// InternalHeartbeatTxn updates the transaction status and heartbeat
// timestamp after receiving transaction heartbeat messages from
// coordinator. Returns the udpated transaction.
func (r *Range) InternalHeartbeatTxn(args *proto.InternalHeartbeatTxnRequest, reply *proto.InternalHeartbeatTxnResponse) {
// Create the actual key to the system-local transaction table.
key := engine.MakeKey(engine.KeyLocalTransactionPrefix, args.Key)
var txn proto.Transaction
ok, err := engine.GetProto(r.engine, key, &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.PutProto(r.engine, key, &txn); err != nil {
reply.SetGoError(err)
return
}
}
reply.Txn = &txn
}
// CreateRange allocates a new range ID and stores range metadata.
// On success, returns the new range.
func (s *Store) CreateRange(startKey, endKey engine.Key, replicas []proto.Replica) (*Range, error) {
rangeID, err := engine.Increment(s.engine, engine.KeyLocalRangeIDGenerator, 1)
if err != nil {
return nil, err
}
if ok, _ := engine.GetProto(s.engine, makeRangeKey(rangeID), nil); ok {
return nil, util.Error("newly allocated range ID already in use")
}
// RangeMetadata is stored local to this store only. It is neither
// replicated via raft nor available via the global kv store.
meta := &proto.RangeMetadata{
ClusterID: s.Ident.ClusterID,
RangeID: rangeID,
RangeDescriptor: proto.RangeDescriptor{
StartKey: startKey,
EndKey: endKey,
Replicas: replicas,
},
}
err = engine.PutProto(s.engine, makeRangeKey(rangeID), meta)
if err != nil {
return nil, err
}
rng := NewRange(meta, s.clock, s.engine, s.allocator, s.gossip, s)
rng.Start()
s.mu.Lock()
defer s.mu.Unlock()
s.ranges[rangeID] = rng
return rng, nil
}
// Init starts the engine, sets the GC and reads the StoreIdent.
func (s *Store) Init() error {
// Start engine and set garbage collector.
if err := s.engine.Start(); err != nil {
return err
}
s.engine.SetGCTimeouts(func() (minTxnTS, minRCacheTS int64) {
now := s.clock.Now()
minTxnTS = 0 // disable GC of transactions until we know minimum write intent age
minRCacheTS = now.WallTime - GCResponseCacheExpiration.Nanoseconds()
return
})
// Read store ident and return a not-bootstrapped error if necessary.
ok, err := engine.GetProto(s.engine, engine.KeyLocalIdent, &s.Ident)
if err != nil {
return err
} else if !ok {
return &NotBootstrappedError{}
}
// TODO(spencer): scan through all range metadata and instantiate
// ranges. Right now we just get range ID hardcoded as 1.
var meta proto.RangeMetadata
ok, err = engine.GetProto(s.engine, makeRangeKey(1), &meta)
if err != nil || !ok {
return err
}
rng := NewRange(&meta, s.clock, s.engine, s.allocator, s.gossip, s)
rng.Start()
s.mu.Lock()
defer s.mu.Unlock()
s.ranges[meta.RangeID] = rng
return nil
}
// InternalHeartbeatTxn updates the transaction status and heartbeat
// timestamp after receiving transaction heartbeat messages from
// coordinator. The range will return the current status for this
// transaction to the coordinator.
func (r *Range) InternalHeartbeatTxn(args *proto.InternalHeartbeatTxnRequest, reply *proto.InternalHeartbeatTxnResponse) {
// Create the actual key to the system-local transaction table.
key := engine.MakeKey(engine.KeyLocalTransactionPrefix, args.Key)
var txn proto.Transaction
if _, err := engine.GetProto(r.engine, key, &txn); err != nil {
reply.SetGoError(err)
return
}
if txn.Status == proto.PENDING {
if !args.Header().Timestamp.Less(txn.LastHeartbeat) {
txn.LastHeartbeat = args.Header().Timestamp
}
if err := engine.PutProto(r.engine, key, &txn); err != nil {
reply.SetGoError(err)
return
}
}
reply.Status = txn.Status
}
// GetResponse looks up a response matching the specified cmdID and
// returns true if found. The response is deserialized into the
// supplied reply parameter. If no response is found, returns
// false. If a command is pending already for the cmdID, then this
// method will block until the the command is completed or the
// response cache is cleared.
func (rc *ResponseCache) GetResponse(cmdID proto.ClientCmdID, reply interface{}) (bool, error) {
// Do nothing if command ID is empty.
if cmdID.IsEmpty() {
return false, nil
}
// If the command is inflight, wait for it to complete.
rc.Lock()
for {
if cond, ok := rc.inflight[makeCmdIDKey(cmdID)]; ok {
cond.Wait()
} else {
break
}
}
// Adding inflight here is preemptive; we don't want to hold lock
// while fetching from the on-disk cache. The vast, vast majority of
// calls to GetResponse will be cache misses, so this saves us
// from acquiring the lock twice: once here and once below in the
// event we experience a cache miss.
rc.addInflightLocked(cmdID)
rc.Unlock()
// If the response is in the cache or we experienced an error, return.
rwResp := proto.ReadWriteCmdResponse{}
if ok, err := engine.GetProto(rc.engine, rc.makeKey(cmdID), &rwResp); ok || err != nil {
rc.Lock() // Take lock after fetching response from cache.
defer rc.Unlock()
rc.removeInflightLocked(cmdID)
if err == nil && rwResp.GetValue() != nil {
gogoproto.Merge(reply.(gogoproto.Message), rwResp.GetValue().(gogoproto.Message))
}
return ok, err
}
// There's no command result cached for this ID; but inflight was added above.
return false, nil
}
// Init starts the engine, sets the GC and reads the StoreIdent.
func (s *Store) Init() error {
// Close store for idempotency.
s.Close()
// Start engine and set garbage collector.
if err := s.engine.Start(); err != nil {
return err
}
// Create ID allocators.
s.raftIDAlloc = NewIDAllocator(engine.KeyRaftIDGenerator, s.db, 2, raftIDAllocCount)
s.rangeIDAlloc = NewIDAllocator(engine.KeyRangeIDGenerator, s.db, 2, rangeIDAllocCount)
// GCTimeouts method is called each time an engine compaction is
// underway. It sets minimum timeouts for transaction records and
// response cache entries.
s.engine.SetGCTimeouts(func() (minTxnTS, minRCacheTS int64) {
now := s.clock.Now()
minTxnTS = 0 // disable GC of transactions until we know minimum write intent age
minRCacheTS = now.WallTime - GCResponseCacheExpiration.Nanoseconds()
return
})
// Read store ident and return a not-bootstrapped error if necessary.
ok, err := engine.GetProto(s.engine, engine.KeyLocalIdent, &s.Ident)
if err != nil {
return err
} else if !ok {
return &NotBootstrappedError{}
}
s.mu.Lock()
defer s.mu.Unlock()
start := engine.KeyLocalRangeMetadataPrefix
end := engine.PrefixEndKey(start)
const rows = 64
for {
kvs, err := s.engine.Scan(start, end, rows)
if err != nil {
return err
}
for _, kv := range kvs {
var meta proto.RangeMetadata
if err := gogoproto.Unmarshal(kv.Value, &meta); err != nil {
return err
}
rng := NewRange(&meta, s.clock, s.engine, s.allocator, s.gossip, s)
rng.Start()
s.ranges[meta.RangeID] = rng
s.rangesByKey = append(s.rangesByKey, rng)
}
if len(kvs) < rows {
break
}
start = engine.NextKey(kvs[rows-1].Key)
}
// Ensure that ranges are sorted.
sort.Sort(s.rangesByKey)
return nil
}
// EndTransaction either commits or aborts (rolls back) an extant
// transaction according to the args.Commit parameter.
func (r *Range) EndTransaction(args *proto.EndTransactionRequest, reply *proto.EndTransactionResponse) {
// Create the actual key to the system-local transaction table.
key := engine.MakeKey(engine.KeyLocalTransactionPrefix, args.Key)
// Start with supplied transaction, then possibly load from txn record.
reply.Txn = gogoproto.Clone(args.Txn).(*proto.Transaction)
// Fetch existing transaction if possible.
existTxn := &proto.Transaction{}
ok, err := engine.GetProto(r.engine, key, 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 {
if existTxn.Status == proto.COMMITTED {
reply.SetGoError(proto.NewTransactionStatusError(existTxn, "already committed"))
return
} else if existTxn.Status == proto.ABORTED {
reply.SetGoError(proto.NewTransactionStatusError(existTxn, "already aborted"))
return
} else if args.Txn.Epoch < existTxn.Epoch {
reply.SetGoError(proto.NewTransactionStatusError(existTxn, fmt.Sprintf("epoch regression: %d", args.Txn.Epoch)))
return
} else if existTxn.Timestamp.Less(args.Txn.Timestamp) {
// 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 transaction timestamp.
reply.SetGoError(proto.NewTransactionStatusError(existTxn, fmt.Sprintf("timestamp regression: %+v", args.Txn.Timestamp)))
return
}
// Use the persisted transaction record as final transaction.
gogoproto.Merge(reply.Txn, existTxn)
}
// 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.Timestamp) {
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 timestmap).
if err := engine.PutProto(r.engine, key, reply.Txn); err != nil {
reply.SetGoError(err)
return
}
}
// 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 or aborted,
// depending on value of Request.Abort.
//
// 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.Abort. If
// args.Abort is true, set txn.Status to ABORTED, and priority to one
// less than the pusher's priority and return success. If args.Abort
// is false, set txn.Timestamp to pusher's txn.Timestamp + 1.
//
// Higher Txn Priority: If pushee txn has a higher priority than
// pusher, return TransactionRetryError. Transaction will be retried
// with priority one less than the pushee's higher priority.
func (r *Range) InternalPushTxn(args *proto.InternalPushTxnRequest, reply *proto.InternalPushTxnResponse) {
if !bytes.Equal(args.Key, args.PusheeTxn.ID) {
reply.SetGoError(util.Errorf("request key %q should match pushee's txn ID %q", args.Key, args.PusheeTxn.ID))
return
}
// Create the actual key to the system-local transaction table.
key := engine.MakeKey(engine.KeyLocalTransactionPrefix, args.Key)
// Fetch existing transaction if possible.
existTxn := &proto.Transaction{}
ok, err := engine.GetProto(r.engine, key, 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 and timestamp as necessary.
if reply.PusheeTxn.Epoch < args.PusheeTxn.Epoch {
reply.PusheeTxn.Epoch = args.PusheeTxn.Epoch
}
if reply.PusheeTxn.Timestamp.Less(args.PusheeTxn.Timestamp) {
reply.PusheeTxn.Timestamp = args.PusheeTxn.Timestamp
}
} 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.Abort && args.Timestamp.Less(reply.PusheeTxn.Timestamp) {
// Trivial noop.
return
}
// pusherWins bool is true in the event the pusher prevails.
var pusherWins bool
// Check for txn timeout.
if reply.PusheeTxn.LastHeartbeat == nil {
reply.PusheeTxn.LastHeartbeat = &reply.PusheeTxn.Timestamp
}
// Compute heartbeat expiration.
expiry := r.clock.Now()
expiry.WallTime -= 2 * DefaultHeartbeatInterval.Nanoseconds()
if reply.PusheeTxn.LastHeartbeat.Less(expiry) {
log.V(1).Infof("pushing expired txn %+v", reply.PusheeTxn)
pusherWins = true
} else if args.PusheeTxn.Epoch < reply.PusheeTxn.Epoch {
// Check for an intent from a prior epoch.
log.V(1).Infof("pushing intent from previous epoch for txn %+v", reply.PusheeTxn)
pusherWins = true
} else if reply.PusheeTxn.Priority < args.Txn.Priority ||
(reply.PusheeTxn.Priority == args.Txn.Priority && args.Txn.Timestamp.Less(reply.PusheeTxn.Timestamp)) {
// Finally, choose based on priority; if priorities are equal, order by lower txn timestamp.
log.V(1).Infof("pushing intent from txn with lower priority %+v vs %+v", reply.PusheeTxn, args.Txn)
pusherWins = true
}
if !pusherWins {
log.V(1).Infof("failed to push intent %+v vs %+v", reply.PusheeTxn, args.Txn)
reply.SetGoError(proto.NewTransactionRetryError(reply.PusheeTxn))
return
}
// If aborting transaction, set new status and return success.
if args.Abort {
reply.PusheeTxn.Status = proto.ABORTED
} else {
// 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.
if err := engine.PutProto(r.engine, key, reply.PusheeTxn); err != nil {
reply.SetGoError(err)
return
}
}