scriptoriumadmin/featurebase
1
0
Fork 0
mirror of https://github.com/featurebasedb/featurebase.git synced 2026-09-11 15:21:02 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

shared/generic functionality for iterating holders

Browse source 
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.
This commit is contained in:
Seebs 2020-05-26 16:39:04 -05:00 • committed by Jaden Weiss
parent 2826ecc0b7
commit 4d494f6699
No known key found for this signature in database
GPG key ID: 177F065773634B67
15 changed files with 1206 additions and 525 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
api.go
View file

@ -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")

4
cmd/inspect.go
View file

@ -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
}

221
ctl/inspect.go
View file

@ -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
}

50
fragment.go
View file

@ -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)

9
fragment_internal_test.go
View file

@ -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{}),

28
handler.go
View file

@ -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
}
}

413
holder.go
View file

@ -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
}

430
holder_internal_test.go
View file

@ -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)
}
}

34
http/handler.go
View file

@ -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"`
}

44
roaring/container_stash.go
View file

@ -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 "<nil container>"
}
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("<unknown %s%d container, N=%d>", froze, c.typeID, c.N())
return fmt.Sprintf("<unknown %s%s%d container, N=%d>", 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)]

32
roaring/roaring.go
View file

@ -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

2
roaring/roaring_test.go
View file

@ -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

93
roaring/unmarshal_binary.go
View file

@ -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
}

1
server.go
View file

@ -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")

366
snapshotqueue.go
View file

@ -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
}
}