Merge pull request #6754 from filecoin-project/feat/splitstore-refactor

Splitstore: Some small fixes
This commit is contained in:
Steven Allen 2021-07-14 13:09:59 -07:00 committed by GitHub
commit 44d01712c8
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 99 additions and 56 deletions

View File

@ -94,6 +94,7 @@ type Blockstore struct {
var _ blockstore.Blockstore = (*Blockstore)(nil) var _ blockstore.Blockstore = (*Blockstore)(nil)
var _ blockstore.Viewer = (*Blockstore)(nil) var _ blockstore.Viewer = (*Blockstore)(nil)
var _ blockstore.BlockstoreIterator = (*Blockstore)(nil) var _ blockstore.BlockstoreIterator = (*Blockstore)(nil)
var _ blockstore.BlockstoreGC = (*Blockstore)(nil)
var _ io.Closer = (*Blockstore)(nil) var _ io.Closer = (*Blockstore)(nil)
// Open creates a new badger-backed blockstore, with the supplied options. // Open creates a new badger-backed blockstore, with the supplied options.
@ -167,32 +168,27 @@ func (b *Blockstore) CollectGarbage() error {
} }
defer b.viewers.Done() defer b.viewers.Done()
var err error // compact first to gather the necessary statistics for GC
for err == nil {
err = b.DB.RunValueLogGC(0.125)
}
if err == badger.ErrNoRewrite {
// not really an error in this case
return nil
}
return err
}
// Compact runs a synchronous compaction
func (b *Blockstore) Compact() error {
if err := b.access(); err != nil {
return err
}
defer b.viewers.Done()
nworkers := runtime.NumCPU() / 2 nworkers := runtime.NumCPU() / 2
if nworkers < 2 { if nworkers < 2 {
nworkers = 2 nworkers = 2
} }
return b.DB.Flatten(nworkers) err := b.DB.Flatten(nworkers)
if err != nil {
return err
}
for err == nil {
err = b.DB.RunValueLogGC(0.125)
}
if err == badger.ErrNoRewrite {
// not really an error in this case, it signals the end of GC
return nil
}
return err
} }
// View implements blockstore.Viewer, which leverages zero-copy read-only // View implements blockstore.Viewer, which leverages zero-copy read-only

View File

@ -35,6 +35,11 @@ type BlockstoreIterator interface {
ForEachKey(func(cid.Cid) error) error ForEachKey(func(cid.Cid) error) error
} }
// BlockstoreGC is a trait for blockstores that support online garbage collection
type BlockstoreGC interface {
CollectGarbage() error
}
// WrapIDStore wraps the underlying blockstore in an "identity" blockstore. // WrapIDStore wraps the underlying blockstore in an "identity" blockstore.
// The ID store filters out all puts for blocks with CIDs using the "identity" // The ID store filters out all puts for blocks with CIDs using the "identity"
// hash function. It also extracts inlined blocks from CIDs using the identity // hash function. It also extracts inlined blocks from CIDs using the identity

View File

@ -72,6 +72,9 @@ var (
// this is first computed at warmup and updated in every compaction // this is first computed at warmup and updated in every compaction
markSetSizeKey = dstore.NewKey("/splitstore/markSetSize") markSetSizeKey = dstore.NewKey("/splitstore/markSetSize")
// compactionIndexKey stores the compaction index (serial number)
compactionIndexKey = dstore.NewKey("/splitstore/compactionIndex")
log = logging.Logger("splitstore") log = logging.Logger("splitstore")
// used to signal end of walk // used to signal end of walk
@ -140,6 +143,8 @@ type SplitStore struct {
markSetEnv MarkSetEnv markSetEnv MarkSetEnv
markSetSize int64 markSetSize int64
compactionIndex int64
ctx context.Context ctx context.Context
cancel func() cancel func()
@ -480,6 +485,9 @@ func (s *SplitStore) Start(chain ChainAccessor) error {
s.chain = chain s.chain = chain
curTs := chain.GetHeaviestTipSet() curTs := chain.GetHeaviestTipSet()
// should we warmup
warmup := false
// load base epoch from metadata ds // load base epoch from metadata ds
// if none, then use current epoch because it's a fresh start // if none, then use current epoch because it's a fresh start
bs, err := s.ds.Get(baseEpochKey) bs, err := s.ds.Get(baseEpochKey)
@ -509,11 +517,7 @@ func (s *SplitStore) Start(chain ChainAccessor) error {
s.warmupEpoch = bytesToEpoch(bs) s.warmupEpoch = bytesToEpoch(bs)
case dstore.ErrNotFound: case dstore.ErrNotFound:
// the hotstore hasn't warmed up, start a concurrent warm up warmup = true
err = s.warmup(curTs)
if err != nil {
return xerrors.Errorf("error warming up: %w", err)
}
default: default:
return xerrors.Errorf("error loading warmup epoch: %w", err) return xerrors.Errorf("error loading warmup epoch: %w", err)
@ -530,8 +534,29 @@ func (s *SplitStore) Start(chain ChainAccessor) error {
return xerrors.Errorf("error loading mark set size: %w", err) return xerrors.Errorf("error loading mark set size: %w", err)
} }
// load compactionIndex from metadata ds to provide a hint as to when to perform moving gc
bs, err = s.ds.Get(compactionIndexKey)
switch err {
case nil:
s.compactionIndex = bytesToInt64(bs)
case dstore.ErrNotFound:
// this is potentially an upgrade from splitstore v0; schedule a warmup as v0 has
// some issues with hot references leaking into the coldstore.
warmup = true
default:
return xerrors.Errorf("error loading compaction index: %w", err)
}
log.Infow("starting splitstore", "baseEpoch", s.baseEpoch, "warmupEpoch", s.warmupEpoch) log.Infow("starting splitstore", "baseEpoch", s.baseEpoch, "warmupEpoch", s.warmupEpoch)
if warmup {
err = s.warmup(curTs)
if err != nil {
return xerrors.Errorf("error starting warmup: %w", err)
}
}
// watch the chain // watch the chain
chain.SubscribeHeadChanges(s.HeadChange) chain.SubscribeHeadChanges(s.HeadChange)
@ -653,7 +678,7 @@ func (s *SplitStore) viewDone() {
s.txnViews-- s.txnViews--
if s.txnViews == 0 && s.txnViewsWaiting { if s.txnViews == 0 && s.txnViewsWaiting {
s.txnViewsCond.Signal() s.txnViewsCond.Broadcast()
} }
} }
@ -717,7 +742,7 @@ func (s *SplitStore) trackTxnRefMany(cids []cid.Cid) {
quiet = true quiet = true
log.Warnf("error checking markset: %s", err) log.Warnf("error checking markset: %s", err)
} }
continue // track it anyways
} }
if mark { if mark {
@ -943,6 +968,12 @@ func (s *SplitStore) doWarmup(curTs *types.TipSet) error {
s.warmupEpoch = epoch s.warmupEpoch = epoch
s.mx.Unlock() s.mx.Unlock()
// also save the compactionIndex, as this is used as an indicator of warmup for upgraded nodes
err = s.ds.Put(compactionIndexKey, int64ToBytes(s.compactionIndex))
if err != nil {
return xerrors.Errorf("error saving compaction index: %w", err)
}
return nil return nil
} }
@ -977,7 +1008,7 @@ func (s *SplitStore) doCompact(curTs *types.TipSet) error {
currentEpoch := curTs.Height() currentEpoch := curTs.Height()
boundaryEpoch := currentEpoch - CompactionBoundary boundaryEpoch := currentEpoch - CompactionBoundary
log.Infow("running compaction", "currentEpoch", currentEpoch, "baseEpoch", s.baseEpoch, "boundaryEpoch", boundaryEpoch) log.Infow("running compaction", "currentEpoch", currentEpoch, "baseEpoch", s.baseEpoch, "boundaryEpoch", boundaryEpoch, "compactionIndex", s.compactionIndex)
markSet, err := s.markSetEnv.Create("live", s.markSetSize) markSet, err := s.markSetEnv.Create("live", s.markSetSize)
if err != nil { if err != nil {
@ -994,7 +1025,7 @@ func (s *SplitStore) doCompact(curTs *types.TipSet) error {
s.beginTxnMarking(markSet) s.beginTxnMarking(markSet)
// 1. mark reachable objects by walking the chain from the current epoch; we keep state roots // 1. mark reachable objects by walking the chain from the current epoch; we keep state roots
// and messages until the boundary epoch. // and messages until the boundary epoch.nn
log.Info("marking reachable objects") log.Info("marking reachable objects")
startMark := time.Now() startMark := time.Now()
@ -1147,6 +1178,12 @@ func (s *SplitStore) doCompact(curTs *types.TipSet) error {
return xerrors.Errorf("error saving mark set size: %w", err) return xerrors.Errorf("error saving mark set size: %w", err)
} }
s.compactionIndex++
err = s.ds.Put(compactionIndexKey, int64ToBytes(s.compactionIndex))
if err != nil {
return xerrors.Errorf("error saving compaction index: %w", err)
}
return nil return nil
} }
@ -1402,8 +1439,7 @@ func (s *SplitStore) has(c cid.Cid) (bool, error) {
func (s *SplitStore) checkClosing() error { func (s *SplitStore) checkClosing() error {
if atomic.LoadInt32(&s.closing) == 1 { if atomic.LoadInt32(&s.closing) == 1 {
log.Info("splitstore is closing; aborting compaction") return xerrors.Errorf("splitstore is closing")
return xerrors.Errorf("compaction aborted")
} }
return nil return nil
@ -1697,30 +1733,6 @@ func (s *SplitStore) waitForMissingRefs(markSet MarkSet) {
} }
} }
func (s *SplitStore) gcHotstore() {
if compact, ok := s.hot.(interface{ Compact() error }); ok {
log.Infof("compacting hotstore")
startCompact := time.Now()
err := compact.Compact()
if err != nil {
log.Warnf("error compacting hotstore: %s", err)
return
}
log.Infow("hotstore compaction done", "took", time.Since(startCompact))
}
if gc, ok := s.hot.(interface{ CollectGarbage() error }); ok {
log.Infof("garbage collecting hotstore")
startGC := time.Now()
err := gc.CollectGarbage()
if err != nil {
log.Warnf("error garbage collecting hotstore: %s", err)
return
}
log.Infow("hotstore garbage collection done", "took", time.Since(startGC))
}
}
func (s *SplitStore) setBaseEpoch(epoch abi.ChainEpoch) error { func (s *SplitStore) setBaseEpoch(epoch abi.ChainEpoch) error {
s.baseEpoch = epoch s.baseEpoch = epoch
return s.ds.Put(baseEpochKey, epochToBytes(epoch)) return s.ds.Put(baseEpochKey, epochToBytes(epoch))

View File

@ -0,0 +1,30 @@
package splitstore
import (
"fmt"
"time"
bstore "github.com/filecoin-project/lotus/blockstore"
)
func (s *SplitStore) gcHotstore() {
if err := s.gcBlockstoreOnline(s.hot); err != nil {
log.Warnf("error garbage collecting hostore: %s", err)
}
}
func (s *SplitStore) gcBlockstoreOnline(b bstore.Blockstore) error {
if gc, ok := b.(bstore.BlockstoreGC); ok {
log.Info("garbage collecting blockstore")
startGC := time.Now()
if err := gc.CollectGarbage(); err != nil {
return err
}
log.Infow("garbage collecting hotstore done", "took", time.Since(startGC))
return nil
}
return fmt.Errorf("blockstore doesn't support online gc: %T", b)
}