func (db *testMetadataDB) splitRange(t *testing.T, key engine.Key) {
metadataKey := engine.RangeMetaKey(key)
v := db.data.Ceil(&testMetadataNode{endKey: metadataKey})
if v == nil {
t.Fatalf("Error splitting range at key %s, range to split not found", string(key))
}
val := v.(*testMetadataNode)
if bytes.Compare(val.desc.EndKey, key) == 0 {
t.Fatalf("Attempt to split existing range at Endkey: %s", string(key))
}
db.data.Insert(&testMetadataNode{
endKey: metadataKey,
desc: &storage.RangeDescriptor{
StartKey: val.desc.StartKey,
EndKey: key,
},
})
db.data.Insert(&testMetadataNode{
endKey: val.endKey,
desc: &storage.RangeDescriptor{
StartKey: key,
EndKey: val.desc.EndKey,
},
})
}
// TestRangeCache is a simple test which verifies that metadata ranges are being
// cached and retrieved properly. It sets up a fake backing store for the
// cache, and measures how often that backing store is accessed when looking up
// metadata keys through the cache.
func TestRangeCache(t *testing.T) {
db := newTestMetadataDB()
db.splitRange(t, engine.Key("a"))
db.splitRange(t, engine.Key("b"))
db.splitRange(t, engine.Key("c"))
db.splitRange(t, engine.Key("d"))
db.splitRange(t, engine.Key("e"))
db.splitRange(t, engine.Key("f"))
db.splitRange(t, engine.RangeMetaKey(engine.Key("d")))
db.hitCount = 0
rangeCache := NewRangeMetadataCache(db)
db.cache = rangeCache
doLookup(t, rangeCache, "ba")
db.assertHitCount(t, 2)
doLookup(t, rangeCache, "bb")
db.assertHitCount(t, 0)
doLookup(t, rangeCache, "ca")
db.assertHitCount(t, 1)
// Different metadata one range
doLookup(t, rangeCache, "da")
db.assertHitCount(t, 2)
doLookup(t, rangeCache, "fa")
db.assertHitCount(t, 1)
// Evict clears both level 1 and level 2 cache for a key
rangeCache.EvictCachedRangeMetadata(engine.Key("da"))
doLookup(t, rangeCache, "fa")
db.assertHitCount(t, 0)
doLookup(t, rangeCache, "da")
db.assertHitCount(t, 2)
}
func (db *testMetadataDB) getMetadata(key engine.Key) (engine.Key, *storage.RangeDescriptor, error) {
metadataKey := engine.RangeMetaKey(key)
v := db.data.Ceil(&testMetadataNode{endKey: metadataKey})
if v == nil {
return nil, nil, util.Errorf("Range for key %s not found", key)
}
val := v.(*testMetadataNode)
db.hitCount++
return val.endKey, val.desc, nil
}
func (db *testMetadataDB) getRangeMetadata(key engine.Key) ([]proto.RangeDescriptor, error) {
db.hitCount++
metadataKey := engine.RangeMetaKey(key)
// Recursively call into cache as the real DB would, terminating recursion
// when a meta1key is encountered.
if len(metadataKey) > 0 && !bytes.HasPrefix(metadataKey, engine.KeyMeta1Prefix) {
db.cache.LookupRangeMetadata(metadataKey)
}
return db.getMetadata(key), nil
}
func (db *testMetadataDB) LookupRangeMetadata(key engine.Key) (engine.Key, *storage.RangeDescriptor, error) {
metadataKey := engine.RangeMetaKey(key)
// Recursively call into cache as the real DB would, terminating when the
// initial key is encountered.
if len(metadataKey) == 0 {
return nil, nil, nil
}
db.cache.LookupRangeMetadata(metadataKey)
return db.getMetadata(key)
}
// TestRangeCache is a simple test which verifies that metadata ranges
// are being cached and retrieved properly. It sets up a fake backing
// store for the cache, and measures how often that backing store is
// accessed when looking up metadata keys through the cache.
func TestRangeCache(t *testing.T) {
db := newTestMetadataDB()
for i, char := range "abcdefghijklmnopqrstuvwx" {
db.splitRange(t, engine.Key(string(char)))
if i > 0 && i%6 == 0 {
db.splitRange(t, engine.RangeMetaKey(engine.Key(string(char))))
}
}
rangeCache := NewRangeMetadataCache(db)
db.cache = rangeCache
doLookup(t, rangeCache, "aa")
db.assertHitCount(t, 2)
// Metadata for the following ranges should be cached
doLookup(t, rangeCache, "ab")
db.assertHitCount(t, 0)
doLookup(t, rangeCache, "ba")
db.assertHitCount(t, 0)
doLookup(t, rangeCache, "cz")
db.assertHitCount(t, 0)
// Metadata two ranges weren't cached, same metadata 1 range
doLookup(t, rangeCache, "d")
db.assertHitCount(t, 1)
doLookup(t, rangeCache, "fa")
db.assertHitCount(t, 0)
// Metadata two ranges weren't cached, metadata 1 was aggressively cached
doLookup(t, rangeCache, "ij")
db.assertHitCount(t, 1)
doLookup(t, rangeCache, "jk")
db.assertHitCount(t, 0)
doLookup(t, rangeCache, "pn")
db.assertHitCount(t, 1)
// Totally uncached ranges
doLookup(t, rangeCache, "vu")
db.assertHitCount(t, 2)
doLookup(t, rangeCache, "xx")
db.assertHitCount(t, 0)
// Evict clears one level 1 and one level 2 cache
rangeCache.EvictCachedRangeMetadata(engine.Key("da"))
doLookup(t, rangeCache, "fa")
db.assertHitCount(t, 0)
doLookup(t, rangeCache, "da")
db.assertHitCount(t, 2)
}
// EvictCachedRangeDescriptor will evict any cached range descriptors
// for the given key. It is intended that this method be called from a
// consumer of RangeDescriptorCache if the returned range descriptor is
// discovered to be stale.
func (rmc *RangeDescriptorCache) EvictCachedRangeDescriptor(key proto.Key) {
for {
k, _ := rmc.getCachedRangeDescriptor(key)
if k != nil {
rmc.rangeCacheMu.Lock()
rmc.rangeCache.Del(k)
rmc.rangeCacheMu.Unlock()
}
// Retrieve the metadata range key for the next level of metadata, and
// evict that key as well. This loop ends after the meta1 range, which
// returns KeyMin as its metadata key.
key = engine.RangeMetaKey(key)
if len(key) == 0 {
break
}
}
}
// getCachedRangeDescriptor is a helper function to retrieve the
// descriptor of the range which contains the given key, if present in
// the cache.
func (rmc *RangeDescriptorCache) getCachedRangeDescriptor(key proto.Key) (
rangeCacheKey, *proto.RangeDescriptor) {
metaKey := engine.RangeMetaKey(key)
rmc.rangeCacheMu.RLock()
defer rmc.rangeCacheMu.RUnlock()
k, v, ok := rmc.rangeCache.Ceil(rangeCacheKey(metaKey))
if !ok {
return nil, nil
}
metaEndKey := k.(rangeCacheKey)
rd := v.(*proto.RangeDescriptor)
// Check that key actually belongs to range
if !rd.ContainsKey(engine.KeyAddress(key)) {
return nil, nil
}
return metaEndKey, rd
}
// getRangeMetadata retrieves metadata for the range containing the given key
// from storage. This function returns a sorted slice of RangeDescriptors for a
// set of consecutive ranges, the first which must contain the requested key.
// The additional RangeDescriptors are returned with the intent of pre-caching
// subsequent ranges which are likely to be requested soon by the current
// workload.
func (kv *DistKV) getRangeMetadata(key engine.Key) ([]proto.RangeDescriptor, error) {
var (
// metadataKey is sent to InternalRangeLookup to find the
// RangeDescriptor which contains key.
metadataKey = engine.RangeMetaKey(key)
// metadataRange is the RangeDescriptor for the range which contains
// metadataKey.
metadataRange *proto.RangeDescriptor
err error
)
if len(metadataKey) == 0 {
// In this case, the requested key is stored in the cluster's first
// range. Return the first range, which is always gossiped and not
// queried from the datastore.
rd, err := kv.getFirstRangeDescriptor()
if err != nil {
return nil, err
}
return []proto.RangeDescriptor{*rd}, nil
}
if bytes.HasPrefix(metadataKey, engine.KeyMeta1Prefix) {
// In this case, metadataRange is the cluster's first range.
if metadataRange, err = kv.getFirstRangeDescriptor(); err != nil {
return nil, err
}
} else {
// Look up metadataRange from the cache, which will recursively call
// into kv.getRangeMetadata if it is not cached.
metadataRange, err = kv.rangeCache.LookupRangeMetadata(metadataKey)
if err != nil {
return nil, err
}
}
return kv.internalRangeLookup(metadataKey, metadataRange)
}
// TestUpdateRangeAddressing verifies range addressing records are
// correctly updated on creation of new range descriptors.
func TestUpdateRangeAddressing(t *testing.T) {
store := createTestStore(t)
// When split is false, merging treats the right range as the merged
// range. With merging, expNewLeft indicates the addressing keys we
// expect to be removed.
testCases := []struct {
split bool
leftStart, leftEnd proto.Key
rightStart, rightEnd proto.Key
leftExpNew, rightExpNew []proto.Key
}{
// Start out with whole range.
{false, engine.KeyMin, engine.KeyMax, engine.KeyMin, engine.KeyMax,
[]proto.Key{}, []proto.Key{meta1Key(engine.KeyMax), meta2Key(engine.KeyMax)}},
// Split KeyMin-KeyMax at key "a".
{true, engine.KeyMin, proto.Key("a"), proto.Key("a"), engine.KeyMax,
[]proto.Key{meta1Key(engine.KeyMax), meta2Key(proto.Key("a"))}, []proto.Key{meta2Key(engine.KeyMax)}},
// Split "a"-KeyMax at key "z".
{true, proto.Key("a"), proto.Key("z"), proto.Key("z"), engine.KeyMax,
[]proto.Key{meta2Key(proto.Key("z"))}, []proto.Key{meta2Key(engine.KeyMax)}},
// Split "a"-"z" at key "m".
{true, proto.Key("a"), proto.Key("m"), proto.Key("m"), proto.Key("z"),
[]proto.Key{meta2Key(proto.Key("m"))}, []proto.Key{meta2Key(proto.Key("z"))}},
// Split KeyMin-"a" at meta2(m).
{true, engine.KeyMin, engine.RangeMetaKey(proto.Key("m")), engine.RangeMetaKey(proto.Key("m")), proto.Key("a"),
[]proto.Key{meta1Key(proto.Key("m"))}, []proto.Key{meta1Key(engine.KeyMax), meta2Key(proto.Key("a"))}},
// Split meta2(m)-"a" at meta2(z).
{true, engine.RangeMetaKey(proto.Key("m")), engine.RangeMetaKey(proto.Key("z")), engine.RangeMetaKey(proto.Key("z")), proto.Key("a"),
[]proto.Key{meta1Key(proto.Key("z"))}, []proto.Key{meta1Key(engine.KeyMax), meta2Key(proto.Key("a"))}},
// Split meta2(m)-meta2(z) at meta2(r).
{true, engine.RangeMetaKey(proto.Key("m")), engine.RangeMetaKey(proto.Key("r")), engine.RangeMetaKey(proto.Key("r")), engine.RangeMetaKey(proto.Key("z")),
[]proto.Key{meta1Key(proto.Key("r"))}, []proto.Key{meta1Key(proto.Key("z"))}},
// Now, merge all of our splits backwards...
// Merge meta2(m)-meta2(z).
{false, engine.RangeMetaKey(proto.Key("m")), engine.RangeMetaKey(proto.Key("r")), engine.RangeMetaKey(proto.Key("m")), engine.RangeMetaKey(proto.Key("z")),
[]proto.Key{meta1Key(proto.Key("r"))}, []proto.Key{meta1Key(proto.Key("z"))}},
// Merge meta2(m)-"a".
{false, engine.RangeMetaKey(proto.Key("m")), engine.RangeMetaKey(proto.Key("z")), engine.RangeMetaKey(proto.Key("m")), proto.Key("a"),
[]proto.Key{meta1Key(proto.Key("z"))}, []proto.Key{meta1Key(engine.KeyMax), meta2Key(proto.Key("a"))}},
// Merge KeyMin-"a".
{false, engine.KeyMin, engine.RangeMetaKey(proto.Key("m")), engine.KeyMin, proto.Key("a"),
[]proto.Key{meta1Key(proto.Key("m"))}, []proto.Key{meta1Key(engine.KeyMax), meta2Key(proto.Key("a"))}},
// Merge "a"-"z".
{false, proto.Key("a"), proto.Key("m"), proto.Key("a"), proto.Key("z"),
[]proto.Key{meta2Key(proto.Key("m"))}, []proto.Key{meta2Key(proto.Key("z"))}},
// Merge "a"-KeyMax.
{false, proto.Key("a"), proto.Key("z"), proto.Key("a"), engine.KeyMax,
[]proto.Key{meta2Key(proto.Key("z"))}, []proto.Key{meta2Key(engine.KeyMax)}},
// Merge KeyMin-KeyMax.
{false, engine.KeyMin, proto.Key("a"), engine.KeyMin, engine.KeyMax,
[]proto.Key{meta2Key(proto.Key("a"))}, []proto.Key{meta1Key(engine.KeyMax), meta2Key(engine.KeyMax)}},
}
expMetas := metaSlice{}
for i, test := range testCases {
left := &proto.RangeDescriptor{RaftID: int64(i * 2), StartKey: test.leftStart, EndKey: test.leftEnd}
right := &proto.RangeDescriptor{RaftID: int64(i*2 + 1), StartKey: test.rightStart, EndKey: test.rightEnd}
if test.split {
if err := storage.SplitRangeAddressing(store.DB(), left, right); err != nil {
t.Fatal(err)
}
} else {
if err := storage.MergeRangeAddressing(store.DB(), left, right); err != nil {
t.Fatal(err)
}
}
store.DB().Flush()
// Scan meta keys directly from engine.
kvs, err := engine.MVCCScan(store.Engine(), engine.KeyMetaPrefix, engine.KeyMetaMax, 0, proto.MaxTimestamp, nil)
if err != nil {
t.Fatal(err)
}
metas := metaSlice{}
for _, kv := range kvs {
scannedDesc := &proto.RangeDescriptor{}
if err := gogoproto.Unmarshal(kv.Value.Bytes, scannedDesc); err != nil {
t.Fatal(err)
}
metas = append(metas, metaRecord{key: kv.Key, desc: scannedDesc})
}
// Continue to build up the expected metas slice, replacing any earlier
// version of same key.
addOrRemoveNew := func(keys []proto.Key, desc *proto.RangeDescriptor, add bool) {
for _, n := range keys {
found := -1
for i := range expMetas {
if expMetas[i].key.Equal(n) {
found = i
expMetas[i].desc = desc
break
}
}
if found == -1 && add {
expMetas = append(expMetas, metaRecord{key: n, desc: desc})
} else if found != -1 && !add {
expMetas = append(expMetas[:found], expMetas[found+1:]...)
}
}
}
addOrRemoveNew(test.leftExpNew, left, test.split /* on split, add; on merge, remove */)
addOrRemoveNew(test.rightExpNew, right, true)
sort.Sort(expMetas)
if test.split {
log.V(1).Infof("test case %d: split %q-%q at %q", i, left.StartKey, right.EndKey, left.EndKey)
} else {
log.V(1).Infof("test case %d: merge %q-%q + %q-%q", i, left.StartKey, left.EndKey, left.EndKey, right.EndKey)
}
for _, meta := range metas {
log.V(1).Infof("%q", meta.key)
}
log.V(1).Infof("")
if !reflect.DeepEqual(expMetas, metas) {
t.Errorf("expected metas don't match")
if len(expMetas) != len(metas) {
t.Errorf("len(expMetas) != len(metas); %d != %d", len(expMetas), len(metas))
} else {
for i, meta := range expMetas {
if !meta.key.Equal(metas[i].key) {
fmt.Printf("%d: expected %q vs %q\n", i, meta.key, metas[i].key)
}
if !reflect.DeepEqual(meta.desc, metas[i].desc) {
fmt.Printf("%d: expected %q vs %q and %s vs %s\n", i, meta.key, metas[i].key, meta.desc, metas[i].desc)
}
}
}
}
}
}