// Test that leases held before a restart are not used after the restart.
// See replica.mu.minLeaseProposedTS for the reasons why this isn't allowed.
func TestLeaseNotUsedAfterRestart(t *testing.T) {
defer leaktest.AfterTest(t)()
sc := storage.TestStoreConfig(nil)
var leaseAcquisitionTrap atomic.Value
// Disable the split queue so that no ranges are split. This makes it easy
// below to trap any lease request and infer that it refers to the range we're
// interested in.
sc.TestingKnobs.DisableSplitQueue = true
sc.TestingKnobs.LeaseRequestEvent = func(ts hlc.Timestamp) {
val := leaseAcquisitionTrap.Load()
if val == nil {
return
}
trapCallback := val.(func(ts hlc.Timestamp))
if trapCallback != nil {
trapCallback(ts)
}
}
mtc := &multiTestContext{storeConfig: &sc}
mtc.Start(t, 1)
defer mtc.Stop()
// Send a read, to acquire a lease.
getArgs := getArgs([]byte("a"))
if _, err := client.SendWrapped(context.Background(), rg1(mtc.stores[0]), &getArgs); err != nil {
t.Fatal(err)
}
// Restart the mtc. Before we do that, we're installing a callback used to
// assert that a new lease has been requested. The callback is installed
// before the restart, as the lease might be requested at any time and for
// many reasons by background processes, even before we send the read below.
leaseAcquisitionCh := make(chan error)
var once sync.Once
leaseAcquisitionTrap.Store(func(_ hlc.Timestamp) {
once.Do(func() {
close(leaseAcquisitionCh)
})
})
mtc.restart()
// Send another read and check that the pre-existing lease has not been used.
// Concretely, we check that a new lease is requested.
if _, err := client.SendWrapped(context.Background(), rg1(mtc.stores[0]), &getArgs); err != nil {
t.Fatal(err)
}
// Check that the Send above triggered a lease acquisition.
select {
case <-leaseAcquisitionCh:
case <-time.After(time.Second):
t.Fatalf("read did not acquire a new lease")
}
}
// TestStoreRangeMergeStats starts by splitting a range, then writing random data
// to both sides of the split. It then merges the ranges and verifies the merged
// range has stats consistent with recomputations.
func TestStoreRangeMergeStats(t *testing.T) {
defer leaktest.AfterTest(t)()
manual := hlc.NewManualClock(123)
storeCfg := storage.TestStoreConfig(hlc.NewClock(manual.UnixNano, time.Nanosecond))
storeCfg.TestingKnobs.DisableSplitQueue = true
store, stopper := createTestStoreWithConfig(t, storeCfg)
defer stopper.Stop()
// Split the range.
aDesc, bDesc, pErr := createSplitRanges(store)
if pErr != nil {
t.Fatal(pErr)
}
// Write some values left and right of the proposed split key.
writeRandomDataToRange(t, store, aDesc.RangeID, []byte("aaa"))
writeRandomDataToRange(t, store, bDesc.RangeID, []byte("ccc"))
// Get the range stats for both ranges now that we have data.
snap := store.Engine().NewSnapshot()
defer snap.Close()
msA, err := engine.MVCCGetRangeStats(context.Background(), snap, aDesc.RangeID)
if err != nil {
t.Fatal(err)
}
msB, err := engine.MVCCGetRangeStats(context.Background(), snap, bDesc.RangeID)
if err != nil {
t.Fatal(err)
}
// Stats should agree with recomputation.
if err := verifyRecomputedStats(snap, aDesc, msA, manual.UnixNano()); err != nil {
t.Fatalf("failed to verify range A's stats before split: %v", err)
}
if err := verifyRecomputedStats(snap, bDesc, msB, manual.UnixNano()); err != nil {
t.Fatalf("failed to verify range B's stats before split: %v", err)
}
manual.Increment(100)
// Merge the b range back into the a range.
args := adminMergeArgs(roachpb.KeyMin)
if _, err := client.SendWrapped(context.Background(), rg1(store), &args); err != nil {
t.Fatal(err)
}
replMerged := store.LookupReplica(aDesc.StartKey, nil)
// Get the range stats for the merged range and verify.
snap = store.Engine().NewSnapshot()
defer snap.Close()
msMerged, err := engine.MVCCGetRangeStats(context.Background(), snap, replMerged.RangeID)
if err != nil {
t.Fatal(err)
}
// Merged stats should agree with recomputation.
if err := verifyRecomputedStats(snap, replMerged.Desc(), msMerged, manual.UnixNano()); err != nil {
t.Errorf("failed to verify range's stats after merge: %v", err)
}
}
// TestStoreRangeMergeTwoEmptyRanges tries to merge two empty ranges together.
func TestStoreRangeMergeTwoEmptyRanges(t *testing.T) {
defer leaktest.AfterTest(t)()
storeCfg := storage.TestStoreConfig(nil)
storeCfg.TestingKnobs.DisableSplitQueue = true
store, stopper := createTestStoreWithConfig(t, storeCfg)
defer stopper.Stop()
if _, _, err := createSplitRanges(store); err != nil {
t.Fatal(err)
}
// Merge the b range back into the a range.
args := adminMergeArgs(roachpb.KeyMin)
_, err := client.SendWrapped(context.Background(), rg1(store), &args)
if err != nil {
t.Fatal(err)
}
// Verify the merge by looking up keys from both ranges.
replicaA := store.LookupReplica([]byte("a"), nil)
replicaB := store.LookupReplica([]byte("c"), nil)
if !reflect.DeepEqual(replicaA, replicaB) {
t.Fatalf("ranges were not merged %s!=%s", replicaA, replicaB)
}
}
// TestStoreRangeMergeNonCollocated attempts to merge two ranges
// that are not on the same stores.
func TestStoreRangeMergeNonCollocated(t *testing.T) {
defer leaktest.AfterTest(t)()
mtc := startMultiTestContext(t, 4)
defer mtc.Stop()
store := mtc.stores[0]
// Split into 3 ranges
argsSplit := adminSplitArgs(roachpb.KeyMin, []byte("d"))
if _, pErr := client.SendWrapped(context.Background(), rg1(store), &argsSplit); pErr != nil {
t.Fatalf("Can't split range %s", pErr)
}
argsSplit = adminSplitArgs(roachpb.KeyMin, []byte("b"))
if _, pErr := client.SendWrapped(context.Background(), rg1(store), &argsSplit); pErr != nil {
t.Fatalf("Can't split range %s", pErr)
}
rangeA := store.LookupReplica([]byte("a"), nil)
rangeADesc := rangeA.Desc()
rangeB := store.LookupReplica([]byte("c"), nil)
rangeBDesc := rangeB.Desc()
rangeC := store.LookupReplica([]byte("e"), nil)
rangeCDesc := rangeC.Desc()
if bytes.Equal(rangeADesc.StartKey, rangeBDesc.StartKey) {
log.Errorf(context.TODO(), "split ranges keys are equal %q!=%q", rangeADesc.StartKey, rangeBDesc.StartKey)
}
if bytes.Equal(rangeBDesc.StartKey, rangeCDesc.StartKey) {
log.Errorf(context.TODO(), "split ranges keys are equal %q!=%q", rangeBDesc.StartKey, rangeCDesc.StartKey)
}
if bytes.Equal(rangeADesc.StartKey, rangeCDesc.StartKey) {
log.Errorf(context.TODO(), "split ranges keys are equal %q!=%q", rangeADesc.StartKey, rangeCDesc.StartKey)
}
// Replicate the ranges to different sets of stores. Ranges A and C
// are collocated, but B is different.
mtc.replicateRange(rangeA.RangeID, 1, 2)
mtc.replicateRange(rangeB.RangeID, 1, 3)
mtc.replicateRange(rangeC.RangeID, 1, 2)
// Attempt to merge.
rangeADesc = rangeA.Desc()
argsMerge := adminMergeArgs(roachpb.Key(rangeADesc.StartKey))
if _, pErr := rangeA.AdminMerge(context.Background(), argsMerge, rangeADesc); !testutils.IsPError(pErr, "ranges not collocated") {
t.Fatalf("did not got expected error; got %s", pErr)
}
}
// TestStoreRangeMergeLastRange verifies that merging the last range
// fails.
func TestStoreRangeMergeLastRange(t *testing.T) {
defer leaktest.AfterTest(t)()
storeCfg := storage.TestStoreConfig(nil)
storeCfg.TestingKnobs.DisableSplitQueue = true
store, stopper := createTestStoreWithConfig(t, storeCfg)
defer stopper.Stop()
// Merge last range.
args := adminMergeArgs(roachpb.KeyMin)
if _, pErr := client.SendWrapped(context.Background(), rg1(store), &args); !testutils.IsPError(pErr, "cannot merge final range") {
t.Fatalf("expected 'cannot merge final range' error; got %s", pErr)
}
}
func BenchmarkStoreRangeMerge(b *testing.B) {
defer tracing.Disable()()
storeCfg := storage.TestStoreConfig(nil)
storeCfg.TestingKnobs.DisableSplitQueue = true
stopper := stop.NewStopper()
defer stopper.Stop()
store := createTestStoreWithConfig(b, stopper, storeCfg)
// Perform initial split of ranges.
sArgs := adminSplitArgs(roachpb.KeyMin, []byte("b"))
if _, err := client.SendWrapped(context.Background(), rg1(store), sArgs); err != nil {
b.Fatal(err)
}
// Write some values left and right of the proposed split key.
aDesc := store.LookupReplica([]byte("a"), nil).Desc()
bDesc := store.LookupReplica([]byte("c"), nil).Desc()
writeRandomDataToRange(b, store, aDesc.RangeID, []byte("aaa"))
writeRandomDataToRange(b, store, bDesc.RangeID, []byte("ccc"))
// Create args to merge the b range back into the a range.
mArgs := adminMergeArgs(roachpb.KeyMin)
b.ResetTimer()
for i := 0; i < b.N; i++ {
// Merge the ranges.
b.StartTimer()
if _, err := client.SendWrapped(context.Background(), rg1(store), mArgs); err != nil {
b.Fatal(err)
}
// Split the range.
b.StopTimer()
if _, err := client.SendWrapped(context.Background(), rg1(store), sArgs); err != nil {
b.Fatal(err)
}
}
}
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,
}
sender := c.Clients[nodeIdx].GetSender()
resp, pErr := client.SendWrapped(context.Background(), sender, req)
if pErr != nil {
return nil, errors.Errorf("%s: lookup range: %s", key, pErr)
}
return &resp.(*roachpb.RangeLookupResponse).Ranges[0], nil
}
// LookupRange returns the descriptor of the range containing key.
func (ts *TestServer) LookupRange(key roachpb.Key) (roachpb.RangeDescriptor, error) {
rangeLookupReq := roachpb.RangeLookupRequest{
Span: roachpb.Span{
Key: keys.RangeMetaKey(keys.MustAddr(key)),
},
MaxRanges: 1,
}
resp, pErr := client.SendWrapped(context.Background(), ts.DistSender(), &rangeLookupReq)
if pErr != nil {
return roachpb.RangeDescriptor{}, errors.Errorf(
"%q: lookup range unexpected error: %s", key, pErr)
}
return resp.(*roachpb.RangeLookupResponse).Ranges[0], nil
}
// SplitRange splits the range containing splitKey.
// The right range created by the split starts at the split key and extends to the
// original range's end key.
// Returns the new descriptors of the left and right ranges.
//
// splitKey must correspond to a SQL table key (it must end with a family ID /
// col ID).
func (ts *TestServer) SplitRange(
splitKey roachpb.Key,
) (roachpb.RangeDescriptor, roachpb.RangeDescriptor, error) {
splitRKey, err := keys.Addr(splitKey)
if err != nil {
return roachpb.RangeDescriptor{}, roachpb.RangeDescriptor{}, err
}
origRangeDesc, err := ts.LookupRange(splitKey)
if err != nil {
return roachpb.RangeDescriptor{}, roachpb.RangeDescriptor{}, err
}
if origRangeDesc.StartKey.Equal(splitRKey) {
return roachpb.RangeDescriptor{}, roachpb.RangeDescriptor{},
errors.Errorf(
"cannot split range %+v at start key %q", origRangeDesc, splitKey)
}
splitReq := roachpb.AdminSplitRequest{
Span: roachpb.Span{
Key: splitKey,
},
SplitKey: splitKey,
}
_, pErr := client.SendWrapped(context.Background(), ts.DistSender(), &splitReq)
if pErr != nil {
return roachpb.RangeDescriptor{}, roachpb.RangeDescriptor{},
errors.Errorf(
"%q: split unexpected error: %s", splitReq.SplitKey, pErr)
}
var leftRangeDesc, rightRangeDesc roachpb.RangeDescriptor
if err := ts.DB().GetProto(context.TODO(),
keys.RangeDescriptorKey(origRangeDesc.StartKey), &leftRangeDesc); err != nil {
return roachpb.RangeDescriptor{}, roachpb.RangeDescriptor{},
errors.Wrap(err, "could not look up left-hand side descriptor")
}
// The split point might not be exactly the one we requested (it can be
// adjusted slightly so we don't split in the middle of SQL rows). Update it
// to the real point.
splitRKey = leftRangeDesc.EndKey
if err := ts.DB().GetProto(context.TODO(),
keys.RangeDescriptorKey(splitRKey), &rightRangeDesc); err != nil {
return roachpb.RangeDescriptor{}, roachpb.RangeDescriptor{},
errors.Wrap(err, "could not look up right-hand side descriptor")
}
return leftRangeDesc, rightRangeDesc, nil
}
// TestProactiveRaftLogTruncate verifies that we proactively truncate the raft
// log even when replica scanning is disabled.
func TestProactiveRaftLogTruncate(t *testing.T) {
defer leaktest.AfterTest(t)()
t.Skip("#9772")
store, _, stopper := createTestStore(t)
defer stopper.Stop()
store.SetReplicaScannerActive(false)
r, err := store.GetReplica(1)
if err != nil {
t.Fatal(err)
}
r.mu.Lock()
oldFirstIndex, err := r.FirstIndex()
r.mu.Unlock()
if err != nil {
t.Fatal(err)
}
// Write a few keys to the range. While writing these keys, the raft log
// should be proactively truncated even though replica scanning is disabled.
for i := 0; i < 2*RaftLogQueueStaleThreshold; i++ {
key := roachpb.Key(fmt.Sprintf("key%02d", i))
args := putArgs(key, []byte(fmt.Sprintf("value%02d", i)))
if _, err := client.SendWrapped(context.Background(), store.testSender(), &args); err != nil {
t.Fatal(err)
}
}
// Wait for any asynchronous tasks to finish.
stopper.Quiesce()
r.mu.Lock()
newFirstIndex, err := r.FirstIndex()
r.mu.Unlock()
if err != nil {
t.Fatal(err)
}
if newFirstIndex <= oldFirstIndex {
t.Errorf("log was not correctly truncated, old first index:%d, current first index:%d",
oldFirstIndex, newFirstIndex)
}
}
func createSplitRanges(
store *storage.Store,
) (*roachpb.RangeDescriptor, *roachpb.RangeDescriptor, *roachpb.Error) {
args := adminSplitArgs(roachpb.KeyMin, []byte("b"))
if _, err := client.SendWrapped(context.Background(), rg1(store), &args); err != nil {
return nil, nil, err
}
rangeADesc := store.LookupReplica([]byte("a"), nil).Desc()
rangeBDesc := store.LookupReplica([]byte("c"), nil).Desc()
if bytes.Equal(rangeADesc.StartKey, rangeBDesc.StartKey) {
log.Errorf(context.TODO(), "split ranges keys are equal %q!=%q", rangeADesc.StartKey, rangeBDesc.StartKey)
}
return rangeADesc, rangeBDesc, nil
}
// TestServerStartClock tests that a server's clock is not pushed out of thin
// air. This used to happen - the simple act of starting was causing a server's
// clock to be pushed because we were introducing bogus future timestamps into
// our system.
func TestServerStartClock(t *testing.T) {
defer leaktest.AfterTest(t)()
// Set a high max-offset so that, if the server's clock is pushed by
// MaxOffset, we don't hide that under the latency of the Start operation
// which would allow the physical clock to catch up to the pushed one.
params := base.TestServerArgs{
Knobs: base.TestingKnobs{
Store: &storage.StoreTestingKnobs{
MaxOffset: time.Second,
},
},
}
s, _, _ := serverutils.StartServer(t, params)
defer s.Stopper().Stop()
// Run a command so that we are sure to touch the timestamp cache. This is
// actually not needed because other commands run during server
// initialization, but we cannot guarantee that's going to stay that way.
get := &roachpb.GetRequest{
Span: roachpb.Span{Key: roachpb.Key("a")},
}
if _, err := client.SendWrapped(
context.Background(), s.KVClient().(*client.DB).GetSender(), get,
); err != nil {
t.Fatal(err)
}
now := s.Clock().Now()
// We rely on s.Clock() having been initialized from hlc.UnixNano(), which is a
// bit fragile.
physicalNow := hlc.UnixNano()
serverClockWasPushed := (now.Logical > 0) || (now.WallTime > physicalNow)
if serverClockWasPushed {
t.Fatalf("time: server %s vs actual %d", now, physicalNow)
}
}
func TestStoreMetrics(t *testing.T) {
defer leaktest.AfterTest(t)()
t.Skip("TODO(mrtracy): #9204")
mtc := &multiTestContext{}
defer mtc.Stop()
mtc.Start(t, 3)
// Flush RocksDB memtables, so that RocksDB begins using block-based tables.
// This is useful, because most of the stats we track don't apply to
// memtables.
if err := mtc.stores[0].Engine().Flush(); err != nil {
t.Fatal(err)
}
if err := mtc.stores[1].Engine().Flush(); err != nil {
t.Fatal(err)
}
// Disable the raft log truncation which confuses this test.
for _, s := range mtc.stores {
s.SetRaftLogQueueActive(false)
}
// Perform a split, which has special metrics handling.
splitArgs := adminSplitArgs(roachpb.KeyMin, roachpb.Key("m"))
if _, err := client.SendWrapped(context.Background(), rg1(mtc.stores[0]), splitArgs); err != nil {
t.Fatal(err)
}
// Verify range count is as expected
checkCounter(t, mtc.stores[0].Metrics().ReplicaCount, 2)
// Verify all stats on store0 after split.
verifyStats(t, mtc, 0)
// Replicate the "right" range to the other stores.
replica := mtc.stores[0].LookupReplica(roachpb.RKey("z"), nil)
mtc.replicateRange(replica.RangeID, 1, 2)
// Verify stats on store1 after replication.
verifyStats(t, mtc, 1)
// Add some data to the "right" range.
dataKey := []byte("z")
if _, err := mtc.dbs[0].Inc(context.TODO(), dataKey, 5); err != nil {
t.Fatal(err)
}
mtc.waitForValues(roachpb.Key("z"), []int64{5, 5, 5})
// Verify all stats on store 0 and 1 after addition.
verifyStats(t, mtc, 0, 1)
// Create a transaction statement that fails, but will add an entry to the
// sequence cache. Regression test for #4969.
if err := mtc.dbs[0].Txn(context.TODO(), func(txn *client.Txn) error {
b := txn.NewBatch()
b.CPut(dataKey, 7, 6)
return txn.Run(b)
}); err == nil {
t.Fatal("Expected transaction error, but none received")
}
// Verify stats after sequence cache addition.
verifyStats(t, mtc, 0)
checkCounter(t, mtc.stores[0].Metrics().ReplicaCount, 2)
// Unreplicate range from the first store.
mtc.unreplicateRange(replica.RangeID, 0)
// Force GC Scan on store 0 in order to fully remove range.
mtc.stores[1].ForceReplicaGCScanAndProcess()
mtc.waitForValues(roachpb.Key("z"), []int64{0, 5, 5})
// Verify range count is as expected.
checkCounter(t, mtc.stores[0].Metrics().ReplicaCount, 1)
checkCounter(t, mtc.stores[1].Metrics().ReplicaCount, 1)
// Verify all stats on store0 and store1 after range is removed.
verifyStats(t, mtc, 0, 1)
verifyRocksDBStats(t, mtc.stores[0])
verifyRocksDBStats(t, mtc.stores[1])
}
// TestGetTruncatableIndexes verifies that old raft log entries are correctly
// removed.
func TestGetTruncatableIndexes(t *testing.T) {
defer leaktest.AfterTest(t)()
store, _, stopper := createTestStore(t)
defer stopper.Stop()
store.SetRaftLogQueueActive(false)
r, err := store.GetReplica(1)
if err != nil {
t.Fatal(err)
}
getIndexes := func() (uint64, uint64, uint64, error) {
r.mu.Lock()
firstIndex, err := r.FirstIndex()
r.mu.Unlock()
if err != nil {
return 0, 0, 0, err
}
truncatableIndexes, oldestIndex, err := getTruncatableIndexes(context.Background(), r)
if err != nil {
return 0, 0, 0, err
}
return firstIndex, truncatableIndexes, oldestIndex, nil
}
aFirst, aTruncatable, aOldest, err := getIndexes()
if err != nil {
t.Fatal(err)
}
if aFirst == 0 {
t.Errorf("expected first index to be greater than 0, got %d", aFirst)
}
// Write a few keys to the range.
for i := 0; i < RaftLogQueueStaleThreshold+1; i++ {
key := roachpb.Key(fmt.Sprintf("key%02d", i))
args := putArgs(key, []byte(fmt.Sprintf("value%02d", i)))
if _, err := client.SendWrapped(context.Background(), store.testSender(), &args); err != nil {
t.Fatal(err)
}
}
bFirst, bTruncatable, bOldest, err := getIndexes()
if err != nil {
t.Fatal(err)
}
if aFirst != bFirst {
t.Errorf("expected firstIndex to not change, instead it changed from %d -> %d", aFirst, bFirst)
}
if aTruncatable >= bTruncatable {
t.Errorf("expected truncatableIndexes to increase, instead it changed from %d -> %d", aTruncatable, bTruncatable)
}
if aOldest >= bOldest {
t.Errorf("expected oldestIndex to increase, instead it changed from %d -> %d", aOldest, bOldest)
}
// Enable the raft log scanner and and force a truncation.
store.SetRaftLogQueueActive(true)
store.ForceRaftLogScanAndProcess()
store.SetRaftLogQueueActive(false)
// There can be a delay from when the truncation command is issued and the
// indexes updating.
var cFirst, cTruncatable, cOldest uint64
util.SucceedsSoon(t, func() error {
var err error
cFirst, cTruncatable, cOldest, err = getIndexes()
if err != nil {
t.Fatal(err)
}
if bFirst == cFirst {
return errors.Errorf("truncation did not occur, expected firstIndex to change, instead it remained at %d", cFirst)
}
return nil
})
if bTruncatable < cTruncatable {
t.Errorf("expected truncatableIndexes to decrease, instead it changed from %d -> %d", bTruncatable, cTruncatable)
}
if bOldest >= cOldest {
t.Errorf("expected oldestIndex to increase, instead it changed from %d -> %d", bOldest, cOldest)
}
// Again, enable the raft log scanner and and force a truncation. This time
// we expect no truncation to occur.
store.SetRaftLogQueueActive(true)
store.ForceRaftLogScanAndProcess()
store.SetRaftLogQueueActive(false)
// Unlike the last iteration, where we expect a truncation and can wait on
// it with succeedsSoon, we can't do that here. This check is fragile in
// that the truncation triggered here may lose the race against the call to
// GetFirstIndex or getTruncatableIndexes, giving a false negative. Fixing
// this requires additional instrumentation of the queues, which was deemed
// to require too much work at the time of this writing.
dFirst, dTruncatable, dOldest, err := getIndexes()
if err != nil {
t.Fatal(err)
}
if cFirst != dFirst {
t.Errorf("truncation should not have occurred, but firstIndex changed from %d -> %d", cFirst, dFirst)
}
if cTruncatable != dTruncatable {
t.Errorf("truncation should not have occurred, but truncatableIndexes changed from %d -> %d", cTruncatable, dTruncatable)
}
if cOldest != dOldest {
t.Errorf("truncation should not have occurred, but oldestIndex changed from %d -> %d", cOldest, dOldest)
}
}
// TestStoreRangeMergeMetadataCleanup tests that all metadata of a
// subsumed range is cleaned up on merge.
func TestStoreRangeMergeMetadataCleanup(t *testing.T) {
defer leaktest.AfterTest(t)()
storeCfg := storage.TestStoreConfig(nil)
storeCfg.TestingKnobs.DisableSplitQueue = true
store, stopper := createTestStoreWithConfig(t, storeCfg)
defer stopper.Stop()
scan := func(f func(roachpb.KeyValue) (bool, error)) {
if _, err := engine.MVCCIterate(context.Background(), store.Engine(), roachpb.KeyMin, roachpb.KeyMax, hlc.ZeroTimestamp, true, nil, false, f); err != nil {
t.Fatal(err)
}
}
content := roachpb.Key("testing!")
// Write some values left of the proposed split key.
pArgs := putArgs([]byte("aaa"), content)
if _, err := client.SendWrapped(context.Background(), rg1(store), &pArgs); err != nil {
t.Fatal(err)
}
// Collect all the keys.
preKeys := make(map[string]struct{})
scan(func(kv roachpb.KeyValue) (bool, error) {
preKeys[string(kv.Key)] = struct{}{}
return false, nil
})
// Split the range.
_, bDesc, err := createSplitRanges(store)
if err != nil {
t.Fatal(err)
}
// Write some values right of the split key.
pArgs = putArgs([]byte("ccc"), content)
if _, err := client.SendWrappedWith(context.Background(), rg1(store), roachpb.Header{
RangeID: bDesc.RangeID,
}, &pArgs); err != nil {
t.Fatal(err)
}
// Merge the b range back into the a range.
args := adminMergeArgs(roachpb.KeyMin)
if _, err := client.SendWrapped(context.Background(), rg1(store), &args); err != nil {
t.Fatal(err)
}
// Collect all the keys again.
postKeys := make(map[string]struct{})
scan(func(kv roachpb.KeyValue) (bool, error) {
postKeys[string(kv.Key)] = struct{}{}
return false, nil
})
// Compute the new keys.
for k := range preKeys {
delete(postKeys, k)
}
// Keep only the subsumed range's local keys.
localRangeKeyPrefix := string(keys.MakeRangeIDPrefix(bDesc.RangeID))
for k := range postKeys {
if !strings.HasPrefix(k, localRangeKeyPrefix) {
delete(postKeys, k)
}
}
if numKeys := len(postKeys); numKeys > 0 {
var buf bytes.Buffer
fmt.Fprintf(&buf, "%d keys were not cleaned up:\n", numKeys)
for k := range postKeys {
fmt.Fprintf(&buf, "%q\n", k)
}
t.Fatal(buf.String())
}
}
// TestLeaseMetricsOnSplitAndTransfer verifies that lease-related metrics
// are updated after splitting a range and then initiating one successful
// and one failing lease transfer.
func TestLeaseMetricsOnSplitAndTransfer(t *testing.T) {
defer leaktest.AfterTest(t)()
var injectLeaseTransferError atomic.Value
sc := storage.TestStoreConfig(nil)
sc.TestingKnobs.DisableSplitQueue = true
sc.TestingKnobs.TestingCommandFilter =
func(filterArgs storagebase.FilterArgs) *roachpb.Error {
if args, ok := filterArgs.Req.(*roachpb.TransferLeaseRequest); ok {
if val := injectLeaseTransferError.Load(); val != nil && val.(bool) {
// Note that we can't just return an error here as we only
// end up counting failures in the metrics if the command
// makes it through to being executed. So use a fake store ID.
args.Lease.Replica.StoreID = roachpb.StoreID(1000)
}
}
return nil
}
mtc := &multiTestContext{storeConfig: &sc}
defer mtc.Stop()
mtc.Start(t, 2)
// Up-replicate to two replicas.
keyMinReplica0 := mtc.stores[0].LookupReplica(roachpb.RKeyMin, nil)
mtc.replicateRange(keyMinReplica0.RangeID, 1)
// Split the key space at key "a".
splitKey := roachpb.RKey("a")
splitArgs := adminSplitArgs(splitKey.AsRawKey(), splitKey.AsRawKey())
if _, pErr := client.SendWrapped(
context.Background(), rg1(mtc.stores[0]), splitArgs,
); pErr != nil {
t.Fatal(pErr)
}
// Now, a successful transfer from LHS replica 0 to replica 1.
injectLeaseTransferError.Store(false)
if err := mtc.dbs[0].AdminTransferLease(
context.TODO(), keyMinReplica0.Desc().StartKey.AsRawKey(), mtc.stores[1].StoreID(),
); err != nil {
t.Fatalf("unable to transfer lease to replica 1: %s", err)
}
// Wait for all replicas to process.
testutils.SucceedsSoon(t, func() error {
for i := 0; i < 2; i++ {
r := mtc.stores[i].LookupReplica(roachpb.RKeyMin, nil)
if l, _ := r.GetLease(); l.Replica.StoreID != mtc.stores[1].StoreID() {
return errors.Errorf("expected lease to transfer to replica 2: got %s", l)
}
}
return nil
})
// Next a failed transfer from RHS replica 0 to replica 1.
injectLeaseTransferError.Store(true)
keyAReplica0 := mtc.stores[0].LookupReplica(splitKey, nil)
if err := mtc.dbs[0].AdminTransferLease(
context.TODO(), keyAReplica0.Desc().StartKey.AsRawKey(), mtc.stores[1].StoreID(),
); err == nil {
t.Fatal("expected an error transferring to an unknown store ID")
}
metrics := mtc.stores[0].Metrics()
if a, e := metrics.LeaseTransferSuccessCount.Count(), int64(1); a != e {
t.Errorf("expected %d lease transfer successes; got %d", e, a)
}
if a, e := metrics.LeaseTransferErrorCount.Count(), int64(1); a != e {
t.Errorf("expected %d lease transfer errors; got %d", e, a)
}
// Expire current leases and put a key to RHS of split to request
// an epoch-based lease.
testutils.SucceedsSoon(t, func() error {
mtc.expireLeases(context.TODO())
if err := mtc.stores[0].DB().Put(context.TODO(), "a", "foo"); err != nil {
return err
}
// Update replication gauges on store 1 and verify we have 1 each of
// expiration and epoch leases. These values are counted from store 1
// because it will have the higher replica IDs. Expire leases to make
// sure that epoch-based leases are used for the split range.
if err := mtc.stores[1].ComputeMetrics(context.Background(), 0); err != nil {
return err
}
metrics = mtc.stores[1].Metrics()
if a, e := metrics.LeaseExpirationCount.Value(), int64(1); a != e {
return errors.Errorf("expected %d expiration lease count; got %d", e, a)
}
if a, e := metrics.LeaseEpochCount.Value(), int64(1); a != e {
return errors.Errorf("expected %d epoch lease count; got %d", e, a)
}
return nil
})
}
func TestRangeInfo(t *testing.T) {
defer leaktest.AfterTest(t)()
mtc := startMultiTestContext(t, 2)
defer mtc.Stop()
// Up-replicate to two replicas.
mtc.replicateRange(mtc.stores[0].LookupReplica(roachpb.RKeyMin, nil).RangeID, 1)
// Split the key space at key "a".
splitKey := roachpb.RKey("a")
splitArgs := adminSplitArgs(splitKey.AsRawKey(), splitKey.AsRawKey())
if _, pErr := client.SendWrapped(
context.Background(), rg1(mtc.stores[0]), &splitArgs,
); pErr != nil {
t.Fatal(pErr)
}
// Get the replicas for each side of the split. This is done within
// a SucceedsSoon loop to ensure the split completes.
var lhsReplica0, lhsReplica1, rhsReplica0, rhsReplica1 *storage.Replica
util.SucceedsSoon(t, func() error {
lhsReplica0 = mtc.stores[0].LookupReplica(roachpb.RKeyMin, nil)
lhsReplica1 = mtc.stores[1].LookupReplica(roachpb.RKeyMin, nil)
rhsReplica0 = mtc.stores[0].LookupReplica(splitKey, nil)
rhsReplica1 = mtc.stores[1].LookupReplica(splitKey, nil)
if lhsReplica0 == rhsReplica0 || lhsReplica1 == rhsReplica1 {
return errors.Errorf("replicas not post-split %v, %v, %v, %v",
lhsReplica0, rhsReplica0, rhsReplica0, rhsReplica1)
}
return nil
})
lhsLease, _ := lhsReplica0.GetLease()
rhsLease, _ := rhsReplica0.GetLease()
// Verify range info is not set if unrequested.
getArgs := getArgs(splitKey.AsRawKey())
reply, pErr := client.SendWrapped(context.Background(), mtc.distSenders[0], &getArgs)
if pErr != nil {
t.Fatal(pErr)
}
if len(reply.Header().RangeInfos) > 0 {
t.Errorf("expected empty range infos if unrequested; got %v", reply.Header().RangeInfos)
}
// Verify range info on a get request.
h := roachpb.Header{
ReturnRangeInfo: true,
}
reply, pErr = client.SendWrappedWith(context.Background(), mtc.distSenders[0], h, &getArgs)
if pErr != nil {
t.Fatal(pErr)
}
expRangeInfos := []roachpb.RangeInfo{
{
Desc: *rhsReplica0.Desc(),
Lease: *rhsLease,
},
}
if !reflect.DeepEqual(reply.Header().RangeInfos, expRangeInfos) {
t.Errorf("on get reply, expected %+v; got %+v", expRangeInfos, reply.Header().RangeInfos)
}
// Verify range info on a put request.
putArgs := putArgs(splitKey.AsRawKey(), []byte("foo"))
reply, pErr = client.SendWrappedWith(context.Background(), mtc.distSenders[0], h, &putArgs)
if pErr != nil {
t.Fatal(pErr)
}
if !reflect.DeepEqual(reply.Header().RangeInfos, expRangeInfos) {
t.Errorf("on put reply, expected %+v; got %+v", expRangeInfos, reply.Header().RangeInfos)
}
// Verify multiple range infos on a scan request.
scanArgs := roachpb.ScanRequest{
Span: roachpb.Span{
Key: keys.SystemMax,
EndKey: roachpb.KeyMax,
},
}
h.Txn = roachpb.NewTransaction("test", roachpb.KeyMin, 1, enginepb.SERIALIZABLE, mtc.clock.Now(), 0)
reply, pErr = client.SendWrappedWith(context.Background(), mtc.distSenders[0], h, &scanArgs)
if pErr != nil {
t.Fatal(pErr)
}
expRangeInfos = []roachpb.RangeInfo{
{
Desc: *lhsReplica0.Desc(),
Lease: *lhsLease,
},
{
Desc: *rhsReplica0.Desc(),
Lease: *rhsLease,
},
}
if !reflect.DeepEqual(reply.Header().RangeInfos, expRangeInfos) {
t.Errorf("on scan reply, expected %+v; got %+v", expRangeInfos, reply.Header().RangeInfos)
}
// Verify multiple range infos and order on a reverse scan request.
revScanArgs := roachpb.ReverseScanRequest{
Span: roachpb.Span{
Key: keys.SystemMax,
EndKey: roachpb.KeyMax,
},
}
reply, pErr = client.SendWrappedWith(context.Background(), mtc.distSenders[0], h, &revScanArgs)
if pErr != nil {
t.Fatal(pErr)
}
expRangeInfos = []roachpb.RangeInfo{
{
Desc: *rhsReplica0.Desc(),
Lease: *rhsLease,
},
{
Desc: *lhsReplica0.Desc(),
Lease: *lhsLease,
},
}
if !reflect.DeepEqual(reply.Header().RangeInfos, expRangeInfos) {
t.Errorf("on reverse scan reply, expected %+v; got %+v", expRangeInfos, reply.Header().RangeInfos)
}
// Change lease holders for both ranges and re-scan.
for _, r := range []*storage.Replica{lhsReplica1, rhsReplica1} {
replDesc, err := r.GetReplicaDescriptor()
if err != nil {
t.Fatal(err)
}
if err = mtc.dbs[0].AdminTransferLease(context.TODO(),
r.Desc().StartKey.AsRawKey(), replDesc.StoreID); err != nil {
t.Fatalf("unable to transfer lease to replica %s: %s", r, err)
}
}
reply, pErr = client.SendWrappedWith(context.Background(), mtc.distSenders[0], h, &scanArgs)
if pErr != nil {
t.Fatal(pErr)
}
lhsLease, _ = lhsReplica1.GetLease()
rhsLease, _ = rhsReplica1.GetLease()
expRangeInfos = []roachpb.RangeInfo{
{
Desc: *lhsReplica1.Desc(),
Lease: *lhsLease,
},
{
Desc: *rhsReplica1.Desc(),
Lease: *rhsLease,
},
}
if !reflect.DeepEqual(reply.Header().RangeInfos, expRangeInfos) {
t.Errorf("on scan reply, expected %+v; got %+v", expRangeInfos, reply.Header().RangeInfos)
}
}
func TestRangeTransferLease(t *testing.T) {
defer leaktest.AfterTest(t)()
cfg := storage.TestStoreConfig(nil)
var filterMu syncutil.Mutex
var filter func(filterArgs storagebase.FilterArgs) *roachpb.Error
cfg.TestingKnobs.TestingCommandFilter =
func(filterArgs storagebase.FilterArgs) *roachpb.Error {
filterMu.Lock()
filterCopy := filter
filterMu.Unlock()
if filterCopy != nil {
return filterCopy(filterArgs)
}
return nil
}
var waitForTransferBlocked atomic.Value
waitForTransferBlocked.Store(false)
transferBlocked := make(chan struct{})
cfg.TestingKnobs.LeaseTransferBlockedOnExtensionEvent = func(
_ roachpb.ReplicaDescriptor) {
if waitForTransferBlocked.Load().(bool) {
transferBlocked <- struct{}{}
waitForTransferBlocked.Store(false)
}
}
mtc := &multiTestContext{}
mtc.storeConfig = &cfg
mtc.Start(t, 2)
defer mtc.Stop()
// First, do a write; we'll use it to determine when the dust has settled.
leftKey := roachpb.Key("a")
incArgs := incrementArgs(leftKey, 1)
if _, pErr := client.SendWrapped(context.Background(), mtc.distSenders[0], &incArgs); pErr != nil {
t.Fatal(pErr)
}
// Get the left range's ID.
rangeID := mtc.stores[0].LookupReplica(roachpb.RKey("a"), nil).RangeID
// Replicate the left range onto node 1.
mtc.replicateRange(rangeID, 1)
replica0 := mtc.stores[0].LookupReplica(roachpb.RKey("a"), nil)
replica1 := mtc.stores[1].LookupReplica(roachpb.RKey("a"), nil)
gArgs := getArgs(leftKey)
replica0Desc, err := replica0.GetReplicaDescriptor()
if err != nil {
t.Fatal(err)
}
// Check that replica0 can serve reads OK.
if _, pErr := client.SendWrappedWith(
context.Background(),
mtc.senders[0],
roachpb.Header{Replica: replica0Desc},
&gArgs,
); pErr != nil {
t.Fatal(pErr)
}
{
// Transferring the lease to ourself should be a no-op.
origLeasePtr, _ := replica0.GetLease()
origLease := *origLeasePtr
if err := replica0.AdminTransferLease(replica0Desc.StoreID); err != nil {
t.Fatal(err)
}
newLeasePtr, _ := replica0.GetLease()
if origLeasePtr != newLeasePtr || origLease != *newLeasePtr {
t.Fatalf("expected %+v, but found %+v", origLeasePtr, newLeasePtr)
}
}
{
// An invalid target should result in an error.
const expected = "unable to find store .* in range"
if err := replica0.AdminTransferLease(1000); !testutils.IsError(err, expected) {
t.Fatalf("expected %s, but found %v", expected, err)
}
}
// Move the lease to store 1.
var newHolderDesc roachpb.ReplicaDescriptor
util.SucceedsSoon(t, func() error {
var err error
newHolderDesc, err = replica1.GetReplicaDescriptor()
return err
})
if err := replica0.AdminTransferLease(newHolderDesc.StoreID); err != nil {
t.Fatal(err)
}
// Check that replica0 doesn't serve reads any more.
replica0Desc, err = replica0.GetReplicaDescriptor()
if err != nil {
t.Fatal(err)
}
_, pErr := client.SendWrappedWith(
context.Background(),
mtc.senders[0],
roachpb.Header{Replica: replica0Desc},
&gArgs,
)
nlhe, ok := pErr.GetDetail().(*roachpb.NotLeaseHolderError)
if !ok {
t.Fatalf("expected %T, got %s", &roachpb.NotLeaseHolderError{}, pErr)
}
if *(nlhe.LeaseHolder) != newHolderDesc {
t.Fatalf("expected lease holder %+v, got %+v",
newHolderDesc, nlhe.LeaseHolder)
}
// Check that replica1 now has the lease (or gets it soon).
util.SucceedsSoon(t, func() error {
if _, pErr := client.SendWrappedWith(
context.Background(),
mtc.senders[1],
roachpb.Header{Replica: replica0Desc},
&gArgs,
); pErr != nil {
return pErr.GoError()
}
return nil
})
replica1Lease, _ := replica1.GetLease()
// Verify the timestamp cache low water. Because we executed a transfer lease
// request, the low water should be set to the new lease start time which is
// less than the previous lease's expiration time.
if lowWater := replica1.GetTimestampCacheLowWater(); lowWater != replica1Lease.Start {
t.Fatalf("expected timestamp cache low water %s, but found %s",
replica1Lease.Start, lowWater)
}
// Make replica1 extend its lease and transfer the lease immediately after
// that. Test that the transfer still happens (it'll wait until the extension
// is done).
extensionSem := make(chan struct{})
filterMu.Lock()
filter = func(filterArgs storagebase.FilterArgs) *roachpb.Error {
if filterArgs.Sid != mtc.stores[1].Ident.StoreID {
return nil
}
llReq, ok := filterArgs.Req.(*roachpb.RequestLeaseRequest)
if !ok {
return nil
}
if llReq.Lease.Replica == newHolderDesc {
// Notify the main thread that the extension is in progress and wait for
// the signal to proceed.
filterMu.Lock()
filter = nil
filterMu.Unlock()
extensionSem <- struct{}{}
<-extensionSem
}
return nil
}
filterMu.Unlock()
// Initiate an extension.
var wg sync.WaitGroup
wg.Add(1)
go func() {
defer wg.Done()
shouldRenewTS := replica1Lease.StartStasis.Add(-1, 0)
mtc.manualClock.Set(shouldRenewTS.WallTime + 1)
if _, pErr := client.SendWrappedWith(
context.Background(),
mtc.senders[1],
roachpb.Header{Replica: replica0Desc},
&gArgs,
); pErr != nil {
panic(pErr)
}
}()
<-extensionSem
waitForTransferBlocked.Store(true)
// Initiate a transfer.
wg.Add(1)
go func() {
defer wg.Done()
// Transfer back from replica1 to replica0.
if err := replica1.AdminTransferLease(replica0Desc.StoreID); err != nil {
panic(err)
}
}()
// Wait for the transfer to be blocked by the extension.
<-transferBlocked
// Now unblock the extension.
extensionSem <- struct{}{}
// Check that the transfer to replica1 eventually happens.
util.SucceedsSoon(t, func() error {
if _, pErr := client.SendWrappedWith(
context.Background(),
mtc.senders[0],
roachpb.Header{Replica: replica0Desc},
&gArgs,
); pErr != nil {
return pErr.GoError()
}
return nil
})
filterMu.Lock()
filter = nil
filterMu.Unlock()
wg.Wait()
}
// TestStoreRangeMergeWithData attempts to merge two collocate ranges
// each containing data.
func TestStoreRangeMergeWithData(t *testing.T) {
defer leaktest.AfterTest(t)()
storeCfg := storage.TestStoreConfig(nil)
storeCfg.TestingKnobs.DisableSplitQueue = true
store, stopper := createTestStoreWithConfig(t, storeCfg)
defer stopper.Stop()
content := roachpb.Key("testing!")
aDesc, bDesc, err := createSplitRanges(store)
if err != nil {
t.Fatal(err)
}
// Write some values left and right of the proposed split key.
pArgs := putArgs([]byte("aaa"), content)
if _, err := client.SendWrapped(context.Background(), rg1(store), &pArgs); err != nil {
t.Fatal(err)
}
pArgs = putArgs([]byte("ccc"), content)
if _, err := client.SendWrappedWith(context.Background(), rg1(store), roachpb.Header{
RangeID: bDesc.RangeID,
}, &pArgs); err != nil {
t.Fatal(err)
}
// Confirm the values are there.
gArgs := getArgs([]byte("aaa"))
if reply, err := client.SendWrapped(context.Background(), rg1(store), &gArgs); err != nil {
t.Fatal(err)
} else if replyBytes, err := reply.(*roachpb.GetResponse).Value.GetBytes(); err != nil {
t.Fatal(err)
} else if !bytes.Equal(replyBytes, content) {
t.Fatalf("actual value %q did not match expected value %q", replyBytes, content)
}
gArgs = getArgs([]byte("ccc"))
if reply, err := client.SendWrappedWith(context.Background(), rg1(store), roachpb.Header{
RangeID: bDesc.RangeID,
}, &gArgs); err != nil {
t.Fatal(err)
} else if replyBytes, err := reply.(*roachpb.GetResponse).Value.GetBytes(); err != nil {
t.Fatal(err)
} else if !bytes.Equal(replyBytes, content) {
t.Fatalf("actual value %q did not match expected value %q", replyBytes, content)
}
// Merge the b range back into the a range.
args := adminMergeArgs(roachpb.KeyMin)
if _, err := client.SendWrapped(context.Background(), rg1(store), &args); err != nil {
t.Fatal(err)
}
// Verify no intents remains on range descriptor keys.
for _, key := range []roachpb.Key{keys.RangeDescriptorKey(aDesc.StartKey), keys.RangeDescriptorKey(bDesc.StartKey)} {
if _, _, err := engine.MVCCGet(context.Background(), store.Engine(), key, store.Clock().Now(), true, nil); err != nil {
t.Fatal(err)
}
}
// Verify the merge by looking up keys from both ranges.
rangeA := store.LookupReplica([]byte("a"), nil)
rangeB := store.LookupReplica([]byte("c"), nil)
rangeADesc := rangeA.Desc()
rangeBDesc := rangeB.Desc()
if !reflect.DeepEqual(rangeA, rangeB) {
t.Fatalf("ranges were not merged %+v=%+v", rangeADesc, rangeBDesc)
}
if !bytes.Equal(rangeADesc.StartKey, roachpb.RKeyMin) {
t.Fatalf("The start key is not equal to KeyMin %q=%q", rangeADesc.StartKey, roachpb.RKeyMin)
}
if !bytes.Equal(rangeADesc.EndKey, roachpb.RKeyMax) {
t.Fatalf("The end key is not equal to KeyMax %q=%q", rangeADesc.EndKey, roachpb.RKeyMax)
}
// Try to get values from after the merge.
gArgs = getArgs([]byte("aaa"))
if reply, err := client.SendWrapped(context.Background(), rg1(store), &gArgs); err != nil {
t.Fatal(err)
} else if replyBytes, err := reply.(*roachpb.GetResponse).Value.GetBytes(); err != nil {
t.Fatal(err)
} else if !bytes.Equal(replyBytes, content) {
t.Fatalf("actual value %q did not match expected value %q", replyBytes, content)
}
gArgs = getArgs([]byte("ccc"))
if reply, err := client.SendWrappedWith(context.Background(), rg1(store), roachpb.Header{
RangeID: rangeB.RangeID,
}, &gArgs); err != nil {
t.Fatal(err)
} else if replyBytes, err := reply.(*roachpb.GetResponse).Value.GetBytes(); err != nil {
t.Fatal(err)
} else if !bytes.Equal(replyBytes, content) {
t.Fatalf("actual value %q did not match expected value %q", replyBytes, content)
}
// Put new values after the merge on both sides.
pArgs = putArgs([]byte("aaaa"), content)
if _, err := client.SendWrapped(context.Background(), rg1(store), &pArgs); err != nil {
t.Fatal(err)
}
pArgs = putArgs([]byte("cccc"), content)
if _, err := client.SendWrappedWith(context.Background(), rg1(store), roachpb.Header{
RangeID: rangeB.RangeID,
}, &pArgs); err != nil {
t.Fatal(err)
}
// Try to get the newly placed values.
gArgs = getArgs([]byte("aaaa"))
if reply, err := client.SendWrapped(context.Background(), rg1(store), &gArgs); err != nil {
t.Fatal(err)
} else if replyBytes, err := reply.(*roachpb.GetResponse).Value.GetBytes(); err != nil {
t.Fatal(err)
} else if !bytes.Equal(replyBytes, content) {
t.Fatalf("actual value %q did not match expected value %q", replyBytes, content)
}
gArgs = getArgs([]byte("cccc"))
if reply, err := client.SendWrapped(context.Background(), rg1(store), &gArgs); err != nil {
t.Fatal(err)
} else if replyBytes, err := reply.(*roachpb.GetResponse).Value.GetBytes(); err != nil {
t.Fatal(err)
} else if !bytes.Equal(replyBytes, content) {
t.Fatalf("actual value %q did not match expected value %q", replyBytes, content)
}
}
// TestStoreRangeLease verifies that ranges after range 0 get
// epoch-based range leases if enabled and expiration-based
// otherwise.
func TestStoreRangeLease(t *testing.T) {
defer leaktest.AfterTest(t)()
for _, enableEpoch := range []bool{true, false} {
t.Run(fmt.Sprintf("epoch-based leases? %t", enableEpoch), func(t *testing.T) {
sc := storage.TestStoreConfig(nil)
sc.EnableEpochRangeLeases = enableEpoch
mtc := &multiTestContext{storeConfig: &sc}
defer mtc.Stop()
mtc.Start(t, 1)
splitKeys := []roachpb.Key{roachpb.Key("a"), roachpb.Key("b"), roachpb.Key("c")}
for _, splitKey := range splitKeys {
splitArgs := adminSplitArgs(splitKey, splitKey)
if _, pErr := client.SendWrapped(context.Background(), mtc.distSenders[0], splitArgs); pErr != nil {
t.Fatal(pErr)
}
}
rLeft := mtc.stores[0].LookupReplica(roachpb.RKeyMin, nil)
lease, _ := rLeft.GetLease()
if lt := lease.Type(); lt != roachpb.LeaseExpiration {
t.Fatalf("expected lease type expiration; got %d", lt)
}
// After the split, expect an expiration lease for other ranges.
for _, key := range splitKeys {
repl := mtc.stores[0].LookupReplica(roachpb.RKey(key), nil)
lease, _ = repl.GetLease()
if lt := lease.Type(); lt != roachpb.LeaseExpiration {
t.Fatalf("%s: expected lease type epoch; got %d", key, lt)
}
}
// Allow leases to expire and send commands to ensure we
// re-acquire, then check types again.
mtc.expireLeases()
for _, key := range splitKeys {
if _, err := mtc.dbs[0].Inc(context.TODO(), key, 1); err != nil {
t.Fatalf("%s failed to increment: %s", key, err)
}
}
// After the expiration, expect an epoch lease for the RHS if
// we've enabled epoch based range leases.
for _, key := range splitKeys {
repl := mtc.stores[0].LookupReplica(roachpb.RKey(key), nil)
lease, _ = repl.GetLease()
if enableEpoch {
if lt := lease.Type(); lt != roachpb.LeaseEpoch {
t.Fatalf("expected lease type epoch; got %d", lt)
}
} else {
if lt := lease.Type(); lt != roachpb.LeaseExpiration {
t.Fatalf("expected lease type expiration; got %d", lt)
}
}
}
})
}
}
// TestMultiRangeScanDeleteRange tests that commands which access multiple
// ranges are carried out properly.
func TestMultiRangeScanDeleteRange(t *testing.T) {
defer leaktest.AfterTest(t)()
s, _, _ := serverutils.StartServer(t, base.TestServerArgs{})
defer s.Stopper().Stop()
ts := s.(*TestServer)
retryOpts := base.DefaultRetryOptions()
retryOpts.Closer = ts.stopper.ShouldQuiesce()
ds := kv.NewDistSender(kv.DistSenderConfig{
Clock: s.Clock(),
RPCContext: s.RPCContext(),
RPCRetryOptions: &retryOpts,
}, ts.Gossip())
ambient := log.AmbientContext{Tracer: tracing.NewTracer()}
tds := kv.NewTxnCoordSender(
ambient,
ds,
s.Clock(),
ts.Cfg.Linearizable,
ts.stopper,
kv.MakeTxnMetrics(metric.TestSampleInterval),
)
if err := ts.node.storeCfg.DB.AdminSplit(context.TODO(), "m"); err != nil {
t.Fatal(err)
}
writes := []roachpb.Key{roachpb.Key("a"), roachpb.Key("z")}
get := &roachpb.GetRequest{
Span: roachpb.Span{Key: writes[0]},
}
get.EndKey = writes[len(writes)-1]
if _, err := client.SendWrapped(context.Background(), tds, get); err == nil {
t.Errorf("able to call Get with a key range: %v", get)
}
var delTS hlc.Timestamp
for i, k := range writes {
put := roachpb.NewPut(k, roachpb.MakeValueFromBytes(k))
if _, err := client.SendWrapped(context.Background(), tds, put); err != nil {
t.Fatal(err)
}
scan := roachpb.NewScan(writes[0], writes[len(writes)-1].Next())
reply, err := client.SendWrapped(context.Background(), tds, scan)
if err != nil {
t.Fatal(err)
}
sr := reply.(*roachpb.ScanResponse)
if sr.Txn != nil {
// This was the other way around at some point in the past.
// Same below for Delete, etc.
t.Errorf("expected no transaction in response header")
}
if rows := sr.Rows; len(rows) != i+1 {
t.Fatalf("expected %d rows, but got %d", i+1, len(rows))
}
}
del := &roachpb.DeleteRangeRequest{
Span: roachpb.Span{
Key: writes[0],
EndKey: roachpb.Key(writes[len(writes)-1]).Next(),
},
ReturnKeys: true,
}
reply, err := client.SendWrappedWith(context.Background(), tds, roachpb.Header{Timestamp: delTS}, del)
if err != nil {
t.Fatal(err)
}
dr := reply.(*roachpb.DeleteRangeResponse)
if dr.Txn != nil {
t.Errorf("expected no transaction in response header")
}
if !reflect.DeepEqual(dr.Keys, writes) {
t.Errorf("expected %d keys to be deleted, but got %d instead", writes, dr.Keys)
}
scan := roachpb.NewScan(writes[0], writes[len(writes)-1].Next())
txn := &roachpb.Transaction{Name: "MyTxn"}
reply, err = client.SendWrappedWith(context.Background(), tds, roachpb.Header{Txn: txn}, scan)
if err != nil {
t.Fatal(err)
}
sr := reply.(*roachpb.ScanResponse)
if txn := sr.Txn; txn == nil || txn.Name != "MyTxn" {
t.Errorf("wanted Txn to persist, but it changed to %v", txn)
}
if rows := sr.Rows; len(rows) > 0 {
t.Fatalf("scan after delete returned rows: %v", rows)
}
}
// TestRaftLogQueue verifies that the raft log queue correctly truncates the
// raft log.
func TestRaftLogQueue(t *testing.T) {
defer leaktest.AfterTest(t)()
mtc := &multiTestContext{}
// Set maxBytes to something small so we can trigger the raft log truncation
// without adding 64MB of logs.
const maxBytes = 1 << 16
defer config.TestingSetDefaultZoneConfig(config.ZoneConfig{
RangeMaxBytes: maxBytes,
})()
// Turn off raft elections so the raft leader won't change out from under
// us in this test.
sc := storage.TestStoreConfig(nil)
sc.RaftTickInterval = math.MaxInt32
sc.RaftElectionTimeoutTicks = 1000000
mtc.storeConfig = &sc
defer mtc.Stop()
mtc.Start(t, 3)
// Write a single value to ensure we have a leader.
pArgs := putArgs([]byte("key"), []byte("value"))
if _, err := client.SendWrapped(context.Background(), rg1(mtc.stores[0]), pArgs); err != nil {
t.Fatal(err)
}
// Get the raft leader (and ensure one exists).
rangeID := mtc.stores[0].LookupReplica([]byte("a"), nil).RangeID
raftLeaderRepl := mtc.getRaftLeader(rangeID)
if raftLeaderRepl == nil {
t.Fatalf("could not find raft leader replica for range %d", rangeID)
}
originalIndex, err := raftLeaderRepl.GetFirstIndex()
if err != nil {
t.Fatal(err)
}
// Disable splits since we're increasing the raft log with puts.
for _, store := range mtc.stores {
store.SetSplitQueueActive(false)
}
// Write a collection of values to increase the raft log.
value := bytes.Repeat([]byte("a"), 1000) // 1KB
for size := int64(0); size < 2*maxBytes; size += int64(len(value)) {
pArgs = putArgs([]byte(fmt.Sprintf("key-%d", size)), value)
if _, err := client.SendWrapped(context.Background(), rg1(mtc.stores[0]), pArgs); err != nil {
t.Fatal(err)
}
}
// Sadly, occasionally the queue has a race with the force processing so
// this succeeds within will captures those rare cases.
var afterTruncationIndex uint64
testutils.SucceedsSoon(t, func() error {
// Force a truncation check.
for _, store := range mtc.stores {
store.ForceRaftLogScanAndProcess()
}
// Ensure that firstIndex has increased indicating that the log
// truncation has occurred.
var err error
afterTruncationIndex, err = raftLeaderRepl.GetFirstIndex()
if err != nil {
t.Fatal(err)
}
if afterTruncationIndex <= originalIndex {
return errors.Errorf("raft log has not been truncated yet, afterTruncationIndex:%d originalIndex:%d",
afterTruncationIndex, originalIndex)
}
return nil
})
// Force a truncation check again to ensure that attempting to truncate an
// already truncated log has no effect. This check, unlike in the last
// iteration, cannot use a succeedsSoon. This check is fragile in that the
// truncation triggered here may lose the race against the call to
// GetFirstIndex, giving a false negative. Fixing this requires additional
// instrumentation of the queues, which was deemed to require too much work
// at the time of this writing.
for _, store := range mtc.stores {
store.ForceRaftLogScanAndProcess()
}
after2ndTruncationIndex, err := raftLeaderRepl.GetFirstIndex()
if err != nil {
t.Fatal(err)
}
if afterTruncationIndex > after2ndTruncationIndex {
t.Fatalf("second truncation destroyed state: afterTruncationIndex:%d after2ndTruncationIndex:%d",
afterTruncationIndex, after2ndTruncationIndex)
}
}
// TestStoreRangeLeaseSwitcheroo verifies that ranges can be switched
// between expiration and epoch and back.
func TestStoreRangeLeaseSwitcheroo(t *testing.T) {
defer leaktest.AfterTest(t)()
sc := storage.TestStoreConfig(nil)
sc.EnableEpochRangeLeases = true
mtc := &multiTestContext{storeConfig: &sc}
defer mtc.Stop()
mtc.Start(t, 1)
splitKey := roachpb.Key("a")
splitArgs := adminSplitArgs(splitKey, splitKey)
if _, pErr := client.SendWrapped(context.Background(), mtc.distSenders[0], splitArgs); pErr != nil {
t.Fatal(pErr)
}
// Allow leases to expire and send commands to ensure we
// re-acquire, then check types again.
mtc.expireLeases()
if _, err := mtc.dbs[0].Inc(context.TODO(), splitKey, 1); err != nil {
t.Fatalf("failed to increment: %s", err)
}
// We started with epoch ranges enabled, so verify we have an epoch lease.
repl := mtc.stores[0].LookupReplica(roachpb.RKey(splitKey), nil)
lease, _ := repl.GetLease()
if lt := lease.Type(); lt != roachpb.LeaseEpoch {
t.Fatalf("expected lease type epoch; got %d", lt)
}
// Stop the store and reverse the epoch range lease setting.
mtc.stopStore(0)
sc.EnableEpochRangeLeases = false
mtc.restartStore(0)
mtc.expireLeases()
if _, err := mtc.dbs[0].Inc(context.TODO(), splitKey, 1); err != nil {
t.Fatalf("failed to increment: %s", err)
}
// Verify we end up with an expiration lease on restart.
repl = mtc.stores[0].LookupReplica(roachpb.RKey(splitKey), nil)
lease, _ = repl.GetLease()
if lt := lease.Type(); lt != roachpb.LeaseExpiration {
t.Fatalf("expected lease type expiration; got %d", lt)
}
// Now, one more time, switch back to epoch-based.
mtc.stopStore(0)
sc.EnableEpochRangeLeases = true
mtc.restartStore(0)
mtc.expireLeases()
if _, err := mtc.dbs[0].Inc(context.TODO(), splitKey, 1); err != nil {
t.Fatalf("failed to increment: %s", err)
}
// Verify we end up with an epoch lease on restart.
repl = mtc.stores[0].LookupReplica(roachpb.RKey(splitKey), nil)
lease, _ = repl.GetLease()
if lt := lease.Type(); lt != roachpb.LeaseEpoch {
t.Fatalf("expected lease type epoch; got %d", lt)
}
}
// 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)()
storeCfg := storage.TestStoreConfig(nil)
storeCfg.TestingKnobs.DisableSplitQueue = true
store, stopper := createTestStoreWithConfig(t, storeCfg)
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(context.Background(), rg1(store), &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(context.Background(), rg1(store), &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(context.Background(), rg1(store), 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 rsCount, preRSCount := len(rlReply.Ranges), len(rlReply.PrefetchedRanges); int32(rsCount+preRSCount) != test.request.MaxRanges {
t.Fatalf("%d: returned results count, expected %d, but got %d+%d", testIdx, test.request.MaxRanges, rsCount, preRSCount)
}
// 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])
}
}
}
}
}
// TestMultiRangeScanWithMaxResults tests that commands which access multiple
// ranges with MaxResults parameter are carried out properly.
func TestMultiRangeScanWithMaxResults(t *testing.T) {
defer leaktest.AfterTest(t)()
testCases := []struct {
splitKeys []roachpb.Key
keys []roachpb.Key
}{
{[]roachpb.Key{roachpb.Key("m")},
[]roachpb.Key{roachpb.Key("a"), roachpb.Key("z")}},
{[]roachpb.Key{roachpb.Key("h"), roachpb.Key("q")},
[]roachpb.Key{roachpb.Key("b"), roachpb.Key("f"), roachpb.Key("k"),
roachpb.Key("r"), roachpb.Key("w"), roachpb.Key("y")}},
}
for i, tc := range testCases {
s, _, _ := serverutils.StartServer(t, base.TestServerArgs{})
defer s.Stopper().Stop()
ts := s.(*TestServer)
retryOpts := base.DefaultRetryOptions()
retryOpts.Closer = ts.stopper.ShouldQuiesce()
ds := kv.NewDistSender(kv.DistSenderConfig{
Clock: s.Clock(),
RPCContext: s.RPCContext(),
RPCRetryOptions: &retryOpts,
}, ts.Gossip())
ambient := log.AmbientContext{Tracer: tracing.NewTracer()}
tds := kv.NewTxnCoordSender(
ambient,
ds,
ts.Clock(),
ts.Cfg.Linearizable,
ts.stopper,
kv.MakeTxnMetrics(metric.TestSampleInterval),
)
for _, sk := range tc.splitKeys {
if err := ts.node.storeCfg.DB.AdminSplit(context.TODO(), sk); err != nil {
t.Fatal(err)
}
}
for _, k := range tc.keys {
put := roachpb.NewPut(k, roachpb.MakeValueFromBytes(k))
if _, err := client.SendWrapped(context.Background(), tds, put); err != nil {
t.Fatal(err)
}
}
// Try every possible ScanRequest startKey.
for start := 0; start < len(tc.keys); start++ {
// Try every possible maxResults, from 1 to beyond the size of key array.
for maxResults := 1; maxResults <= len(tc.keys)-start+1; maxResults++ {
scan := roachpb.NewScan(tc.keys[start], tc.keys[len(tc.keys)-1].Next())
reply, err := client.SendWrappedWith(
context.Background(), tds, roachpb.Header{MaxSpanRequestKeys: int64(maxResults)}, scan,
)
if err != nil {
t.Fatal(err)
}
rows := reply.(*roachpb.ScanResponse).Rows
if start+maxResults <= len(tc.keys) && len(rows) != maxResults {
t.Errorf("%d: start=%s: expected %d rows, but got %d", i, tc.keys[start], maxResults, len(rows))
} else if start+maxResults == len(tc.keys)+1 && len(rows) != maxResults-1 {
t.Errorf("%d: expected %d rows, but got %d", i, maxResults-1, len(rows))
}
}
}
}
}
// Test that a lease extension (a RequestLeaseRequest that doesn't change the
// lease holder) is not blocked by ongoing reads.
// The test relies on two things:
// 1) Lease extensions, unlike lease transfers, are not blocked by reads through their
// PostCommitTrigger.noConcurrentReads.
// 2) Requests with the non-KV flag, such as RequestLeaseRequest, do not
// go through the command queue.
func TestLeaseExtensionNotBlockedByRead(t *testing.T) {
defer leaktest.AfterTest(t)()
readBlocked := make(chan struct{})
cmdFilter := func(fArgs storagebase.FilterArgs) *roachpb.Error {
if fArgs.Hdr.UserPriority == 42 {
// Signal that the read is blocked.
readBlocked <- struct{}{}
// Wait for read to be unblocked.
<-readBlocked
}
return nil
}
srv, _, _ := serverutils.StartServer(t,
base.TestServerArgs{
Knobs: base.TestingKnobs{
Store: &storage.StoreTestingKnobs{
TestingCommandFilter: cmdFilter,
},
},
})
s := srv.(*server.TestServer)
defer s.Stopper().Stop()
// Start a read and wait for it to block.
key := roachpb.Key("a")
errChan := make(chan error)
go func() {
getReq := roachpb.GetRequest{
Span: roachpb.Span{
Key: key,
},
}
if _, pErr := client.SendWrappedWith(context.Background(), s.DistSender(),
roachpb.Header{UserPriority: 42},
&getReq); pErr != nil {
errChan <- pErr.GoError()
}
}()
select {
case err := <-errChan:
t.Fatal(err)
case <-readBlocked:
// Send the lease request.
rKey, err := keys.Addr(key)
if err != nil {
t.Fatal(err)
}
_, repDesc, err := s.Stores().LookupReplica(rKey, nil)
if err != nil {
t.Fatal(err)
}
leaseReq := roachpb.RequestLeaseRequest{
Span: roachpb.Span{
Key: key,
},
Lease: roachpb.Lease{
Start: s.Clock().Now(),
StartStasis: s.Clock().Now().Add(time.Second.Nanoseconds(), 0),
Expiration: s.Clock().Now().Add(2*time.Second.Nanoseconds(), 0),
Replica: repDesc,
},
}
if _, pErr := client.SendWrapped(context.Background(), s.DistSender(), &leaseReq); pErr != nil {
t.Fatal(pErr)
}
// Unblock the read.
readBlocked <- struct{}{}
}
}