scriptoriumadmin/featurebase
1
0
Fork 0
mirror of https://github.com/featurebasedb/featurebase.git synced 2026-08-28 10:54:59 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

remove a bunch of roaring backend stuff

Browse source 
snapshotQueue, op tracking, roaring-only tests
This commit is contained in:
Matthew Jaffee 2022-02-01 14:46:00 -06:00
parent fa4855c887
commit 69c00a92ad
26 changed files with 197 additions and 2069 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

12
Makefile
View file

@ -79,9 +79,6 @@ testvsub-race:
cd ..; \
done
tour:
./tournament.sh
bench:
$(GO) test ./... -bench=. -run=NoneZ -timeout=127m $(TESTFLAGS)
@ -349,14 +346,5 @@ install-gometalinter:
GO111MODULE=off gometalinter --install
GO111MODULE=off $(GO) get github.com/remyoudompheng/go-misc/deadcode
test-txstore-rbf:
PILOSA_STORAGE_BACKEND=rbf $(MAKE) testv-race
# WARNING: This feature is no longer being tested regularly in CI. The test is
# very slow and very expensive, and we're not sure it actually provides useful
# information now.
test-txstore-rbf_bolt:
PILOSA_STORAGE_BACKEND=rbf_bolt $(MAKE) testv-race
test-external-lookup:
$(GO) test . -tags='$(BUILD_TAGS) $(TEST_TAGS)' $(TESTFLAGS) -run ^TestExternalLookup$$ -externalLookupDSN $(EXTERNAL_LOOKUP_DSN)

8
ctl/server.go
View file

@ -75,13 +75,7 @@ func BuildServerFlags(cmd *cobra.Command, srv *server.Command) {
flags.IntVar(&srv.Config.Profile.BlockRate, "profile.block-rate", srv.Config.Profile.BlockRate, "Sampling rate for goroutine blocking profiler. One sample per <rate> ns.")
flags.IntVar(&srv.Config.Profile.MutexFraction, "profile.mutex-fraction", srv.Config.Profile.MutexFraction, "Sampling fraction for mutex contention profiling. Sample 1/<rate> of events.")
// Storage
// Note: the default for --storage.backend must be kept "" empty string.
// Otherwise we cannot detect and honor the PILOSA_STORAGE_BACKEND env var
// over-ride.
// TODO: the comment above was carried over from the PILOSA_TXSRC flag, but
// we should confirm that this still applies.
flags.StringVar(&srv.Config.Storage.Backend, "storage.backend", storage.DefaultBackend, fmt.Sprintf("transaction/storage to use: one of roaring or rbf. The default is: %v. The env var PILOSA_STORAGE_BACKEND is over-ridden by --storage.backend option on the command line.", storage.DefaultBackend))
flags.StringVar(&srv.Config.Storage.Backend, "storage.backend", storage.DefaultBackend, fmt.Sprintf("transaction/storage to use: 'rbf' is only supported value.", storage.DefaultBackend))
flags.BoolVar(&srv.Config.Storage.FsyncEnabled, "storage.fsync", true, "enable fsync fully safe flush-to-disk")
// RBF specific flags. See pilosa/rbf/cfg/cfg.go for definitions.

55
dbshard.go
View file

@ -67,9 +67,8 @@ type DBShard struct {
Shard uint64
Open bool
typ txtype
styp string
hasRoaring bool // if either of the types is roaringTxn
typ txtype
styp string
W DBWrapper
ParentDBIndex *DBIndex
@ -131,8 +130,7 @@ type DBPerShard struct {
// Easily see how many we have.
Flatmap map[flatkey]*DBShard
typ txtype
hasRoaring bool
typ txtype
txf *TxFactory
holder *Holder
@ -269,11 +267,6 @@ func (txf *TxFactory) NewDBPerShard(typ txtype, holderDir string, holder *Holder
vprint.PanicOn("must have holder.cfg.RBFConfig and holder.cfg.StorageConfig set here")
}
hasRoaring := false
if typ == roaringTxn {
hasRoaring = true
}
d = &DBPerShard{
typ: typ,
HolderDir: holderDir,
@ -281,7 +274,6 @@ func (txf *TxFactory) NewDBPerShard(typ txtype, holderDir string, holder *Holder
dbh: NewDBHolder(),
Flatmap: make(map[flatkey]*DBShard),
txf: txf,
hasRoaring: hasRoaring,
index2shards: newIndex2Shards(),
StorageConfig: holder.cfg.StorageConfig,
RBFConfig: holder.cfg.RBFConfig,
@ -407,10 +399,7 @@ func (per *DBPerShard) unprotectedGetDBShard(index string, shard uint64, idx *In
}
dbs, ok = dbi.Shard[shard]
if dbs != nil && dbs.closed {
// roaring txn are nil/fake anyway. Don't freak out.
if per.typ != roaringTxn {
vprint.PanicOn(fmt.Sprintf("cannot retain closed dbs across holder ReOpen dbs='%p'; per.typ='%v'", dbs, per.typ))
}
vprint.PanicOn(fmt.Sprintf("cannot retain closed dbs across holder ReOpen dbs='%p'; per.typ='%v'", dbs, per.typ))
}
if !ok {
dbs = &DBShard{
@ -421,7 +410,6 @@ func (per *DBPerShard) unprotectedGetDBShard(index string, shard uint64, idx *In
HolderPath: per.HolderDir,
idx: idx,
per: per,
hasRoaring: per.hasRoaring,
}
dbs.styp = per.typ.String()
dbi.Shard[shard] = dbs
@ -430,8 +418,6 @@ func (per *DBPerShard) unprotectedGetDBShard(index string, shard uint64, idx *In
if !dbs.Open {
var registry DBRegistry
switch dbs.typ {
case roaringTxn:
registry = globalRoaringReg
case rbfTxn:
registry = globalRbfDBReg
registry.(*rbfDBRegistrar).SetRBFConfig(per.RBFConfig)
@ -470,8 +456,6 @@ func (f *TxFactory) GetShardsForIndex(idx *Index, roaringViewPath string, requir
return f.dbPerShard.TypedDBPerShardGetShardsForIndex(f.typ, idx, roaringViewPath, requireData)
}
// if roaringViewPath is "" then for ty == roaringTxn we go to disk to discover
// all the view paths under idx for type ty.
// requireData means open the database file and verify that at least one key is set.
// The returned sliceOfShards should not be modified. We will cache it for subsequent
// queries.
@ -485,14 +469,6 @@ func (per *DBPerShard) TypedDBPerShardGetShardsForIndex(ty txtype, idx *Index, r
per.Mu.Lock()
defer per.Mu.Unlock()
if ty == roaringTxn && roaringViewPath != "" {
shardMap, err := roaringMapOfShards(roaringViewPath)
if err != nil {
return nil, err
}
return shardMap, nil
}
i2ss := per.index2shards
ss, ok := i2ss[idx.name]
@ -507,27 +483,6 @@ func (per *DBPerShard) TypedDBPerShardGetShardsForIndex(ty txtype, idx *Index, r
// Upon return, cache the setOfShards value and reuse it next time
if ty == roaringTxn {
// INVAR: roaringViewPath == "", because the other case is
// handled above.
fields := idx.Fields()
for _, field := range fields {
for _, view := range field.views() {
shardMap, err := roaringMapOfShards(view.path)
if err != nil {
return nil,
errors.Wrap(err, fmt.Sprintf(
"TypedDBPerShardGetLocalShardsForIndex roaringTxn view.path='%v'", view.path))
}
for shard := range shardMap {
setOfShards.add(shard)
}
}
}
return setOfShards.CloneMaybe(), nil
}
// INVAR: not-roaring.
path := per.prefixForType(idx, ty)
ignoreEmpty := false
@ -730,8 +685,6 @@ func (per *DBPerShard) GetFieldView2ShardsMapForIndex(idx *Index) (vs *FieldView
ty := per.typ
switch ty {
case roaringTxn:
return roaringGetFieldView2Shards(idx)
default:
vs = NewFieldView2Shards()

71
dbshard_internal_test.go
View file

@ -71,7 +71,7 @@ func Test_DBPerShard_GetShardsForIndex_LocalOnly(t *testing.T) {
v2s.addViewShardSet(txkey.FieldView{Field: field, View: "standard"}, stdShardSet)
}
for _, src := range []string{"roaring", "rbf"} {
for _, src := range []string{"rbf"} {
cfg := mustHolderConfig()
cfg.StorageConfig.Backend = src
holder := NewHolder(tmpdir, cfg)
@ -82,7 +82,6 @@ func Test_DBPerShard_GetShardsForIndex_LocalOnly(t *testing.T) {
idx, err = NewIndex(holder, filepath.Join(tmpdir, index), index)
PanicOn(err)
}
estd := "rick/fields/_exists/views/standard"
std := "rick/fields/f/views/standard"
shards, err := holder.txf.GetShardsForIndex(idx, tmpdir+sep+std, false)
@ -93,65 +92,23 @@ func Test_DBPerShard_GetShardsForIndex_LocalOnly(t *testing.T) {
panic(fmt.Sprintf("missing shard=%v from shards='%#v'", shard, shards))
}
}
if src == "roaring" {
// check estd too
shards, err = holder.txf.GetShardsForIndex(idx, tmpdir+sep+estd, false)
for _, shard := range []uint64{93, 223, 221, 215, 219, 217} {
tx := idx.holder.txf.NewTx(Txo{Write: !writable, Index: idx, Shard: shard})
fvs, err := tx.GetSortedFieldViewList(idx, shard)
PanicOn(err)
for _, shard := range []uint64{93, 223, 221, 215, 219, 217} {
if !shards[shard] {
panic(fmt.Sprintf("missing shard=%v from shards='%#v'", shard, shards))
}
// expect these same two field/views for all 6 shards
expect0 := txkey.FieldView{Field: "_exists", View: "standard"}
expect1 := txkey.FieldView{Field: "f", View: "standard"}
if len(fvs) != 2 {
panic(fmt.Sprintf("fvs should be len 2, got '%#v' (%s)", fvs, src))
}
// check GetSortedFieldViewList() and roaringGetFieldView2Shards()
vs, err := roaringGetFieldView2Shards(idx)
PanicOn(err)
for _, shard := range []uint64{93, 223, 221, 215, 219, 217} {
tx := idx.holder.txf.NewTx(Txo{Write: !writable, Index: idx, Shard: shard})
fvs, err := tx.GetSortedFieldViewList(idx, shard)
PanicOn(err)
// expect these same two field/views for all 6 shards
expect0 := txkey.FieldView{Field: "_exists", View: "standard"}
expect1 := txkey.FieldView{Field: "f", View: "standard"}
if len(fvs) != 2 {
panic(fmt.Sprintf("fvs should be len 2, got '%#v' (%s)", fvs, src))
}
if fvs[0] != expect0 {
panic(fmt.Sprintf("expected fvs[0]='%#v', but got '%#v'", expect0, fvs[0]))
}
if fvs[1] != expect1 {
panic(fmt.Sprintf("expected fvs[1]='%#v', but got '%#v'", expect1, fvs[1]))
}
for _, fv := range fvs {
if !vs.has(fv.Field, fv.View, shard) {
panic(fmt.Sprintf("vs did not contain fv='%#v' for shard %v", fv, shard))
}
}
tx.Rollback()
if fvs[0] != expect0 {
panic(fmt.Sprintf("expected fvs[0]='%#v', but got '%#v'", expect0, fvs[0]))
}
} else {
// non-roaring: rbf
for _, shard := range []uint64{93, 223, 221, 215, 219, 217} {
tx := idx.holder.txf.NewTx(Txo{Write: !writable, Index: idx, Shard: shard})
fvs, err := tx.GetSortedFieldViewList(idx, shard)
PanicOn(err)
// expect these same two field/views for all 6 shards
expect0 := txkey.FieldView{Field: "_exists", View: "standard"}
expect1 := txkey.FieldView{Field: "f", View: "standard"}
if len(fvs) != 2 {
panic(fmt.Sprintf("fvs should be len 2, got '%#v' (%s)", fvs, src))
}
if fvs[0] != expect0 {
panic(fmt.Sprintf("expected fvs[0]='%#v', but got '%#v'", expect0, fvs[0]))
}
if fvs[1] != expect1 {
panic(fmt.Sprintf("expected fvs[1]='%#v', but got '%#v'", expect1, fvs[1]))
}
tx.Rollback()
if fvs[1] != expect1 {
panic(fmt.Sprintf("expected fvs[1]='%#v', but got '%#v'", expect1, fvs[1]))
}
tx.Rollback()
}
holder.Close()
}

10
executor_test.go
View file

@ -33,7 +33,6 @@ import (
"github.com/molecula/featurebase/v3/pql"
"github.com/molecula/featurebase/v3/proto"
"github.com/molecula/featurebase/v3/server"
"github.com/molecula/featurebase/v3/storage"
"github.com/molecula/featurebase/v3/test"
"github.com/molecula/featurebase/v3/testhook"
. "github.com/molecula/featurebase/v3/vprint" // nolint:staticcheck
@ -1277,15 +1276,6 @@ func TestExecutor_Execute_Count(t *testing.T) {
}
func roaringOnlyTest(t *testing.T) {
src := pilosa.CurrentBackend()
if src == pilosa.RoaringTxn || (storage.DefaultBackend == pilosa.RoaringTxn && src == "") {
// okay to run, we are under roaring only
} else {
t.Skip("skip for everything but roaring")
}
}
// Ensure a set query can be executed.
func TestExecutor_Execute_Set(t *testing.T) {
t.Run("RowIDColumnID", func(t *testing.T) {

1
field_internal_test.go
View file

@ -247,7 +247,6 @@ func NewTestField(t testing.TB, opts FieldOption) *TestField {
}
cfg := DefaultHolderConfig()
cfg.StorageConfig.Backend = CurrentBackendOrDefault()
cfg.StorageConfig.FsyncEnabled = false
cfg.RBFConfig.FsyncEnabled = false
h := NewHolder(path, cfg)

287
fragment.go
View file

@ -3,7 +3,6 @@ package pilosa
import (
"archive/tar"
"bufio"
"bytes"
"container/heap"
"context"
@ -16,7 +15,6 @@ import (
"math/bits"
"os"
"path/filepath"
"runtime/debug"
"sort"
"strconv"
"strings"
@ -59,18 +57,12 @@ const (
// width of roaring containers is 2^16
containerWidth = 1 << 16
// snapshotExt is the file extension used for an in-process snapshot.
snapshotExt = ".snapshotting"
// cacheExt is the file extension for persisted cache ids.
cacheExt = ".cache"
// HashBlockSize is the number of rows in a merkle hash block.
HashBlockSize = 100
// defaultFragmentMaxOpN is the default value for Fragment.MaxOpN.
defaultFragmentMaxOpN = 10000
// Row ids used for boolean fields.
falseRowID = uint64(0)
trueRowID = uint64(1)
@ -132,22 +124,11 @@ type fragment struct {
// idx cached to avoid repeatedly looking it up everywhere.
idx *Index
// parent holder, used to find snapshot queue, etc.
// parent holder
holder *Holder
// debugging tool: addresses of current and previous maps
prevdata, currdata struct{ from, to uintptr }
// File-backed storage
flags byte // user-defined flags passed to roaring
storage *roaring.Bitmap
opN int // number of ops since snapshot (may be approximate for imports)
ops int // number of higher-level operations, as opposed to bit changes
snapshotPending bool // set to true when requesting a snapshot, set to false after snapshot completes
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?
storage *roaring.Bitmap
// Cache for row counts.
CacheType string // passed in by field
@ -164,11 +145,6 @@ type fragment struct {
// Cached checksums for each block.
checksums map[int][]byte
// Number of operations performed before performing a snapshot.
// This limits the size of fragments on the heap and flushes them to disk
// so that they can be mmapped and heap utilization can be kept low.
MaxOpN int
// Logger used for out-of-band log entries.
Logger logger.Logger
@ -194,18 +170,15 @@ func newFragment(holder *Holder, spec fragSpec, shard uint64, flags byte) *fragm
fieldstr: spec.fieldstr,
fld: spec.field,
shard: shard,
flags: flags,
idx: idx,
CacheType: DefaultCacheType,
CacheSize: DefaultCacheSize,
holder: holder,
MaxOpN: defaultFragmentMaxOpN,
stats: stats.NopStatsClient,
}
f.snapshotCond = sync.Cond{L: &f.mu}
return f
}
@ -259,12 +232,6 @@ func (f *fragment) Open() error {
defer f.mu.Unlock()
if err := func() error {
// Initialize storage in a function so we can close if anything goes wrong.
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.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 {
@ -278,24 +245,12 @@ func (f *fragment) Open() error {
f.close()
return err
}
f.open = true
_ = testhook.Opened(f.holder.Auditor, f, nil)
f.holder.Logger.Debugf("successfully opened index/field/view/fragment: %s/%s/%s/%d", f.index(), f.field(), f.view(), f.shard)
return nil
}
// openStorage opens the storage bitmap. Does nothing in RBF-world and will be removed soon.
func (f *fragment) openStorage(unmarshalData bool) error {
if !f.idx.NeedsSnapshot() {
f.currdata = struct{ from, to uintptr }{}
f.prevdata = f.currdata
return nil // openStorage becomes a noop under RBF, Badger, etc.
}
return nil
}
// openCache initializes the cache from row ids persisted to disk.
func (f *fragment) openCache() error {
// Determine cache type from field name.
@ -351,12 +306,6 @@ func (f *fragment) Close() error {
defer func() {
_ = testhook.Closed(f.holder.Auditor, f, nil)
}()
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()
}
@ -367,28 +316,12 @@ func (f *fragment) close() error {
return errors.Wrap(err, "flushing cache")
}
// Close underlying storage.
if err := f.closeStorage(); err != nil {
f.holder.Logger.Errorf("fragment: error closing storage: err=%s, path=%s", err, f.path())
return errors.Wrap(err, "closing storage")
}
// Remove checksums.
f.checksums = nil
return nil
}
// closeStorage is essentially a no-op and will go away soon.
func (f *fragment) closeStorage() error {
// opN is determined by how many bit set/clear operations are in the storage
// write log, so once the storage is closed it should be 0. Opening new
// storage will set opN appropriately.
f.opN = 0
return nil
}
// mutexCheck checks for any entries in fragment which violate the mutex
// property of having only one value set for a given column ID.
func (f *fragment) mutexCheck(tx Tx, details bool, limit int) (map[uint64][]uint64, error) {
@ -511,9 +444,6 @@ func (f *fragment) unprotectedSetBit(tx Tx, rowID, columnID uint64) (changed boo
// Invalidate block checksum.
delete(f.checksums, int(rowID/HashBlockSize))
// Increment number of operations until snapshot is required.
f.incrementOpN(1)
// If we're using a cache, update it. Otherwise skip the
// possibly-expensive count operation.
if f.CacheType != CacheTypeNone {
@ -563,9 +493,6 @@ func (f *fragment) unprotectedClearBit(tx Tx, rowID, columnID uint64) (changed b
// Invalidate block checksum.
delete(f.checksums, int(rowID/HashBlockSize))
// Increment number of operations until snapshot is required.
f.incrementOpN(1)
// If we're using a cache, update it. Otherwise skip the
// possibly-expensive count operation.
if f.CacheType != CacheTypeNone {
@ -632,8 +559,6 @@ func (f *fragment) unprotectedSetRow(tx Tx, row *Row, rowID uint64) (changed boo
}
}
// Snapshot storage.
f.holder.SnapshotQueue.Enqueue(f)
f.stats.Count("setRow", 1, 1.0)
return changed, nil
@ -672,9 +597,6 @@ func (f *fragment) unprotectedClearRow(tx Tx, rowID uint64) (changed bool, err e
// Clear the row in cache.
f.cache.Add(rowID, 0)
// Snapshot storage.
f.holder.SnapshotQueue.Enqueue(f)
return changed, nil
}
@ -1929,7 +1851,7 @@ func (f *fragment) mergeBlock(tx Tx, id int, data []pairSet) (sets, clears []pai
return sets[1:], clears[1:], err
}
// bulkImport bulk imports a set of bits and then snapshots the storage.
// bulkImport bulk imports a set of bits.
// The cache is updated to reflect the new data.
func (f *fragment) bulkImport(tx Tx, rowIDs, columnIDs []uint64, options *ImportOptions) error {
// Verify that there are an equal number of row ids and column ids.
@ -2142,68 +2064,48 @@ func (p parallelSlices) Swap(i, j int) {
// snapshot of the fragment or just do in-memory updates while appending
// operations to the op log.
func (f *fragment) importPositions(tx Tx, set, clear []uint64, rowSet map[uint64]struct{}) error {
//tx.AddN()
doFunc := func() error {
if len(set) > 0 {
f.stats.Count(MetricImportingN, int64(len(set)), 1)
if len(set) > 0 {
f.stats.Count(MetricImportingN, int64(len(set)), 1)
// TODO benchmark Add/RemoveN behavior with sorted/unsorted positions
changedN, err := tx.Add(f.index(), f.field(), f.view(), f.shard, set...)
if err != nil {
return errors.Wrap(err, "adding positions")
}
f.stats.Count(MetricImportedN, int64(changedN), 1)
f.incrementOpN(changedN)
// TODO benchmark Add/RemoveN behavior with sorted/unsorted positions
changedN, err := tx.Add(f.index(), f.field(), f.view(), f.shard, set...)
if err != nil {
return errors.Wrap(err, "adding positions")
}
f.stats.Count(MetricImportedN, int64(changedN), 1)
}
if len(clear) > 0 {
f.stats.Count(MetricClearingN, int64(len(clear)), 1)
changedN, err := tx.Remove(f.index(), f.field(), f.view(), f.shard, clear...)
if err != nil {
return errors.Wrap(err, "clearing positions")
}
f.stats.Count(MetricClearedN, int64(changedN), 1)
f.incrementOpN(changedN)
if len(clear) > 0 {
f.stats.Count(MetricClearingN, int64(len(clear)), 1)
changedN, err := tx.Remove(f.index(), f.field(), f.view(), f.shard, clear...)
if err != nil {
return errors.Wrap(err, "clearing positions")
}
f.stats.Count(MetricClearedN, int64(changedN), 1)
}
// Update cache counts for all affected rows.
for rowID := range rowSet {
// Invalidate block checksum.
delete(f.checksums, int(rowID/HashBlockSize))
if f.CacheType != CacheTypeNone {
start := rowID * ShardWidth
end := (rowID + 1) * ShardWidth
n, err := tx.CountRange(f.index(), f.field(), f.view(), f.shard, start, end)
if err != nil {
return errors.Wrap(err, "CountRange")
}
f.cache.BulkAdd(rowID, n)
}
}
// Update cache counts for all affected rows.
for rowID := range rowSet {
// Invalidate block checksum.
delete(f.checksums, int(rowID/HashBlockSize))
if f.CacheType != CacheTypeNone {
f.cache.Invalidate()
}
return nil
}
err := doFunc()
if err != nil && f.storage != nil {
// 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.holder.Logger.Errorf("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.holder.Logger.Errorf("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)
start := rowID * ShardWidth
end := (rowID + 1) * ShardWidth
n, err := tx.CountRange(f.index(), f.field(), f.view(), f.shard, start, end)
if err != nil {
return errors.Wrap(err, "CountRange")
}
f.cache.BulkAdd(rowID, n)
}
}
return err
if f.CacheType != CacheTypeNone {
f.cache.Invalidate()
}
return nil
}
// sliceDifference removes everything from original that's found in remove,
@ -2503,125 +2405,6 @@ func (f *fragment) importRoaringOverwrite(ctx context.Context, tx Tx, data []byt
return f.importRoaring(ctx, tx, data, false)
}
// incrementOpN increase the operation count by one.
// If the count exceeds the maximum allowed then a snapshot is performed.
func (f *fragment) incrementOpN(changed int) {
if changed <= 0 {
return
}
// don't count opN or ops if our index doesn't want snapshots
if !f.idx.NeedsSnapshot() {
return
}
f.opN += changed
f.ops++
if f.opN > f.MaxOpN {
f.holder.SnapshotQueue.Enqueue(f)
}
}
// Snapshot writes the storage bitmap to disk and reopens it. This may
// coexist with existing background-queue snapshotting; it does not remove
// things from the queue. You probably don't want to do this; use
// the snapshotQueue's Enqueue/Await.
func (f *fragment) Snapshot() error {
f.mu.Lock()
defer f.mu.Unlock()
return f.snapshot()
}
func track(start time.Time, message string, stats stats.StatsClient, logger logger.Logger) {
elapsed := time.Since(start)
logger.Debugf("%s took %s", message, elapsed)
stats.Timing(MetricSnapshotDurationSeconds, elapsed, 1.0)
}
// snapshot does the actual snapshot operation. it does not check or care
// about f.snapshotPending.
func (f *fragment) snapshot() (err error) {
if !f.idx.NeedsSnapshot() {
return nil
}
if !f.open {
return errors.New("snapshot request on closed fragment")
}
wouldPanic := debug.SetPanicOnFault(true)
defer func() {
debug.SetPanicOnFault(wouldPanic)
if r := recover(); r != nil {
if e2, ok := r.(error); ok {
err = e2
// special case: if we caught a page fault, we diagnose that directly. sadly,
// 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.holder.Logger.Errorf("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 {
err = fmt.Errorf("non-error PanicOn: %v", r)
}
}
}()
_, err = unprotectedWriteToFragment(f, f.storage)
if err == nil {
f.snapshotStamp = time.Now()
}
return err
}
// unprotectedWriteToFragment writes the fragment f with bm as the data. It is unprotected, and
// f.mu must be locked when calling it.
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.holder.Logger)
// Create a temporary file to snapshot to.
snapshotPath := f.path() + snapshotExt
file, err := os.Create(snapshotPath)
if err != nil {
return n, fmt.Errorf("create snapshot file: %s", err)
}
// No deferred close, because we want to close it sooner than the
// end of this function.
// Write storage to snapshot.
bw := bufio.NewWriter(file)
if n, err = bm.WriteTo(bw); err != nil {
file.Close()
return n, fmt.Errorf("snapshot write to: %s", err)
}
if err := bw.Flush(); err != nil {
file.Close()
return n, fmt.Errorf("flush: %s", err)
}
// we close the file here so we don't still have it open when trying
// to open it in a moment.
file.Close()
// Move snapshot to data file location.
if err := os.Rename(snapshotPath, f.path()); err != nil {
return n, fmt.Errorf("rename snapshot: %s", err)
}
// if we reloaded from the file, we'd end up with this bitmap
// as our storage. so... let's use this bitmap. as our storage.
f.storage = bm
// Reopen storage.
if err := f.openStorage(false); err != nil {
return n, fmt.Errorf("open storage: %s", err)
}
// Reset operation count.
f.opN = 0
return n, nil
}
// RecalculateCache rebuilds the cache regardless of invalidate time delay.
func (f *fragment) RecalculateCache() {
f.mu.Lock()

974
fragment_internal_test.go
View file

File diff suppressed because it is too large Load diff

28
holder.go
View file

@ -85,8 +85,7 @@ 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
// Instantiates new translation stores
OpenTranslateStore OpenTranslateStoreFunc
@ -271,8 +270,6 @@ func NewHolder(path string, cfg *HolderConfig) *Holder {
Logger: cfg.Logger,
Opts: HolderOpts{StorageBackend: cfg.StorageConfig.Backend},
SnapshotQueue: defaultSnapshotQueue,
Auditor: NewAuditor(),
path: path,
@ -734,16 +731,15 @@ func (h *Holder) maybeSpool(msg Message) bool {
return true
}
// 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)
// Activate runs the background tasks relevant to keeping a holder in
// a stable state, such as 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)
h.wg.Add(1)
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
@ -791,7 +787,6 @@ func (h *Holder) Close() error {
// Notify goroutines of closing and wait for completion.
close(h.closing)
h.wg.Wait()
for _, index := range h.Indexes() {
if err := index.Close(); err != nil {
return errors.Wrap(err, "closing index")
@ -809,10 +804,6 @@ func (h *Holder) Close() error {
h.opened.mu.Lock()
h.opened.ch = make(chan struct{})
h.opened.mu.Unlock()
if h.SnapshotQueue != nil {
h.SnapshotQueue.Stop()
h.SnapshotQueue = nil
}
if h.lookupDB != nil {
err := h.lookupDB.Close()
@ -827,13 +818,6 @@ func (h *Holder) Close() error {
return nil
}
func (h *Holder) NeedsSnapshot() bool {
h.mu.RLock()
defer h.mu.RUnlock()
return h.txf.NeedsSnapshot()
}
// HasData returns true if Holder contains at least one index.
// This is used to determine if the rebalancing of data is necessary
// when a node joins the cluster.

10
holder_internal_test.go
View file

@ -174,17 +174,9 @@ func TestHolderOperatorCancel(t *testing.T) {
}
}
// mustHolderConfig is meant to help minimize the number of places in the code
// where we're reading the PILOSA_STORAGE_BACKEND environment variable for
// testing purposes. Ideally we would handle this differently, but this is a
// first attempt at improving things. Note: the actual os.Getenv() call was
// moved to the CurrentBackend() function.
// mustHolderConfig sets up a default holder config for tests.
func mustHolderConfig() *HolderConfig {
cfg := DefaultHolderConfig()
if backend := CurrentBackend(); backend != "" {
_ = MustBackendToTxtype(backend)
cfg.StorageConfig.Backend = backend
}
cfg.StorageConfig.FsyncEnabled = false
cfg.RBFConfig.FsyncEnabled = false
cfg.Schemator = disco.InMemSchemator

111
holder_test.go
View file

@ -5,13 +5,12 @@ import (
"context"
"math"
"os"
"path/filepath"
"reflect"
"strings"
"testing"
"time"
"github.com/molecula/featurebase/v3"
pilosa "github.com/molecula/featurebase/v3"
"github.com/molecula/featurebase/v3/disco"
"github.com/molecula/featurebase/v3/pql"
"github.com/molecula/featurebase/v3/test"
@ -21,10 +20,7 @@ import (
// mustHolderConfig provides a default test-friendly holder config.
func mustHolderConfig() *pilosa.HolderConfig {
cfg := pilosa.DefaultHolderConfig()
if backend := pilosa.CurrentBackend(); backend != "" {
_ = pilosa.MustBackendToTxtype(backend)
cfg.StorageConfig.Backend = backend
}
cfg.StorageConfig.Backend = "rbf"
cfg.StorageConfig.FsyncEnabled = false
cfg.RBFConfig.FsyncEnabled = false
cfg.Schemator = disco.InMemSchemator
@ -55,109 +51,6 @@ func TestHolder_Open(t *testing.T) {
t.Fatalf("unexpected error: %v", err)
}
})
t.Run("ErrFragmentStoragePermission", func(t *testing.T) {
roaringOnlyTest(t)
if os.Geteuid() == 0 {
t.Skip("Skipping permissions test since user is root.")
}
h := test.MustOpenHolder(t)
defer h.Close()
var idx *pilosa.Index
var err error
if idx, err = h.CreateIndex("foo", pilosa.IndexOptions{}); err != nil {
t.Fatal(err)
}
var shard uint64
tx := idx.Txf().NewTx(pilosa.Txo{Write: writable, Index: idx, Shard: shard})
defer tx.Rollback()
if field, err := idx.CreateField("bar", pilosa.OptFieldTypeDefault()); err != nil {
t.Fatal(err)
} else if _, err := field.SetBit(tx, 0, 0, nil); err != nil {
t.Fatal(err)
} else if err := tx.Commit(); 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"), 0000); err != nil {
t.Fatal(err)
}
defer func() {
_ = os.Chmod(filepath.Join(h.Path(), "foo", "bar", "views", "standard", "fragments", "0"), 0644)
}()
if err := h.Reopen(); err == nil || !strings.Contains(err.Error(), "permission denied") {
t.Fatalf("unexpected error: %s", err)
}
})
t.Run("ErrFragmentStorageCorrupt", func(t *testing.T) {
roaringOnlyTest(t)
h := test.MustOpenHolder(t)
defer h.Close()
var idx *pilosa.Index
var err error
if idx, err = h.CreateIndex("foo", pilosa.IndexOptions{}); err != nil {
t.Fatal(err)
}
var shard uint64
tx := idx.Txf().NewTx(pilosa.Txo{Write: writable, Index: idx, Shard: shard})
if err != nil {
t.Fatal(err)
}
defer tx.Rollback()
if field, err := idx.CreateField("bar", pilosa.OptFieldTypeDefault()); err != nil {
t.Fatal(err)
} else if _, err := field.SetBit(tx, 0, 0, nil); err != nil {
t.Fatal(err)
} else if err := tx.Commit(); err != nil {
t.Fatal(err)
} else if err := h.Holder.Close(); err != nil {
t.Fatal(err)
} else if err := os.Truncate(filepath.Join(h.Path(), "foo", "bar", "views", "standard", "fragments", "0"), 2); err != nil {
t.Fatal(err)
}
if err := h.Reopen(); err == nil || !strings.Contains(err.Error(), "open fragment: shard=0, err=opening storage: unmarshal storage") {
t.Fatalf("unexpected error: %s", err)
}
})
t.Run("ErrFragmentStorageRecoverable", func(t *testing.T) {
roaringOnlyTest(t)
h := test.MustOpenHolder(t)
defer h.Close()
idx, err := h.CreateIndex("foo", pilosa.IndexOptions{})
if err != nil {
t.Fatal(err)
}
var shard uint64
tx := idx.Txf().NewTx(pilosa.Txo{Write: writable, Index: idx, Shard: shard})
defer tx.Rollback()
if field, err := idx.CreateField("bar", pilosa.OptFieldTypeDefault()); err != nil {
t.Fatal(err)
} else if _, err := field.SetBit(tx, 0, 0, nil); err != nil {
t.Fatal(err)
} else if err := tx.Commit(); err != nil {
t.Fatal(err)
} else if err := h.Holder.Close(); err != nil {
t.Fatal(err)
} else if err := os.Truncate(filepath.Join(h.IndexesPath(), "foo", "bar", "views", "standard", "fragments", "0"), 20); err != nil {
t.Fatal(err)
}
if err := h.Reopen(); err != nil {
t.Fatalf("unexpected error: %s", err)
}
})
t.Run("ForeignIndex", func(t *testing.T) {
t.Run("ErrForeignIndexNotFound", func(t *testing.T) {
h := test.MustOpenHolder(t)

3
http/handler.go
View file

@ -37,6 +37,7 @@ import (
"github.com/molecula/featurebase/v3/logger"
"github.com/molecula/featurebase/v3/pql"
"github.com/molecula/featurebase/v3/rbf"
"github.com/molecula/featurebase/v3/storage"
"github.com/molecula/featurebase/v3/topology"
"github.com/molecula/featurebase/v3/tracing"
"github.com/pkg/errors"
@ -1087,7 +1088,7 @@ func (h *Handler) handlePostQuery(w http.ResponseWriter, r *http.Request) {
req, ok := qreq.(*pilosa.QueryRequest)
if DoPerQueryProfiling {
backend := pilosa.CurrentBackend()
backend := storage.DefaultBackend
reqHash := hash(req.Query)
qlen := len(req.Query)

4
index.go
View file

@ -93,10 +93,6 @@ func (i *Index) NewTx(txo Txo) Tx {
return i.holder.txf.NewTx(txo)
}
func (i *Index) NeedsSnapshot() bool {
return i.holder.txf.NeedsSnapshot()
}
// CreatedAt is an timestamp for a specific version of an index.
func (i *Index) CreatedAt() int64 {
i.mu.RLock()

68
mmap_test.go
View file

@ -1,68 +0,0 @@
// Copyright 2021 Molecula Corp. All rights reserved.
package pilosa
import (
"math/rand"
"runtime"
"testing"
"github.com/molecula/featurebase/v3/logger"
)
type cv struct {
cols []uint64
vals []int64
}
func forceSnapshotsCheckMapping(t *testing.T) {
depth := uint64(6)
f, idx, tx := mustOpenBSIFragment(t, "i", "f", viewStandard, 0)
tx.Rollback()
f.Logger = logger.NewLogfLogger(t)
defer f.Clean(t)
tx = idx.holder.txf.NewTx(Txo{Write: writable, Index: idx, Fragment: f, Shard: f.shard})
defer tx.Rollback()
for i := 0; i < f.MaxOpN; i++ {
_, _ = f.setBit(tx, 0, uint64(32*i))
}
// force snapshot so we get a mmapped row...
err := f.Snapshot()
if err != nil {
t.Fatalf("initial snapshot error: %v", err)
}
values := make([]cv, 1024)
for i := range values {
cols := make([]uint64, 128)
vals := make([]int64, 128)
for j := range cols {
// pick values in the first 16 cols of each of the 16
// shards in a default shardwidth, so each set will
// probably change some values from the previous one.
cols[j] = uint64(((rand.Int63n(16) & int64(i>>2)) << 16) + rand.Int63n(16))
vals[j] = int64(rand.Int63n(1 << depth))
}
values[i] = cv{cols, vals}
}
// modify the original bitmap, until it causes a snapshot, which
// then invalidates the other map...
for i := 0; i < 32; i++ {
cv := values[i%len(values)]
// periodically force gc, so if we have a small pool of maps
// we'll go in and out of mapping mode
if i%5 == 0 {
runtime.GC()
}
err := f.importValue(tx, cv.cols, cv.vals, depth, (i%3 == 1))
if err != nil {
t.Fatalf("importValue[%d]: %v", i, err)
}
err = f.Snapshot()
if err != nil {
t.Fatalf("snapshot[%d]: %v", i, err)
}
}
}

19
pilosa.go
View file

@ -2,13 +2,11 @@
package pilosa
import (
"os"
"regexp"
"time"
"github.com/molecula/featurebase/v3/disco"
pnet "github.com/molecula/featurebase/v3/net"
"github.com/molecula/featurebase/v3/storage"
"github.com/pkg/errors"
)
@ -157,20 +155,3 @@ func AddressWithDefaults(addr string) (*pnet.URI, error) {
}
return pnet.NewURIFromAddress(addr)
}
// CurrentBackend is one step in an attempt to centralize (and either minimize
// or completely remove), the calls to environment variables throughout the
// tests. Ideally we could get rid of this and rely completely on the
// configuration parameters.
func CurrentBackend() string {
return os.Getenv("PILOSA_STORAGE_BACKEND")
}
// CurrentBackendOrDefault tries the environment variable first, but falls back
// to the default backend if the environment variable is empty.
func CurrentBackendOrDefault() string {
if backend := os.Getenv("PILOSA_STORAGE_BACKEND"); backend != "" {
return backend
}
return storage.DefaultBackend
}

10
pprof.go
View file

@ -19,10 +19,7 @@ import (
// commented out—in holder.go.
func CPUProfileForDur(dur time.Duration, outpath string) {
// per-query pprof output:
backend := CurrentBackend()
if backend == "" {
backend = storage.DefaultBackend
}
backend := storage.DefaultBackend
path := outpath + "." + backend
f, err := os.Create(path)
vprint.PanicOn(err)
@ -45,10 +42,7 @@ func CPUProfileForDur(dur time.Duration, outpath string) {
// commented out—in holder.go.
func MemProfileForDur(dur time.Duration, outpath string) {
// per-query pprof output:
backend := CurrentBackend()
if backend == "" {
backend = storage.DefaultBackend
}
backend := storage.DefaultBackend
path := outpath + "." + backend
f, err := os.Create(path)
vprint.PanicOn(err)

22
server.go
View file

@ -64,11 +64,10 @@ type Server struct { // nolint: maligned
schemator disco.Schemator
// External
systemInfo SystemInfo
gcNotifier GCNotifier
logger logger.Logger
queryLogger logger.Logger
snapshotQueue SnapshotQueue
systemInfo SystemInfo
gcNotifier GCNotifier
logger logger.Logger
queryLogger logger.Logger
nodeID string
uri pnet.URI
@ -544,13 +543,6 @@ func (s *Server) UpAndDown() error {
func (s *Server) Open() error {
s.logger.Infof("open server. PID %v", os.Getpid())
if s.holder.NeedsSnapshot() {
// Start background monitoring.
s.snapshotQueue = newSnapshotQueue(10, 2, s.logger)
} else {
s.snapshotQueue = defaultSnapshotQueue //TODO (twg) rethink this
}
// Log startup
err := s.holder.logStartup()
if err != nil {
@ -612,7 +604,6 @@ 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 we joined existing cluster then broadcast "resize on add" message
if initState == disco.InitialClusterStateExisting {
@ -743,11 +734,6 @@ 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

18
server_internal_test.go
View file

@ -2,7 +2,6 @@
package pilosa
import (
"runtime"
"testing"
"time"
@ -10,23 +9,6 @@ import (
"github.com/molecula/featurebase/v3/testhook"
)
// Ensure the file handle count is working
func TestCountOpenFiles(t *testing.T) {
roaringOnlyTest(t)
// Windows is not supported yet
if runtime.GOOS == "windows" {
t.Skip("Skipping unsupported countOpenFiles test on Windows.")
}
count, err := countOpenFiles()
if err != nil {
t.Errorf("countOpenFiles failed: %s", err)
}
if count == 0 {
t.Error("countOpenFiles returned invalid value 0.")
}
}
func TestMonitorAntiEntropyZero(t *testing.T) {
td, err := testhook.TempDirInDir(t, *TempDir, "")

495
snapshotqueue.go
View file

@ -1,495 +0,0 @@
// Copyright 2021 Molecula Corp. All rights reserved.
package pilosa
import (
"context"
"fmt"
"io"
"math/bits"
"os"
"sync"
"sync/atomic"
"time"
"github.com/molecula/featurebase/v3/logger"
"github.com/molecula/featurebase/v3/testhook"
"github.com/pkg/errors"
)
// snapshotQueue is a thing which can handle enqueuing snapshots. A snapshot
// queue distinguishes between high-priority requests, which get satisfied
// by the next available worker, and regular requests, which get enqueued
// if there's space in the queue, and otherwise dropped. There's also a
// separate background task to scan a holder for fragments which may need
// snapshots, but which is processed only when the queue is empty, and only
// slowly. "Await" awaits an existing snapshot if one is already enqueued.
// "Immediate" tries to do one right away. (If one's already enqueued, this
// can leave it in the queue, which will ignore anything that shows up with
// the request flag cleared.)
//
// Await, Enqueue, and Immediate should be called only with the fragment lock
// held.
//
// 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, chan struct{})
Stop()
}
// queuelessSnapshotQueue isn't a snapshot queue, but it satisfies the
// interface.
type queuelessSnapshotQueue struct{}
func (q *queuelessSnapshotQueue) Enqueue(f *fragment) {
// 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 {
return nil
}
func (q *queuelessSnapshotQueue) Immediate(f *fragment) error {
return f.snapshot()
}
func (q *queuelessSnapshotQueue) ScanHolder(h *Holder, done chan struct{}) {
}
func (q *queuelessSnapshotQueue) Stop() {
}
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 {
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)
}
_ = testhook.Opened(NewAuditor(), sq, nil)
sq.spawnWorkers(w)
return sq
}
type snapshotRequest struct {
frag *fragment
when time.Time
}
// prioritySnapshotQueue gives preference to "immediate" requests, and
// dispreference to "background" requests from ScanHolder. It timestamps
// requests, so it can discard a request if the most recent snapshot is
// newer than the request. The snapshotPending flag in the fragment is
// used to track that a given fragment thinks it has been successfully
// enqueued. Background requests are not considered enqueued, since
// they'll never get processed if there's anything else. In normal workloads,
// immediate/urgent snapshots should be rare, but we'll happily drop
// most requests on the floor; the scanner should pick them up once things
// are quiet.
type prioritySnapshotQueue struct {
logger logger.Logger
urgent chan snapshotRequest
normal chan snapshotRequest
background chan snapshotRequest
ctx context.Context
cancel context.CancelFunc
mu sync.RWMutex
scanWG, workerWG sync.WaitGroup
maxOpN int
observedOpN [16]uint32
stats struct {
enqueued uint32
skipped uint32
}
stopped bool
}
func (sq *prioritySnapshotQueue) spawnWorkers(w int) {
sq.mu.Lock()
defer sq.mu.Unlock()
if sq.ctx.Err() != nil {
sq.logger.Infof("prioritySnapshotQueue worker: already done")
return
}
sq.workerWG.Add(w)
for i := 0; i < w; i++ {
go sq.worker(sq.ctx, sq.urgent, sq.normal, sq.background)
}
}
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:
}
}
}
if req.frag != nil {
sq.process(req)
}
}
}
// process actually runs a fragment. it will do this if either the fragment
// has a pending snapshot, or the force flag is set.
func (sq *prioritySnapshotQueue) process(req snapshotRequest) {
f := req.frag
f.mu.Lock()
defer f.mu.Unlock()
if f.snapshotStamp.Before(req.when) {
f.snapshotErr = f.snapshot()
if f.snapshotErr != nil {
fmt.Printf("ERROR: snapshot error: %v\n", f.snapshotErr)
sq.logger.Errorf("snapshot error: %v", f.snapshotErr)
}
f.snapshotPending = false
f.snapshotCond.Broadcast()
}
}
// Stop shuts down the snapshot queue. It first marks it as done, causing
// the background scanner(s), if any, to shut down, then waits for them, then
// closes and nils the queues. The background scanner has to get stopped
// because otherwise it might try to write to those closed queues.
func (sq *prioritySnapshotQueue) Stop() {
sq.mu.Lock()
defer sq.mu.Unlock()
if sq.stopped {
return
}
sq.stopped = true
sq.cancel()
// scanners need to be done before we close the other channels.
sq.scanWG.Wait()
close(sq.normal)
sq.normal = nil
close(sq.urgent)
sq.urgent = nil
close(sq.background)
sq.background = nil
_ = testhook.Closed(NewAuditor(), sq, nil)
enqueued := atomic.LoadUint32(&sq.stats.enqueued)
skipped := atomic.LoadUint32(&sq.stats.skipped)
if skipped > 0 || enqueued > 1 {
sq.logger.Infof("snapshot queue: enqueued %d, skipped %d\n", sq.stats.enqueued, sq.stats.skipped)
}
}
// Enqueue tries to add a fragment to the queue, if the fragment is not already
// enqueued. You should hold a lock on the fragment when calling this.
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 {
sq.logger.Infof("requested snapshot after snapshot queue was closed")
return
}
// we have to set this before enqueing, because it's
// otherwise possible that we're at the head of the queue,
// and the recipient gets the fragment before we execute the
// line after the send.
f.snapshotPending = true
// try to enqueue snapshot
select {
case sq.normal <- snapshotRequest{frag: f, when: time.Now()}:
atomic.AddUint32(&sq.stats.enqueued, 1)
return
default:
atomic.AddUint32(&sq.stats.skipped, 1)
f.snapshotPending = false
return
}
}
// Await returns when f is not pending a snapshot. Call with the fragment lock
// 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()
}
err, f.snapshotErr = f.snapshotErr, nil
return err
}
// Immediate forces an immediate snapshot of the given fragment. Call with
// the fragment locked. If the queue is already closing, the fragment does
// not get snapshotted.
func (sq *prioritySnapshotQueue) Immediate(f *fragment) error {
sq.mu.RLock()
// no deferred unlock, because we want to unlock this before calling Await.
// Not because that needs this lock, but because once we're that far, we
// *don't* need this lock anymore so someone else should have it.
if sq.urgent == nil {
sq.mu.RUnlock()
sq.logger.Errorf("requested immediate snapshot after snapshot queue was closed")
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
// is now waiting on getting the fragment's lock, so it can run
// a snapshot. So we let go of the lock on the fragment, send the
// request, then request the fragment lock again, because Await will
// be sleeping on the condition variable associated with the lock,
// which means it needs to hold the lock so it can let it go during
// the wait... No, really, this made sense.
f.mu.Unlock()
sq.urgent <- req
sq.mu.RUnlock()
f.mu.Lock()
return sq.Await(f)
}
// 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.
// It then dumps these in the low priority background queue.
func (sq *prioritySnapshotQueue) ScanHolder(h *Holder, done chan struct{}) {
sq.mu.Lock()
sq.scanWG.Add(1)
go sq.scanHolderWorker(h, sq.background, done)
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.Infof("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{}) {
defer sq.scanWG.Done()
ctx, cancel := contextMergedWithStructChan(sq.ctx, done)
defer cancel()
scanner := &prioritySnapshotQueueScanner{
sq: sq,
holder: h,
queue: background,
ctx: sq.ctx,
maxOpN: sq.maxOpN,
}
for {
err := h.Process(ctx, scanner)
if err != nil {
return
}
if scanner.hits > 0 {
sq.logger.Infof("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 <-ctx.Done():
return
}
}
scanner.seen = 0
sq.computeMaxOpN()
scanner.maxOpN = sq.maxOpN
}
}

3
stattx.go
View file

@ -12,6 +12,7 @@ import (
"github.com/molecula/featurebase/v3/debugstats"
"github.com/molecula/featurebase/v3/roaring"
txkey "github.com/molecula/featurebase/v3/short_txkey"
"github.com/molecula/featurebase/v3/storage"
"github.com/molecula/featurebase/v3/vprint"
)
@ -56,7 +57,7 @@ func (w *callStats) reset() {
}
func (c *callStats) report() (r string) {
backend := CurrentBackend()
backend := storage.DefaultBackend
r = fmt.Sprintf("callStats: (%v)\n", backend)
c.mu.Lock()
defer c.mu.Unlock()

4
storage/config.go
View file

@ -3,9 +3,7 @@ package storage
// public strings that pilosa/server/config.go can reference
const (
RoaringBackend string = "roaring"
RBFBackend string = "rbf"
BoltBackend string = "bolt"
RBFBackend string = "rbf"
)
// DefaultBackend is set here. pilosa/server/config.go references it

2
test/cluster.go
View file

@ -593,7 +593,7 @@ func prependTestServerOpts(opts []server.CommandOption) []server.CommandOption {
pilosa.OptServerOpenTranslateStore(pilosa.OpenInMemTranslateStore),
pilosa.OptServerNodeDownRetries(5, 100*time.Millisecond),
pilosa.OptServerStorageConfig(&storage.Config{
Backend: pilosa.CurrentBackendOrDefault(),
Backend: storage.DefaultBackend,
FsyncEnabled: false,
}),
),

13
tournament.sh
View file

@ -1,13 +0,0 @@
#!/bin/bash
## tournament.sh runs a sequence of duels between greens and blues.
## Each test run changes the PILOSA_STORAGE_BACKEND and runs either
## one or two backends through the rigors of make testv-race.
## logs are saved to the tourna.log.${i} files.
for i in rbf roaring bolt rbf_roaring roaring_rbf roaring_bolt; do
echo "$(date) starting ${i}, output to tourna.log.${i}"
echo "***=== ${i} ====================*** $(date)" &> tourna.log.${i}
PILOSA_STORAGE_BACKEND=${i} make testv-race 2>&1 > tourna.log.${i}
done

10
tx_test.go
View file

@ -4,13 +4,11 @@ package pilosa_test
import (
"context"
"fmt"
"strings"
"testing"
pilosa "github.com/molecula/featurebase/v3"
"github.com/molecula/featurebase/v3/http"
"github.com/molecula/featurebase/v3/server"
"github.com/molecula/featurebase/v3/storage"
"github.com/molecula/featurebase/v3/test"
. "github.com/molecula/featurebase/v3/vprint" // nolint:staticcheck
)
@ -47,15 +45,7 @@ func queryBalances(m0api *pilosa.API, acctOwnerID uint64, fldAcct0, fldAcct1, in
return
}
func skipForRoaring(t *testing.T) {
src := pilosa.CurrentBackend()
if (storage.DefaultBackend == pilosa.RoaringTxn) || strings.Contains(src, "roaring") {
t.Skip("skip if roaring pseudo-txn involved -- won't show transactional rollback")
}
}
func TestAPI_ImportAtomicRecord(t *testing.T) {
skipForRoaring(t)
c := test.MustRunCluster(t, 1,
[]server.CommandOption{
server.OptCommandServerOptions(

24
txfactory.go
View file

@ -17,8 +17,7 @@ import (
// public strings that pilosa/server/config.go can reference
const (
RoaringTxn string = "roaring"
RBFTxn string = "rbf"
RBFTxn string = "rbf"
)
// DetectMemAccessPastTx true helps us catch places in api and executor
@ -377,9 +376,8 @@ type TxFactory struct {
type txtype int
const (
noneTxn txtype = 0
roaringTxn txtype = 1 // these don't really have any transactions
rbfTxn txtype = 2
noneTxn txtype = 0
rbfTxn txtype = 2
)
// DirectoryName just returns a string version of the transaction type. We
@ -388,8 +386,6 @@ const (
// replaced/removed) during that refactor.
func (ty txtype) DirectoryName() string {
switch ty {
case roaringTxn:
return "roaring"
case rbfTxn:
return "rbf"
}
@ -397,18 +393,12 @@ func (ty txtype) DirectoryName() string {
return ""
}
func (txf *TxFactory) NeedsSnapshot() (b bool) {
return txf.typ == roaringTxn
}
func MustBackendToTxtype(backend string) (typ txtype) {
if strings.Contains(backend, "_") {
panic("blue-green comparisons removed")
}
switch backend {
case RoaringTxn: // "roaring"
return roaringTxn
case RBFTxn: // "rbf"
return rbfTxn
}
@ -839,8 +829,6 @@ func (ty txtype) String() string {
switch ty {
case noneTxn:
return "noneTxn"
case roaringTxn:
return "roaring"
case rbfTxn:
return "rbf"
}
@ -946,16 +934,10 @@ func anyGlobalDBWrappersStillOpen() bool {
return false
}
func (f *TxFactory) hasRoaring() bool {
return f.typ == roaringTxn
}
func (f *TxFactory) hasRBF() bool {
return f.typ == rbfTxn
}
var _ = (&TxFactory{}).hasRoaring // happy linter
func (f *TxFactory) GetDBShardPath(index string, shard uint64, idx *Index, ty txtype, write bool) (shardPath string, err error) {
dbs, err := f.dbPerShard.GetDBShard(index, shard, idx)
if err != nil {

4
txfactory_internal_test.go
View file

@ -8,8 +8,8 @@ import (
func Test_TxFactory_verifyStringConstantsMatch(t *testing.T) {
// txtype.String() method MUST return strings that match
// our const definitions at the top of txfactory.go.
check := []txtype{roaringTxn, rbfTxn}
expect := []string{RoaringTxn, RBFTxn}
check := []txtype{rbfTxn}
expect := []string{RBFTxn}
for i, chk := range check {
obs := chk.String()
if obs != expect[i] {