// TestRocksDBCompaction verifies that a garbage collector can be
// installed on a RocksDB engine and will properly compact response
// cache and transaction entries.
func TestRocksDBCompaction(t *testing.T) {
defer leaktest.AfterTest(t)
gob.Register(proto.Timestamp{})
rocksdb := newMemRocksDB(proto.Attributes{Attrs: []string{"ssd"}}, testCacheSize)
err := rocksdb.Open()
if err != nil {
t.Fatalf("could not create new in-memory rocksdb db instance: %v", err)
}
rocksdb.SetGCTimeouts(1, 2)
defer rocksdb.Close()
cmdID := &proto.ClientCmdID{WallTime: 1, Random: 1}
// Write two transaction values and two response cache values such
// that exactly one of each should be GC'd based on our GC timeouts.
kvs := []proto.KeyValue{
{
Key: keys.ResponseCacheKey(1, cmdID),
Value: proto.Value{Bytes: encodePutResponse(makeTS(2, 0), t)},
},
{
Key: keys.ResponseCacheKey(2, cmdID),
Value: proto.Value{Bytes: encodePutResponse(makeTS(3, 0), t)},
},
{
Key: keys.TransactionKey(proto.Key("a"), proto.Key(uuid.NewUUID4())),
Value: proto.Value{Bytes: encodeTransaction(makeTS(1, 0), t)},
},
{
Key: keys.TransactionKey(proto.Key("b"), proto.Key(uuid.NewUUID4())),
Value: proto.Value{Bytes: encodeTransaction(makeTS(2, 0), t)},
},
}
for _, kv := range kvs {
if err := MVCCPut(rocksdb, nil, kv.Key, proto.ZeroTimestamp, kv.Value, nil); err != nil {
t.Fatal(err)
}
}
// Compact range and scan remaining values to compare.
rocksdb.CompactRange(nil, nil)
actualKVs, _, err := MVCCScan(rocksdb, proto.KeyMin, proto.KeyMax,
0, proto.ZeroTimestamp, true, nil)
if err != nil {
t.Fatalf("could not run scan: %v", err)
}
var keys []proto.Key
for _, kv := range actualKVs {
keys = append(keys, kv.Key)
}
expKeys := []proto.Key{
kvs[1].Key,
kvs[3].Key,
}
if !reflect.DeepEqual(expKeys, keys) {
t.Errorf("expected keys %+v, got keys %+v", expKeys, keys)
}
}
// InternalHeartbeatTxn updates the transaction status and heartbeat
// timestamp after receiving transaction heartbeat messages from
// coordinator. Returns the updated transaction.
func (r *Range) InternalHeartbeatTxn(batch engine.Engine, ms *engine.MVCCStats, args proto.InternalHeartbeatTxnRequest) (proto.InternalHeartbeatTxnResponse, error) {
var reply proto.InternalHeartbeatTxnResponse
key := keys.TransactionKey(args.Txn.Key, args.Txn.ID)
var txn proto.Transaction
if ok, err := engine.MVCCGetProto(batch, key, proto.ZeroTimestamp, true, nil, &txn); err != nil {
return reply, err
} else if !ok {
// If no existing transaction record was found, initialize to a
// shallow copy of the transaction in the request header. We copy
// to avoid mutating the original below.
txn = *args.Txn
}
if txn.Status == proto.PENDING {
if txn.LastHeartbeat == nil {
txn.LastHeartbeat = &proto.Timestamp{}
}
if txn.LastHeartbeat.Less(args.Header().Timestamp) {
*txn.LastHeartbeat = args.Header().Timestamp
}
if err := engine.MVCCPutProto(batch, ms, key, proto.ZeroTimestamp, nil, &txn); err != nil {
return reply, err
}
}
reply.Txn = &txn
return reply, nil
}
// InternalHeartbeatTxn updates the transaction status and heartbeat
// timestamp after receiving transaction heartbeat messages from
// coordinator. Returns the updated transaction.
func (r *Range) InternalHeartbeatTxn(batch engine.Engine, ms *engine.MVCCStats,
args *proto.InternalHeartbeatTxnRequest, reply *proto.InternalHeartbeatTxnResponse) {
key := keys.TransactionKey(args.Txn.Key, args.Txn.ID)
var txn proto.Transaction
ok, err := engine.MVCCGetProto(batch, key, proto.ZeroTimestamp, true, nil, &txn)
if err != nil {
reply.SetGoError(err)
return
}
// If no existing transaction record was found, initialize
// to the transaction in the request header.
if !ok {
gogoproto.Merge(&txn, args.Txn)
}
if txn.Status == proto.PENDING {
if txn.LastHeartbeat == nil {
txn.LastHeartbeat = &proto.Timestamp{}
}
if txn.LastHeartbeat.Less(args.Header().Timestamp) {
*txn.LastHeartbeat = args.Header().Timestamp
}
if err := engine.MVCCPutProto(batch, ms, key, proto.ZeroTimestamp, nil, &txn); err != nil {
reply.SetGoError(err)
return
}
}
reply.Txn = &txn
}
// TestRocksDBCompaction verifies that a garbage collector can be
// installed on a RocksDB engine and will properly compact transaction
// entries.
func TestRocksDBCompaction(t *testing.T) {
defer leaktest.AfterTest(t)
stopper := stop.NewStopper()
defer stopper.Stop()
rocksdb := newMemRocksDB(roachpb.Attributes{}, testCacheSize, stopper)
err := rocksdb.Open()
if err != nil {
t.Fatalf("could not create new in-memory rocksdb db instance: %v", err)
}
rocksdb.SetGCTimeouts(1)
// Write two transaction values such that exactly one should be GC'd based
// on our GC timeouts.
kvs := []roachpb.KeyValue{
{
Key: keys.TransactionKey(roachpb.Key("a"), roachpb.Key(uuid.NewUUID4())),
Value: roachpb.MakeValueFromBytes(encodeTransaction(makeTS(1, 0), t)),
},
{
Key: keys.TransactionKey(roachpb.Key("b"), roachpb.Key(uuid.NewUUID4())),
Value: roachpb.MakeValueFromBytes(encodeTransaction(makeTS(2, 0), t)),
},
}
for _, kv := range kvs {
if err := MVCCPut(rocksdb, nil, kv.Key, roachpb.ZeroTimestamp, kv.Value, nil); err != nil {
t.Fatal(err)
}
}
// Compact range and scan remaining values to compare.
rocksdb.CompactRange(nil, nil)
actualKVs, _, err := MVCCScan(rocksdb, keyMin, keyMax, 0, roachpb.ZeroTimestamp, true, nil)
if err != nil {
t.Fatalf("could not run scan: %v", err)
}
var keys []roachpb.Key
for _, kv := range actualKVs {
keys = append(keys, kv.Key)
}
expKeys := []roachpb.Key{
kvs[1].Key,
}
if !reflect.DeepEqual(expKeys, keys) {
t.Errorf("expected keys %+v, got keys %+v", expKeys, keys)
}
}
// 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.LeaseAppliedIndexKey(r.RangeID), ts0},
{keys.RangeStatsKey(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.RangeLastVerificationTimestampKey(r.RangeID), ts0},
{keys.RangeDescriptorKey(desc.StartKey), ts},
{keys.TransactionKey(roachpb.Key(desc.StartKey), uuid.NewV4()), ts0},
{keys.TransactionKey(roachpb.Key(desc.StartKey.Next()), uuid.NewV4()), ts0},
{keys.TransactionKey(fakePrevKey(desc.EndKey), uuid.NewV4()), 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
}
// TestStoreResolveWriteIntent adds write intent and then verifies
// that a put returns success and aborts intent's txn in the event the
// pushee has lower priority. Othwerise, verifies that a
// TransactionPushError is returned.
func TestStoreResolveWriteIntent(t *testing.T) {
defer leaktest.AfterTest(t)
store, _, stopper := createTestStore(t)
defer stopper.Stop()
for i, resolvable := range []bool{true, false} {
key := proto.Key(fmt.Sprintf("key-%d", i))
pusher := newTransaction("test", key, 1, proto.SERIALIZABLE, store.ctx.Clock)
pushee := newTransaction("test", key, 1, proto.SERIALIZABLE, store.ctx.Clock)
if resolvable {
pushee.Priority = 1
pusher.Priority = 2 // Pusher will win.
} else {
pushee.Priority = 2
pusher.Priority = 1 // Pusher will lose.
}
// First lay down intent using the pushee's txn.
pArgs := putArgs(key, []byte("value"), 1, store.StoreID())
pArgs.Timestamp = store.ctx.Clock.Now()
pArgs.Txn = pushee
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &pArgs, Reply: pArgs.CreateReply()}); err != nil {
t.Fatal(err)
}
// Now, try a put using the pusher's txn.
pArgs.Timestamp = store.ctx.Clock.Now()
pArgs.Txn = pusher
err := store.ExecuteCmd(context.Background(), proto.Call{Args: &pArgs, Reply: pArgs.CreateReply()})
if resolvable {
if err != nil {
t.Errorf("expected intent resolved; got unexpected error: %s", err)
}
txnKey := keys.TransactionKey(pushee.Key, pushee.ID)
var txn proto.Transaction
ok, err := engine.MVCCGetProto(store.Engine(), txnKey, proto.ZeroTimestamp, true, nil, &txn)
if !ok || err != nil {
t.Fatalf("not found or err: %s", err)
}
if txn.Status != proto.ABORTED {
t.Errorf("expected pushee to be aborted; got %s", txn.Status)
}
} else {
if rErr, ok := err.(*proto.TransactionPushError); !ok {
t.Errorf("expected txn push error; got %s", err)
} else if !bytes.Equal(rErr.PusheeTxn.ID, pushee.ID) {
t.Errorf("expected txn to match pushee %q; got %s", pushee.ID, rErr)
}
// Trying again should fail again.
if err = store.ExecuteCmd(context.Background(), proto.Call{Args: &pArgs, Reply: pArgs.CreateReply()}); err == nil {
t.Errorf("expected another error on latent write intent but succeeded")
}
}
}
}
// 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(r *Replica, t *testing.T) []roachpb.EncodedKey {
ts0 := roachpb.ZeroTimestamp
ts := roachpb.Timestamp{WallTime: 1}
keyTSs := []struct {
key roachpb.Key
ts roachpb.Timestamp
}{
{keys.ResponseCacheKey(r.Desc().RangeID, &roachpb.ClientCmdID{WallTime: 1, Random: 1}), ts0},
{keys.ResponseCacheKey(r.Desc().RangeID, &roachpb.ClientCmdID{WallTime: 2, Random: 2}), ts0},
{keys.RaftHardStateKey(r.Desc().RangeID), ts0},
{keys.RaftLogKey(r.Desc().RangeID, 1), ts0},
{keys.RaftLogKey(r.Desc().RangeID, 2), ts0},
{keys.RangeGCMetadataKey(r.Desc().RangeID), ts0},
{keys.RangeLastVerificationTimestampKey(r.Desc().RangeID), ts0},
{keys.RangeStatsKey(r.Desc().RangeID), ts0},
{keys.RangeDescriptorKey(r.Desc().StartKey), ts},
{keys.TransactionKey(roachpb.Key(r.Desc().StartKey), []byte("1234")), ts0},
{keys.TransactionKey(roachpb.Key(r.Desc().StartKey.Next()), []byte("5678")), ts0},
{keys.TransactionKey(fakePrevKey(r.Desc().EndKey), []byte("2468")), 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{}, r.Desc().StartKey...), '\x01'), ts},
{fakePrevKey(r.Desc().EndKey), ts},
}
keys := []roachpb.EncodedKey{}
for _, keyTS := range keyTSs {
if err := engine.MVCCPut(r.store.Engine(), nil, keyTS.key, keyTS.ts, roachpb.MakeValueFromString("value"), nil); err != nil {
t.Fatal(err)
}
keys = append(keys, engine.MVCCEncodeKey(keyTS.key))
if !keyTS.ts.Equal(ts0) {
keys = append(keys, engine.MVCCEncodeVersionKey(keyTS.key, keyTS.ts))
}
}
return keys
}
// processTransactionTable scans the transaction table and updates txnMap with
// those transactions which are old and either PENDING or with intents
// registered. In the first case we want to push the transaction so that it is
// aborted, and in the second case we may have to resolve the intents success-
// fully before GCing the entry. The transaction records which can be gc'ed are
// returned separately and are not added to txnMap nor intentSpanMap.
func (gcq *gcQueue) processTransactionTable(r *Replica, txnMap map[uuid.UUID]*roachpb.Transaction, cutoff roachpb.Timestamp) ([]roachpb.GCRequest_GCKey, error) {
snap := r.store.Engine().NewSnapshot()
defer snap.Close()
var numResolveAttempts, numQueuedPushes int
var gcKeys []roachpb.GCRequest_GCKey
defer func() {
gcq.eventLog.Infof(true, "attempted to resolve %d intents of %d gc'able transactions; queued %d txns for push", numResolveAttempts, len(gcKeys), numQueuedPushes)
}()
handleOne := func(kv roachpb.KeyValue) error {
var txn roachpb.Transaction
if err := kv.Value.GetProto(&txn); err != nil {
return err
}
ts := txn.Timestamp
if heartbeatTS := txn.LastHeartbeat; heartbeatTS != nil {
ts.Forward(*heartbeatTS)
}
if !ts.Less(cutoff) {
return nil
}
txnID := *txn.ID
// The transaction record should be considered for removal.
switch txn.Status {
case roachpb.PENDING:
// Marked as running, so we need to push it to abort it but won't
// try to GC it in this cycle (for convenience).
// TODO(tschottdorf): refactor so that we can GC PENDING entries
// in the same cycle, but keeping the calls to pushTxn in a central
// location (keeping it easy to batch them up in the future).
numQueuedPushes++
txnMap[txnID] = &txn
return nil
case roachpb.ABORTED:
// If we remove this transaction, it effectively still counts as
// ABORTED (by design). So this can be GC'ed even if we can't
// resolve the intents.
// Note: Most aborted transaction weren't aborted by their client,
// but instead by the coordinator - those will not have any intents
// persisted, though they still might exist in the system.
numResolveAttempts += len(txn.Intents)
if err := r.store.intentResolver.resolveIntents(r.context(), r,
roachpb.AsIntents(txn.Intents, &txn), true /* wait */, false /* !poison */); err != nil {
log.Warningf("failed to resolve intents of aborted txn on gc: %s", err)
}
case roachpb.COMMITTED:
// It's committed, so it doesn't need a push but we can only
// GC it after its intents are resolved.
numResolveAttempts += len(txn.Intents)
if err := r.store.intentResolver.resolveIntents(r.context(), r,
roachpb.AsIntents(txn.Intents, &txn), true /* wait */, false /* !poison */); err != nil {
log.Warningf("unable to resolve intents of committed txn on gc: %s", err)
// Returning the error here would abort the whole GC run, and
// we don't want that. Instead, we simply don't GC this entry.
return nil
}
default:
panic(fmt.Sprintf("invalid transaction state: %s", txn))
}
gcKeys = append(gcKeys, roachpb.GCRequest_GCKey{Key: kv.Key}) // zero timestamp
return nil
}
startKey := keys.TransactionKey(roachpb.KeyMin, uuid.EmptyUUID)
endKey := keys.TransactionKey(roachpb.KeyMax, uuid.EmptyUUID)
_, err := engine.MVCCIterate(snap, startKey, endKey, roachpb.ZeroTimestamp, true /* consistent */, nil /* txn */, false /* !reverse */, func(kv roachpb.KeyValue) (bool, error) {
return false, handleOne(kv)
})
return gcKeys, err
}
// TestStoreVerifyKeys checks that key length is enforced and
// that end keys must sort >= start.
func TestStoreVerifyKeys(t *testing.T) {
defer leaktest.AfterTest(t)
store, _, stopper := createTestStore(t)
defer stopper.Stop()
tooLongKey := proto.Key(strings.Repeat("x", proto.KeyMaxLength+1))
// Start with a too-long key on a get.
gArgs := getArgs(tooLongKey, 1, store.StoreID())
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &gArgs, Reply: gArgs.CreateReply()}); err == nil {
t.Fatal("expected error for key too long")
}
// Try a start key == KeyMax.
gArgs.Key = proto.KeyMax
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &gArgs, Reply: gArgs.CreateReply()}); err == nil {
t.Fatal("expected error for start key == KeyMax")
}
// Try a get with an end key specified (get requires only a start key and should fail).
gArgs.EndKey = proto.KeyMax
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &gArgs, Reply: gArgs.CreateReply()}); err == nil {
t.Fatal("expected error for end key specified on a non-range-based operation")
}
// Try a scan with too-long EndKey.
sArgs := scanArgs(proto.KeyMin, tooLongKey, 1, store.StoreID())
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &sArgs, Reply: sArgs.CreateReply()}); err == nil {
t.Fatal("expected error for end key too long")
}
// Try a scan with end key < start key.
sArgs.Key = []byte("b")
sArgs.EndKey = []byte("a")
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &sArgs, Reply: sArgs.CreateReply()}); err == nil {
t.Fatal("expected error for end key < start")
}
// Try a scan with start key == end key.
sArgs.Key = []byte("a")
sArgs.EndKey = sArgs.Key
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &sArgs, Reply: sArgs.CreateReply()}); err == nil {
t.Fatal("expected error for start == end key")
}
// Try a put to meta2 key which would otherwise exceed maximum key
// length, but is accepted because of the meta prefix.
meta2KeyMax := keys.MakeKey(keys.Meta2Prefix, proto.KeyMax)
pArgs := putArgs(meta2KeyMax, []byte("value"), 1, store.StoreID())
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &pArgs, Reply: pArgs.CreateReply()}); err != nil {
t.Fatalf("unexpected error on put to meta2 value: %s", err)
}
// Try to put a range descriptor record for a start key which is
// maximum length.
key := append([]byte{}, proto.KeyMax...)
key[len(key)-1] = 0x01
pArgs = putArgs(keys.RangeDescriptorKey(key), []byte("value"), 1, store.StoreID())
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &pArgs, Reply: pArgs.CreateReply()}); err != nil {
t.Fatalf("unexpected error on put to range descriptor for KeyMax value: %s", err)
}
// Try a put to txn record for a meta2 key (note that this doesn't
// actually happen in practice, as txn records are not put directly,
// but are instead manipulated only through txn methods).
pArgs = putArgs(keys.TransactionKey(meta2KeyMax, []byte(uuid.NewUUID4())),
[]byte("value"), 1, store.StoreID())
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &pArgs, Reply: pArgs.CreateReply()}); err != nil {
t.Fatalf("unexpected error on put to txn meta2 value: %s", err)
}
}
// TestStoreResolveWriteIntentNoTxn verifies that reads and writes
// which are not part of a transaction can push intents.
func TestStoreResolveWriteIntentNoTxn(t *testing.T) {
defer leaktest.AfterTest(t)
store, _, stopper := createTestStore(t)
defer stopper.Stop()
key := proto.Key("a")
pushee := newTransaction("test", key, 1, proto.SERIALIZABLE, store.ctx.Clock)
pushee.Priority = 0 // pushee should lose all conflicts
// First, lay down intent from pushee.
args := putArgs(key, []byte("value1"), 1, store.StoreID())
reply := args.CreateReply()
args.Timestamp = pushee.Timestamp
args.Txn = pushee
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &args, Reply: reply}); err != nil {
t.Fatal(err)
}
// Now, try to read outside a transaction.
gArgs := getArgs(key, 1, store.StoreID())
gReply := gArgs.CreateReply().(*proto.GetResponse)
gArgs.Timestamp = store.ctx.Clock.Now()
gArgs.UserPriority = gogoproto.Int32(math.MaxInt32)
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &gArgs, Reply: gReply}); err != nil {
t.Errorf("expected read to succeed: %s", err)
} else if gReply.Value != nil {
t.Errorf("expected value to be nil, got %+v", gReply.Value)
}
// Next, try to write outside of a transaction. We will succeed in pushing txn.
args.Timestamp = store.ctx.Clock.Now()
args.Value.Bytes = []byte("value2")
args.Txn = nil
args.UserPriority = gogoproto.Int32(math.MaxInt32)
if err := store.ExecuteCmd(context.Background(), proto.Call{Args: &args, Reply: reply}); err != nil {
t.Errorf("expected success aborting pushee's txn; got %s", err)
}
// Read pushee's txn.
txnKey := keys.TransactionKey(pushee.Key, pushee.ID)
var txn proto.Transaction
if ok, err := engine.MVCCGetProto(store.Engine(), txnKey, proto.ZeroTimestamp, true, nil, &txn); !ok || err != nil {
t.Fatalf("not found or err: %s", err)
}
if txn.Status != proto.ABORTED {
t.Errorf("expected pushee to be aborted; got %s", txn.Status)
}
// Verify that the pushee's timestamp was moved forward on
// former read, since we have it available in write intent error.
expTS := gArgs.Timestamp
expTS.Logical++
if !txn.Timestamp.Equal(expTS) {
t.Errorf("expected pushee timestamp pushed to %s; got %s", expTS, txn.Timestamp)
}
// Similarly, verify that pushee's priority was moved from 0
// to math.MaxInt32-1 during push.
if txn.Priority != math.MaxInt32-1 {
t.Errorf("expected pushee priority to be pushed to %d; got %d", math.MaxInt32-1, txn.Priority)
}
// Finally, try to end the pushee's transaction; it should have
// been aborted.
etArgs := endTxnArgs(pushee, true, 1, store.StoreID())
etArgs.Timestamp = pushee.Timestamp
err := store.ExecuteCmd(context.Background(), proto.Call{Args: &etArgs, Reply: etArgs.CreateReply()})
if err == nil {
t.Errorf("unexpected success committing transaction")
}
if _, ok := err.(*proto.TransactionAbortedError); !ok {
t.Errorf("expected transaction aborted error; got %s", err)
}
}
func TestGCQueueTransactionTable(t *testing.T) {
defer leaktest.AfterTest(t)
const now time.Duration = 3 * 24 * time.Hour
const tTxnThreshold = now - txnCleanupThreshold
type spec struct {
status roachpb.TransactionStatus
ts time.Duration
heartbeatTS time.Duration
newStatus roachpb.TransactionStatus // -1 for GCed
failResolve bool // do we want to fail resolves in this trial?
expResolve bool // expect attempt at removing txn-persisted intents?
expSeqGC bool // expect sequence cache entries removed?
}
// Describes the state of the Txn table before the test.
testCases := map[string]spec{
// Too young, should not touch.
"a": {roachpb.PENDING, tTxnThreshold + 1, 0, roachpb.PENDING, false, false, false},
// Old and pending, but still heartbeat (so no Push attempted; it would succeed).
// No GC.
"b": {roachpb.PENDING, 0, tTxnThreshold + 1, roachpb.PENDING, false, false, false},
// Old, pending and abandoned. Should push and abort it successfully,
// but not GC it just yet (this is an artifact of the implementation).
// The sequence cache gets cleaned up though.
"c": {roachpb.PENDING, tTxnThreshold - 1, 0, roachpb.ABORTED, false, false, true},
// Old and aborted, should delete.
"d": {roachpb.ABORTED, tTxnThreshold - 1, 0, -1, false, true, true},
// Committed and fresh, so no action.
"e": {roachpb.COMMITTED, tTxnThreshold + 1, 0, roachpb.COMMITTED, false, false, false},
// Committed and old. It has an intent (like all tests here), which is
// resolvable and hence we can GC.
"f": {roachpb.COMMITTED, tTxnThreshold - 1, 0, -1, false, true, true},
// Same as the previous one, but we've rigged things so that the intent
// resolution here will fail and consequently no GC is expected.
"g": {roachpb.COMMITTED, tTxnThreshold - 1, 0, roachpb.COMMITTED, true, true, true},
}
resolved := map[string][]roachpb.Span{}
TestingCommandFilter = func(_ roachpb.StoreID, req roachpb.Request, _ roachpb.Header) error {
if resArgs, ok := req.(*roachpb.ResolveIntentRequest); ok {
id := string(resArgs.IntentTxn.Key)
resolved[id] = append(resolved[id], roachpb.Span{
Key: resArgs.Key,
EndKey: resArgs.EndKey,
})
// We've special cased one test case. Note that the intent is still
// counted in `resolved`.
if testCases[id].failResolve {
return util.Errorf("boom")
}
}
return nil
}
defer func() { TestingCommandFilter = nil }()
tc := testContext{}
tc.Start(t)
defer tc.Stop()
tc.manualClock.Set(int64(now))
testIntents := []roachpb.Span{{Key: roachpb.Key("intent")}}
txns := map[string]roachpb.Transaction{}
var epo uint32
for strKey, test := range testCases {
epo++
baseKey := roachpb.Key(strKey)
txnClock := hlc.NewClock(hlc.NewManualClock(int64(test.ts)).UnixNano)
txn := newTransaction("txn1", baseKey, 1, roachpb.SERIALIZABLE, txnClock)
txn.Status = test.status
txn.Intents = testIntents
txn.LastHeartbeat = &roachpb.Timestamp{WallTime: int64(test.heartbeatTS)}
txns[strKey] = *txn
key := keys.TransactionKey(baseKey, txn.ID)
if err := engine.MVCCPutProto(tc.engine, nil, key, roachpb.ZeroTimestamp, nil, txn); err != nil {
t.Fatal(err)
}
seqTS := txn.Timestamp
seqTS.Forward(*txn.LastHeartbeat)
if err := tc.rng.sequence.Put(tc.engine, txn.ID, epo, 2*epo, txn.Key, seqTS, nil /* err */); err != nil {
t.Fatal(err)
}
}
// Run GC.
gcQ := newGCQueue(tc.gossip)
cfg := tc.gossip.GetSystemConfig()
if cfg == nil {
t.Fatal("nil config")
}
if err := gcQ.process(tc.clock.Now(), tc.rng, cfg); err != nil {
t.Fatal(err)
}
util.SucceedsWithin(t, time.Second, func() error {
for strKey, sp := range testCases {
txn := &roachpb.Transaction{}
key := keys.TransactionKey(roachpb.Key(strKey), txns[strKey].ID)
ok, err := engine.MVCCGetProto(tc.engine, key, roachpb.ZeroTimestamp, true, nil, txn)
if err != nil {
return err
}
if expGC := (sp.newStatus == -1); expGC {
if expGC != !ok {
return fmt.Errorf("%s: expected gc: %t, but found %s\n%s", strKey, expGC, txn, roachpb.Key(strKey))
}
} else if sp.newStatus != txn.Status {
return fmt.Errorf("%s: expected status %s, but found %s", strKey, sp.newStatus, txn.Status)
}
var expIntents []roachpb.Span
if sp.expResolve {
expIntents = testIntents
}
if !reflect.DeepEqual(resolved[strKey], expIntents) {
return fmt.Errorf("%s: unexpected intent resolutions:\nexpected: %s\nobserved: %s",
strKey, expIntents, resolved[strKey])
}
if kvs, err := tc.rng.sequence.GetAllTransactionID(tc.store.Engine(), txns[strKey].ID); err != nil {
t.Fatal(err)
} else if (len(kvs) != 0) == sp.expSeqGC {
return fmt.Errorf("%s: expected sequence cache gc: %t, found %+v", strKey, sp.expSeqGC, kvs)
}
}
return nil
})
}
// EndTransaction either commits or aborts (rolls back) an extant
// transaction according to the args.Commit parameter.
func (r *Range) EndTransaction(batch engine.Engine, ms *engine.MVCCStats, args *proto.EndTransactionRequest, reply *proto.EndTransactionResponse) {
if args.Txn == nil {
reply.SetGoError(util.Errorf("no transaction specified to EndTransaction"))
return
}
key := keys.TransactionKey(args.Txn.Key, args.Txn.ID)
// Fetch existing transaction if possible.
existTxn := &proto.Transaction{}
ok, err := engine.MVCCGetProto(batch, key, proto.ZeroTimestamp, true, nil, existTxn)
if err != nil {
reply.SetGoError(err)
return
}
// If the transaction record already exists, verify that we can either
// commit it or abort it (according to args.Commit), and also that the
// Timestamp and Epoch have not suffered regression.
if ok {
// Use the persisted transaction record as final transaction.
reply.Txn = gogoproto.Clone(existTxn).(*proto.Transaction)
if existTxn.Status == proto.COMMITTED {
reply.SetGoError(proto.NewTransactionStatusError(existTxn, "already committed"))
return
} else if existTxn.Status == proto.ABORTED {
reply.SetGoError(proto.NewTransactionAbortedError(existTxn))
return
} else if args.Txn.Epoch < existTxn.Epoch {
reply.SetGoError(proto.NewTransactionStatusError(existTxn, fmt.Sprintf("epoch regression: %d", args.Txn.Epoch)))
return
} else if args.Txn.Epoch == existTxn.Epoch && existTxn.Timestamp.Less(args.Txn.OrigTimestamp) {
// The transaction record can only ever be pushed forward, so it's an
// error if somehow the transaction record has an earlier timestamp
// than the original transaction timestamp.
reply.SetGoError(proto.NewTransactionStatusError(existTxn, fmt.Sprintf("timestamp regression: %s", args.Txn.OrigTimestamp)))
return
}
// Take max of requested epoch and existing epoch. The requester
// may have incremented the epoch on retries.
if reply.Txn.Epoch < args.Txn.Epoch {
reply.Txn.Epoch = args.Txn.Epoch
}
// Take max of requested priority and existing priority. This isn't
// terribly useful, but we do it for completeness.
if reply.Txn.Priority < args.Txn.Priority {
reply.Txn.Priority = args.Txn.Priority
}
} else {
// The transaction doesn't exist yet on disk; use the supplied version.
reply.Txn = gogoproto.Clone(args.Txn).(*proto.Transaction)
}
// Take max of requested timestamp and possibly "pushed" txn
// record timestamp as the final commit timestamp.
if reply.Txn.Timestamp.Less(args.Timestamp) {
reply.Txn.Timestamp = args.Timestamp
}
// Set transaction status to COMMITTED or ABORTED as per the
// args.Commit parameter.
if args.Commit {
// If the isolation level is SERIALIZABLE, return a transaction
// retry error if the commit timestamp isn't equal to the txn
// timestamp.
if args.Txn.Isolation == proto.SERIALIZABLE && !reply.Txn.Timestamp.Equal(args.Txn.OrigTimestamp) {
reply.SetGoError(proto.NewTransactionRetryError(reply.Txn))
return
}
reply.Txn.Status = proto.COMMITTED
} else {
reply.Txn.Status = proto.ABORTED
}
// Persist the transaction record with updated status (& possibly timestamp).
if err := engine.MVCCPutProto(batch, ms, key, proto.ZeroTimestamp, nil, reply.Txn); err != nil {
reply.SetGoError(err)
return
}
// Run triggers if successfully committed. Any failures running
// triggers will set an error and prevent the batch from committing.
if ct := args.InternalCommitTrigger; ct != nil {
// Resolve any explicit intents.
for _, key := range ct.Intents {
if log.V(1) {
log.Infof("resolving intent at %s on end transaction [%s]", key, reply.Txn.Status)
}
if err := engine.MVCCResolveWriteIntent(batch, ms, key, reply.Txn.Timestamp, reply.Txn); err != nil {
reply.SetGoError(err)
return
}
reply.Resolved = append(reply.Resolved, key)
}
// Run appropriate trigger.
if reply.Txn.Status == proto.COMMITTED {
if ct.SplitTrigger != nil {
*ms = engine.MVCCStats{} // clear stats, as split will recompute from scratch.
reply.SetGoError(r.splitTrigger(batch, ct.SplitTrigger))
} else if ct.MergeTrigger != nil {
*ms = engine.MVCCStats{} // clear stats, as merge will recompute from scratch.
reply.SetGoError(r.mergeTrigger(batch, ct.MergeTrigger))
} else if ct.ChangeReplicasTrigger != nil {
reply.SetGoError(r.changeReplicasTrigger(ct.ChangeReplicasTrigger))
}
}
}
}
// processTransactionTable scans the transaction table and updates txnMap with
// those transactions which are old and either PENDING or with intents
// registered. In the first case we want to push the transaction so that it is
// aborted, and in the second case we may have to resolve the intents success-
// fully before GCing the entry. The transaction records which can be gc'ed are
// returned separately and are not added to txnMap nor intentSpanMap.
func processTransactionTable(
ctx context.Context,
snap engine.Engine,
desc *roachpb.RangeDescriptor,
txnMap map[uuid.UUID]*roachpb.Transaction,
cutoff roachpb.Timestamp,
infoMu *lockableGCInfo,
resolveIntents resolveFunc,
) ([]roachpb.GCRequest_GCKey, error) {
infoMu.Lock()
defer infoMu.Unlock()
var gcKeys []roachpb.GCRequest_GCKey
handleOne := func(kv roachpb.KeyValue) error {
var txn roachpb.Transaction
if err := kv.Value.GetProto(&txn); err != nil {
return err
}
infoMu.TransactionSpanTotal++
if !txn.LastActive().Less(cutoff) {
return nil
}
txnID := *txn.ID
// The transaction record should be considered for removal.
switch txn.Status {
case roachpb.PENDING:
// Marked as running, so we need to push it to abort it but won't
// try to GC it in this cycle (for convenience).
// TODO(tschottdorf): refactor so that we can GC PENDING entries
// in the same cycle, but keeping the calls to pushTxn in a central
// location (keeping it easy to batch them up in the future).
infoMu.TransactionSpanGCPending++
txnMap[txnID] = &txn
return nil
case roachpb.ABORTED:
// If we remove this transaction, it effectively still counts as
// ABORTED (by design). So this can be GC'ed even if we can't
// resolve the intents.
// Note: Most aborted transaction weren't aborted by their client,
// but instead by the coordinator - those will not have any intents
// persisted, though they still might exist in the system.
infoMu.TransactionSpanGCAborted++
func() {
infoMu.Unlock() // intentional
defer infoMu.Lock()
if err := resolveIntents(roachpb.AsIntents(txn.Intents, &txn),
true /* wait */, false /* !poison */); err != nil {
log.Warningf("failed to resolve intents of aborted txn on gc: %s", err)
}
}()
case roachpb.COMMITTED:
// It's committed, so it doesn't need a push but we can only
// GC it after its intents are resolved.
if err := func() error {
infoMu.Unlock() // intentional
defer infoMu.Lock()
return resolveIntents(roachpb.AsIntents(txn.Intents, &txn), true /* wait */, false /* !poison */)
}(); err != nil {
log.Warningf("unable to resolve intents of committed txn on gc: %s", err)
// Returning the error here would abort the whole GC run, and
// we don't want that. Instead, we simply don't GC this entry.
return nil
}
infoMu.TransactionSpanGCCommitted++
default:
panic(fmt.Sprintf("invalid transaction state: %s", txn))
}
gcKeys = append(gcKeys, roachpb.GCRequest_GCKey{Key: kv.Key}) // zero timestamp
return nil
}
startKey := keys.TransactionKey(desc.StartKey.AsRawKey(), uuid.EmptyUUID)
endKey := keys.TransactionKey(desc.EndKey.AsRawKey(), uuid.EmptyUUID)
_, err := engine.MVCCIterate(ctx, snap, startKey, endKey,
roachpb.ZeroTimestamp, true /* consistent */, nil, /* txn */
false /* !reverse */, func(kv roachpb.KeyValue) (bool, error) {
return false, handleOne(kv)
})
return gcKeys, err
}
// TestStoreVerifyKeys checks that key length is enforced and
// that end keys must sort >= start.
func TestStoreVerifyKeys(t *testing.T) {
defer leaktest.AfterTest(t)
store, _, stopper := createTestStore(t)
defer stopper.Stop()
tooLongKey := roachpb.Key(strings.Repeat("x", roachpb.KeyMaxLength+1))
// Start with a too-long key on a get.
gArgs := getArgs(tooLongKey, 1, store.StoreID())
if _, err := client.SendWrapped(store, nil, &gArgs); !testutils.IsError(err, "exceeded") {
t.Fatalf("unexpected error for key too long: %v", err)
}
// Try a start key == KeyMax.
gArgs.Key = roachpb.KeyMax
if _, err := client.SendWrapped(store, nil, &gArgs); !testutils.IsError(err, "must be less than KeyMax") {
t.Fatalf("expected error for start key == KeyMax: %v", err)
}
// Try a get with an end key specified (get requires only a start key and should fail).
gArgs.EndKey = roachpb.KeyMax
if _, err := client.SendWrapped(store, nil, &gArgs); !testutils.IsError(err, "must be less than KeyMax") {
t.Fatalf("unexpected error for end key specified on a non-range-based operation: %v", err)
}
// Try a scan with too-long EndKey.
sArgs := scanArgs(roachpb.KeyMin, tooLongKey, 1, store.StoreID())
if _, err := client.SendWrapped(store, nil, &sArgs); !testutils.IsError(err, "length exceeded") {
t.Fatalf("unexpected error for end key too long: %v", err)
}
// Try a scan with end key < start key.
sArgs.Key = []byte("b")
sArgs.EndKey = []byte("a")
if _, err := client.SendWrapped(store, nil, &sArgs); !testutils.IsError(err, "must be greater than") {
t.Fatalf("unexpected error for end key < start: %v", err)
}
// Try a scan with start key == end key.
sArgs.Key = []byte("a")
sArgs.EndKey = sArgs.Key
if _, err := client.SendWrapped(store, nil, &sArgs); !testutils.IsError(err, "must be greater than") {
t.Fatalf("unexpected error for start == end key: %v", err)
}
// Try a scan with range-local start key, but "regular" end key.
sArgs.Key = keys.MakeRangeKey([]byte("test"), []byte("sffx"), nil)
sArgs.EndKey = []byte("z")
if _, err := client.SendWrapped(store, nil, &sArgs); !testutils.IsError(err, "range-local") {
t.Fatalf("unexpected error for local start, non-local end key: %v", err)
}
// Try a put to meta2 key which would otherwise exceed maximum key
// length, but is accepted because of the meta prefix.
meta2KeyMax := keys.MakeKey(keys.Meta2Prefix, roachpb.KeyMax)
pArgs := putArgs(meta2KeyMax, []byte("value"), 1, store.StoreID())
if _, err := client.SendWrapped(store, nil, &pArgs); err != nil {
t.Fatalf("unexpected error on put to meta2 value: %s", err)
}
// Try to put a range descriptor record for a start key which is
// maximum length.
key := append([]byte{}, roachpb.KeyMax...)
key[len(key)-1] = 0x01
pArgs = putArgs(keys.RangeDescriptorKey(key), []byte("value"), 1, store.StoreID())
if _, err := client.SendWrapped(store, nil, &pArgs); err != nil {
t.Fatalf("unexpected error on put to range descriptor for KeyMax value: %s", err)
}
// Try a put to txn record for a meta2 key (note that this doesn't
// actually happen in practice, as txn records are not put directly,
// but are instead manipulated only through txn methods).
pArgs = putArgs(keys.TransactionKey(meta2KeyMax, []byte(uuid.NewUUID4())),
[]byte("value"), 1, store.StoreID())
if _, err := client.SendWrapped(store, nil, &pArgs); err != nil {
t.Fatalf("unexpected error on put to txn meta2 value: %s", err)
}
}
// processIntentsAsync asynchronously processes intents which were
// encountered during another command but did not interfere with the
// execution of that command. This occurs in two cases: inconsistent
// reads and EndTransaction (which queues its own external intents for
// processing via this method). The two cases are handled somewhat
// differently and would be better served by different entry points,
// but combining them simplifies the plumbing necessary in Replica.
func (ir *intentResolver) processIntentsAsync(r *Replica, intents []intentsWithArg) {
if len(intents) == 0 {
return
}
now := r.store.Clock().Now()
ctx := context.TODO()
stopper := r.store.Stopper()
for _, item := range intents {
if item.args.Method() != roachpb.EndTransaction {
if err := stopper.RunLimitedAsyncTask(ir.sem, func() {
// Everything here is best effort; give up rather than waiting
// too long (helps avoid deadlocks during test shutdown,
// although this is imperfect due to the use of an
// uninterruptible WaitGroup.Wait in beginCmds).
ctxWithTimeout, cancel := context.WithTimeout(ctx, base.NetworkTimeout)
defer cancel()
h := roachpb.Header{Timestamp: now}
resolveIntents, pushErr := ir.maybePushTransactions(ctxWithTimeout,
item.intents, h, roachpb.PUSH_TOUCH, true /* skipInFlight */)
// resolveIntents with poison=true because we're resolving
// intents outside of the context of an EndTransaction.
//
// Naively, it doesn't seem like we need to poison the abort
// cache since we're pushing with PUSH_TOUCH - meaning that
// the primary way our Push leads to aborting intents is that
// of the transaction having timed out (and thus presumably no
// client being around any more, though at the time of writing
// we don't guarantee that). But there's another path in which
// the Push comes back successful, namely that of the
// transaction already having been aborted by someone else, in
// which case the client may still be running. Thus, we must
// poison.
if err := ir.resolveIntents(ctxWithTimeout, resolveIntents,
true /* wait */, true /* poison */); err != nil {
log.Warningf(context.TODO(), "%s: failed to resolve intents: %s", r, err)
return
}
if pushErr != nil {
log.Warningf(context.TODO(), "%s: failed to push during intent resolution: %s", r, pushErr)
return
}
}); err != nil {
log.Warningf(context.TODO(), "failed to resolve intents: %s", err)
return
}
} else { // EndTransaction
if err := stopper.RunLimitedAsyncTask(ir.sem, func() {
ctxWithTimeout, cancel := context.WithTimeout(ctx, base.NetworkTimeout)
defer cancel()
// For EndTransaction, we know the transaction is finalized so
// we can skip the push and go straight to the resolve.
//
// This mechanism assumes that when an EndTransaction fails,
// the client makes no assumptions about the result. For
// example, an attempt to explicitly rollback the transaction
// may succeed (triggering this code path), but the result may
// not make it back to the client.
if err := ir.resolveIntents(ctxWithTimeout, item.intents,
true /* wait */, false /* !poison */); err != nil {
log.Warningf(context.TODO(), "%s: failed to resolve intents: %s", r, err)
return
}
// We successfully resolved the intents, so we're able to GC from
// the txn span directly.
b := &client.Batch{}
txn := item.intents[0].Txn
txnKey := keys.TransactionKey(txn.Key, txn.ID)
// This is pretty tricky. Transaction keys are range-local and
// so they are encoded specially. The key range addressed by
// (txnKey, txnKey.Next()) might be empty (since Next() does
// not imply monotonicity on the address side). Instead, we
// send this request to a range determined using the resolved
// transaction anchor, i.e. if the txn is anchored on
// /Local/RangeDescriptor/"a"/uuid, the key range below would
// be ["a", "a\x00"). However, the first range is special again
// because the above procedure results in KeyMin, but we need
// at least KeyLocalMax.
//
// #7880 will address this by making GCRequest less special and
// thus obviating the need to cook up an artificial range here.
var gcArgs roachpb.GCRequest
{
key := keys.MustAddr(txn.Key)
if localMax := keys.MustAddr(keys.LocalMax); key.Less(localMax) {
key = localMax
}
endKey := key.Next()
gcArgs.Span = roachpb.Span{
Key: key.AsRawKey(),
EndKey: endKey.AsRawKey(),
}
}
gcArgs.Keys = append(gcArgs.Keys, roachpb.GCRequest_GCKey{
Key: txnKey,
})
b.AddRawRequest(&gcArgs)
if err := ir.store.db.Run(b); err != nil {
log.Warningf(
context.TODO(),
"could not GC completed transaction anchored at %s: %s",
roachpb.Key(txn.Key), err,
)
return
}
}); err != nil {
log.Warningf(context.TODO(), "failed to resolve intents: %s", err)
return
}
}
}
}
// processIntentsAsync asynchronously processes intents which were
// encountered during another command but did not interfere with the
// execution of that command. This occurs in two cases: inconsistent
// reads and EndTransaction (which queues its own external intents for
// processing via this method). The two cases are handled somewhat
// differently and would be better served by different entry points,
// but combining them simplifies the plumbing necessary in Replica.
func (ir *intentResolver) processIntentsAsync(r *Replica, intents []intentsWithArg) {
if len(intents) == 0 {
return
}
now := r.store.Clock().Now()
ctx := r.context(context.TODO())
stopper := r.store.Stopper()
for _, item := range intents {
if item.args.Method() != roachpb.EndTransaction {
stopper.RunLimitedAsyncTask(ir.sem, func() {
// Everything here is best effort; give up rather than waiting
// too long (helps avoid deadlocks during test shutdown,
// although this is imperfect due to the use of an
// uninterruptible WaitGroup.Wait in beginCmds).
ctxWithTimeout, cancel := context.WithTimeout(ctx, base.NetworkTimeout)
defer cancel()
h := roachpb.Header{Timestamp: now}
resolveIntents, pushErr := ir.maybePushTransactions(ctxWithTimeout,
item.intents, h, roachpb.PUSH_TOUCH, true /* skipInFlight */)
// resolveIntents with poison=true because we're resolving
// intents outside of the context of an EndTransaction.
//
// Naively, it doesn't seem like we need to poison the abort
// cache since we're pushing with PUSH_TOUCH - meaning that
// the primary way our Push leads to aborting intents is that
// of the transaction having timed out (and thus presumably no
// client being around any more, though at the time of writing
// we don't guarantee that). But there's another path in which
// the Push comes back successful, namely that of the
// transaction already having been aborted by someone else, in
// which case the client may still be running. Thus, we must
// poison.
if err := ir.resolveIntents(ctxWithTimeout, r, resolveIntents,
true /* wait */, true /* poison */); err != nil {
log.Warningc(ctxWithTimeout, "failed to resolve intents: %s", err)
return
}
if pushErr != nil {
log.Warningc(ctxWithTimeout, "failed to push during intent resolution: %s", pushErr)
return
}
})
} else { // EndTransaction
stopper.RunLimitedAsyncTask(ir.sem, func() {
ctxWithTimeout, cancel := context.WithTimeout(ctx, base.NetworkTimeout)
defer cancel()
// For EndTransaction, we know the transaction is finalized so
// we can skip the push and go straight to the resolve.
//
// This mechanism assumes that when an EndTransaction fails,
// the client makes no assumptions about the result. For
// example, an attempt to explicitly rollback the transaction
// may succeed (triggering this code path), but the result may
// not make it back to the client.
if err := ir.resolveIntents(ctxWithTimeout, r, item.intents,
true /* wait */, false /* !poison */); err != nil {
log.Warningc(ctxWithTimeout, "failed to resolve intents: %s", err)
return
}
// We successfully resolved the intents, so we're able to GC from
// the txn span directly.
var ba roachpb.BatchRequest
ba.Timestamp = now
txn := item.intents[0].Txn
gcArgs := roachpb.GCRequest{
Span: roachpb.Span{
Key: r.Desc().StartKey.AsRawKey(),
EndKey: r.Desc().EndKey.AsRawKey(),
},
}
gcArgs.Keys = append(gcArgs.Keys, roachpb.GCRequest_GCKey{
Key: keys.TransactionKey(txn.Key, txn.ID),
})
ba.Add(&gcArgs)
if _, pErr := r.addWriteCmd(ctxWithTimeout, ba, nil /* nil */); pErr != nil {
log.Warningf("could not GC completed transaction: %s", pErr)
}
})
}
}
}
// InternalPushTxn resolves conflicts between concurrent txns (or
// between a non-transactional reader or writer and a txn) in several
// ways depending on the statuses and priorities of the conflicting
// transactions. The InternalPushTxn operation is invoked by a
// "pusher" (the writer trying to abort a conflicting txn or the
// reader trying to push a conflicting txn's commit timestamp
// forward), who attempts to resolve a conflict with a "pushee"
// (args.PushTxn -- the pushee txn whose intent(s) caused the
// conflict).
//
// Txn already committed/aborted: If pushee txn is committed or
// aborted return success.
//
// Txn Timeout: If pushee txn entry isn't present or its LastHeartbeat
// timestamp isn't set, use PushTxn.Timestamp as LastHeartbeat. If
// current time - LastHeartbeat > 2 * DefaultHeartbeatInterval, then
// the pushee txn should be either pushed forward, aborted, or
// confirmed not pending, depending on value of Request.PushType.
//
// Old Txn Epoch: If persisted pushee txn entry has a newer Epoch than
// PushTxn.Epoch, return success, as older epoch may be removed.
//
// Lower Txn Priority: If pushee txn has a lower priority than pusher,
// adjust pushee's persisted txn depending on value of
// args.PushType. If args.PushType is ABORT_TXN, set txn.Status to
// ABORTED, and priority to one less than the pusher's priority and
// return success. If args.PushType is PUSH_TIMESTAMP, set
// txn.Timestamp to pusher's Timestamp + 1 (note that we use the
// pusher's Args.Timestamp, not Txn.Timestamp because the args
// timestamp can advance during the txn).
//
// Higher Txn Priority: If pushee txn has a higher priority than
// pusher, return TransactionPushError. Transaction will be retried
// with priority one less than the pushee's higher priority.
func (r *Range) InternalPushTxn(batch engine.Engine, ms *engine.MVCCStats, args *proto.InternalPushTxnRequest, reply *proto.InternalPushTxnResponse) {
if !bytes.Equal(args.Key, args.PusheeTxn.Key) {
reply.SetGoError(util.Errorf("request key %s should match pushee's txn key %s", args.Key, args.PusheeTxn.Key))
return
}
key := keys.TransactionKey(args.PusheeTxn.Key, args.PusheeTxn.ID)
// Fetch existing transaction if possible.
existTxn := &proto.Transaction{}
ok, err := engine.MVCCGetProto(batch, key, proto.ZeroTimestamp,
true /* consistent */, nil /* txn */, existTxn)
if err != nil {
reply.SetGoError(err)
return
}
if ok {
// Start with the persisted transaction record as final transaction.
reply.PusheeTxn = gogoproto.Clone(existTxn).(*proto.Transaction)
// Upgrade the epoch, timestamp and priority as necessary.
if reply.PusheeTxn.Epoch < args.PusheeTxn.Epoch {
reply.PusheeTxn.Epoch = args.PusheeTxn.Epoch
}
reply.PusheeTxn.Timestamp.Forward(args.PusheeTxn.Timestamp)
if reply.PusheeTxn.Priority < args.PusheeTxn.Priority {
reply.PusheeTxn.Priority = args.PusheeTxn.Priority
}
} else {
// Some sanity checks for case where we don't find a transaction record.
if args.PusheeTxn.LastHeartbeat != nil {
reply.SetGoError(proto.NewTransactionStatusError(&args.PusheeTxn,
"no txn persisted, yet intent has heartbeat"))
return
} else if args.PusheeTxn.Status != proto.PENDING {
reply.SetGoError(proto.NewTransactionStatusError(&args.PusheeTxn,
fmt.Sprintf("no txn persisted, yet intent has status %s", args.PusheeTxn.Status)))
return
}
// The transaction doesn't exist yet on disk; use the supplied version.
reply.PusheeTxn = gogoproto.Clone(&args.PusheeTxn).(*proto.Transaction)
}
// If already committed or aborted, return success.
if reply.PusheeTxn.Status != proto.PENDING {
// Trivial noop.
return
}
// If we're trying to move the timestamp forward, and it's already
// far enough forward, return success.
if args.PushType == proto.PUSH_TIMESTAMP && args.Timestamp.Less(reply.PusheeTxn.Timestamp) {
// Trivial noop.
return
}
// pusherWins bool is true in the event the pusher prevails.
var pusherWins bool
// If there's no incoming transaction, the pusher is non-transactional.
// We make a random priority, biased by specified
// args.Header().UserPriority in this case.
var priority int32
if args.Txn != nil {
priority = args.Txn.Priority
} else {
// Make sure we have a deterministic random number when generating
// a priority for this txn-less request, so all replicas see same priority.
randGen := rand.New(rand.NewSource(int64(reply.PusheeTxn.Priority) ^ args.Timestamp.WallTime))
priority = proto.MakePriority(randGen, args.GetUserPriority())
}
// Check for txn timeout.
if reply.PusheeTxn.LastHeartbeat == nil {
reply.PusheeTxn.LastHeartbeat = &reply.PusheeTxn.Timestamp
}
if args.Now.Equal(proto.ZeroTimestamp) {
reply.SetGoError(util.Error("the field Now must be provided"))
return
}
// Compute heartbeat expiration (all replicas must see the same result).
expiry := args.Now
expiry.Forward(args.Timestamp) // if Timestamp is ahead, use that
expiry.WallTime -= 2 * DefaultHeartbeatInterval.Nanoseconds()
if reply.PusheeTxn.LastHeartbeat.Less(expiry) {
if log.V(1) {
log.Infof("pushing expired txn %s", reply.PusheeTxn)
}
pusherWins = true
} else if reply.PusheeTxn.Isolation == proto.SNAPSHOT && args.PushType == proto.PUSH_TIMESTAMP {
if log.V(1) {
log.Infof("pushing timestamp for snapshot isolation txn")
}
pusherWins = true
} else if args.PushType == proto.CLEANUP_TXN {
// If just attempting to cleanup old or already-committed txns, don't push.
pusherWins = false
} else if reply.PusheeTxn.Priority < priority ||
(reply.PusheeTxn.Priority == priority && args.Txn != nil &&
args.Txn.Timestamp.Less(reply.PusheeTxn.Timestamp)) {
// Pusher wins based on priority; if priorities are equal, order
// by lower txn timestamp.
if log.V(1) {
log.Infof("pushing intent from txn with lower priority %s vs %d", reply.PusheeTxn, priority)
}
pusherWins = true
}
if !pusherWins {
err := proto.NewTransactionPushError(args.Txn, reply.PusheeTxn)
if log.V(1) {
log.Info(err)
}
reply.SetGoError(err)
return
}
// Upgrade priority of pushed transaction to one less than pusher's.
reply.PusheeTxn.UpgradePriority(priority - 1)
// If aborting transaction, set new status and return success.
if args.PushType == proto.ABORT_TXN {
reply.PusheeTxn.Status = proto.ABORTED
} else if args.PushType == proto.PUSH_TIMESTAMP {
// Otherwise, update timestamp to be one greater than the request's timestamp.
reply.PusheeTxn.Timestamp = args.Timestamp
reply.PusheeTxn.Timestamp.Logical++
}
// Persist the pushed transaction using zero timestamp for inline value.
if err := engine.MVCCPutProto(batch, ms, key, proto.ZeroTimestamp, nil, reply.PusheeTxn); err != nil {
reply.SetGoError(err)
return
}
}
func TestGCQueueTransactionTable(t *testing.T) {
defer leaktest.AfterTest(t)()
const now time.Duration = 3 * 24 * time.Hour
const gcTxnAndAC = now - txnCleanupThreshold
const gcACOnly = now - abortCacheAgeThreshold
if gcTxnAndAC >= gcACOnly {
t.Fatalf("test assumption violated due to changing constants; needs adjustment")
}
type spec struct {
status roachpb.TransactionStatus
orig time.Duration
hb time.Duration // last heartbeat (none if ZeroTimestamp)
newStatus roachpb.TransactionStatus // -1 for GCed
failResolve bool // do we want to fail resolves in this trial?
expResolve bool // expect attempt at removing txn-persisted intents?
expAbortGC bool // expect abort cache entries removed?
}
// Describes the state of the Txn table before the test.
// Many of the abort cache entries deleted wouldn't even be there, so don't
// be confused by that.
testCases := map[string]spec{
// Too young, should not touch.
"aa": {
status: roachpb.PENDING,
orig: gcACOnly + 1,
newStatus: roachpb.PENDING,
},
// A little older, so the AbortCache gets cleaned up.
"ab": {
status: roachpb.PENDING,
orig: gcTxnAndAC + 1,
newStatus: roachpb.PENDING,
expAbortGC: true,
},
// Old and pending, but still heartbeat (so no Push attempted; it would succeed).
// It's old enough to delete the abort cache entry though.
"ba": {
status: roachpb.PENDING,
hb: gcTxnAndAC + 1,
newStatus: roachpb.PENDING,
expAbortGC: true,
},
// Not old enough for Txn GC, but old enough to remove the abort cache entry.
"bb": {
status: roachpb.ABORTED,
orig: gcACOnly - 1,
newStatus: roachpb.ABORTED,
expAbortGC: true,
},
// Old, pending and abandoned. Should push and abort it successfully,
// but not GC it just yet (this is an artifact of the implementation).
// The abort cache gets cleaned up though.
"c": {
status: roachpb.PENDING,
orig: gcTxnAndAC - 1,
newStatus: roachpb.ABORTED,
expAbortGC: true,
},
// Old and aborted, should delete.
"d": {
status: roachpb.ABORTED,
orig: gcTxnAndAC - 1,
newStatus: -1,
expResolve: true,
expAbortGC: true,
},
// Committed and fresh, so no action. But the abort cache entry is old
// enough to be discarded.
"e": {
status: roachpb.COMMITTED,
orig: gcTxnAndAC + 1,
newStatus: roachpb.COMMITTED,
expAbortGC: true,
},
// Committed and old. It has an intent (like all tests here), which is
// resolvable and hence we can GC.
"f": {
status: roachpb.COMMITTED,
orig: gcTxnAndAC - 1,
newStatus: -1,
expResolve: true,
expAbortGC: true,
},
// Same as the previous one, but we've rigged things so that the intent
// resolution here will fail and consequently no GC is expected.
"g": {
status: roachpb.COMMITTED,
orig: gcTxnAndAC - 1,
newStatus: roachpb.COMMITTED,
failResolve: true,
expResolve: true,
expAbortGC: true,
},
}
resolved := map[string][]roachpb.Span{}
tc := testContext{}
tsc := TestStoreContext()
tsc.TestingKnobs.TestingCommandFilter =
func(filterArgs storagebase.FilterArgs) *roachpb.Error {
if resArgs, ok := filterArgs.Req.(*roachpb.ResolveIntentRequest); ok {
id := string(resArgs.IntentTxn.Key)
resolved[id] = append(resolved[id], roachpb.Span{
Key: resArgs.Key,
EndKey: resArgs.EndKey,
})
// We've special cased one test case. Note that the intent is still
// counted in `resolved`.
if testCases[id].failResolve {
return roachpb.NewErrorWithTxn(util.Errorf("boom"), filterArgs.Hdr.Txn)
}
}
return nil
}
tc.StartWithStoreContext(t, tsc)
defer tc.Stop()
tc.manualClock.Set(int64(now))
outsideKey := tc.rng.Desc().EndKey.Next().AsRawKey()
testIntents := []roachpb.Span{{Key: roachpb.Key("intent")}}
txns := map[string]roachpb.Transaction{}
for strKey, test := range testCases {
baseKey := roachpb.Key(strKey)
txnClock := hlc.NewClock(hlc.NewManualClock(int64(test.orig)).UnixNano)
txn := newTransaction("txn1", baseKey, 1, enginepb.SERIALIZABLE, txnClock)
txn.Status = test.status
txn.Intents = testIntents
if test.hb > 0 {
txn.LastHeartbeat = &hlc.Timestamp{WallTime: int64(test.hb)}
}
// Set a high Timestamp to make sure it does not matter. Only
// OrigTimestamp (and heartbeat) are used for GC decisions.
txn.Timestamp.Forward(hlc.MaxTimestamp)
txns[strKey] = *txn
for _, addrKey := range []roachpb.Key{baseKey, outsideKey} {
key := keys.TransactionKey(addrKey, txn.ID)
if err := engine.MVCCPutProto(context.Background(), tc.engine, nil, key, hlc.ZeroTimestamp, nil, txn); err != nil {
t.Fatal(err)
}
}
entry := roachpb.AbortCacheEntry{Key: txn.Key, Timestamp: txn.LastActive()}
if err := tc.rng.abortCache.Put(context.Background(), tc.engine, nil, txn.ID, &entry); err != nil {
t.Fatal(err)
}
}
// Run GC.
gcQ := newGCQueue(tc.gossip)
cfg, ok := tc.gossip.GetSystemConfig()
if !ok {
t.Fatal("config not set")
}
if err := gcQ.process(tc.clock.Now(), tc.rng, cfg); err != nil {
t.Fatal(err)
}
util.SucceedsSoon(t, func() error {
for strKey, sp := range testCases {
txn := &roachpb.Transaction{}
key := keys.TransactionKey(roachpb.Key(strKey), txns[strKey].ID)
ok, err := engine.MVCCGetProto(context.Background(), tc.engine, key, hlc.ZeroTimestamp, true, nil, txn)
if err != nil {
return err
}
if expGC := (sp.newStatus == -1); expGC {
if expGC != !ok {
return fmt.Errorf("%s: expected gc: %t, but found %s\n%s", strKey, expGC, txn, roachpb.Key(strKey))
}
} else if sp.newStatus != txn.Status {
return fmt.Errorf("%s: expected status %s, but found %s", strKey, sp.newStatus, txn.Status)
}
var expIntents []roachpb.Span
if sp.expResolve {
expIntents = testIntents
}
if !reflect.DeepEqual(resolved[strKey], expIntents) {
return fmt.Errorf("%s: unexpected intent resolutions:\nexpected: %s\nobserved: %s",
strKey, expIntents, resolved[strKey])
}
entry := &roachpb.AbortCacheEntry{}
abortExists, err := tc.rng.abortCache.Get(context.Background(), tc.store.Engine(), txns[strKey].ID, entry)
if err != nil {
t.Fatal(err)
}
if (abortExists == false) != sp.expAbortGC {
return fmt.Errorf("%s: expected abort cache gc: %t, found %+v", strKey, sp.expAbortGC, entry)
}
}
return nil
})
outsideTxnPrefix := keys.TransactionKey(outsideKey, uuid.EmptyUUID)
outsideTxnPrefixEnd := keys.TransactionKey(outsideKey.Next(), uuid.EmptyUUID)
var count int
if _, err := engine.MVCCIterate(context.Background(), tc.store.Engine(), outsideTxnPrefix, outsideTxnPrefixEnd, hlc.ZeroTimestamp,
true, nil, false, func(roachpb.KeyValue) (bool, error) {
count++
return false, nil
}); err != nil {
t.Fatal(err)
}
if exp := len(testCases); exp != count {
t.Fatalf("expected the %d external transaction entries to remain untouched, "+
"but only %d are left", exp, count)
}
}
// processIntentsAsync asynchronously processes intents which were
// encountered during another command but did not interfere with the
// execution of that command. This occurs in two cases: inconsistent
// reads and EndTransaction (which queues its own external intents for
// processing via this method). The two cases are handled somewhat
// differently and would be better served by different entry points,
// but combining them simplifies the plumbing necessary in Replica.
func (ir *intentResolver) processIntentsAsync(r *Replica, intents []intentsWithArg) {
if len(intents) == 0 {
return
}
now := r.store.Clock().Now()
ctx := r.context()
stopper := r.store.Stopper()
for _, item := range intents {
if item.args.Method() != roachpb.EndTransaction {
stopper.RunLimitedAsyncTask(ir.sem, func() {
// Everything here is best effort; give up rather than waiting
// too long (helps avoid deadlocks during test shutdown,
// although this is imperfect due to the use of an
// uninterruptible WaitGroup.Wait in beginCmds).
ctxWithTimeout, cancel := context.WithTimeout(ctx, base.NetworkTimeout)
defer cancel()
h := roachpb.Header{Timestamp: now}
resolveIntents, pushErr := ir.maybePushTransactions(ctxWithTimeout,
item.intents, h, roachpb.PUSH_TOUCH, true /* skipInFlight */)
if pErr := ir.resolveIntents(ctxWithTimeout, r, resolveIntents,
true /* wait */, false /* TODO(tschottdorf): #5088 */); pErr != nil {
log.Warningc(ctxWithTimeout, "failed to resolve intents: %s", pErr)
return
}
if pushErr != nil {
log.Warningc(ctxWithTimeout, "failed to push during intent resolution: %s", pushErr)
return
}
})
} else { // EndTransaction
stopper.RunLimitedAsyncTask(ir.sem, func() {
ctxWithTimeout, cancel := context.WithTimeout(ctx, base.NetworkTimeout)
defer cancel()
// For EndTransaction, we know the transaction is finalized so
// we can skip the push and go straight to the resolve.
if pErr := ir.resolveIntents(ctxWithTimeout, r, item.intents,
true /* wait */, false /* TODO(tschottdorf): #5088 */); pErr != nil {
log.Warningc(ctxWithTimeout, "failed to resolve intents: %s", pErr)
return
}
// We successfully resolved the intents, so we're able to GC from
// the txn span directly. Note that the sequence cache was cleared
// out synchronously with EndTransaction (see comments within for
// an explanation of why that is kosher).
//
// Note that we poisoned the sequence caches on the external ranges
// above. This may seem counter-intuitive, but it's actually
// necessary: Assume a transaction has committed here, with two
// external intents, and assume that we did not poison. Normally,
// these two intents would be resolved in the same batch, but that
// is not guaranteed (for example, if DistSender has a stale
// descriptor after a Merge). When resolved separately, the first
// ResolveIntent would clear out the sequence cache; an individual
// write on the second (still present) intent could then be
// replayed and would resolve to a real value (at least for a
// window of time unless we delete the local txn entry). That's not
// OK for non-idempotent commands such as Increment.
// TODO(tschottdorf): We should have another side effect on
// MVCCResolveIntent (on commit/abort): If it were able to remove
// the txn from its corresponding entries in the timestamp cache,
// no more replays at the same timestamp would be possible. This
// appears to be a useful performance optimization; we could then
// not poison on EndTransaction. In fact, the above mechanism
// could be an effective alternative to sequence-cache based
// poisoning (or the whole sequence cache?) itself.
//
// TODO(tschottdorf): down the road, can probably unclog the system
// here by batching up a bunch of those GCRequests before proposing.
var ba roachpb.BatchRequest
txn := item.intents[0].Txn
gcArgs := roachpb.GCRequest{
Span: roachpb.Span{
Key: r.Desc().StartKey.AsRawKey(),
EndKey: r.Desc().EndKey.AsRawKey(),
},
}
gcArgs.Keys = append(gcArgs.Keys, roachpb.GCRequest_GCKey{
Key: keys.TransactionKey(txn.Key, txn.ID),
})
ba.Add(&gcArgs)
if _, pErr := r.addWriteCmd(ctxWithTimeout, ba, nil /* nil */); pErr != nil {
log.Warningf("could not GC completed transaction: %s", pErr)
}
})
}
}
}
