func setAppliedIndex(
ctx context.Context,
eng engine.ReadWriter,
ms *enginepb.MVCCStats,
rangeID roachpb.RangeID,
appliedIndex,
leaseAppliedIndex uint64,
) error {
var value roachpb.Value
value.SetInt(int64(appliedIndex))
if err := engine.MVCCPut(ctx, eng, ms,
keys.RaftAppliedIndexKey(rangeID),
hlc.ZeroTimestamp,
value,
nil /* txn */); err != nil {
return err
}
value.SetInt(int64(leaseAppliedIndex))
return engine.MVCCPut(ctx, eng, ms,
keys.LeaseAppliedIndexKey(rangeID),
hlc.ZeroTimestamp,
value,
nil /* txn */)
}
func setLastIndex(
ctx context.Context, eng engine.ReadWriter, rangeID roachpb.RangeID, lastIndex uint64,
) error {
var value roachpb.Value
value.SetInt(int64(lastIndex))
return engine.MVCCPut(ctx, eng, nil, keys.RaftLastIndexKey(rangeID),
hlc.ZeroTimestamp,
value,
nil /* txn */)
}
func setFrozenStatus(
ctx context.Context,
eng engine.ReadWriter,
ms *enginepb.MVCCStats,
rangeID roachpb.RangeID,
frozen storagebase.ReplicaState_FrozenEnum,
) error {
if frozen == storagebase.ReplicaState_FROZEN_UNSPECIFIED {
return errors.New("cannot persist unspecified FrozenStatus")
}
var val roachpb.Value
val.SetBool(frozen == storagebase.ReplicaState_FROZEN)
return engine.MVCCPut(ctx, eng, ms,
keys.RangeFrozenStatusKey(rangeID), hlc.ZeroTimestamp, val, nil)
}
// append the given entries to the raft log. Takes the previous values of
// r.mu.lastIndex and r.mu.raftLogSize, and returns new values. We do this
// rather than modifying them directly because these modifications need to be
// atomic with the commit of the batch.
func (r *Replica) append(
ctx context.Context,
batch engine.ReadWriter,
prevLastIndex uint64,
prevRaftLogSize int64,
entries []raftpb.Entry,
) (uint64, int64, error) {
if len(entries) == 0 {
return prevLastIndex, prevRaftLogSize, nil
}
var diff enginepb.MVCCStats
var value roachpb.Value
for i := range entries {
ent := &entries[i]
key := keys.RaftLogKey(r.RangeID, ent.Index)
if err := value.SetProto(ent); err != nil {
return 0, 0, err
}
value.InitChecksum(key)
var err error
if ent.Index > prevLastIndex {
err = engine.MVCCBlindPut(ctx, batch, &diff, key, hlc.ZeroTimestamp, value, nil /* txn */)
} else {
err = engine.MVCCPut(ctx, batch, &diff, key, hlc.ZeroTimestamp, value, nil /* txn */)
}
if err != nil {
return 0, 0, err
}
}
// Delete any previously appended log entries which never committed.
lastIndex := entries[len(entries)-1].Index
for i := lastIndex + 1; i <= prevLastIndex; i++ {
err := engine.MVCCDelete(ctx, batch, &diff, keys.RaftLogKey(r.RangeID, i),
hlc.ZeroTimestamp, nil /* txn */)
if err != nil {
return 0, 0, err
}
}
if err := setLastIndex(ctx, batch, r.RangeID, lastIndex); err != nil {
return 0, 0, err
}
raftLogSize := prevRaftLogSize + diff.SysBytes
return lastIndex, raftLogSize, nil
}
// createRangeData creates sample range data in all possible areas of
// the key space. Returns a slice of the encoded keys of all created
// data.
func createRangeData(t *testing.T, r *Replica) []engine.MVCCKey {
ts0 := hlc.ZeroTimestamp
ts := hlc.Timestamp{WallTime: 1}
desc := r.Desc()
keyTSs := []struct {
key roachpb.Key
ts hlc.Timestamp
}{
{keys.AbortCacheKey(r.RangeID, testTxnID), ts0},
{keys.AbortCacheKey(r.RangeID, testTxnID2), ts0},
{keys.RangeFrozenStatusKey(r.RangeID), ts0},
{keys.RangeLastGCKey(r.RangeID), ts0},
{keys.RaftAppliedIndexKey(r.RangeID), ts0},
{keys.RaftTruncatedStateKey(r.RangeID), ts0},
{keys.RangeLeaseKey(r.RangeID), ts0},
{keys.LeaseAppliedIndexKey(r.RangeID), ts0},
{keys.RangeStatsKey(r.RangeID), ts0},
{keys.RangeTxnSpanGCThresholdKey(r.RangeID), ts0},
{keys.RaftHardStateKey(r.RangeID), ts0},
{keys.RaftLastIndexKey(r.RangeID), ts0},
{keys.RaftLogKey(r.RangeID, 1), ts0},
{keys.RaftLogKey(r.RangeID, 2), ts0},
{keys.RangeLastReplicaGCTimestampKey(r.RangeID), ts0},
{keys.RangeLastVerificationTimestampKeyDeprecated(r.RangeID), ts0},
{keys.RangeDescriptorKey(desc.StartKey), ts},
{keys.TransactionKey(roachpb.Key(desc.StartKey), uuid.MakeV4()), ts0},
{keys.TransactionKey(roachpb.Key(desc.StartKey.Next()), uuid.MakeV4()), ts0},
{keys.TransactionKey(fakePrevKey(desc.EndKey), uuid.MakeV4()), ts0},
// TODO(bdarnell): KeyMin.Next() results in a key in the reserved system-local space.
// Once we have resolved https://github.com/cockroachdb/cockroach/issues/437,
// replace this with something that reliably generates the first valid key in the range.
//{r.Desc().StartKey.Next(), ts},
// The following line is similar to StartKey.Next() but adds more to the key to
// avoid falling into the system-local space.
{append(append([]byte{}, desc.StartKey...), '\x02'), ts},
{fakePrevKey(r.Desc().EndKey), ts},
}
keys := []engine.MVCCKey{}
for _, keyTS := range keyTSs {
if err := engine.MVCCPut(context.Background(), r.store.Engine(), nil, keyTS.key, keyTS.ts, roachpb.MakeValueFromString("value"), nil); err != nil {
t.Fatal(err)
}
keys = append(keys, engine.MVCCKey{Key: keyTS.key, Timestamp: keyTS.ts})
}
return keys
}
// TestUncertaintyObservedTimestampForwarding checks that when receiving an
// uncertainty restart on a node, the next attempt to read (at the increased
// timestamp) is free from uncertainty. See roachpb.Transaction for details.
func TestUncertaintyMaxTimestampForwarding(t *testing.T) {
defer leaktest.AfterTest(t)()
dbCtx := client.DefaultDBContext()
s := &localtestcluster.LocalTestCluster{
// Large offset so that any value in the future is an uncertain read. Also
// makes sure that the values we write in the future below don't actually
// wind up in the past.
Clock: hlc.NewClock(hlc.UnixNano, 50*time.Second),
DBContext: &dbCtx,
}
s.Start(t, testutils.NewNodeTestBaseContext(), InitSenderForLocalTestCluster)
defer s.Stop()
disableOwnNodeCertain(t, s)
offsetNS := int64(100)
keySlow := roachpb.Key("slow")
keyFast := roachpb.Key("fast")
valSlow := []byte("wols")
valFast := []byte("tsaf")
// Write keySlow at now+offset, keyFast at now+2*offset
futureTS := s.Clock.Now()
futureTS.WallTime += offsetNS
val := roachpb.MakeValueFromBytes(valSlow)
if err := engine.MVCCPut(context.Background(), s.Eng, nil, keySlow, futureTS, val, nil); err != nil {
t.Fatal(err)
}
futureTS.WallTime += offsetNS
val.SetBytes(valFast)
if err := engine.MVCCPut(context.Background(), s.Eng, nil, keyFast, futureTS, val, nil); err != nil {
t.Fatal(err)
}
i := 0
if tErr := s.DB.Txn(context.TODO(), func(txn *client.Txn) error {
i++
// The first command serves to start a Txn, fixing the timestamps.
// There will be a restart, but this is idempotent.
if _, err := txn.Scan("t", roachpb.Key("t").Next(), 0); err != nil {
t.Fatal(err)
}
// This is a bit of a hack for the sake of this test: By visiting the
// node above, we've made a note of its clock, which allows us to
// prevent the restart. But we want to catch the restart, so reset the
// observed timestamps.
txn.Proto.ResetObservedTimestamps()
// The server's clock suddenly jumps ahead of keyFast's timestamp.
s.Manual.Increment(2*offsetNS + 1)
// Now read slowKey first. It should read at 0, catch an uncertainty error,
// and get keySlow's timestamp in that error, but upgrade it to the larger
// node clock (which is ahead of keyFast as well). If the last part does
// not happen, the read of keyFast should fail (i.e. read nothing).
// There will be exactly one restart here.
if gr, err := txn.Get(keySlow); err != nil {
if i != 1 {
t.Fatalf("unexpected transaction error: %s", err)
}
return err
} else if !gr.Exists() || !bytes.Equal(gr.ValueBytes(), valSlow) {
t.Fatalf("read of %q returned %v, wanted value %q", keySlow, gr.Value, valSlow)
}
// The node should already be certain, so we expect no restart here
// and to read the correct key.
if gr, err := txn.Get(keyFast); err != nil {
t.Fatalf("second Get failed with %s", err)
} else if !gr.Exists() || !bytes.Equal(gr.ValueBytes(), valFast) {
t.Fatalf("read of %q returned %v, wanted value %q", keyFast, gr.Value, valFast)
}
return nil
}); tErr != nil {
t.Fatal(tErr)
}
}