// TestReplicateRange verifies basic replication functionality by creating two stores
// and a range, replicating the range to the second store, and reading its data there.
func TestReplicateRange(t *testing.T) {
defer leaktest.AfterTest(t)
mtc := multiTestContext{}
mtc.Start(t, 2)
defer mtc.Stop()
// Issue a command on the first node before replicating.
incArgs, incResp := incrementArgs([]byte("a"), 5, 1, mtc.stores[0].StoreID())
if err := mtc.stores[0].ExecuteCmd(context.Background(), proto.Call{Args: incArgs, Reply: incResp}); err != nil {
t.Fatal(err)
}
rng, err := mtc.stores[0].GetRange(1)
if err != nil {
t.Fatal(err)
}
if err := rng.ChangeReplicas(proto.ADD_REPLICA,
proto.Replica{
NodeID: mtc.stores[1].Ident.NodeID,
StoreID: mtc.stores[1].Ident.StoreID,
}); err != nil {
t.Fatal(err)
}
// Verify no intent remains on range descriptor key.
key := keys.RangeDescriptorKey(rng.Desc().StartKey)
desc := proto.RangeDescriptor{}
if ok, err := engine.MVCCGetProto(mtc.stores[0].Engine(), key, mtc.stores[0].Clock().Now(), true, nil, &desc); !ok || err != nil {
t.Fatalf("fetching range descriptor yielded %t, %s", ok, err)
}
// Verify that in time, no intents remain on meta addressing
// keys, and that range descriptor on the meta records is correct.
util.SucceedsWithin(t, 1*time.Second, func() error {
meta2 := keys.RangeMetaKey(proto.KeyMax)
meta1 := keys.RangeMetaKey(meta2)
for _, key := range []proto.Key{meta2, meta1} {
metaDesc := proto.RangeDescriptor{}
if ok, err := engine.MVCCGetProto(mtc.stores[0].Engine(), key, mtc.stores[0].Clock().Now(), true, nil, &metaDesc); !ok || err != nil {
return util.Errorf("failed to resolve %s", key)
}
if !reflect.DeepEqual(metaDesc, desc) {
return util.Errorf("descs not equal: %+v != %+v", metaDesc, desc)
}
}
return nil
})
// Verify that the same data is available on the replica.
util.SucceedsWithin(t, 1*time.Second, func() error {
getArgs, getResp := getArgs([]byte("a"), 1, mtc.stores[1].StoreID())
getArgs.ReadConsistency = proto.INCONSISTENT
if err := mtc.stores[1].ExecuteCmd(context.Background(), proto.Call{Args: getArgs, Reply: getResp}); err != nil {
return util.Errorf("failed to read data")
}
if v := mustGetInteger(getResp.Value); v != 5 {
return util.Errorf("failed to read correct data: %d", v)
}
return nil
})
}
func TestRangeCacheAssumptions(t *testing.T) {
defer leaktest.AfterTest(t)
expKeyMin := keys.RangeMetaKey(keys.RangeMetaKey(keys.RangeMetaKey(proto.Key("test"))))
if !bytes.Equal(expKeyMin, proto.KeyMin) {
t.Fatalf("RangeCache relies on RangeMetaKey returning KeyMin after two levels, but got %s", expKeyMin)
}
}
// TestRangeCacheClearOverlappingMeta prevents regression of a bug which caused
// a panic when clearing overlapping descriptors for [KeyMin, Meta2Key). The
// issue was that when attempting to clear out descriptors which were subsumed
// by the above range, an iteration over the corresponding meta keys was
// performed, with the left endpoint excluded. This exclusion was incorrect: it
// first incremented the start key (KeyMin) and then formed the meta key; for
// KeyMin this leads to Meta2Prefix\x00. For the above EndKey, the meta key is
// a Meta1key which sorts before Meta2Prefix\x00, causing a panic. The fix was
// simply to increment the meta key for StartKey, not StartKey itself.
func TestRangeCacheClearOverlappingMeta(t *testing.T) {
defer leaktest.AfterTest(t)
firstDesc := &proto.RangeDescriptor{
StartKey: proto.KeyMin,
EndKey: proto.Key("zzz"),
}
restDesc := &proto.RangeDescriptor{
StartKey: firstDesc.StartKey,
EndKey: proto.KeyMax,
}
cache := newRangeDescriptorCache(nil, 2<<10)
cache.rangeCache.Add(rangeCacheKey(keys.RangeMetaKey(firstDesc.EndKey)),
firstDesc)
cache.rangeCache.Add(rangeCacheKey(keys.RangeMetaKey(restDesc.EndKey)),
restDesc)
// Add new range, corresponding to splitting the first range at a meta key.
metaSplitDesc := &proto.RangeDescriptor{
StartKey: proto.KeyMin,
EndKey: proto.Key(keys.RangeMetaKey(proto.Key("foo"))),
}
func() {
defer func() {
if r := recover(); r != nil {
t.Fatalf("invocation of clearOverlappingCachedRangeDescriptors panicked: %v", r)
}
}()
cache.clearOverlappingCachedRangeDescriptors(metaSplitDesc.EndKey, keys.RangeMetaKey(metaSplitDesc.EndKey), metaSplitDesc)
}()
}
// clearOverlappingCachedRangeDescriptors looks up and clears any
// cache entries which overlap the specified key or descriptor.
func (rdc *rangeDescriptorCache) clearOverlappingCachedRangeDescriptors(key, metaKey proto.Key, desc *proto.RangeDescriptor) {
if desc.StartKey.Equal(desc.EndKey) { // True for some unittests.
return
}
// Clear out any descriptors which subsume the key which we're going
// to cache. For example, if an existing KeyMin->KeyMax descriptor
// should be cleared out in favor of a KeyMin->"m" descriptor.
k, v, ok := rdc.rangeCache.Ceil(rangeCacheKey(metaKey))
if ok {
descriptor := v.(*proto.RangeDescriptor)
addrKey := keys.KeyAddress(key)
if !addrKey.Less(descriptor.StartKey) && !descriptor.EndKey.Less(addrKey) {
if log.V(1) {
log.Infof("clearing overlapping descriptor: key=%s desc=%s", k, descriptor)
}
rdc.rangeCache.Del(k.(rangeCacheKey))
}
}
// Also clear any descriptors which are subsumed by the one we're
// going to cache. This could happen on a merge (and also happens
// when there's a lot of concurrency). Iterate from StartKey.Next().
rdc.rangeCache.DoRange(func(k, v interface{}) {
if log.V(1) {
log.Infof("clearing subsumed descriptor: key=%s desc=%s", k, v.(*proto.RangeDescriptor))
}
rdc.rangeCache.Del(k.(rangeCacheKey))
}, rangeCacheKey(keys.RangeMetaKey(desc.StartKey.Next())),
rangeCacheKey(keys.RangeMetaKey(desc.EndKey)))
}
// getCachedRangeDescriptorLocked is a helper function to retrieve the
// descriptor of the range which contains the given key, if present in the
// cache. It is assumed that the caller holds a read lock on rdc.rangeCacheMu.
func (rdc *rangeDescriptorCache) getCachedRangeDescriptorLocked(key proto.Key, inclusive bool) (
rangeCacheKey, *proto.RangeDescriptor) {
// The cache is indexed using the end-key of the range, but the
// end-key is non-inclusive by default.
var metaKey proto.Key
if !inclusive {
metaKey = keys.RangeMetaKey(key.Next())
} else {
metaKey = keys.RangeMetaKey(key)
}
k, v, ok := rdc.rangeCache.Ceil(rangeCacheKey(metaKey))
if !ok {
return nil, nil
}
metaEndKey := k.(rangeCacheKey)
rd := v.(*proto.RangeDescriptor)
// Check that key actually belongs to the range.
if !rd.ContainsKey(key) {
// The key is the EndKey and we're inclusive, so just return the range descriptor.
if inclusive && key.Equal(rd.EndKey) {
return metaEndKey, rd
}
return nil, nil
}
// The key is the StartKey, but we're inclusive and thus need to return the
// previous range descriptor, but it is not in the cache yet.
if inclusive && key.Equal(rd.StartKey) {
return nil, nil
}
return metaEndKey, rd
}
// TestRangeSplitMeta executes various splits (including at meta addressing)
// and checks that all created intents are resolved. This includes both intents
// which are resolved synchronously with EndTransaction and via RPC.
func TestRangeSplitMeta(t *testing.T) {
defer leaktest.AfterTest(t)
s := createTestDB(t)
defer s.Stop()
splitKeys := []roachpb.Key{roachpb.Key("G"), keys.RangeMetaKey(roachpb.Key("F")),
keys.RangeMetaKey(roachpb.Key("K")), keys.RangeMetaKey(roachpb.Key("H"))}
// Execute the consecutive splits.
for _, splitKey := range splitKeys {
log.Infof("starting split at key %q...", splitKey)
if err := s.DB.AdminSplit(splitKey); err != nil {
t.Fatal(err)
}
log.Infof("split at key %q complete", splitKey)
}
if err := util.IsTrueWithin(func() bool {
if _, _, err := engine.MVCCScan(s.Eng, keys.LocalMax, roachpb.KeyMax, 0, roachpb.MaxTimestamp, true, nil); err != nil {
log.Infof("mvcc scan should be clean: %s", err)
return false
}
return true
}, 500*time.Millisecond); err != nil {
t.Error("failed to verify no dangling intents within 500ms")
}
}
// TestReplicateRange verifies basic replication functionality by creating two stores
// and a range, replicating the range to the second store, and reading its data there.
func TestReplicateRange(t *testing.T) {
defer leaktest.AfterTest(t)
mtc := startMultiTestContext(t, 2)
defer mtc.Stop()
// Issue a command on the first node before replicating.
incArgs := incrementArgs([]byte("a"), 5)
if _, err := client.SendWrapped(rg1(mtc.stores[0]), nil, &incArgs); err != nil {
t.Fatal(err)
}
rng, err := mtc.stores[0].GetReplica(1)
if err != nil {
t.Fatal(err)
}
if err := rng.ChangeReplicas(roachpb.ADD_REPLICA,
roachpb.ReplicaDescriptor{
NodeID: mtc.stores[1].Ident.NodeID,
StoreID: mtc.stores[1].Ident.StoreID,
}, rng.Desc()); err != nil {
t.Fatal(err)
}
// Verify no intent remains on range descriptor key.
key := keys.RangeDescriptorKey(rng.Desc().StartKey)
desc := roachpb.RangeDescriptor{}
if ok, err := engine.MVCCGetProto(mtc.stores[0].Engine(), key, mtc.stores[0].Clock().Now(), true, nil, &desc); !ok || err != nil {
t.Fatalf("fetching range descriptor yielded %t, %s", ok, err)
}
// Verify that in time, no intents remain on meta addressing
// keys, and that range descriptor on the meta records is correct.
util.SucceedsWithin(t, 1*time.Second, func() error {
meta2 := keys.Addr(keys.RangeMetaKey(roachpb.RKeyMax))
meta1 := keys.Addr(keys.RangeMetaKey(meta2))
for _, key := range []roachpb.RKey{meta2, meta1} {
metaDesc := roachpb.RangeDescriptor{}
if ok, err := engine.MVCCGetProto(mtc.stores[0].Engine(), key.AsRawKey(), mtc.stores[0].Clock().Now(), true, nil, &metaDesc); !ok || err != nil {
return util.Errorf("failed to resolve %s", key.AsRawKey())
}
if !reflect.DeepEqual(metaDesc, desc) {
return util.Errorf("descs not equal: %+v != %+v", metaDesc, desc)
}
}
return nil
})
// Verify that the same data is available on the replica.
util.SucceedsWithin(t, replicaReadTimeout, func() error {
getArgs := getArgs([]byte("a"))
if reply, err := client.SendWrappedWith(rg1(mtc.stores[1]), nil, roachpb.Header{
ReadConsistency: roachpb.INCONSISTENT,
}, &getArgs); err != nil {
return util.Errorf("failed to read data: %s", err)
} else if e, v := int64(5), mustGetInt(reply.(*roachpb.GetResponse).Value); v != e {
return util.Errorf("failed to read correct data: expected %d, got %d", e, v)
}
return nil
})
}
// 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.
// seenDesc should always be passed in and is used as the basis of a
// compare-and-evict (as pointers); if it is nil, eviction is unconditional
// but a warning will be logged.
func (rdc *rangeDescriptorCache) EvictCachedRangeDescriptor(descKey proto.Key, seenDesc *proto.RangeDescriptor) {
if seenDesc == nil {
log.Warningf("compare-and-evict for key %s with nil descriptor; clearing unconditionally", descKey)
}
rdc.rangeCacheMu.Lock()
defer rdc.rangeCacheMu.Unlock()
rngKey, cachedDesc := rdc.getCachedRangeDescriptorLocked(descKey)
// Note that we're doing a "compare-and-erase": If seenDesc is not nil,
// we want to clean the cache only if it equals the cached range
// descriptor as a pointer. If not, then likely some other caller
// already evicted previously, and we can save work by not doing it
// again (which would prompt another expensive lookup).
if seenDesc != nil && seenDesc != cachedDesc {
return
}
for !bytes.Equal(descKey, proto.KeyMin) {
if log.V(2) {
log.Infof("evict cached descriptor: key=%s desc=%s\n%s", descKey, cachedDesc, rdc.stringLocked())
} else if log.V(1) {
log.Infof("evict cached descriptor: key=%s desc=%s", descKey, cachedDesc)
}
rdc.rangeCache.Del(rngKey)
// 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.
descKey = keys.RangeMetaKey(descKey)
rngKey, cachedDesc = rdc.getCachedRangeDescriptorLocked(descKey)
}
}
func metaKey(key roachpb.RKey) []byte {
rk, err := keys.Addr(keys.RangeMetaKey(key))
if err != nil {
panic(err)
}
return rk
}
func runLsRanges(cmd *cobra.Command, args []string) {
if len(args) > 1 {
mustUsage(cmd)
return
}
var startKey proto.Key
if len(args) >= 1 {
startKey = keys.RangeMetaKey(proto.Key(args[0]))
} else {
startKey = keys.Meta2Prefix
}
kvDB, stopper := makeDBClient()
defer stopper.Stop()
rows, err := kvDB.Scan(startKey, keys.Meta2Prefix.PrefixEnd(), maxResults)
if err != nil {
fmt.Fprintf(os.Stderr, "scan failed: %s\n", err)
osExit(1)
return
}
for _, row := range rows {
desc := &proto.RangeDescriptor{}
if err := row.ValueProto(desc); err != nil {
fmt.Fprintf(os.Stderr, "%s: unable to unmarshal range descriptor\n", row.Key)
continue
}
fmt.Printf("%s-%s [%d]\n", desc.StartKey, desc.EndKey, desc.RangeID)
for i, replica := range desc.Replicas {
fmt.Printf("\t%d: node-id=%d store-id=%d\n",
i, replica.NodeID, replica.StoreID)
}
}
fmt.Printf("%d result(s)\n", len(rows))
}
// rangeAddressing updates or deletes the range addressing metadata
// for the range specified by desc. The action to take is specified by
// the supplied metaAction function.
//
// The rules for meta1 and meta2 records are as follows:
//
// 1. If desc.StartKey or desc.EndKey is meta1:
// - ERROR
// 2. If desc.EndKey is meta2:
// - meta1(desc.EndKey)
// 3. If desc.EndKey is normal user key:
// - meta2(desc.EndKey)
// 3a. If desc.StartKey is KeyMin or meta2:
// - meta1(KeyMax)
func rangeAddressing(b *client.Batch, desc *roachpb.RangeDescriptor, action metaAction) error {
// 1. handle illegal case of start or end key being meta1.
if bytes.HasPrefix(desc.EndKey, keys.Meta1Prefix) ||
bytes.HasPrefix(desc.StartKey, keys.Meta1Prefix) {
return errors.Errorf("meta1 addressing records cannot be split: %+v", desc)
}
// Note that both cases 2 and 3 are handled by keys.RangeMetaKey.
//
// 2. the case of the range ending with a meta2 prefix. This means
// the range is full of meta2. We must update the relevant meta1
// entry pointing to the end of this range.
//
// 3. the range ends with a normal user key, so we must update the
// relevant meta2 entry pointing to the end of this range.
action(b, keys.RangeMetaKey(desc.EndKey), desc)
if !bytes.HasPrefix(desc.EndKey, keys.Meta2Prefix) {
// 3a. the range starts with KeyMin or a meta2 addressing record,
// update the meta1 entry for KeyMax.
if bytes.Equal(desc.StartKey, roachpb.RKeyMin) ||
bytes.HasPrefix(desc.StartKey, keys.Meta2Prefix) {
action(b, keys.Meta1KeyMax, desc)
}
}
return nil
}
func (db *testDescriptorDB) getRangeDescriptors(key proto.Key,
options lookupOptions) ([]proto.RangeDescriptor, error) {
db.lookupCount++
metadataKey := keys.RangeMetaKey(key)
var err error
// Recursively call into cache as the real DB would, terminating recursion
// when a meta1key is encountered.
if len(metadataKey) > 0 && !bytes.HasPrefix(metadataKey, keys.Meta1Prefix) {
_, err = db.cache.LookupRangeDescriptor(metadataKey, options)
}
return db.getDescriptor(key), err
}
// getCachedRangeDescriptorLocked is a helper function to retrieve the
// descriptor of the range which contains the given key, if present in the
// cache. It is assumed that the caller holds a read lock on rdc.rangeCacheMu.
func (rdc *rangeDescriptorCache) getCachedRangeDescriptorLocked(key proto.Key, isReverse bool) (
rangeCacheKey, *proto.RangeDescriptor) {
// The cache is indexed using the end-key of the range, but the
// end-key is non-inclusive.
var metaKey proto.Key
if !isReverse {
// If it is not reverse scan, we access the cache using key.Next().
metaKey = keys.RangeMetaKey(key.Next())
} else {
// Because reverse scan request is begining at end key(exclusive),so we
// access the cache using key directly.
metaKey = keys.RangeMetaKey(key)
}
k, v, ok := rdc.rangeCache.Ceil(rangeCacheKey(metaKey))
if !ok {
return nil, nil
}
metaEndKey := k.(rangeCacheKey)
rd := v.(*proto.RangeDescriptor)
// Check that key actually belongs to the range.
if !rd.ContainsKey(keys.KeyAddress(key)) {
// The key is the EndKey of the range in reverse scan, just return the range descriptor.
if isReverse && key.Equal(rd.EndKey) {
return metaEndKey, rd
}
return nil, nil
}
// The key is the StartKey of the range in reverse scan. We need to return the previous range
// descriptor, but it is not in the cache yet.
if isReverse && key.Equal(rd.StartKey) {
return nil, nil
}
return metaEndKey, rd
}
// LookupRange returns the descriptor of the range containing key.
func (tc *TestCluster) LookupRange(key roachpb.Key) (roachpb.RangeDescriptor, error) {
rangeLookupReq := roachpb.RangeLookupRequest{
Span: roachpb.Span{
Key: keys.RangeMetaKey(keys.MustAddr(key)),
},
MaxRanges: 1,
ConsiderIntents: false,
}
resp, pErr := client.SendWrapped(tc.Servers[0].GetDistSender(), nil, &rangeLookupReq)
if pErr != nil {
return roachpb.RangeDescriptor{}, errors.Errorf(
"%q: lookup range unexpected error: %s", key, pErr)
}
return resp.(*roachpb.RangeLookupResponse).Ranges[0], nil
}
func (c *cluster) lookupRange(nodeIdx int, key roachpb.Key) (*roachpb.RangeDescriptor, error) {
req := &roachpb.RangeLookupRequest{
Span: roachpb.Span{
Key: keys.RangeMetaKey(keys.MustAddr(key)),
},
MaxRanges: 1,
ConsiderIntents: false,
}
sender := c.clients[nodeIdx].GetSender()
resp, pErr := client.SendWrapped(sender, nil, req)
if pErr != nil {
return nil, errors.Errorf("%s: lookup range: %s", key, pErr)
}
return &resp.(*roachpb.RangeLookupResponse).Ranges[0], nil
}
// LookupRangeDescriptor attempts to locate a descriptor for the range
// containing the given Key. This is done by querying the two-level
// lookup table of range descriptors which cockroach maintains.
//
// This method first looks up the specified key in the first level of
// range metadata, which returns the location of the key within the
// second level of range metadata. This second level location is then
// queried to retrieve a descriptor for the range where the key's
// value resides. Range descriptors retrieved during each search are
// cached for subsequent lookups.
//
// This method returns the RangeDescriptor for the range containing
// the key's data, or an error if any occurred.
func (rdc *rangeDescriptorCache) LookupRangeDescriptor(key proto.Key,
options lookupOptions) (*proto.RangeDescriptor, error) {
if _, r := rdc.getCachedRangeDescriptor(key); r != nil {
return r, nil
}
if log.V(1) {
log.Infof("lookup range descriptor: key=%s", key)
} else if log.V(2) {
log.Infof("lookup range descriptor: key=%s\n%s", key, rdc)
}
rs, err := rdc.db.getRangeDescriptors(key, options)
if err != nil {
return nil, err
}
// TODO(tamird): there is a race here; multiple readers may experience cache
// misses and concurrently attempt to refresh the cache, duplicating work.
// Locking over the getRangeDescriptors call is even worse though, because
// that blocks the cache completely for the duration of a slow query to the
// cluster.
rdc.rangeCacheMu.Lock()
for i := range rs {
// Note: we append the end key of each range to meta[12] records
// so that calls to rdc.rangeCache.Ceil() for a key will return
// the correct range. Using the start key would require using
// Floor() which is a possibility for our llrb-based OrderedCache
// but not possible for RocksDB.
// Before adding a new descriptor, make sure we clear out any
// pre-existing, overlapping descriptor which might have been
// re-inserted due to concurrent range lookups.
rangeKey := keys.RangeMetaKey(rs[i].EndKey)
if log.V(1) {
log.Infof("adding descriptor: key=%s desc=%s", rangeKey, &rs[i])
}
rdc.clearOverlappingCachedRangeDescriptors(rs[i].EndKey, rangeKey, &rs[i])
rdc.rangeCache.Add(rangeCacheKey(rangeKey), &rs[i])
}
if len(rs) == 0 {
log.Fatalf("no range descriptors returned for %s", key)
}
rdc.rangeCacheMu.Unlock()
return &rs[0], nil
}
// getCachedRangeDescriptorLocked is a helper function to retrieve the
// descriptor of the range which contains the given key, if present in the
// cache. It is assumed that the caller holds a read lock on rdc.rangeCacheMu.
func (rdc *rangeDescriptorCache) getCachedRangeDescriptorLocked(key proto.Key) (
rangeCacheKey, *proto.RangeDescriptor) {
// The cache is indexed using the end-key of the range, but the
// end-key is non-inclusive. If inclusive is false, we access the
// cache using key.Next().
metaKey := keys.RangeMetaKey(key.Next())
k, v, ok := rdc.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(keys.KeyAddress(key)) {
return nil, nil
}
return metaEndKey, rd
}
// 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.
// seenDesc should always be passed in and is used as the basis of a
// compare-and-evict (as pointers); if it is nil, eviction is unconditional
// but a warning will be logged.
func (rdc *rangeDescriptorCache) EvictCachedRangeDescriptor(descKey proto.Key,
seenDesc *proto.RangeDescriptor, inclusive bool) {
if seenDesc == nil {
log.Warningf("compare-and-evict for key %s with nil descriptor; clearing unconditionally", descKey)
}
rdc.rangeCacheMu.Lock()
defer rdc.rangeCacheMu.Unlock()
rngKey, cachedDesc := rdc.getCachedRangeDescriptorLocked(descKey, inclusive)
// Note that we're doing a "compare-and-erase": If seenDesc is not nil,
// we want to clean the cache only if it equals the cached range
// descriptor as a pointer. If not, then likely some other caller
// already evicted previously, and we can save work by not doing it
// again (which would prompt another expensive lookup).
if seenDesc != nil && seenDesc != cachedDesc {
return
}
for {
if log.V(2) {
log.Infof("evict cached descriptor: key=%s desc=%s\n%s", descKey, cachedDesc, rdc.stringLocked())
} else if log.V(1) {
log.Infof("evict cached descriptor: key=%s desc=%s", descKey, cachedDesc)
}
rdc.rangeCache.Del(rngKey)
// 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.
descKey = keys.RangeMetaKey(descKey)
rngKey, cachedDesc = rdc.getCachedRangeDescriptorLocked(descKey, inclusive)
// TODO(tschottdorf): write a test that verifies that the first descriptor
// can also be evicted. This is necessary since the initial range
// [KeyMin,KeyMax) may turn into [KeyMin, "something"), after which
// larger ranges don't fit into it any more.
if bytes.Equal(descKey, proto.KeyMin) {
break
}
}
}
func runLsRanges(cmd *cobra.Command, args []string) {
if len(args) > 1 {
mustUsage(cmd)
return
}
var startKey roachpb.Key
{
k := roachpb.KeyMin.Next()
if len(args) > 0 {
k = roachpb.Key(args[0])
}
rk, err := keys.Addr(k)
if err != nil {
panic(err)
}
startKey = keys.RangeMetaKey(rk)
}
endKey := keys.Meta2Prefix.PrefixEnd()
kvDB, stopper := makeDBClient()
defer stopper.Stop()
rows, err := kvDB.Scan(startKey, endKey, maxResults)
if err != nil {
panicf("scan failed: %s\n", err)
}
for _, row := range rows {
desc := &roachpb.RangeDescriptor{}
if err := row.ValueProto(desc); err != nil {
panicf("%s: unable to unmarshal range descriptor\n", row.Key)
continue
}
fmt.Printf("%s-%s [%d]\n", desc.StartKey, desc.EndKey, desc.RangeID)
for i, replica := range desc.Replicas {
fmt.Printf("\t%d: node-id=%d store-id=%d\n",
i, replica.NodeID, replica.StoreID)
}
}
fmt.Printf("%d result(s)\n", len(rows))
}
// getRangeMetadata retrieves the current range descriptor for the target
// range.
func getRangeMetadata(key roachpb.RKey, mtc *multiTestContext, t *testing.T) roachpb.RangeDescriptor {
// Calls to RangeLookup typically use inconsistent reads, but we
// want to do a consistent read here. This is important when we are
// considering one of the metadata ranges: we must not do an
// inconsistent lookup in our own copy of the range.
b := &client.Batch{}
b.InternalAddRequest(&roachpb.RangeLookupRequest{
Span: roachpb.Span{
Key: keys.RangeMetaKey(key),
},
MaxRanges: 1,
})
var reply *roachpb.RangeLookupResponse
if br, err := mtc.db.RunWithResponse(b); err != nil {
t.Fatalf("error getting range metadata: %s", err)
} else {
reply = br.Responses[0].GetInner().(*roachpb.RangeLookupResponse)
}
if a, e := len(reply.Ranges), 1; a != e {
t.Fatalf("expected %d range descriptor, got %d", e, a)
}
return reply.Ranges[0]
}
// TestRangeCacheClearOverlapping verifies that existing, overlapping
// cached entries are cleared when adding a new entry.
func TestRangeCacheClearOverlapping(t *testing.T) {
defer leaktest.AfterTest(t)
defDesc := &proto.RangeDescriptor{
StartKey: proto.KeyMin,
EndKey: proto.KeyMax,
}
cache := newRangeDescriptorCache(nil, 2<<10)
cache.rangeCache.Add(rangeCacheKey(keys.RangeMetaKey(proto.KeyMax)), defDesc)
// Now, add a new, overlapping set of descriptors.
minToBDesc := &proto.RangeDescriptor{
StartKey: proto.KeyMin,
EndKey: proto.Key("b"),
}
bToMaxDesc := &proto.RangeDescriptor{
StartKey: proto.Key("b"),
EndKey: proto.KeyMax,
}
cache.clearOverlappingCachedRangeDescriptors(proto.Key("b"), keys.RangeMetaKey(proto.Key("b")), minToBDesc)
cache.rangeCache.Add(rangeCacheKey(keys.RangeMetaKey(proto.Key("b"))), minToBDesc)
if _, desc := cache.getCachedRangeDescriptor(proto.Key("b"), false); desc != nil {
t.Errorf("descriptor unexpectedly non-nil: %s", desc)
}
cache.clearOverlappingCachedRangeDescriptors(proto.KeyMax, keys.RangeMetaKey(proto.KeyMax), bToMaxDesc)
cache.rangeCache.Add(rangeCacheKey(keys.RangeMetaKey(proto.KeyMax)), bToMaxDesc)
if _, desc := cache.getCachedRangeDescriptor(proto.Key("b"), false); desc != bToMaxDesc {
t.Errorf("expected descriptor %s; got %s", bToMaxDesc, desc)
}
// Add default descriptor back which should remove two split descriptors.
cache.clearOverlappingCachedRangeDescriptors(proto.KeyMax, keys.RangeMetaKey(proto.KeyMax), defDesc)
cache.rangeCache.Add(rangeCacheKey(keys.RangeMetaKey(proto.KeyMax)), defDesc)
for _, key := range []proto.Key{proto.Key("a"), proto.Key("b")} {
if _, desc := cache.getCachedRangeDescriptor(key, false); desc != defDesc {
t.Errorf("expected descriptor %s for key %s; got %s", defDesc, key, desc)
}
}
}
// getRangeDescriptors returns a sorted slice of RangeDescriptors for a set of
// consecutive ranges, the first of 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 (ds *DistSender) getRangeDescriptors(key proto.Key, options lookupOptions) ([]proto.RangeDescriptor, error) {
var (
// metadataKey is sent to internalRangeLookup to find the
// RangeDescriptor which contains key.
metadataKey = keys.RangeMetaKey(key)
// desc is the RangeDescriptor for the range which contains
// metadataKey.
desc *proto.RangeDescriptor
err error
)
if bytes.Equal(metadataKey, proto.KeyMin) {
// 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 := ds.getFirstRangeDescriptor()
if err != nil {
return nil, err
}
return []proto.RangeDescriptor{*rd}, nil
}
if bytes.HasPrefix(metadataKey, keys.Meta1Prefix) {
// In this case, desc is the cluster's first range.
if desc, err = ds.getFirstRangeDescriptor(); err != nil {
return nil, err
}
} else {
// Look up desc from the cache, which will recursively call into
// ds.getRangeDescriptors if it is not cached.
desc, err = ds.rangeCache.LookupRangeDescriptor(metadataKey, options)
if err != nil {
return nil, err
}
}
return ds.internalRangeLookup(metadataKey, options, desc)
}
// getRangeMetadata retrieves the current range descriptor for the target
// range.
func getRangeMetadata(key proto.Key, mtc *multiTestContext, t *testing.T) proto.RangeDescriptor {
// Calls to RangeLookup typically use inconsistent reads, but we
// want to do a consistent read here. This is important when we are
// considering one of the metadata ranges: we must not do an
// inconsistent lookup in our own copy of the range.
reply := proto.RangeLookupResponse{}
b := &client.Batch{}
b.InternalAddCall(proto.Call{
Args: &proto.RangeLookupRequest{
RequestHeader: proto.RequestHeader{
Key: keys.RangeMetaKey(key),
},
MaxRanges: 1,
},
Reply: &reply,
})
if err := mtc.db.Run(b).GoError(); err != nil {
t.Fatalf("error getting range metadata: %s", err)
}
if a, e := len(reply.Ranges), 1; a != e {
t.Fatalf("expected %d range descriptor, got %d", e, a)
}
return reply.Ranges[0]
}
// process performs a consistent lookup on the range descriptor to see if we are
// still a member of the range.
func (q *rangeGCQueue) process(now proto.Timestamp, rng *Range) error {
// Calls to InternalRangeLookup typically use inconsistent reads, but we
// want to do a consistent read here. This is important when we are
// considering one of the metadata ranges: we must not do an inconsistent
// lookup in our own copy of the range.
reply := proto.InternalRangeLookupResponse{}
b := &client.Batch{}
b.InternalAddCall(proto.Call{
Args: &proto.InternalRangeLookupRequest{
RequestHeader: proto.RequestHeader{
Key: keys.RangeMetaKey(rng.Desc().StartKey),
},
MaxRanges: 1,
},
Reply: &reply,
})
if err := q.db.Run(b); err != nil {
return err
}
if len(reply.Ranges) != 1 {
return util.Errorf("expected 1 range descriptor, got %d", len(reply.Ranges))
}
desc := reply.Ranges[0]
currentMember := false
if me := rng.GetReplica(); me != nil {
for _, rep := range desc.Replicas {
if rep.StoreID == me.StoreID {
currentMember = true
break
}
}
}
if !currentMember {
// We are no longer a member of this range; clean up our local data.
if log.V(1) {
log.Infof("destroying local data from range %d", rng.Desc().RaftID)
}
if err := rng.rm.RemoveRange(rng); err != nil {
return err
}
// TODO(bdarnell): update Destroy to leave tombstones for removed ranges (#768)
// TODO(bdarnell): add some sort of locking to prevent the range
// from being recreated while the underlying data is being destroyed.
if err := rng.Destroy(); err != nil {
return err
}
} else if desc.RaftID != rng.Desc().RaftID {
// If we get a different raft ID back, then the range has been merged
// away. But currentMember is true, so we are still a member of the
// subsuming range. Shut down raft processing for the former range
// and delete any remaining metadata, but do not delete the data.
if log.V(1) {
log.Infof("removing merged range %d", rng.Desc().RaftID)
}
if err := rng.rm.RemoveRange(rng); err != nil {
return err
}
// TODO(bdarnell): remove raft logs and other metadata (while leaving a
// tombstone). Add tests for GC of merged ranges.
}
return nil
}
// TestRangeLookupUseReverse tests whether the results and the results count
// are correct when scanning in reverse order.
func TestRangeLookupUseReverse(t *testing.T) {
defer leaktest.AfterTest(t)()
sCtx := storage.TestStoreContext()
sCtx.TestingKnobs.DisableSplitQueue = true
store, stopper, _ := createTestStoreWithContext(t, sCtx)
defer stopper.Stop()
// Init test ranges:
// ["","a"), ["a","c"), ["c","e"), ["e","g") and ["g","\xff\xff").
splits := []roachpb.AdminSplitRequest{
adminSplitArgs(roachpb.Key("g"), roachpb.Key("g")),
adminSplitArgs(roachpb.Key("e"), roachpb.Key("e")),
adminSplitArgs(roachpb.Key("c"), roachpb.Key("c")),
adminSplitArgs(roachpb.Key("a"), roachpb.Key("a")),
}
for _, split := range splits {
_, pErr := client.SendWrapped(rg1(store), nil, &split)
if pErr != nil {
t.Fatalf("%q: split unexpected error: %s", split.SplitKey, pErr)
}
}
// Resolve the intents.
scanArgs := roachpb.ScanRequest{
Span: roachpb.Span{
Key: keys.RangeMetaKey(roachpb.RKeyMin.Next()),
EndKey: keys.RangeMetaKey(roachpb.RKeyMax),
},
}
util.SucceedsSoon(t, func() error {
_, pErr := client.SendWrapped(rg1(store), nil, &scanArgs)
return pErr.GoError()
})
revScanArgs := func(key []byte, maxResults int32) *roachpb.RangeLookupRequest {
return &roachpb.RangeLookupRequest{
Span: roachpb.Span{
Key: key,
},
MaxRanges: maxResults,
Reverse: true,
}
}
// Test cases.
testCases := []struct {
request *roachpb.RangeLookupRequest
expected []roachpb.RangeDescriptor
expectedPre []roachpb.RangeDescriptor
}{
// Test key in the middle of the range.
{
request: revScanArgs(keys.RangeMetaKey(roachpb.RKey("f")), 2),
// ["e","g") and ["c","e").
expected: []roachpb.RangeDescriptor{
{StartKey: roachpb.RKey("e"), EndKey: roachpb.RKey("g")},
},
expectedPre: []roachpb.RangeDescriptor{
{StartKey: roachpb.RKey("c"), EndKey: roachpb.RKey("e")},
},
},
// Test key in the end key of the range.
{
request: revScanArgs(keys.RangeMetaKey(roachpb.RKey("g")), 3),
// ["e","g"), ["c","e") and ["a","c").
expected: []roachpb.RangeDescriptor{
{StartKey: roachpb.RKey("e"), EndKey: roachpb.RKey("g")},
},
expectedPre: []roachpb.RangeDescriptor{
{StartKey: roachpb.RKey("c"), EndKey: roachpb.RKey("e")},
{StartKey: roachpb.RKey("a"), EndKey: roachpb.RKey("c")},
},
},
{
request: revScanArgs(keys.RangeMetaKey(roachpb.RKey("e")), 2),
// ["c","e") and ["a","c").
expected: []roachpb.RangeDescriptor{
{StartKey: roachpb.RKey("c"), EndKey: roachpb.RKey("e")},
},
expectedPre: []roachpb.RangeDescriptor{
{StartKey: roachpb.RKey("a"), EndKey: roachpb.RKey("c")},
},
},
// Test Meta2KeyMax.
{
request: revScanArgs(keys.Meta2KeyMax, 2),
// ["e","g") and ["g","\xff\xff")
expected: []roachpb.RangeDescriptor{
{StartKey: roachpb.RKey("g"), EndKey: roachpb.RKey("\xff\xff")},
},
expectedPre: []roachpb.RangeDescriptor{
{StartKey: roachpb.RKey("e"), EndKey: roachpb.RKey("g")},
},
},
// Test Meta1KeyMax.
{
request: revScanArgs(keys.Meta1KeyMax, 1),
// ["","a")
expected: []roachpb.RangeDescriptor{
{StartKey: roachpb.RKeyMin, EndKey: roachpb.RKey("a")},
},
},
}
for testIdx, test := range testCases {
resp, pErr := client.SendWrappedWith(rg1(store), nil, roachpb.Header{
ReadConsistency: roachpb.INCONSISTENT,
}, test.request)
if pErr != nil {
t.Fatalf("%d: RangeLookup error: %s", testIdx, pErr)
}
rlReply := resp.(*roachpb.RangeLookupResponse)
// Checks the results count.
if int32(len(rlReply.Ranges))+int32(len(rlReply.PrefetchedRanges)) != test.request.MaxRanges {
t.Fatalf("%d: returned results count, expected %d,but got %d", testIdx, test.request.MaxRanges, len(rlReply.Ranges))
}
// Checks the range descriptors.
for _, rngSlice := range []struct {
expect, reply []roachpb.RangeDescriptor
}{
{test.expected, rlReply.Ranges},
{test.expectedPre, rlReply.PrefetchedRanges},
} {
for i, rng := range rngSlice.expect {
if !(rng.StartKey.Equal(rngSlice.reply[i].StartKey) && rng.EndKey.Equal(rngSlice.reply[i].EndKey)) {
t.Fatalf("%d: returned range is not correct, expected %v ,but got %v", testIdx, rng, rngSlice.reply[i])
}
}
}
}
}
func meta(k roachpb.RKey) roachpb.RKey {
return keys.Addr(keys.RangeMetaKey(k))
}
// TestGetCachedRangeDescriptorInclusive verifies the correctness of the result
// that is returned by getCachedRangeDescriptor with inclusive=true.
func TestGetCachedRangeDescriptorInclusive(t *testing.T) {
defer leaktest.AfterTest(t)
testData := []*proto.RangeDescriptor{
{StartKey: proto.Key("a"), EndKey: proto.Key("c")},
{StartKey: proto.Key("c"), EndKey: proto.Key("e")},
{StartKey: proto.Key("g"), EndKey: proto.Key("z")},
}
cache := newRangeDescriptorCache(nil, 2<<10)
for _, rd := range testData {
cache.rangeCache.Add(rangeCacheKey(keys.RangeMetaKey(rd.EndKey)), rd)
}
testCases := []struct {
queryKey proto.Key
cacheKey rangeCacheKey
rng *proto.RangeDescriptor
}{
{
// Check range start key.
queryKey: proto.Key("a"),
cacheKey: nil,
rng: nil,
},
{
// Check range end key.
queryKey: proto.Key("c"),
cacheKey: rangeCacheKey(keys.RangeMetaKey(proto.Key("c"))),
rng: &proto.RangeDescriptor{StartKey: proto.Key("a"), EndKey: proto.Key("c")},
},
{
// Check range middle key.
queryKey: proto.Key("d"),
cacheKey: rangeCacheKey(keys.RangeMetaKey(proto.Key("e"))),
rng: &proto.RangeDescriptor{StartKey: proto.Key("c"), EndKey: proto.Key("e")},
},
{
// Check miss range key.
queryKey: proto.Key("f"),
cacheKey: nil,
rng: nil,
},
{
// Check range start key with previous range miss.
queryKey: proto.Key("g"),
cacheKey: nil,
rng: nil,
},
}
for _, test := range testCases {
cacheKey, targetRange := cache.getCachedRangeDescriptor(test.queryKey, true /* inclusive */)
if !reflect.DeepEqual(targetRange, test.rng) {
t.Fatalf("expect range %v, actual get %v", test.rng, targetRange)
}
if !reflect.DeepEqual(cacheKey, test.cacheKey) {
t.Fatalf("expect cache key %v, actual get %v", test.cacheKey, cacheKey)
}
}
}
// 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
// lookuped when looking up metadata keys through the cache.
func TestRangeCache(t *testing.T) {
defer leaktest.AfterTest(t)
db := newTestDescriptorDB()
for i, char := range "abcdefghijklmnopqrstuvwx" {
db.splitRange(t, proto.Key(string(char)))
if i > 0 && i%6 == 0 {
db.splitRange(t, keys.RangeMetaKey(proto.Key(string(char))))
}
}
db.cache = newRangeDescriptorCache(db, 2<<10)
doLookup(t, db.cache, "aa")
db.assertLookupCount(t, 2, "aa")
// Descriptors for the following ranges should be cached.
doLookup(t, db.cache, "ab")
db.assertLookupCount(t, 0, "ab")
doLookup(t, db.cache, "ba")
db.assertLookupCount(t, 0, "ba")
doLookup(t, db.cache, "cz")
db.assertLookupCount(t, 0, "cz")
// Metadata two ranges weren't cached, same metadata 1 range.
doLookup(t, db.cache, "d")
db.assertLookupCount(t, 1, "d")
doLookup(t, db.cache, "fa")
db.assertLookupCount(t, 0, "fa")
// Metadata two ranges weren't cached, metadata 1 was aggressively cached
doLookup(t, db.cache, "ij")
db.assertLookupCount(t, 1, "ij")
doLookup(t, db.cache, "jk")
db.assertLookupCount(t, 0, "jk")
doLookup(t, db.cache, "pn")
db.assertLookupCount(t, 1, "pn")
// Totally uncached ranges
doLookup(t, db.cache, "vu")
db.assertLookupCount(t, 2, "vu")
doLookup(t, db.cache, "xx")
db.assertLookupCount(t, 0, "xx")
// Evict clears one level 1 and one level 2 cache
db.cache.EvictCachedRangeDescriptor(proto.Key("da"), nil, false)
doLookup(t, db.cache, "fa")
db.assertLookupCount(t, 0, "fa")
doLookup(t, db.cache, "da")
db.assertLookupCount(t, 2, "da")
// Looking up a descriptor that lands on an end-key should work
// without a cache miss.
doLookup(t, db.cache, "a")
db.assertLookupCount(t, 0, "a")
// Attempt to compare-and-evict with a descriptor that is not equal to the
// cached one; it should not alter the cache.
db.cache.EvictCachedRangeDescriptor(proto.Key("cz"), &proto.RangeDescriptor{}, false)
doLookup(t, db.cache, "cz")
db.assertLookupCount(t, 0, "cz")
// Now evict with the actual descriptor. The cache should clear the
// descriptor and the cached meta key.
db.cache.EvictCachedRangeDescriptor(proto.Key("cz"), doLookup(t, db.cache, "cz"), false)
doLookup(t, db.cache, "cz")
db.assertLookupCount(t, 2, "cz")
}
// TestRangeCacheClearOverlapping verifies that existing, overlapping
// cached entries are cleared when adding a new entry.
func TestRangeCacheClearOverlapping(t *testing.T) {
defer leaktest.AfterTest(t)()
defDesc := &roachpb.RangeDescriptor{
StartKey: roachpb.RKeyMin,
EndKey: roachpb.RKeyMax,
}
cache := newRangeDescriptorCache(nil, 2<<10)
cache.rangeCache.cache.Add(rangeCacheKey(keys.RangeMetaKey(roachpb.RKeyMax)), defDesc)
// Now, add a new, overlapping set of descriptors.
minToBDesc := &roachpb.RangeDescriptor{
StartKey: roachpb.RKeyMin,
EndKey: roachpb.RKey("b"),
}
bToMaxDesc := &roachpb.RangeDescriptor{
StartKey: roachpb.RKey("b"),
EndKey: roachpb.RKeyMax,
}
if err := cache.clearOverlappingCachedRangeDescriptors(minToBDesc); err != nil {
t.Fatal(err)
}
cache.rangeCache.cache.Add(rangeCacheKey(mustMeta(roachpb.RKey("b"))), minToBDesc)
if _, desc, err := cache.getCachedRangeDescriptor(roachpb.RKey("b"), false); err != nil {
t.Fatal(err)
} else if desc != nil {
t.Errorf("descriptor unexpectedly non-nil: %s", desc)
}
if err := cache.clearOverlappingCachedRangeDescriptors(bToMaxDesc); err != nil {
t.Fatal(err)
}
cache.rangeCache.cache.Add(rangeCacheKey(mustMeta(roachpb.RKeyMax)), bToMaxDesc)
if _, desc, err := cache.getCachedRangeDescriptor(roachpb.RKey("b"), false); err != nil {
t.Fatal(err)
} else if desc != bToMaxDesc {
t.Errorf("expected descriptor %s; got %s", bToMaxDesc, desc)
}
// Add default descriptor back which should remove two split descriptors.
if err := cache.clearOverlappingCachedRangeDescriptors(defDesc); err != nil {
t.Fatal(err)
}
cache.rangeCache.cache.Add(rangeCacheKey(keys.RangeMetaKey(roachpb.RKeyMax)), defDesc)
for _, key := range []roachpb.RKey{roachpb.RKey("a"), roachpb.RKey("b")} {
if _, desc, err := cache.getCachedRangeDescriptor(key, false); err != nil {
t.Fatal(err)
} else if desc != defDesc {
t.Errorf("expected descriptor %s for key %s; got %s", defDesc, key, desc)
}
}
// Insert ["b", "c") and then insert ["a", b"). Verify that the former is not evicted by the latter.
bToCDesc := &roachpb.RangeDescriptor{
StartKey: roachpb.RKey("b"),
EndKey: roachpb.RKey("c"),
}
if err := cache.clearOverlappingCachedRangeDescriptors(bToCDesc); err != nil {
t.Fatal(err)
}
cache.rangeCache.cache.Add(rangeCacheKey(mustMeta(roachpb.RKey("c"))), bToCDesc)
if _, desc, err := cache.getCachedRangeDescriptor(roachpb.RKey("c"), true); err != nil {
t.Fatal(err)
} else if desc != bToCDesc {
t.Errorf("expected descriptor %s; got %s", bToCDesc, desc)
}
aToBDesc := &roachpb.RangeDescriptor{
StartKey: roachpb.RKey("a"),
EndKey: roachpb.RKey("b"),
}
if err := cache.clearOverlappingCachedRangeDescriptors(aToBDesc); err != nil {
t.Fatal(err)
}
cache.rangeCache.cache.Add(rangeCacheKey(mustMeta(roachpb.RKey("b"))), aToBDesc)
if _, desc, err := cache.getCachedRangeDescriptor(roachpb.RKey("c"), true); err != nil {
t.Fatal(err)
} else if desc != bToCDesc {
t.Errorf("expected descriptor %s; got %s", bToCDesc, desc)
}
}
// process performs a consistent lookup on the range descriptor to see if we are
// still a member of the range.
func (q *replicaGCQueue) process(now roachpb.Timestamp, rng *Replica, _ config.SystemConfig) error {
// Note that the Replicas field of desc is probably out of date, so
// we should only use `desc` for its static fields like RangeID and
// StartKey (and avoid rng.GetReplica() for the same reason).
desc := rng.Desc()
// Calls to RangeLookup typically use inconsistent reads, but we
// want to do a consistent read here. This is important when we are
// considering one of the metadata ranges: we must not do an
// inconsistent lookup in our own copy of the range.
b := &client.Batch{}
b.InternalAddRequest(&roachpb.RangeLookupRequest{
Span: roachpb.Span{
Key: keys.RangeMetaKey(desc.StartKey),
},
MaxRanges: 1,
})
br, err := q.db.RunWithResponse(b)
if err != nil {
return err
}
reply := br.Responses[0].GetInner().(*roachpb.RangeLookupResponse)
if len(reply.Ranges) != 1 {
return util.Errorf("expected 1 range descriptor, got %d", len(reply.Ranges))
}
replyDesc := reply.Ranges[0]
currentMember := false
storeID := rng.store.StoreID()
for _, rep := range replyDesc.Replicas {
if rep.StoreID == storeID {
currentMember = true
break
}
}
if !currentMember {
// We are no longer a member of this range; clean up our local data.
if log.V(1) {
log.Infof("destroying local data from range %d", desc.RangeID)
}
if err := rng.store.RemoveReplica(rng, replyDesc, true); err != nil {
return err
}
} else if desc.RangeID != replyDesc.RangeID {
// If we get a different range ID back, then the range has been merged
// away. But currentMember is true, so we are still a member of the
// subsuming range. Shut down raft processing for the former range
// and delete any remaining metadata, but do not delete the data.
if log.V(1) {
log.Infof("removing merged range %d", desc.RangeID)
}
if err := rng.store.RemoveReplica(rng, replyDesc, false); err != nil {
return err
}
// TODO(bdarnell): remove raft logs and other metadata (while leaving a
// tombstone). Add tests for GC of merged ranges.
} else {
// This range is a current member of the raft group. Set the last replica
// GC check time to avoid re-processing for another check interval.
if err := rng.setLastReplicaGCTimestamp(now); err != nil {
return err
}
}
return nil
}
