err := findPartitionPath(name, len(topology.NodeIDs)) + if err != nil { + fmt.Fprintf(cmd.Stdout, "%s: [can't find partition: %v]\n", name, err) + } else { + if partition == myPartition { + fmt.Fprintf(cmd.Stdout, "%s:\n", name) + } else { + fmt.Fprintf(cmd.Stdout, "%s: [primary node %q]\n", name, topology.NodeIDs[partition]) + } + } + details := holderInfo.FragmentInfo[name] + cmd.DisplayInfo(details.BitmapInfo) + if details.BlockChecksums != nil { + fmt.Fprintf(cmd.Stdout, " Checksums [%d total]:\n", len(details.BlockChecksums)) + for _, block := range details.BlockChecksums { + fmt.Fprintf(cmd.Stdout, " %8d: %x\n", block.ID, block.Checksum) + } + } + } + return nil +} + +func (cmd *InspectCommand) InspectFile(f *os.File, fi os.FileInfo) error { // Memory map the file. data, err := syscall.Mmap(int(f.Fd()), 0, int(fi.Size()), syscall.PROT_READ, syscall.MAP_SHARED) if err != nil { @@ -184,22 +364,27 @@ func (cmd *InspectCommand) Run(_ context.Context) error { // Attach the mmap file to the bitmap. t := time.Now() fmt.Fprintf(cmd.Stderr, "inspecting bitmap...") - info, err := roaring.InspectBinary(data) + var info roaring.BitmapInfo + _, _, err = roaring.InspectBinary(data, true, &info) + fmt.Fprintf(cmd.Stderr, " (%s)\n", time.Since(t)) + cmd.DisplayInfo(info) if err != nil { return errors.Wrap(err, "inspecting") } - fmt.Fprintf(cmd.Stderr, " (%s)\n", time.Since(t)) + return nil +} +func (cmd *InspectCommand) DisplayInfo(info roaring.BitmapInfo) { pC := pointerContext{ - from: uintptr(unsafe.Pointer(&data[0])), + from: info.From, + to: info.To, } - pC.to = pC.from + uintptr(len(data)) // Print top-level info. - fmt.Fprintf(cmd.Stdout, "== Bitmap Info ==\n") - fmt.Fprintf(cmd.Stdout, "Bits: %d\n", info.BitCount) - fmt.Fprintf(cmd.Stdout, "Containers: %d (%d roaring)\n", info.ContainerCount, len(info.Containers)) - fmt.Fprintf(cmd.Stdout, "Operations: %d (%d bits)\n", info.Ops, info.OpN) + fmt.Fprintf(cmd.Stdout, " Bitmap Info:\n") + fmt.Fprintf(cmd.Stdout, " Bits: %d\n", info.BitCount) + fmt.Fprintf(cmd.Stdout, " Containers: %d (%d roaring)\n", info.ContainerCount, len(info.Containers)) + fmt.Fprintf(cmd.Stdout, " Operations: %d (%d bits)\n", info.Ops, info.OpN) fmt.Fprintln(cmd.Stdout, "") // Print info for each container. @@ -211,6 +396,4 @@ func (cmd *InspectCommand) Run(_ context.Context) error { cmd.PrintOps(info) } } - - return nil } diff --git a/fragment.go b/fragment.go index 9aeceb79a..61ac78313 100644 --- a/fragment.go +++ b/fragment.go @@ -124,6 +124,7 @@ type fragment struct { snapshotCond sync.Cond snapshotErr error // error yielded by the last snapshot operation snapshotStamp time.Time // timestamp of last snapshot + open bool // is this fragment actually open? // Cache for row counts. CacheType string // passed in by field @@ -156,6 +157,8 @@ type fragment struct { mutexVector vector stats stats.StatsClient + + bitmapInfo *roaring.BitmapInfo } // newFragment returns a new instance of Fragment. @@ -182,6 +185,23 @@ func newFragment(holder *Holder, path, index, field, view string, shard uint64, // cachePath returns the path to the fragment's cache data. func (f *fragment) cachePath() string { return f.path + cacheExt } +type FragmentInfo struct { + BitmapInfo roaring.BitmapInfo + BlockChecksums []FragmentBlock `json:"BlockChecksums,omitempty"` +} + +func (f *fragment) inspect(params InspectRequestParams) (fi FragmentInfo) { + if f.bitmapInfo == nil { + fi.BitmapInfo = f.storage.Info(params.Containers) + } else { + fi.BitmapInfo = *f.bitmapInfo + } + if params.Checksum { + fi.BlockChecksums = f.Blocks() + } + return fi +} + // Open opens the underlying storage. func (f *fragment) Open() error { f.mu.Lock() @@ -214,6 +234,7 @@ func (f *fragment) Open() error { f.close() return err } + f.open = true f.holder.Logger.Debugf("successfully opened index/field/view/fragment: %s/%s/%s/%d", f.index, f.field, f.view, f.shard) return nil @@ -223,7 +244,7 @@ func (f *fragment) Open() error { // get no data. It tries to write the current storage to the provided file, // which is assumed to be the file they didn't get any data from. func (f *fragment) emptyStorage(file *os.File) (bool, error) { - if f.holder.ReadOnly { + if f.holder.Opts.ReadOnly { return false, errors.New("can't flush/create storage for read-only holder") } // No data. We'll mark this for no mapping, clear any existing @@ -279,7 +300,7 @@ func (f *fragment) importStorage(data []byte, file *os.File, newGen generation, } f.holder.Logger.Printf("warning: unmarshal storage, file=%s, err=%v", file.Name(), err) trunc, ok := cause.(roaring.FileShouldBeTruncatedError) - if ok && !f.holder.ReadOnly { + if ok && !f.holder.Opts.ReadOnly { // if the holder is ReadOnly, we silently ignore the "advisory" // error. This may be a bad idea. @@ -342,6 +363,12 @@ func (f *fragment) applyStorage(data []byte, file *os.File, newGen generation, m return mapped, err } +func (f *fragment) inspectStorage(data []byte, file *os.File, newGen generation, mapped bool) (didMap bool, err error) { + f.bitmapInfo = &roaring.BitmapInfo{} + f.storage, didMap, err = roaring.InspectBinary(data, mapped, f.bitmapInfo) + return didMap, err +} + // openStorage opens the storage bitmap. // // This has been massively reworked recently, and now hands a lot of @@ -360,10 +387,17 @@ func (f *fragment) openStorage(unmarshalData bool) error { } f.rowCache = &simpleCache{make(map[uint64]*Row)} var storageOp func([]byte, *os.File, generation, bool) (bool, error) - if unmarshalData { - storageOp = f.importStorage + if f.holder.Opts.Inspect { + // note that this will unmarshal even if we already have + // storage; when Inspect is on for a holder, we actually want + // to be able to report this. + storageOp = f.inspectStorage } else { - storageOp = f.applyStorage + if unmarshalData { + storageOp = f.importStorage + } else { + storageOp = f.applyStorage + } } var err error f.gen, err = newGeneration(f.gen, f.path, unmarshalData, storageOp, f.holder.Logger) @@ -436,6 +470,9 @@ func (f *fragment) Close() error { for f.snapshotPending { f.snapshotCond.Wait() } + // Note: snapshots won't progress on a closed fragment, so we + // wait until after a possible pending snapshot to close. + f.open = false return f.close() } @@ -2283,6 +2320,9 @@ func track(start time.Time, message string, stats stats.StatsClient, logger logg // snapshot does the actual snapshot operation. it does not check or care // about f.snapshotPending. func (f *fragment) snapshot() (err error) { + if !f.open { + return errors.New("snapshot request on closed fragment") + } wouldPanic := debug.SetPanicOnFault(true) defer func() { debug.SetPanicOnFault(wouldPanic) diff --git a/fragment_internal_test.go b/fragment_internal_test.go index 2d8c27b73..fa1ffbeec 100644 --- a/fragment_internal_test.go +++ b/fragment_internal_test.go @@ -2666,6 +2666,12 @@ func mustOpenBSIFragment(index, field, view string, shard uint64) *fragment { return mustOpenFragmentFlags(index, field, view, shard, "", 1) } +var testHolder = NewHolder(DefaultPartitionN) + +func init() { + testHolder.SnapshotQueue = newSnapshotQueue(1, 1, nil) +} + // mustOpenFragment returns a new instance of Fragment with a temporary path. func mustOpenFragmentFlags(index, field, view string, shard uint64, cacheType string, flags byte) *fragment { file, err := ioutil.TempFile(*TempDir, "pilosa-fragment-") @@ -2678,7 +2684,8 @@ func mustOpenFragmentFlags(index, field, view string, shard uint64, cacheType st cacheType = DefaultCacheType } - f := newFragment(NewHolder(DefaultPartitionN), file.Name(), index, field, view, shard, flags) + f := newFragment(testHolder, file.Name(), index, field, view, shard, flags) + f.CacheType = cacheType f.RowAttrStore = &memAttrStore{ store: make(map[uint64]map[string]interface{}), diff --git a/handler.go b/handler.go index 17156245b..bb5aa245d 100644 --- a/handler.go +++ b/handler.go @@ -287,3 +287,31 @@ type TranslateIDsRequest struct { type TranslateIDsResponse struct { Keys []string } + +// InspectRequestParams represents the parts of an InspectRequest that +// aren't generic holder filtering attributes. +type InspectRequestParams struct { + Containers bool // include container details + Checksum bool // perform checksums +} + +// InspectRequest represents a request for a possibly-partial +// holder inspection, using a provided holder filter and inspect-specific +// parameters. +type InspectRequest struct { + HolderFilterParams + InspectRequestParams +} + +// InspectResponse contains the structured results for an InspectRequest. +// It may some day be expanded to include metadata about views or indexes. +type InspectResponse struct { + Fragments []struct { + Index string + Field string + View string + Shard int64 + Path string + Info *FragmentInfo + } +} diff --git a/holder.go b/holder.go index 90df16860..f8ac267f9 100644 --- a/holder.go +++ b/holder.go @@ -21,7 +21,9 @@ import ( "os" "path" "path/filepath" + "regexp" "sort" + "strconv" "strings" "sync" "syscall" @@ -102,10 +104,17 @@ type Holder struct { // have opened. opening bool + Opts HolderOpts +} + +type HolderOpts struct { // ReadOnly indicates that this holder's contents should not produce // disk writes under any circumstances. It must be set before Open // is called, and changing it is not supported. ReadOnly bool + // If Inspect is set, we'll try to obtain additional information + // about fragments when opening them. + Inspect bool } func (h *Holder) StartTransaction(ctx context.Context, id string, timeout time.Duration, exclusive bool) (*Transaction, error) { @@ -178,6 +187,280 @@ func NewHolder(partitionN int) *Holder { } } +type HolderInfo struct { + FragmentInfo map[string]FragmentInfo + FragmentNames []string +} + +type regexpList []*regexp.Regexp + +func newRegexpList(regexes string) (results regexpList, err error) { + if regexes == "" { + return nil, nil + } + for _, sub := range strings.Split(regexes, ",") { + re, err := regexp.Compile(sub) + if err != nil { + return nil, err + } + results = append(results, re) + } + return results, nil +} + +func (rl regexpList) Match(haystack string) bool { + if rl == nil { + return true + } + for _, re := range rl { + if re.MatchString(haystack) { + return true + } + } + return false +} + +// shardRange represents a series of shards +type shardRange struct { + min, max uint64 +} + +type shardRangeList []shardRange + +func newShardRangeList(shards string) (results shardRangeList, err error) { + if shards == "" { + return nil, nil + } + for _, sub := range strings.Split(shards, ",") { + var sr shardRange + minMax := strings.Split(sub, "-") + if len(minMax) > 2 { + return nil, fmt.Errorf("invalid range %q", sub) + } + sr.min, err = strconv.ParseUint(minMax[0], 10, 64) + if err != nil { + return nil, err + } + sr.max = sr.min + if len(minMax) == 2 { + sr.max, err = strconv.ParseUint(minMax[0], 10, 64) + if err != nil { + return nil, err + } + } + if sr.max < sr.min { + return nil, fmt.Errorf("invalid range %q: max < min", sub) + } + results = append(results, sr) + } + return results, nil +} + +func (sl shardRangeList) Match(shard uint64) bool { + if sl == nil { + return true + } + for _, sr := range sl { + if shard >= sr.min && shard <= sr.max { + return true + } + } + return false +} + +// HolderFilter represents something that potentially filters out +// parts of a holder, indicating whether or not to process them, +// or recurse into them. It is permissible to recurse a thing +// without processing it, or process it without recursing it. +// For instance, something looking to accumulate statistics +// about views might return (true, false) from CheckView, +// while a fragment scanning operation would return (false, true) +// from everything above CheckFrag. +type HolderFilter interface { + CheckIndex(iname string) (process bool, recurse bool) + CheckField(iname, fname string) (process bool, recurse bool) + CheckView(iname, fname, vname string) (process bool, recurse bool) + CheckFragment(iname, fname, vname string, shard uint64) (process bool) +} + +// HolderFilterAll is a placeholder type which always returns true for the +// check functions. You can embed it to make a HolderOperator which processes +// everything. +type HolderFilterAll struct{} + +func (HolderFilterAll) CheckIndex(string) (bool, bool) { + return true, true +} + +func (HolderFilterAll) CheckField(string, string) (bool, bool) { + return true, true +} + +func (HolderFilterAll) CheckView(string, string, string) (bool, bool) { + return true, true +} + +func (HolderFilterAll) CheckFragment(string, string, string, uint64) bool { + return true +} + +// HolderProcessNone is a placeholder type which does nothing for the +// process functions. You can embed it to make a HolderOperator which +// does nothing, or embed it and provide your own ProcessFragment to +// do just that. +type HolderProcessNone struct{} + +func (HolderProcessNone) ProcessIndex(*Index) error { + return nil +} + +func (HolderProcessNone) ProcessField(*Field) error { + return nil +} + +func (HolderProcessNone) ProcessView(*view) error { + return nil +} + +func (HolderProcessNone) ProcessFragment(*fragment) error { + return nil +} + +// HolderProcess represents something that has operations which can be +// performed on indexes, fields, views, and/or fragments. +type HolderProcess interface { + ProcessIndex(*Index) error + ProcessField(*Field) error + ProcessView(*view) error + ProcessFragment(*fragment) error +} + +// HolderOperator is both a filter and a process. This is the general +// form of "I want to do something to some part of a holder." +type HolderOperator interface { + HolderFilter + HolderProcess +} + +var _ HolderOperator = (*holderInspector)(nil) + +type HolderFilterParams struct { + Indexes string + Fields string + Views string + Shards string +} + +type holderFilterFull struct { + HolderFilterParams + indexRegexps regexpList + fieldRegexps regexpList + viewRegexps regexpList + shardRanges shardRangeList +} + +type inspectRequestFull struct { + HolderFilter + params InspectRequestParams +} + +func (i *holderFilterFull) CheckIndex(iname string) (process, recurse bool) { + return true, i.indexRegexps.Match(iname) +} + +func (i *holderFilterFull) CheckField(iname, fname string) (process, recurse bool) { + return true, i.fieldRegexps.Match(fname) +} + +func (i *holderFilterFull) CheckView(iname, fname, vname string) (process, recurse bool) { + return true, i.viewRegexps.Match(vname) +} + +func (i *holderFilterFull) CheckFragment(iname, fname, vname string, shard uint64) (process bool) { + return i.shardRanges.Match(shard) +} + +func NewHolderFilter(params HolderFilterParams) (result HolderFilter, err error) { + filter := &holderFilterFull{ + HolderFilterParams: params, + } + filter.indexRegexps, err = newRegexpList(params.Indexes) + if err != nil { + return nil, err + } + filter.fieldRegexps, err = newRegexpList(params.Fields) + if err != nil { + return nil, err + } + filter.viewRegexps, err = newRegexpList(params.Views) + if err != nil { + return nil, err + } + filter.shardRanges, err = newShardRangeList(params.Shards) + if err != nil { + return nil, err + } + return filter, nil +} + +func expandInspectRequest(req *InspectRequest) (*inspectRequestFull, error) { + filter, err := NewHolderFilter(req.HolderFilterParams) + if err != nil { + return nil, err + } + irf := &inspectRequestFull{ + HolderFilter: filter, + params: req.InspectRequestParams, + } + return irf, nil +} + +type holderInspector struct { + *inspectRequestFull + pathParts [3]string + path string + hi *HolderInfo +} + +func (h *holderInspector) ProcessIndex(i *Index) error { + h.pathParts[0] = i.name + return nil +} + +func (h *holderInspector) ProcessField(f *Field) error { + h.pathParts[1] = f.name + return nil +} + +func (h *holderInspector) ProcessView(v *view) error { + h.pathParts[2] = v.name + h.path = strings.Join(h.pathParts[:], "/") + return nil +} + +func (h *holderInspector) ProcessFragment(f *fragment) error { + path := h.path + "/" + strconv.FormatUint(f.shard, 10) + h.hi.FragmentInfo[path] = f.inspect(h.inspectRequestFull.params) + h.hi.FragmentNames = append(h.hi.FragmentNames, path) + return nil +} + +func (h *Holder) Inspect(ctx context.Context, req *InspectRequest) (*HolderInfo, error) { + fullReq, err := expandInspectRequest(req) + if err != nil { + return nil, err + } + inspector := &holderInspector{ + inspectRequestFull: fullReq, + hi: &HolderInfo{ + FragmentInfo: make(map[string]FragmentInfo), + }, + } + err = h.Process(ctx, inspector) + sort.Strings(inspector.hi.FragmentNames) + return inspector.hi, err +} + // Open initializes the root data directory for the holder. func (h *Holder) Open() error { h.opening = true @@ -1369,3 +1652,133 @@ func uint64InSlice(i uint64, s []uint64) bool { } return false } + +// Process loops through a holder based on the Check functions in op, calling +// the Process functions in op when indicated. +func (h *Holder) Process(ctx context.Context, op HolderOperator) (err error) { + var indexNames, fieldNames, viewNames []string + var fragNums []uint64 + + h.mu.Lock() + indexNames = indexNames[:0] + for indexName := range h.indexes { + indexNames = append(indexNames, indexName) + } + h.mu.Unlock() + for _, indexName := range indexNames { + if err = ctx.Err(); err != nil { + return err + } + process, recurse := op.CheckIndex(indexName) + if !process && !recurse { + continue + } + h.mu.Lock() + index := h.indexes[indexName] + h.mu.Unlock() + if index == nil { + continue + } + if err = ctx.Err(); err != nil { + return err + } + if process { + err = op.ProcessIndex(index) + if err != nil { + return err + } + } + if !recurse { + continue + } + fieldNames = fieldNames[:0] + index.mu.Lock() + for fieldName := range index.fields { + fieldNames = append(fieldNames, fieldName) + } + index.mu.Unlock() + for _, fieldName := range fieldNames { + if err = ctx.Err(); err != nil { + return err + } + process, recurse := op.CheckField(indexName, fieldName) + if !process && !recurse { + continue + } + index.mu.Lock() + field := index.fields[fieldName] + index.mu.Unlock() + if field == nil { + continue + } + if err = ctx.Err(); err != nil { + return err + } + if process { + err = op.ProcessField(field) + if err != nil { + return err + } + } + if !recurse { + continue + } + viewNames = viewNames[:0] + field.mu.Lock() + for viewName := range field.viewMap { + viewNames = append(viewNames, viewName) + } + field.mu.Unlock() + for _, viewName := range viewNames { + if err = ctx.Err(); err != nil { + return err + } + process, recurse := op.CheckView(indexName, fieldName, viewName) + if !process && !recurse { + continue + } + field.mu.Lock() + view := field.viewMap[viewName] + field.mu.Unlock() + if view == nil { + continue + } + if err = ctx.Err(); err != nil { + return err + } + if process { + err = op.ProcessView(view) + if err != nil { + return err + } + } + if !recurse { + continue + } + fragNums := fragNums[:0] + view.mu.Lock() + for fragNum := range view.fragments { + fragNums = append(fragNums, fragNum) + } + view.mu.Unlock() + for _, fragNum := range fragNums { + if err = ctx.Err(); err != nil { + return err + } + process := op.CheckFragment(indexName, fieldName, viewName, fragNum) + if !process { + continue + } + view.mu.Lock() + frag := view.fragments[fragNum] + view.mu.Unlock() + err = op.ProcessFragment(frag) + if err != nil { + return err + } + } + } + } + } + return nil +} diff --git a/holder_internal_test.go b/holder_internal_test.go index ccc52c5b0..5ae365362 100644 --- a/holder_internal_test.go +++ b/holder_internal_test.go @@ -1,4 +1,4 @@ -// Copyright 2017 Pilosa Corp. +// Copyright 2020 Pilosa Corp. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. @@ -15,313 +15,163 @@ package pilosa import ( + "context" "io/ioutil" "os" - "path/filepath" - "reflect" - "strings" "testing" - "time" - - "github.com/pilosa/pilosa/v2/roaring" ) -type tHolder struct { - *Holder +type testHolderOperator struct { + indexSeen, indexProcessed int + fieldSeen, fieldProcessed int + viewSeen, viewProcessed int + fragmentSeen, fragmentProcessed int + waitHere chan struct{} } -// Close closes the holder and removes all underlying data. -func (h *tHolder) Close() error { - defer os.RemoveAll(h.Path) - return h.Holder.Close() +func (t *testHolderOperator) CheckIndex(string) (bool, bool) { + t.indexSeen++ + return true, true } -// Reopen instantiates and opens a new holder. -// Note that the holder must be Closed first. -func (h *tHolder) Reopen() error { - path, logger := h.Path, h.Holder.Logger - h.Holder = NewHolder(DefaultPartitionN) - h.Holder.Path = path - h.Holder.Logger = logger - return h.Holder.Open() +func (t *testHolderOperator) CheckField(string, string) (bool, bool) { + t.fieldSeen++ + return true, true } -func newHolder() *tHolder { - path, err := ioutil.TempDir(*TempDir, "pilosa-") +func (t *testHolderOperator) CheckView(string, string, string) (bool, bool) { + t.viewSeen++ + return true, true +} + +func (t *testHolderOperator) CheckFragment(string, string, string, uint64) bool { + t.fragmentSeen++ + return true +} + +func (t *testHolderOperator) ProcessIndex(*Index) error { + t.indexProcessed++ + return nil +} + +func (t *testHolderOperator) ProcessField(*Field) error { + t.fieldProcessed++ + return nil +} + +func (t *testHolderOperator) ProcessView(*view) error { + t.viewProcessed++ + return nil +} + +func (t *testHolderOperator) ProcessFragment(*fragment) error { + if t.waitHere != nil { + <-t.waitHere + } + t.fragmentProcessed++ + return nil +} + +func makeHolder() (*Holder, string, error) { + path, err := ioutil.TempDir("", "pilosa-") if err != nil { - panic(err) + return nil, "", err } - - h := &tHolder{Holder: NewHolder(DefaultPartitionN)} - h.Path = path - return h -} - -// MustCreateFieldIfNotExists returns a given field. Panic on error. -func (h *tHolder) MustCreateFieldIfNotExists(index, field string) *Field { - f, err := h.MustCreateIndexIfNotExists(index, IndexOptions{}).CreateFieldIfNotExists(field, OptFieldTypeDefault()) - if err != nil { - panic(err) - } - return f -} - -// MustCreateIndexIfNotExists returns a given index. Panic on error. -func (h *tHolder) MustCreateIndexIfNotExists(index string, opt IndexOptions) *Index { - idx, err := h.Holder.CreateIndexIfNotExists(index, opt) - if err != nil { - panic(err) - } - return idx -} - -// SetBit clears a bit on the given field. -func (h *tHolder) SetBit(index, field string, rowID, columnID uint64) { - f := h.MustCreateFieldIfNotExists(index, field) - _, err := f.SetBit(rowID, columnID, nil) - if err != nil { - panic(err) - } -} - -// Row returns a Row for a given field. -func (h *tHolder) Row(index, field string, rowID uint64) *Row { - f := h.MustCreateFieldIfNotExists(index, field) - row, err := f.Row(rowID) - if err != nil { - panic(err) - } - return row -} - -func TestHolder_Optn(t *testing.T) { - t.Run("ErrViewPermission", func(t *testing.T) { - if os.Geteuid() == 0 { - t.Skip("Skipping permissions test since user is root.") - } - availableShardFileFlushDuration.Set(100 * time.Millisecond) - h := newHolder() - defer h.Close() - - if idx, err := h.CreateIndex("foo", IndexOptions{}); err != nil { - t.Fatal(err) - } else if field, err := idx.CreateField("bar", OptFieldTypeDefault()); err != nil { - t.Fatal(err) - } else if _, err := field.createViewIfNotExists(viewStandard); err != nil { - t.Fatal(err) - } else if err := h.Holder.Close(); err != nil { - t.Fatal(err) - } else if err := os.Chmod(filepath.Join(h.Path, "foo", "bar", "views", "standard"), 0000); err != nil { - t.Fatal(err) - } - defer func() { - // we don't care about a failure here - _ = os.Chmod(filepath.Join(h.Path, "foo", "bar", "views", "standard"), 0755) - }() - if err := h.Reopen(); err == nil || !strings.Contains(err.Error(), "permission denied") { - t.Fatalf("unexpected error: %s", err) - } - }) - t.Run("ErrViewFragmentsMkdir", func(t *testing.T) { - if os.Geteuid() == 0 { - t.Skip("Skipping permissions test since user is root.") - } - h := newHolder() - defer h.Close() - - if idx, err := h.CreateIndex("foo", IndexOptions{}); err != nil { - t.Fatal(err) - } else if field, err := idx.CreateField("bar", OptFieldTypeDefault()); err != nil { - t.Fatal(err) - } else if _, err := field.createViewIfNotExists(viewStandard); err != nil { - t.Fatal(err) - } else if err := h.Holder.Close(); err != nil { - t.Fatal(err) - } else if err := os.Chmod(filepath.Join(h.Path, "foo", "bar", "views", "standard", "fragments"), 0000); err != nil { - t.Fatal(err) - } - defer func() { - // we don't care about a failure here - _ = os.Chmod(filepath.Join(h.Path, "foo", "bar", "views", "standard", "fragments"), 0755) - }() - - if err := h.Reopen(); err == nil || !strings.Contains(err.Error(), "permission denied") { - t.Fatalf("unexpected error: %s", err) - } - }) - - t.Run("ErrFragmentCachePermission", func(t *testing.T) { - if os.Geteuid() == 0 { - t.Skip("Skipping permissions test since user is root.") - } - h := newHolder() - defer h.Close() - - if idx, err := h.CreateIndex("foo", IndexOptions{}); err != nil { - t.Fatal(err) - } else if field, err := idx.CreateField("bar", OptFieldTypeDefault()); err != nil { - t.Fatal(err) - } else if view, err := field.createViewIfNotExists(viewStandard); err != nil { - t.Fatal(err) - } else if _, err := field.SetBit(0, 0, nil); err != nil { - t.Fatal(err) - } else if err := view.Fragment(0).FlushCache(); err != nil { - t.Fatal(err) - } else if err := h.Holder.Close(); err != nil { - t.Fatal(err) - } else if err := os.Chmod(filepath.Join(h.Path, "foo", "bar", "views", "standard", "fragments", "0.cache"), 0000); err != nil { - t.Fatal(err) - } - defer func() { - _ = os.Chmod(filepath.Join(h.Path, "foo", "bar", "views", "standard", "fragments", "0.cache"), 0644) - }() - if err := h.Reopen(); err == nil || !strings.Contains(err.Error(), "permission denied") { - t.Fatalf("unexpected error: %s", err) - } - }) - -} - -// Ensure holder can clean up orphaned fragments. -func TestHolderCleaner_CleanHolder(t *testing.T) { - availableShardFileFlushDuration.Set(100 * time.Millisecond) //shorten the default time to force a file write - cluster := NewTestCluster(2) - - // Create a local holder. - hldr0 := newHolder() - defer hldr0.Close() - - // Mock 2-node, fully replicated cluster. - cluster.ReplicaN = 2 - - cluster.nodes[0].URI = NewTestURIFromHostPort("localhost", 0) - - // Create fields on nodes. - for _, hldr := range []*tHolder{hldr0} { - hldr.MustCreateFieldIfNotExists("i", "f") - hldr.MustCreateFieldIfNotExists("i", "f0") - hldr.MustCreateFieldIfNotExists("y", "z") - } - - // Set data on the local holder. - hldr0.SetBit("i", "f", 0, 10) - hldr0.SetBit("i", "f", 0, 4000) - hldr0.SetBit("i", "f", 2, 20) - hldr0.SetBit("i", "f", 3, 10) - hldr0.SetBit("i", "f", 120, 10) - hldr0.SetBit("i", "f", 200, 4) - - hldr0.SetBit("i", "f0", 9, ShardWidth+5) - - hldr0.SetBit("y", "z", 10, (2*ShardWidth)+4) - hldr0.SetBit("y", "z", 10, (2*ShardWidth)+5) - hldr0.SetBit("y", "z", 10, (2*ShardWidth)+7) - - // Set highest shard. - err := hldr0.Field("i", "f").AddRemoteAvailableShards(roaring.NewBitmap(0, 1)) - if err != nil { - t.Fatalf("adding remote shards: %v", err) - } - err = hldr0.Field("y", "z").AddRemoteAvailableShards(roaring.NewBitmap(0, 1, 2)) - if err != nil { - t.Fatalf("adding remote shards: %v", err) - } - time.Sleep(2 * availableShardFileFlushDuration.Get()) - - // Keep replication the same and ensure we get the expected results. - cluster.ReplicaN = 2 - - // Set up cleaner for replication 2. - cleaner2 := holderCleaner{ - Node: cluster.nodes[0], - Holder: hldr0.Holder, - Cluster: cluster, - } - - if err := cleaner2.CleanHolder(); err != nil { - t.Fatal(err) - } - - // Verify data is the same on both nodes. - for i, hldr := range []*tHolder{hldr0} { - if a := hldr.Row("i", "f", 0).Columns(); !reflect.DeepEqual(a, []uint64{10, 4000}) { - t.Fatalf("unexpected columns(%d/0): %+v", i, a) - } else if a := hldr.Row("i", "f", 2).Columns(); !reflect.DeepEqual(a, []uint64{20}) { - t.Fatalf("unexpected columns(%d/2): %+v", i, a) - } else if a := hldr.Row("i", "f", 3).Columns(); !reflect.DeepEqual(a, []uint64{10}) { - t.Fatalf("unexpected columns(%d/3): %+v", i, a) - } else if a := hldr.Row("i", "f", 120).Columns(); !reflect.DeepEqual(a, []uint64{10}) { - t.Fatalf("unexpected columns(%d/120): %+v", i, a) - } else if a := hldr.Row("i", "f", 200).Columns(); !reflect.DeepEqual(a, []uint64{4}) { - t.Fatalf("unexpected columns(%d/200): %+v", i, a) - } - - if a := hldr.Row("i", "f0", 9).Columns(); !reflect.DeepEqual(a, []uint64{ShardWidth + 5}) { - t.Fatalf("unexpected columns(%d/d/f0): %+v", i, a) - } - - if a := hldr.Row("y", "z", 10).Columns(); !reflect.DeepEqual(a, []uint64{(2 * ShardWidth) + 4, (2 * ShardWidth) + 5, (2 * ShardWidth) + 7}) { - t.Fatalf("unexpected columns(%d/y/z): %+v", i, a) - } - } - - // Change replication factor to ensure we have fragments to remove. - cluster.ReplicaN = 1 - - // Set up cleaner for replication 1. - cleaner1 := holderCleaner{ - Node: cluster.nodes[0], - Holder: hldr0.Holder, - Cluster: cluster, - } - - if err := cleaner1.CleanHolder(); err != nil { - t.Fatal(err) - } - - // Verify data is the same on both nodes. - for i, hldr := range []*tHolder{hldr0} { - if a := hldr.Row("i", "f", 0).Columns(); !reflect.DeepEqual(a, []uint64{10, 4000}) { - t.Fatalf("unexpected columns(%d/0): %+v", i, a) - } else if a := hldr.Row("i", "f", 2).Columns(); !reflect.DeepEqual(a, []uint64{20}) { - t.Fatalf("unexpected columns(%d/2): %+v", i, a) - } else if a := hldr.Row("i", "f", 3).Columns(); !reflect.DeepEqual(a, []uint64{10}) { - t.Fatalf("unexpected columns(%d/3): %+v", i, a) - } else if a := hldr.Row("i", "f", 120).Columns(); !reflect.DeepEqual(a, []uint64{10}) { - t.Fatalf("unexpected columns(%d/120): %+v", i, a) - } else if a := hldr.Row("i", "f", 200).Columns(); !reflect.DeepEqual(a, []uint64{4}) { - t.Fatalf("unexpected columns(%d/200): %+v", i, a) - } - - f := hldr.fragment("i", "f0", viewStandard, 1) - if f != nil { - t.Fatalf("expected fragment to be deleted: (%d/i/f0): %+v", i, f) - } - - if a := hldr.Row("y", "z", 10).Columns(); !reflect.DeepEqual(a, []uint64{(2 * ShardWidth) + 4, (2 * ShardWidth) + 5, (2 * ShardWidth) + 7}) { - t.Fatalf("unexpected columns(%d/y/z): %+v", i, a) - } - } -} - -// Ensure holder can reopen. -func TestHolderCleaner_Reopen(t *testing.T) { h := NewHolder(DefaultPartitionN) - h.Path = "path" - err := h.Open() + + return h, path, nil +} + +func testSetBit(t *testing.T, h *Holder, index, field string, rowID, columnID uint64) { + idx, err := h.CreateIndexIfNotExists(index, IndexOptions{}) if err != nil { - t.Fatalf("couldn't open holder: %v", err) + t.Fatalf("creating index: %v", err) } - err = h.Close() + f, err := idx.CreateFieldIfNotExists(field, OptFieldTypeDefault()) if err != nil { - t.Fatalf("couldn't close holder: %v", err) + t.Fatalf("setting bit: %v", err) } - err = h.Open() + _, err = f.SetBit(rowID, columnID, nil) if err != nil { - t.Fatalf("couldn't open holder: %v", err) - } - err = h.Close() - if err != nil { - t.Fatalf("couldn't close holder: %v", err) + t.Fatalf("setting bit: %v", err) + } +} + +func TestHolderOperatorProcess(t *testing.T) { + h, path, err := makeHolder() + if err != nil { + t.Fatalf("creating holder: %v", err) + } + defer os.RemoveAll(path) + defer h.Close() + + // Write bits to separate indexes. + testSetBit(t, h, "i0", "f", 100, 200) + testSetBit(t, h, "i1", "f", 100, 200) + testSetBit(t, h, "i1", "f", 100, 12345678) + + testOp := testHolderOperator{} + ctx := context.Background() + err = h.Process(ctx, &testOp) + if err != nil { + t.Fatalf("processing holder: %v", err) + } + expected := testHolderOperator{ + indexSeen: 2, indexProcessed: 2, + fieldSeen: 2, fieldProcessed: 2, + viewSeen: 2, viewProcessed: 2, + fragmentSeen: 3, fragmentProcessed: 3, + } + if testOp != expected { + t.Fatalf("holder processor did not process as expected. expected %#v, got %#v", expected, testOp) + } +} + +func TestHolderOperatorCancel(t *testing.T) { + h, path, err := makeHolder() + if err != nil { + t.Fatalf("creating holder: %v", err) + } + defer os.RemoveAll(path) + defer h.Close() + + // Write bits to separate indexes. + testSetBit(t, h, "i0", "f", 100, 200) + testSetBit(t, h, "i1", "f", 100, 200) + testSetBit(t, h, "i1", "f", 100, 12345678) + + // Here, we want to ensure that the operation gets cancelled + // successfully. In practice we expect it to process one fragment, then + // end up blocked on the waitHere, then get cancelled... But the + // waitHere blockage isn't really something holder.Process can do + // anything about, so we close the channel, so two fragments are + // processed. But in theory you could end up with only one fragment + // processed if this goroutine managed to cancel before the processor + // gets to the next fragment. Point is, it shouldn't hit all three, + // because the checks against the cancellation should fire before it + // gets there. + testOp := testHolderOperator{waitHere: make(chan struct{})} + ctx, cancel := context.WithCancel(context.Background()) + done := make(chan struct{}) + go func() { + err = h.Process(ctx, &testOp) + close(done) + }() + testOp.waitHere <- struct{}{} + cancel() + close(testOp.waitHere) + <-done + if err != context.Canceled { + t.Fatalf("processing holder: expected context.Canceled, got %v", err) + } + testOp.waitHere = nil + expected := testHolderOperator{ + indexSeen: 2, indexProcessed: 2, + fieldSeen: 2, fieldProcessed: 2, + viewSeen: 2, viewProcessed: 2, + fragmentSeen: 3, fragmentProcessed: 3, + } + if testOp == expected { + t.Fatalf("holder processor did not cancel. expected something other than %#v", expected) } } diff --git a/http/handler.go b/http/handler.go index 67be9e297..06a62ebd7 100644 --- a/http/handler.go +++ b/http/handler.go @@ -221,6 +221,8 @@ func (h *Handler) populateValidators() { h.validators["GetTransaction"] = queryValidationSpecRequired() h.validators["PostTransaction"] = queryValidationSpecRequired() h.validators["PostFinishTransaction"] = queryValidationSpecRequired() + h.validators["Inspect"] = queryValidationSpecRequired().Optional("indexes", "fields", "views", "shards", "checksum", "containers") + } type contextKeyQuery int @@ -352,6 +354,7 @@ func newRouter(handler *Handler) *mux.Router { router.HandleFunc("/index/{index}/field/{field}/import-roaring/{shard}", handler.handlePostImportRoaring).Methods("POST").Name("PostImportRoaring") router.HandleFunc("/index/{index}/query", handler.handlePostQuery).Methods("POST").Name("PostQuery") router.HandleFunc("/info", handler.handleGetInfo).Methods("GET").Name("GetInfo") + router.HandleFunc("/inspect", handler.handleInspect).Methods("GET").Name("Inspect") router.HandleFunc("/recalculate-caches", handler.handleRecalculateCaches).Methods("POST").Name("RecalculateCaches") router.HandleFunc("/schema", handler.handleGetSchema).Methods("GET").Name("GetSchema") router.HandleFunc("/schema", handler.handlePostSchema).Methods("POST").Name("PostSchema") @@ -590,6 +593,37 @@ func (h *Handler) handleGetInfo(w http.ResponseWriter, r *http.Request) { } } +func (h *Handler) handleInspect(w http.ResponseWriter, r *http.Request) { + if !validHeaderAcceptJSON(r.Header) { + http.Error(w, "JSON only acceptable response", http.StatusNotAcceptable) + return + } + q := r.URL.Query() + _, checksum := q["checksum"] + _, containers := q["containers"] + req := pilosa.InspectRequest{ + HolderFilterParams: pilosa.HolderFilterParams{ + Indexes: q.Get("indexes"), + Fields: q.Get("fields"), + Views: q.Get("views"), + Shards: q.Get("shards"), + }, + InspectRequestParams: pilosa.InspectRequestParams{ + Checksum: checksum, + Containers: containers, + }, + } + info, err := h.api.Inspect(r.Context(), &req) + if err != nil { + http.Error(w, fmt.Sprintf("inspect request: %v", err), http.StatusBadRequest) + return + } + w.Header().Set("Content-Type", "application/json") + if err := json.NewEncoder(w).Encode(info); err != nil { + h.logger.Printf("write inspect response error: %s", err) + } +} + type getSchemaResponse struct { Indexes []*pilosa.IndexInfo `json:"indexes"` } diff --git a/roaring/container_stash.go b/roaring/container_stash.go index 64a6d217e..a43b3eee8 100644 --- a/roaring/container_stash.go +++ b/roaring/container_stash.go @@ -45,35 +45,47 @@ type Container struct { type containerFlags uint8 +var containerFlagStrings = [...]string{ + "", + "mapped", + "frozen", + "frozen/mapped", + "pristine", + "pristine/mapped", + "pristine/frozen", + "pristine/frozen/mapped", +} + +func (f containerFlags) String() string { + return containerFlagStrings[f&7] +} + const ( flagMapped = containerFlags(1 << iota) flagFrozen + flagPristine ) func (c *Container) String() string { if c == nil { return "" } - froze := "" - switch c.flags { - case flagFrozen: - froze = "frozen " - case flagMapped: - froze = "mapped " - case flagFrozen | flagMapped: - froze = "frozen/mapped" + var space, froze string + if c.flags != 0 { + space = " " + froze = c.flags.String() } switch c.typeID { case containerArray: - return fmt.Sprintf("<%sarray container, N=%d>", froze, c.N()) + return fmt.Sprintf("<%s%sarray container, N=%d>", froze, space, c.N()) case containerBitmap: - return fmt.Sprintf("<%sbitmap container, N=%d, len %dx uint64>", - froze, c.N(), len(c.bitmap())) + return fmt.Sprintf("<%s%sbitmap container, N=%d, len %dx uint64>", + froze, space, c.N(), len(c.bitmap())) case containerRun: - return fmt.Sprintf("<%srun container, N=%d, len %dx interval>", - froze, c.N(), len(c.runs())) + return fmt.Sprintf("<%s%srun container, N=%d, len %dx interval>", + froze, space, c.N(), len(c.runs())) default: - return fmt.Sprintf("", froze, c.typeID, c.N()) + return fmt.Sprintf("", froze, space, c.typeID, c.N()) } } @@ -292,6 +304,7 @@ func (c *Container) unmapOrClone() *Container { return c.Clone() } c.flags &^= flagMapped + c.flags &^= flagPristine // mapped: we want to unmap the storage. switch c.typeID { case containerArray: @@ -368,6 +381,7 @@ func (c *Container) setArrayMaybeCopy(array []uint16, doCopy bool) { if len(array) > 1<<16 { panic("impossibly large array") } + c.flags &^= flagPristine // array we can fit in data store: if len(array) <= stashedArraySize { copy(c.data[:stashedArraySize], array) @@ -497,6 +511,7 @@ func (c *Container) setBitmap(bitmap []uint64) { panic("illegal bitmap length") } c.pointer, c.len, c.cap = (*uint16)(unsafe.Pointer(&bitmap[0])), bitmapN, bitmapN + c.flags &^= flagPristine } // runs yields the data viewed as a slice of intervals. @@ -531,6 +546,7 @@ func (c *Container) setRunsMaybeCopy(runs []interval16, doCopy bool) { if len(runs) > 1<<15 { panic("impossibly large run set") } + c.flags &^= flagPristine // array we can fit in data store: if len(runs) <= stashedRunSize { newRuns := (*[stashedRunSize]interval16)(unsafe.Pointer(&c.data))[:len(runs)] diff --git a/roaring/roaring.go b/roaring/roaring.go index f79b20c5b..b87f4d6af 100644 --- a/roaring/roaring.go +++ b/roaring/roaring.go @@ -1721,6 +1721,8 @@ func (b *Bitmap) writeToUnoptimized(w io.Writer) (n int64, err error) { // bitmap and yield information about containers, including type, size, and // the location of their data structures. type roaringIterator interface { + // Len reports the number of containers total. + Len() (count int64) // Next yields the information about the next container Next() (key uint64, cType byte, n int, length int, pointer *uint16, err error) // Remaining yields the bytes left over past the end of the roaring data, @@ -1885,6 +1887,11 @@ func (r *baseRoaringIterator) Done(err error) { r.currentDataOffset = 0 } +// Len() indicates the total number of containers the iterator expects to have. +func (r *baseRoaringIterator) Len() int64 { + return r.keys +} + func (r *baseRoaringIterator) Remaining() ([]byte, int64) { if r.lastDataOffset == 0 { return nil, 0 @@ -2436,20 +2443,24 @@ func (b *Bitmap) roaringSize() (int64, int64) { } // Info returns stats for the bitmap. -func (b *Bitmap) Info() BitmapInfo { +func (b *Bitmap) Info(includeContainers bool) BitmapInfo { info := BitmapInfo{ - OpN: b.opN, - Ops: b.ops, - Containers: make([]ContainerInfo, 0, b.Containers.Size()), + OpN: b.opN, + Ops: b.ops, + ContainerCount: b.Containers.Size(), + } + if includeContainers { + info.Containers = make([]ContainerInfo, 0, info.ContainerCount) } - info.ContainerCount = cap(info.Containers) citer, _ := b.Containers.Iterator(0) for citer.Next() { k, c := citer.Value() ci := c.info() ci.Key = k info.BitCount += uint64(c.N()) - info.Containers = append(info.Containers, ci) + if includeContainers { + info.Containers = append(info.Containers, ci) + } } return info } @@ -2510,11 +2521,12 @@ func (b *Bitmap) Flip(start, end uint64) *Bitmap { type BitmapInfo struct { OpN int Ops int - OpDetails []OpInfo + OpDetails []OpInfo `json:"OpDetails,omitempty"` BitCount uint64 ContainerCount int - Containers []ContainerInfo // The containers found in the bitmap originally - OpContainers []ContainerInfo // The containers resulting from ops log changes. + Containers []ContainerInfo `json:"Containers,omitempty"` // The containers found in the bitmap originally + OpContainers []ContainerInfo `json:"OpContainers,omitempty"` // The containers resulting from ops log changes. + From, To uintptr // if set, indicates the address range used when unpacking } // Iterator represents an iterator over a Bitmap. @@ -3718,6 +3730,7 @@ func (c *Container) info() ContainerInfo { info.Alloc = 0 return info } + info.Flags = c.flags.String() if c.isArray() { info.Type = "array" @@ -3799,6 +3812,7 @@ func (c *Container) bitmapRepair() { type ContainerInfo struct { Key uint64 // container key Type string // container type (array, bitmap, or run) + Flags string // flag state N int32 // number of bits Alloc int // memory used Pointer uintptr // address diff --git a/roaring/roaring_test.go b/roaring/roaring_test.go index 55395ea73..7dda0f5c6 100644 --- a/roaring/roaring_test.go +++ b/roaring/roaring_test.go @@ -1733,7 +1733,7 @@ type benchmarkSampleData struct { var sampleData benchmarkSampleData func isAllType(b *roaring.Bitmap, typ string) bool { - bi := b.Info() + bi := b.Info(true) for _, c := range bi.Containers { if c.Type != typ { return false diff --git a/roaring/unmarshal_binary.go b/roaring/unmarshal_binary.go index ee566236a..aeab0ef2f 100644 --- a/roaring/unmarshal_binary.go +++ b/roaring/unmarshal_binary.go @@ -17,6 +17,7 @@ package roaring import ( "errors" "io" + "unsafe" ) // UnmarshalBinary reads Pilosa's format, or upstream roaring (mostly; @@ -99,9 +100,11 @@ func (b *Bitmap) UnmarshalBinary(data []byte) (err error) { // InspectBinary reads a roaring bitmap, plus a possible ops log, // and reports back on the contents, including distinguishing between // the original ops log and the post-ops-log contents. -func InspectBinary(data []byte) (info BitmapInfo, err error) { +func InspectBinary(data []byte, mapped bool, info *BitmapInfo) (b *Bitmap, mappedAny bool, err error) { + b = NewFileBitmap() + b.PreferMapping(mapped) if data == nil { - return info, errors.New("no roaring bitmap provided") + return b, mappedAny, errors.New("no roaring bitmap provided") } var itr roaringIterator var itrKey uint64 @@ -113,13 +116,13 @@ func InspectBinary(data []byte) (info BitmapInfo, err error) { itr, err = newRoaringIterator(data) if err != nil { - return info, err + return b, mappedAny, err } if itr == nil { - return info, errors.New("failed to create roaring iterator, but don't know why") + return b, mappedAny, errors.New("failed to create roaring iterator, but don't know why") } - - b := NewFileBitmap() + keys := itr.Len() + info.Containers = make([]ContainerInfo, 0, keys) itrKey, itrCType, itrN, itrLen, itrPointer, itrErr = itr.Next() for itrErr == nil { @@ -131,22 +134,54 @@ func InspectBinary(data []byte) (info BitmapInfo, err error) { pointer: itrPointer, flags: flagMapped, } + if !mapped { + newC.unmapOrClone() + } + newC.flags |= flagPristine + if newC.flags&flagMapped != 0 { + mappedAny = true + } + var size int b.Containers.Put(itrKey, newC) + switch itrCType { + case containerArray: + size = int(newC.n) * 2 + case containerBitmap: + size = 8192 + case containerRun: + size = itrLen*interval16Size + runCountHeaderSize + } + info.Containers = append(info.Containers, ContainerInfo{ + N: newC.n, + Mapped: newC.flags&flagMapped != 0, + Type: containerTypeNames[itrCType], + Alloc: size, + Pointer: uintptr(unsafe.Pointer(newC.pointer)), + Key: itrKey, + Flags: newC.flags.String(), + }) + info.ContainerCount++ + info.BitCount += uint64(newC.n) itrKey, itrCType, itrN, itrLen, itrPointer, itrErr = itr.Next() } // note: if we get a non-EOF err, it's possible that we made SOME // changes but didn't log them. I don't have a good solution to this. if itrErr != io.EOF { - return info, itrErr + return b, mappedAny, itrErr } - - // gather initial info - info = b.Info() + // stash pointer ranges + info.From = uintptr(unsafe.Pointer(&data[0])) + info.To = info.From + uintptr(len(data)) // Read ops log until the end of the file. b.ops = 0 b.opN = 0 buf, lastValidOffset := itr.Remaining() + // if there's no ops log, we're done and can just return the + // info so far. + if len(buf) == 0 { + return b, mappedAny, err + } for { // Exit when there are no more ops to parse. if len(buf) == 0 { @@ -155,36 +190,56 @@ func InspectBinary(data []byte) (info BitmapInfo, err error) { // Unmarshal the op and apply it. var opr op - if err := opr.UnmarshalBinary(buf); err != nil { - return info, err + if err = opr.UnmarshalBinary(buf); err != nil { + // we break out here, but we continue on to + // return the bitmap as-is, along with data about + // it, and the error. this lets us share the + // "is anything mapped" check with that code. + break } opr.apply(b) // Increase the op count. - info.Ops++ - info.OpN += opr.count() - info.OpDetails = append(info.OpDetails, opr.info()) - + if info != nil { + info.Ops++ + info.OpN += opr.count() + info.OpDetails = append(info.OpDetails, opr.info()) + } // Move the buffer forward. opSize := opr.size() buf = buf[opSize:] lastValidOffset += int64(opSize) } + citer, _ := b.Containers.Iterator(0) + // it's possible the ops log unmapped every mapped container, so we recheck. + mappedAny = false + if info == nil { + for citer.Next() { + _, c := citer.Value() + if c.Mapped() { + mappedAny = true + break + } + } + return b, mappedAny, err + } // now we want to compute the actual container and bit counts after // ops, and create a report of just the containers which got changed. info.ContainerCount = 0 info.BitCount = 0 - citer, _ := b.Containers.Iterator(0) for citer.Next() { k, c := citer.Value() + if c.Mapped() { + mappedAny = true + } info.ContainerCount++ info.BitCount += uint64(c.N()) - if c.Mapped() { + if c.flags&flagPristine != 0 { continue } ci := c.info() ci.Key = k info.OpContainers = append(info.OpContainers, ci) } - return info, nil + return b, mappedAny, err } diff --git a/server.go b/server.go index 75f5b2024..0afd2fa43 100644 --- a/server.go +++ b/server.go @@ -565,6 +565,7 @@ func (s *Server) Open() error { return errors.Wrap(err, "opening Holder") } // bring up the background tasks for the holder. + s.holder.SnapshotQueue = s.snapshotQueue s.holder.Activate() if err := s.cluster.setNodeState(nodeStateReady); err != nil { return errors.Wrap(err, "setting nodeState") diff --git a/snapshotqueue.go b/snapshotqueue.go index 5dcc5dbf8..1bdcdb7a9 100644 --- a/snapshotqueue.go +++ b/snapshotqueue.go @@ -15,7 +15,9 @@ package pilosa import ( + "context" "fmt" + "io" "math/bits" "os" "sync" @@ -64,7 +66,8 @@ type SnapshotQueue interface { type queuelessSnapshotQueue struct{} func (q *queuelessSnapshotQueue) Enqueue(f *fragment) { - _ = f.snapshot() + // We don't actually try to enqueue the snapshot; it breaks things + // if a snapshot gets caused during a transaction. } func (q *queuelessSnapshotQueue) Await(f *fragment) error { @@ -86,7 +89,16 @@ var defaultSnapshotQueue = &queuelessSnapshotQueue{} // newSnapshotQueue makes a new snapshot queue, of depth N, with // w worker threads. func newSnapshotQueue(n int, w int, l logger.Logger) SnapshotQueue { - sq := prioritySnapshotQueue{normal: make(chan snapshotRequest, n), urgent: make(chan snapshotRequest), background: make(chan snapshotRequest), done: make(chan struct{}), maxOpN: 10000, logger: l} + ctx, cancel := context.WithCancel(context.Background()) + sq := prioritySnapshotQueue{ + normal: make(chan snapshotRequest, n), + urgent: make(chan snapshotRequest), + background: make(chan snapshotRequest), + ctx: ctx, + cancel: cancel, + maxOpN: 10000, + logger: l, + } if sq.logger == nil { sq.logger = logger.NewStandardLogger(os.Stderr) } @@ -114,11 +126,12 @@ type prioritySnapshotQueue struct { urgent chan snapshotRequest normal chan snapshotRequest background chan snapshotRequest - done chan struct{} + ctx context.Context + cancel context.CancelFunc mu sync.RWMutex scanWG, workerWG sync.WaitGroup maxOpN int - observedOpN [16]int + observedOpN [16]uint32 stats struct { enqueued uint32 skipped uint32 @@ -128,38 +141,37 @@ type prioritySnapshotQueue struct { func (sq *prioritySnapshotQueue) spawnWorkers(w int) { sq.mu.Lock() defer sq.mu.Unlock() - if sq.done == nil { - sq.logger.Printf("prioritySnapshotQueue worker: no done channel, already done?") + if sq.ctx.Err() != nil { + sq.logger.Printf("prioritySnapshotQueue worker: already done") return } sq.workerWG.Add(w) for i := 0; i < w; i++ { - go sq.worker(sq.urgent, sq.normal, sq.background, sq.done) + go sq.worker(sq.ctx, sq.urgent, sq.normal, sq.background) } } -func (sq *prioritySnapshotQueue) worker(urgent, normal, background chan snapshotRequest, done chan struct{}) { - // We don't want a race condition on these. If they're non-nil when - // we get them, they should get closed at some point. If done is - // already nil, we shouldn't do anything. +func (sq *prioritySnapshotQueue) worker(ctx context.Context, urgent, normal, background chan snapshotRequest) { defer sq.workerWG.Done() + done := ctx.Done() ok := true var req snapshotRequest for ok { req.frag = nil - select { + case _, ok = <-done: case req, ok = <-urgent: default: select { + case _, ok = <-done: case req, ok = <-urgent: case req, ok = <-normal: default: select { + case _, ok = <-done: case req, ok = <-urgent: case req, ok = <-normal: case req, ok = <-background: - case _, ok = <-done: } } } @@ -193,10 +205,9 @@ func (sq *prioritySnapshotQueue) process(req snapshotRequest) { func (sq *prioritySnapshotQueue) Stop() { sq.mu.Lock() defer sq.mu.Unlock() - close(sq.done) + sq.cancel() // scanners need to be done before we close the other channels. sq.scanWG.Wait() - sq.done = nil close(sq.normal) sq.normal = nil close(sq.urgent) @@ -216,6 +227,7 @@ func (sq *prioritySnapshotQueue) Enqueue(f *fragment) { if f.snapshotPending { return } + sq.observeOpN(uint32(f.opN)) sq.mu.RLock() defer sq.mu.RUnlock() if sq.normal == nil { @@ -270,6 +282,7 @@ func (sq *prioritySnapshotQueue) Immediate(f *fragment) error { return errors.New("requested immediate snapshot after snapshot queue was closed") } f.snapshotPending = true + sq.observeOpN(uint32(f.opN)) req := snapshotRequest{frag: f, when: time.Now()} // if the fragment was already in the work queue, it's *possible* // that the only available worker just picked it off the queue, and @@ -286,35 +299,6 @@ func (sq *prioritySnapshotQueue) Immediate(f *fragment) error { return sq.Await(f) } -// needsSnapshot determines whether a fragment probably wants snapshotting. -// Specifically, it looks for fragments not already marked to receive -// snapshots, but which have a high enough opN to justify a snapshot. This -// is only used from the background scan. -func (sq *prioritySnapshotQueue) needsSnapshot(f *fragment) bool { - if f == nil { - return false - } - f.mu.Lock() - defer f.mu.Unlock() - if f.snapshotPending { - return false - } - if f.opN > sq.maxOpN { - return true - } - // aka "log2(n) + 1", or 0 for n==0 - pow2 := 32 - bits.LeadingZeros32(uint32(f.opN)) - // 15 == 16384. we assume that since 16384 is higher than our - // normal maxOpN, it's always a reasonable value. - if pow2 > 15 { - pow2 = 15 - } - // store in inverse order so the lowest slot in the array is the - // highest cardinality - sq.observedOpN[15-pow2]++ - return false -} - // ScanHolder spawns a goroutine which iterates through the holder's // indexes/fields/views/fragments, looking for fragments which have OpN // high enough to justify a snapshot but don't seem to have one pending. @@ -326,143 +310,191 @@ func (sq *prioritySnapshotQueue) ScanHolder(h *Holder, done chan struct{}) { sq.mu.Unlock() } +// observeOpN reports that a given value of opN was "observed", meaning, +// we encountered a fragment which had that value. This happens for every +// enqueue/immediate, including enqueue attempts which fail to actually +// enter the queue, and it also happens for fragments noticed by the background +// scan but which don't have high enough opN to trigger a snapshot. +func (sq *prioritySnapshotQueue) observeOpN(n uint32) { + // aka "log2(n) + 1", or 0 for n==0 + pow2 := 32 - bits.LeadingZeros32(n) + // 15 == 16384. Our usual fragment maxOpN is 10k, so most fragments + // should end up in the 8k-16k bucket, rather than the 16k+ bucket, + // unless we've got a lot of ingests with large batches going on, + // in which case the 16k bucket will win. + if pow2 > 15 { + pow2 = 15 + } + // store in inverse order so the lowest slot in the array is the + // highest cardinality + atomic.AddUint32(&sq.observedOpN[15-pow2], 1) +} + +// computeMaxOpN tries to pick a reasonable new maxOpN for the background +// scan to use. On a quiet system, we want to gradually lower opN, picking +// the fragments with the highest opN values first, because those offer the +// largest benefit. So, whenever we check a fragment in the background, if we +// *don't* snapshot it, we'll "observe" its OpN value, and then we pick a +// value which picks up at least 1/4 of them. +// +// If there's ingest activity, the Immediate and Enqueue operations will +// "observe" the OpN of fragments submitted to them. This can drive OpN back +// up, if those fragments frequently have very high opN values, which reflects +// the fact that we have enough of that activity that we don't need the +// background scanner adding more. +// +// If we have enough ingest activity that the background scanner never actually +// gets to submit work, we'll rarely get here, because the background scanner +// will block until there's no snapshots pending for the normal workload. +// When we do, we'll probably pick a MaxOpN which is dominated by the ingest +// workload's opN values. So for instance, if everything coming in from the +// ingest workload has 10k or more items, because that's the default fragment +// maxOpN, that will probably set the background snapshot queue value to 8k. +func (sq *prioritySnapshotQueue) computeMaxOpN() { + sq.logger.Debugf("observedOpN by power of 2: %d\n", sq.observedOpN[:]) + total := uint32(0) + for i := range sq.observedOpN { + total += atomic.LoadUint32(&sq.observedOpN[i]) + } + target := (total / 4) + 1 + subTotal := uint32(0) + for i := range sq.observedOpN { + v := atomic.LoadUint32(&sq.observedOpN[i]) + subTotal += v + if subTotal >= target { + prevMaxOpN := sq.maxOpN + sq.maxOpN = (1 << (15 - uint(i))) / 2 + if sq.maxOpN > 0 { + sq.maxOpN-- + } + if prevMaxOpN != sq.maxOpN { + sq.logger.Printf("background scan: %d/%d fragments considered have opN %d or higher\n", + subTotal, total, sq.maxOpN) + } + break + } + } + // It's conceptually possible that we'll miss a couple of observations + // here but that's not really important. This is all pretty approximate. + for i := range sq.observedOpN { + atomic.StoreUint32(&sq.observedOpN[i], 0) + } +} + +// prioritySnapshotQueueScanner is the data type that implements HolderOperator +// and represents a single scan of a holder, with a given maxOpN. +type prioritySnapshotQueueScanner struct { + HolderFilterAll + HolderProcessNone + sq *prioritySnapshotQueue + holder *Holder + queue chan snapshotRequest + ctx context.Context + maxOpN int + seen, hits, counter int +} + +func (s *prioritySnapshotQueueScanner) ProcessFragment(f *fragment) error { + if f == nil { + return nil + } + s.seen++ + // we can't defer this reasonably, because otherwise we'll keep + // the fragment locked forever if we end up trying to send it + // to the queue, but the workers are busy on other fragments. + f.mu.Lock() + open := f.open + snapshotPending, opN := f.snapshotPending, f.opN + f.mu.Unlock() + + // a pending snapshot is one that is either in the normal or + // immediate queue, or is trying to get into the normal queue + // and about to fail, but either way, it already got observed + // there, so we don't need to observe it here. A closed fragment + // doesn't matter to us -- it should be a transient state that + // happens during a shutdown, or shouldn't happen, but we don't + // care about it. + if snapshotPending || !open { + return nil + } + if opN <= s.maxOpN { + // observe the value but don't do a snapshot + s.sq.observeOpN(uint32(opN)) + s.counter++ + if s.counter == 1000 { + select { + case <-time.After(1 * time.Second): + case <-s.ctx.Done(): + return io.EOF + } + s.counter = 0 + } + return nil + } + // we don't observe values when we decide to trigger a snapshot, + // because those values will be changing anyway. we could also + // observe them as zero, but that's also sort of wrong. + s.hits++ + select { + case s.queue <- snapshotRequest{frag: f, when: time.Now()}: + s.sq.logger.Debugf("found fragment needing snapshot: %s\n", f.path) + case <-s.ctx.Done(): + return io.EOF + } + return nil + +} + +func contextMergedWithStructChan(ctx context.Context, ch chan struct{}) (context.Context, context.CancelFunc) { + canCancel, cancel := context.WithCancel(ctx) + go func() { + select { + case <-ctx.Done(): + cancel() + case <-ch: + cancel() + case <-canCancel.Done(): + // don't need to cancel, but do need to exit this + // function + } + }() + return canCancel, cancel +} + // scanHolderWorker is a background task that scans a holder looking for // fragments which need snapshots taken. It's the cleanup task for snapshots // that would have been requested by Enqueue, but the queue was full. func (sq *prioritySnapshotQueue) scanHolderWorker(h *Holder, background chan snapshotRequest, done chan struct{}) { - // queueDone is global to this snapshotQueue, done is specific to this holder - // scanner. - queueDone := sq.done defer sq.scanWG.Done() - var indexNames, fieldNames, viewNames []string - var fragNums []uint64 + ctx, cancel := contextMergedWithStructChan(sq.ctx, done) + defer cancel() + scanner := &prioritySnapshotQueueScanner{ + sq: sq, + holder: h, + queue: background, + ctx: sq.ctx, + maxOpN: sq.maxOpN, + } for { - // To avoid abusing things, cap activity rate; every time we finish - // the holder, or every couple hundred fragments considered, we - // pause for a bit. - counter := 0 - hits := 0 - h.mu.Lock() - indexNames = indexNames[:0] - for indexName := range h.indexes { - indexNames = append(indexNames, indexName) + err := h.Process(ctx, scanner) + if err != nil { + return } - h.mu.Unlock() - for _, indexName := range indexNames { - h.mu.Lock() - index := h.indexes[indexName] - h.mu.Unlock() - if index == nil { - continue - } - fieldNames = fieldNames[:0] - index.mu.Lock() - for fieldName := range index.fields { - fieldNames = append(fieldNames, fieldName) - } - index.mu.Unlock() - for _, fieldName := range fieldNames { - index.mu.Lock() - field := index.fields[fieldName] - index.mu.Unlock() - if field == nil { - continue - } - viewNames = viewNames[:0] - field.mu.Lock() - for viewName := range field.viewMap { - viewNames = append(viewNames, viewName) - } - field.mu.Unlock() - for _, viewName := range viewNames { - field.mu.Lock() - view := field.viewMap[viewName] - field.mu.Unlock() - if view == nil { - continue - } - fragNums := fragNums[:0] - view.mu.Lock() - for fragNum := range view.fragments { - fragNums = append(fragNums, fragNum) - } - view.mu.Unlock() - for _, fragNum := range fragNums { - view.mu.Lock() - frag := view.fragments[fragNum] - view.mu.Unlock() - if sq.needsSnapshot(frag) { - hits++ - select { - case background <- snapshotRequest{frag: frag, when: time.Now()}: - sq.logger.Debugf("found fragment needing snapshot: %s\n", frag.path) - case <-done: - return - } - } else { - // Count fragments examined *without* finding anything that - // needed a snapshot. When we find things that need snapshots, - // the time it takes the workers to respond to us is enough - // of a delay to keep us from eating every CPU. So, if a lot - // of things need snapshots, and the workers aren't doing - // anything else, ScanHolder will mostly keep them saturated. - // If they're busy, we'll block forever in the write to the - // background queue. If there's nothing that needs snapshots, - // we pause frequently for a second or so at a time. - counter++ - if counter == 1000 { - select { - case <-time.After(1 * time.Second): - case <-queueDone: - return - case <-done: - return - } - counter = 0 - } - } - } - } - } - } - if hits > 0 { - sq.logger.Printf("background scan: %d fragments needed snapshots\n", hits) - hits = 0 + + if scanner.hits > 0 { + sq.logger.Printf("background scan: %d/%d fragments needed snapshots\n", scanner.hits, scanner.seen) + scanner.hits = 0 } else { sq.logger.Debugf("background scan: no fragments needed snapshots, waiting\n") // No reason to be active if we're not finding anything. select { case <-time.After(60 * time.Second): - case <-queueDone: - return - case <-done: + case <-ctx.Done(): return } } - sq.logger.Debugf("observedOpN by power of 2: %d\n", sq.observedOpN[:]) - total := 0 - for _, v := range sq.observedOpN { - total += v - } - target := total / 4 - subTotal := 0 - for i, v := range sq.observedOpN { - subTotal += v - if subTotal >= target { - prevMaxOpN := sq.maxOpN - sq.maxOpN = (1 << (15 - uint(i))) / 2 - if sq.maxOpN > 0 { - sq.maxOpN-- - } - if prevMaxOpN != sq.maxOpN { - sq.logger.Printf("background scan: %d/%d fragments considered have opN %d or higher\n", - subTotal, total, sq.maxOpN) - } - break - } - } - for i := range sq.observedOpN { - sq.observedOpN[i] = 0 - } + scanner.seen = 0 + sq.computeMaxOpN() + scanner.maxOpN = sq.maxOpN } } From 31014d11e5a2d38f0ef60c67d289fd2722d84b4e Mon Sep 17 00:00:00 2001 From: Jaden Weiss Date: Mon, 29 Jun 2020 15:22:27 -0400 Subject: [PATCH 8/8] remove unnecesary slice operation when processing holder --- holder.go | 1 - 1 file changed, 1 deletion(-) diff --git a/holder.go b/holder.go index f8ac267f9..d1f457371 100644 --- a/holder.go +++ b/holder.go @@ -1660,7 +1660,6 @@ func (h *Holder) Process(ctx context.Context, op HolderOperator) (err error) { var fragNums []uint64 h.mu.Lock() - indexNames = indexNames[:0] for indexName := range h.indexes { indexNames = append(indexNames, indexName) }