Merge pull request #2818 from aawsome/merge-index-chaining

Merge index (based on chaining index implementation)
This commit is contained in:
MichaelEischer 2020-07-22 22:05:36 +02:00 committed by GitHub
commit bcd47ec3a2
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 170 additions and 57 deletions

View File

@ -190,6 +190,8 @@ func (c *Checker) LoadIndex(ctx context.Context) (hints []error, errs []error) {
}
}
c.masterIndex.MergeFinalIndexes()
err = c.repo.SetIndex(c.masterIndex)
if err != nil {
debug.Log("SetIndex returned error: %v", err)

View File

@ -49,8 +49,8 @@ type Index struct {
// only used by Store, StorePacks does not check for already saved packIDs
packIDToIndex map[restic.ID]int
final bool // set to true for all indexes read from the backend ("finalized")
id restic.ID // set to the ID of the index when it's finalized
final bool // set to true for all indexes read from the backend ("finalized")
ids restic.IDs // set to the IDs of the contained finalized indexes
supersedes restic.IDs
created time.Time
}
@ -393,17 +393,17 @@ func (idx *Index) Finalize() {
idx.packIDToIndex = nil
}
// ID returns the ID of the index, if available. If the index is not yet
// ID returns the IDs of the index, if available. If the index is not yet
// finalized, an error is returned.
func (idx *Index) ID() (restic.ID, error) {
func (idx *Index) IDs() (restic.IDs, error) {
idx.m.Lock()
defer idx.m.Unlock()
if !idx.final {
return restic.ID{}, errors.New("index not finalized")
return nil, errors.New("index not finalized")
}
return idx.id, nil
return idx.ids, nil
}
// SetID sets the ID the index has been written to. This requires that
@ -416,12 +416,12 @@ func (idx *Index) SetID(id restic.ID) error {
return errors.New("index is not final")
}
if !idx.id.IsNull() {
if len(idx.ids) > 0 {
return errors.New("ID already set")
}
debug.Log("ID set to %v", id)
idx.id = id
idx.ids = append(idx.ids, id)
return nil
}
@ -462,6 +462,38 @@ func (idx *Index) TreePacks() restic.IDs {
return idx.treePacks
}
// merge() merges indexes, i.e. idx.merge(idx2) merges the contents of idx2 into idx.
// idx2 is not changed by this method.
func (idx *Index) merge(idx2 *Index) error {
idx.m.Lock()
defer idx.m.Unlock()
idx2.m.Lock()
defer idx2.m.Unlock()
if !idx2.final {
return errors.New("index to merge is not final!")
}
packlen := len(idx.packs)
// copy all index entries of idx2 to idx
for typ := range idx2.byType {
m2 := &idx2.byType[typ]
m := &idx.byType[typ]
m2.foreach(func(entry *indexEntry) bool {
// packIndex is changed as idx2.pack is appended to idx.pack, see below
m.add(entry.id, entry.packIndex+packlen, entry.offset, entry.length)
return true
})
}
idx.packs = append(idx.packs, idx2.packs...)
idx.treePacks = append(idx.treePacks, idx2.treePacks...)
idx.ids = append(idx.ids, idx2.ids...)
idx.supersedes = append(idx.supersedes, idx2.supersedes...)
return nil
}
// isErrOldIndex returns true if the error may be caused by an old index
// format.
func isErrOldIndex(err error) bool {
@ -581,7 +613,7 @@ func LoadIndexWithDecoder(ctx context.Context, repo restic.Repository, buf []byt
return nil, buf[:0], err
}
idx.id = id
idx.ids = append(idx.ids, id)
return idx, buf, nil
}

View File

@ -135,10 +135,9 @@ func TestIndexSerialize(t *testing.T) {
id := restic.NewRandomID()
rtest.OK(t, idx.SetID(id))
id2, err := idx.ID()
ids, err := idx.IDs()
rtest.OK(t, err)
rtest.Assert(t, id2.Equal(id),
"wrong ID returned: want %v, got %v", id, id2)
rtest.Equals(t, restic.IDs{id}, ids)
idx3, err := repository.DecodeIndex(wr3.Bytes())
rtest.OK(t, err)
@ -336,7 +335,7 @@ var (
)
func initBenchmarkIndexJSON() {
idx, _ := createRandomIndex(rand.New(rand.NewSource(0)))
idx, _ := createRandomIndex(rand.New(rand.NewSource(0)), 200000)
var buf bytes.Buffer
idx.Encode(&buf)
benchmarkIndexJSON = buf.Bytes()
@ -418,11 +417,11 @@ func NewRandomTestID(rng *rand.Rand) restic.ID {
return id
}
func createRandomIndex(rng *rand.Rand) (idx *repository.Index, lookupID restic.ID) {
func createRandomIndex(rng *rand.Rand, packfiles int) (idx *repository.Index, lookupID restic.ID) {
idx = repository.NewIndex()
// create index with 200k pack files
for i := 0; i < 200000; i++ {
// create index with given number of pack files
for i := 0; i < packfiles; i++ {
packID := NewRandomTestID(rng)
var blobs []restic.Blob
offset := 0
@ -449,7 +448,7 @@ func createRandomIndex(rng *rand.Rand) (idx *repository.Index, lookupID restic.I
}
func BenchmarkIndexHasUnknown(b *testing.B) {
idx, _ := createRandomIndex(rand.New(rand.NewSource(0)))
idx, _ := createRandomIndex(rand.New(rand.NewSource(0)), 200000)
lookupID := restic.NewRandomID()
b.ResetTimer()
@ -460,7 +459,7 @@ func BenchmarkIndexHasUnknown(b *testing.B) {
}
func BenchmarkIndexHasKnown(b *testing.B) {
idx, lookupID := createRandomIndex(rand.New(rand.NewSource(0)))
idx, lookupID := createRandomIndex(rand.New(rand.NewSource(0)), 200000)
b.ResetTimer()
@ -474,7 +473,7 @@ func BenchmarkIndexAlloc(b *testing.B) {
b.ReportAllocs()
for i := 0; i < b.N; i++ {
createRandomIndex(rng)
createRandomIndex(rng, 200000)
}
}
@ -484,7 +483,7 @@ func BenchmarkIndexAllocParallel(b *testing.B) {
b.RunParallel(func(pb *testing.PB) {
rng := rand.New(rand.NewSource(0))
for pb.Next() {
createRandomIndex(rng)
createRandomIndex(rng, 200000)
}
})
}

View File

@ -18,7 +18,12 @@ type MasterIndex struct {
// NewMasterIndex creates a new master index.
func NewMasterIndex() *MasterIndex {
return &MasterIndex{pendingBlobs: restic.NewBlobSet()}
// Always add an empty final index, such that MergeFinalIndexes can merge into this.
// Note that removing this index could lead to a race condition in the rare
// sitation that only two indexes exist which are saved and merged concurrently.
idx := []*Index{NewIndex()}
idx[0].Finalize()
return &MasterIndex{idx: idx, pendingBlobs: restic.NewBlobSet()}
}
// Lookup queries all known Indexes for the ID and returns the first match.
@ -237,6 +242,31 @@ func (mi *MasterIndex) Each(ctx context.Context) <-chan restic.PackedBlob {
return ch
}
// MergeFinalIndexes merges all final indexes together.
// After calling, there will be only one big final index in MasterIndex
// containing all final index contents.
// Indexes that are not final are left untouched.
// This merging can only be called after all index files are loaded - as
// removing of superseded index contents is only possible for unmerged indexes.
func (mi *MasterIndex) MergeFinalIndexes() {
mi.idxMutex.Lock()
defer mi.idxMutex.Unlock()
// The first index is always final and the one to merge into
newIdx := mi.idx[:1]
for i := 1; i < len(mi.idx); i++ {
idx := mi.idx[i]
// clear reference in masterindex as it may become stale
mi.idx[i] = nil
if !idx.Final() {
newIdx = append(newIdx, idx)
} else {
mi.idx[0].merge(idx)
}
}
mi.idx = newIdx
}
// RebuildIndex combines all known indexes to a new index, leaving out any
// packs whose ID is contained in packBlacklist. The new index contains the IDs
// of all known indexes in the "supersedes" field.
@ -267,15 +297,15 @@ func (mi *MasterIndex) RebuildIndex(packBlacklist restic.IDSet) (*Index, error)
continue
}
id, err := idx.ID()
ids, err := idx.IDs()
if err != nil {
debug.Log("index %d does not have an ID: %v", err)
return nil, err
}
debug.Log("adding index id %v to supersedes field", id)
debug.Log("adding index ids %v to supersedes field", ids)
err = newIndex.AddToSupersedes(id)
err = newIndex.AddToSupersedes(ids...)
if err != nil {
return nil, err
}

View File

@ -57,11 +57,75 @@ func TestMasterIndexLookup(t *testing.T) {
rtest.Assert(t, blobs == nil, "Expected no blobs when fetching with a random id")
}
func BenchmarkMasterIndexLookupSingleIndex(b *testing.B) {
idx1, lookupID := createRandomIndex(rand.New(rand.NewSource(0)))
func TestMasterMergeFinalIndexes(t *testing.T) {
idInIdx1 := restic.NewRandomID()
idInIdx2 := restic.NewRandomID()
blob1 := restic.PackedBlob{
PackID: restic.NewRandomID(),
Blob: restic.Blob{
Type: restic.DataBlob,
ID: idInIdx1,
Length: 10,
Offset: 0,
},
}
blob2 := restic.PackedBlob{
PackID: restic.NewRandomID(),
Blob: restic.Blob{
Type: restic.DataBlob,
ID: idInIdx2,
Length: 100,
Offset: 10,
},
}
idx1 := repository.NewIndex()
idx1.Store(blob1)
idx2 := repository.NewIndex()
idx2.Store(blob2)
mIdx := repository.NewMasterIndex()
mIdx.Insert(idx1)
mIdx.Insert(idx2)
finalIndexes := mIdx.FinalizeNotFinalIndexes()
rtest.Equals(t, []*repository.Index{idx1, idx2}, finalIndexes)
mIdx.MergeFinalIndexes()
blobs, found := mIdx.Lookup(idInIdx1, restic.DataBlob)
rtest.Assert(t, found, "Expected to find blob id %v from index 1", idInIdx1)
rtest.Equals(t, []restic.PackedBlob{blob1}, blobs)
blobs, found = mIdx.Lookup(idInIdx2, restic.DataBlob)
rtest.Assert(t, found, "Expected to find blob id %v from index 2", idInIdx2)
rtest.Equals(t, []restic.PackedBlob{blob2}, blobs)
blobs, found = mIdx.Lookup(restic.NewRandomID(), restic.DataBlob)
rtest.Assert(t, !found, "Expected to not find a blob when fetching with a random id")
rtest.Assert(t, blobs == nil, "Expected no blobs when fetching with a random id")
}
func createRandomMasterIndex(rng *rand.Rand, num, size int) (*repository.MasterIndex, restic.ID) {
mIdx := repository.NewMasterIndex()
for i := 0; i < num-1; i++ {
idx, _ := createRandomIndex(rng, size)
mIdx.Insert(idx)
}
idx1, lookupID := createRandomIndex(rng, size)
mIdx.Insert(idx1)
mIdx.FinalizeNotFinalIndexes()
mIdx.MergeFinalIndexes()
return mIdx, lookupID
}
func BenchmarkMasterIndexLookupSingleIndex(b *testing.B) {
mIdx, lookupID := createRandomMasterIndex(rand.New(rand.NewSource(0)), 1, 200000)
b.ResetTimer()
@ -71,16 +135,7 @@ func BenchmarkMasterIndexLookupSingleIndex(b *testing.B) {
}
func BenchmarkMasterIndexLookupMultipleIndex(b *testing.B) {
rng := rand.New(rand.NewSource(0))
mIdx := repository.NewMasterIndex()
for i := 0; i < 5; i++ {
idx, _ := createRandomIndex(rng)
mIdx.Insert(idx)
}
idx1, lookupID := createRandomIndex(rng)
mIdx.Insert(idx1)
mIdx, lookupID := createRandomMasterIndex(rand.New(rand.NewSource(0)), 100, 10000)
b.ResetTimer()
@ -90,11 +145,9 @@ func BenchmarkMasterIndexLookupMultipleIndex(b *testing.B) {
}
func BenchmarkMasterIndexLookupSingleIndexUnknown(b *testing.B) {
lookupID := restic.NewRandomID()
idx1, _ := createRandomIndex(rand.New(rand.NewSource(0)))
mIdx := repository.NewMasterIndex()
mIdx.Insert(idx1)
lookupID := restic.NewRandomID()
mIdx, _ := createRandomMasterIndex(rand.New(rand.NewSource(0)), 1, 200000)
b.ResetTimer()
@ -104,14 +157,8 @@ func BenchmarkMasterIndexLookupSingleIndexUnknown(b *testing.B) {
}
func BenchmarkMasterIndexLookupMultipleIndexUnknown(b *testing.B) {
rng := rand.New(rand.NewSource(0))
lookupID := restic.NewRandomID()
mIdx := repository.NewMasterIndex()
for i := 0; i < 6; i++ {
idx, _ := createRandomIndex(rng)
mIdx.Insert(idx)
}
mIdx, _ := createRandomMasterIndex(rand.New(rand.NewSource(0)), 100, 10000)
b.ResetTimer()
@ -123,16 +170,12 @@ func BenchmarkMasterIndexLookupMultipleIndexUnknown(b *testing.B) {
func BenchmarkMasterIndexLookupParallel(b *testing.B) {
mIdx := repository.NewMasterIndex()
for _, numindices := range []int{5, 10, 20} {
for _, numindices := range []int{25, 50, 100} {
var lookupID restic.ID
b.StopTimer()
rng := rand.New(rand.NewSource(0))
for i := 0; i < numindices; i++ {
var idx *repository.Index
idx, lookupID = createRandomIndex(rng)
mIdx.Insert(idx)
}
mIdx, lookupID = createRandomMasterIndex(rng, numindices, 10000)
b.StartTimer()
name := fmt.Sprintf("known,indices=%d", numindices)

View File

@ -361,12 +361,14 @@ func (r *Repository) SetIndex(i restic.Index) error {
ids := restic.NewIDSet()
for _, idx := range r.idx.All() {
id, err := idx.ID()
indexIDs, err := idx.IDs()
if err != nil {
debug.Log("not using index, ID() returned error %v", err)
continue
}
ids.Insert(id)
for _, id := range indexIDs {
ids.Insert(id)
}
}
return r.PrepareCache(ids)
@ -396,6 +398,7 @@ func (r *Repository) saveIndex(ctx context.Context, indexes ...*Index) error {
debug.Log("Saved index %d as %v", i, sid)
}
r.idx.MergeFinalIndexes()
return nil
}
@ -479,12 +482,16 @@ func (r *Repository) LoadIndex(ctx context.Context) error {
validIndex := restic.NewIDSet()
wg.Go(func() error {
for idx := range indexCh {
id, err := idx.ID()
ids, err := idx.IDs()
if err == nil {
validIndex.Insert(id)
for _, id := range ids {
validIndex.Insert(id)
}
}
r.idx.Insert(idx)
}
r.idx.MergeFinalIndexes()
return nil
})