// Send forwards the call to the single store. This is a poor man's
// version of kv.TxnCoordSender, but it serves the purposes of
// supporting tests in this package. Transactions are not supported.
// Since kv/ depends on storage/, we can't get access to a
// TxnCoordSender from here.
// TODO(tschottdorf): {kv->storage}.LocalSender
func (db *testSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
if et, ok := ba.GetArg(roachpb.EndTransaction); ok {
return nil, roachpb.NewError(util.Errorf("%s method not supported", et.Method()))
}
// Lookup range and direct request.
key, endKey := keys.Range(ba)
rng := db.store.LookupReplica(key, endKey)
if rng == nil {
return nil, roachpb.NewError(roachpb.NewRangeKeyMismatchError(key, endKey, nil))
}
ba.RangeID = rng.Desc().RangeID
replica := rng.GetReplica()
if replica == nil {
return nil, roachpb.NewError(util.Errorf("own replica missing in range"))
}
ba.Replica = *replica
br, pErr := db.store.Send(ctx, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(db.store, br))
}
if pErr != nil {
return nil, pErr
}
return br, nil
}
// sendRPC sends one or more RPCs to replicas from the supplied roachpb.Replica
// slice. Returns an RPC error if the request could not be sent. Note
// that the reply may contain a higher level error and must be checked in
// addition to the RPC error.
// The replicas are assume to have been ordered by preference, closer ones (if
// any) at the front.
func (ds *DistSender) sendRPC(
ctx context.Context,
rangeID roachpb.RangeID,
replicas ReplicaSlice,
ba roachpb.BatchRequest,
) (*roachpb.BatchResponse, error) {
if len(replicas) == 0 {
return nil, noNodeAddrsAvailError{}
}
// TODO(pmattis): This needs to be tested. If it isn't set we'll
// still route the request appropriately by key, but won't receive
// RangeNotFoundErrors.
ba.RangeID = rangeID
// Set RPC opts with stipulation that one of N RPCs must succeed.
rpcOpts := SendOptions{
SendNextTimeout: ds.sendNextTimeout,
Timeout: base.NetworkTimeout,
Context: ctx,
transportFactory: ds.transportFactory,
}
tracing.AnnotateTrace()
defer tracing.AnnotateTrace()
reply, err := ds.sendToReplicas(rpcOpts, rangeID, replicas, ba, ds.rpcContext)
if err != nil {
return nil, err
}
return reply, nil
}
// sendRPC sends one or more RPCs to replicas from the supplied roachpb.Replica
// slice. First, replicas which have gossiped addresses are corralled (and
// rearranged depending on proximity and whether the request needs to go to a
// leader) and then sent via Send, with requirement that one RPC to a server
// must succeed. Returns an RPC error if the request could not be sent. Note
// that the reply may contain a higher level error and must be checked in
// addition to the RPC error.
func (ds *DistSender) sendRPC(trace opentracing.Span, rangeID roachpb.RangeID, replicas ReplicaSlice,
order orderingPolicy, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
if len(replicas) == 0 {
return nil, roachpb.NewError(noNodeAddrsAvailError{})
}
// TODO(pmattis): This needs to be tested. If it isn't set we'll
// still route the request appropriately by key, but won't receive
// RangeNotFoundErrors.
ba.RangeID = rangeID
// Set RPC opts with stipulation that one of N RPCs must succeed.
rpcOpts := SendOptions{
Ordering: order,
SendNextTimeout: defaultSendNextTimeout,
Timeout: rpc.DefaultRPCTimeout,
Trace: trace,
}
tracing.AnnotateTrace()
defer tracing.AnnotateTrace()
reply, err := ds.rpcSend(rpcOpts, replicas, ba, ds.rpcContext)
if err != nil {
return nil, roachpb.NewError(err)
}
return reply.(*roachpb.BatchResponse), nil
}
// Send implements the client.Sender interface. The store is looked up from the
// store map if specified by the request; otherwise, the command is being
// executed locally, and the replica is determined via lookup through each
// store's LookupRange method. The latter path is taken only by unit tests.
func (ls *Stores) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
var store *Store
var err error
// If we aren't given a Replica, then a little bending over
// backwards here. This case applies exclusively to unittests.
if ba.RangeID == 0 || ba.Replica.StoreID == 0 {
var repl *roachpb.ReplicaDescriptor
var rangeID roachpb.RangeID
rs := keys.Range(ba)
rangeID, repl, err = ls.lookupReplica(rs.Key, rs.EndKey)
if err == nil {
ba.RangeID = rangeID
ba.Replica = *repl
}
}
ctx = log.Add(ctx,
log.RangeID, ba.RangeID)
if err == nil {
store, err = ls.GetStore(ba.Replica.StoreID)
}
if err != nil {
return nil, roachpb.NewError(err)
}
sp, cleanupSp := tracing.SpanFromContext(opStores, store.Tracer(), ctx)
defer cleanupSp()
if ba.Txn != nil {
// For calls that read data within a txn, we keep track of timestamps
// observed from the various participating nodes' HLC clocks. If we have
// a timestamp on file for this Node which is smaller than MaxTimestamp,
// we can lower MaxTimestamp accordingly. If MaxTimestamp drops below
// OrigTimestamp, we effectively can't see uncertainty restarts any
// more.
// Note that it's not an issue if MaxTimestamp propagates back out to
// the client via a returned Transaction update - when updating a Txn
// from another, the larger MaxTimestamp wins.
if maxTS, ok := ba.Txn.GetObservedTimestamp(ba.Replica.NodeID); ok && maxTS.Less(ba.Txn.MaxTimestamp) {
// Copy-on-write to protect others we might be sharing the Txn with.
shallowTxn := *ba.Txn
// The uncertainty window is [OrigTimestamp, maxTS), so if that window
// is empty, there won't be any uncertainty restarts.
if !ba.Txn.OrigTimestamp.Less(maxTS) {
sp.LogEvent("read has no clock uncertainty")
}
shallowTxn.MaxTimestamp.Backward(maxTS)
ba.Txn = &shallowTxn
}
}
br, pErr := store.Send(ctx, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(store, br))
}
return br, pErr
}
// Send implements the client.Sender interface. The store is looked up from the
// store map if specified by the request; otherwise, the command is being
// executed locally, and the replica is determined via lookup through each
// store's LookupRange method. The latter path is taken only by unit tests.
func (ls *Stores) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
sp := tracing.SpanFromContext(ctx)
var store *Store
var pErr *roachpb.Error
// If we aren't given a Replica, then a little bending over
// backwards here. This case applies exclusively to unittests.
if ba.RangeID == 0 || ba.Replica.StoreID == 0 {
var repl *roachpb.ReplicaDescriptor
var rangeID roachpb.RangeID
rs := keys.Range(ba)
rangeID, repl, pErr = ls.lookupReplica(rs.Key, rs.EndKey)
if pErr == nil {
ba.RangeID = rangeID
ba.Replica = *repl
}
}
ctx = log.Add(ctx,
log.RangeID, ba.RangeID)
if pErr == nil {
store, pErr = ls.GetStore(ba.Replica.StoreID)
}
var br *roachpb.BatchResponse
if pErr != nil {
return nil, pErr
}
// For calls that read data within a txn, we can avoid uncertainty
// related retries in certain situations. If the node is in
// "CertainNodes", we need not worry about uncertain reads any
// more. Setting MaxTimestamp=OrigTimestamp for the operation
// accomplishes that. See roachpb.Transaction.CertainNodes for details.
if ba.Txn != nil && ba.Txn.CertainNodes.Contains(ba.Replica.NodeID) {
// MaxTimestamp = Timestamp corresponds to no clock uncertainty.
sp.LogEvent("read has no clock uncertainty")
// Copy-on-write to protect others we might be sharing the Txn with.
shallowTxn := *ba.Txn
// We set to OrigTimestamp because that works for both SNAPSHOT and
// SERIALIZABLE: If we used Timestamp instead, we could run into
// unnecessary retries at SNAPSHOT. For example, a SNAPSHOT txn at
// OrigTimestamp = 1000.0, Timestamp = 2000.0, MaxTimestamp = 3000.0
// will always read at 1000, so a MaxTimestamp of 2000 will still let
// it restart with uncertainty when it finds a value in (1000, 2000).
shallowTxn.MaxTimestamp = ba.Txn.OrigTimestamp
ba.Txn = &shallowTxn
}
br, pErr = store.Send(ctx, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(store, br))
}
return br, pErr
}
// process iterates through all keys in a replica's range, calling the garbage
// collector for each key and associated set of values. GC'd keys are batched
// into GC calls. Extant intents are resolved if intents are older than
// intentAgeThreshold. The transaction and abort cache records are also
// scanned and old entries evicted. During normal operation, both of these
// records are cleaned up when their respective transaction finishes, so the
// amount of work done here is expected to be small.
//
// Some care needs to be taken to avoid cyclic recreation of entries during GC:
// * a Push initiated due to an intent may recreate a transaction entry
// * resolving an intent may write a new abort cache entry
// * obtaining the transaction for a abort cache entry requires a Push
//
// The following order is taken below:
// 1) collect all intents with sufficiently old txn record
// 2) collect these intents' transactions
// 3) scan the transaction table, collecting abandoned or completed txns
// 4) push all of these transactions (possibly recreating entries)
// 5) resolve all intents (unless the txn is still PENDING), which will recreate
// abort cache entries (but with the txn timestamp; i.e. likely gc'able)
// 6) scan the abort cache table for old entries
// 7) push these transactions (again, recreating txn entries).
// 8) send a GCRequest.
func (gcq *gcQueue) process(
ctx context.Context,
now hlc.Timestamp,
repl *Replica,
sysCfg config.SystemConfig,
) error {
snap := repl.store.Engine().NewSnapshot()
desc := repl.Desc()
defer snap.Close()
// Lookup the GC policy for the zone containing this key range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return errors.Errorf("could not find zone config for range %s: %s", repl, err)
}
gcKeys, info, err := RunGC(ctx, desc, snap, now, zone.GC,
func(now hlc.Timestamp, txn *roachpb.Transaction, typ roachpb.PushTxnType) {
pushTxn(gcq.store.DB(), now, txn, typ)
},
func(intents []roachpb.Intent, poison bool, wait bool) error {
return repl.store.intentResolver.resolveIntents(ctx, intents, poison, wait)
})
if err != nil {
return err
}
gcq.eventLog.VInfof(true, "completed with stats %+v", info)
var ba roachpb.BatchRequest
var gcArgs roachpb.GCRequest
// TODO(tschottdorf): This is one of these instances in which we want
// to be more careful that the request ends up on the correct Replica,
// and we might have to worry about mixing range-local and global keys
// in a batch which might end up spanning Ranges by the time it executes.
gcArgs.Key = desc.StartKey.AsRawKey()
gcArgs.EndKey = desc.EndKey.AsRawKey()
gcArgs.Keys = gcKeys
gcArgs.Threshold = info.Threshold
// Technically not needed since we're talking directly to the Range.
ba.RangeID = desc.RangeID
ba.Timestamp = now
ba.Add(&gcArgs)
if _, pErr := repl.Send(ctx, ba); pErr != nil {
return pErr.GoError()
}
return nil
}
// Send implements the client.Sender interface. The store is looked up from the
// store map if specified by the request; otherwise, the command is being
// executed locally, and the replica is determined via lookup through each
// store's LookupRange method. The latter path is taken only by unit tests.
func (ls *Stores) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
sp := tracing.SpanFromContext(ctx)
var store *Store
var pErr *roachpb.Error
// If we aren't given a Replica, then a little bending over
// backwards here. This case applies exclusively to unittests.
if ba.RangeID == 0 || ba.Replica.StoreID == 0 {
var repl *roachpb.ReplicaDescriptor
var rangeID roachpb.RangeID
rs := keys.Range(ba)
rangeID, repl, pErr = ls.lookupReplica(rs.Key, rs.EndKey)
if pErr == nil {
ba.RangeID = rangeID
ba.Replica = *repl
}
}
ctx = log.Add(ctx,
log.RangeID, ba.RangeID)
if pErr == nil {
store, pErr = ls.GetStore(ba.Replica.StoreID)
}
var br *roachpb.BatchResponse
if pErr != nil {
return nil, pErr
}
// For calls that read data within a txn, we can avoid uncertainty
// related retries in certain situations. If the node is in
// "CertainNodes", we need not worry about uncertain reads any
// more. Setting MaxTimestamp=Timestamp for the operation
// accomplishes that. See roachpb.Transaction.CertainNodes for details.
if ba.Txn != nil && ba.Txn.CertainNodes.Contains(ba.Replica.NodeID) {
// MaxTimestamp = Timestamp corresponds to no clock uncertainty.
sp.LogEvent("read has no clock uncertainty")
// Copy-on-write to protect others we might be sharing the Txn with.
shallowTxn := *ba.Txn
shallowTxn.MaxTimestamp = ba.Txn.Timestamp
ba.Txn = &shallowTxn
}
br, pErr = store.Send(ctx, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(store, br))
}
return br, pErr
}
// Send implements the client.Sender interface. The store is looked up from the
// store map if specified by the request; otherwise, the command is being
// executed locally, and the replica is determined via lookup through each
// store's LookupRange method. The latter path is taken only by unit tests.
func (ls *LocalSender) Send(ctx context.Context, ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
trace := tracer.FromCtx(ctx)
var store *storage.Store
var err error
// If we aren't given a Replica, then a little bending over
// backwards here. This case applies exclusively to unittests.
if ba.RangeID == 0 || ba.Replica.StoreID == 0 {
var repl *roachpb.ReplicaDescriptor
var rangeID roachpb.RangeID
key, endKey := keys.Range(ba)
rangeID, repl, err = ls.lookupReplica(key, endKey)
if err == nil {
ba.RangeID = rangeID
ba.Replica = *repl
}
}
ctx = log.Add(ctx,
log.RangeID, ba.RangeID)
if err == nil {
store, err = ls.GetStore(ba.Replica.StoreID)
}
var br *roachpb.BatchResponse
if err != nil {
return nil, roachpb.NewError(err)
}
// For calls that read data within a txn, we can avoid uncertainty
// related retries in certain situations. If the node is in
// "CertainNodes", we need not worry about uncertain reads any
// more. Setting MaxTimestamp=Timestamp for the operation
// accomplishes that. See roachpb.Transaction.CertainNodes for details.
if ba.Txn != nil && ba.Txn.CertainNodes.Contains(ba.Replica.NodeID) {
// MaxTimestamp = Timestamp corresponds to no clock uncertainty.
trace.Event("read has no clock uncertainty")
ba.Txn.MaxTimestamp = ba.Txn.Timestamp
}
br, pErr := store.Send(ctx, ba)
if br != nil && br.Error != nil {
panic(roachpb.ErrorUnexpectedlySet(store, br))
}
return br, pErr
}
func TestStoreExecuteNoop(t *testing.T) {
defer leaktest.AfterTest(t)
store, _, stopper := createTestStore(t)
defer stopper.Stop()
ba := roachpb.BatchRequest{}
ba.Key = nil // intentional
ba.RangeID = 1
ba.Replica = roachpb.ReplicaDescriptor{StoreID: store.StoreID()}
ba.Add(&roachpb.GetRequest{RequestHeader: roachpb.RequestHeader{Key: roachpb.Key("a")}})
ba.Add(&roachpb.NoopRequest{})
br, pErr := store.Send(context.Background(), ba)
if pErr != nil {
t.Error(pErr)
}
reply := br.Responses[1].GetInner()
if _, ok := reply.(*roachpb.NoopResponse); !ok {
t.Errorf("expected *roachpb.NoopResponse, got %T", reply)
}
}
// sendRPC sends one or more RPCs to replicas from the supplied roachpb.Replica
// slice. First, replicas which have gossiped addresses are corralled (and
// rearranged depending on proximity and whether the request needs to go to a
// leader) and then sent via rpc.Send, with requirement that one RPC to a
// server must succeed. Returns an RPC error if the request could not be sent.
// Note that the reply may contain a higher level error and must be checked in
// addition to the RPC error.
func (ds *DistSender) sendRPC(trace *tracer.Trace, rangeID roachpb.RangeID, replicas replicaSlice, order rpc.OrderingPolicy,
ba roachpb.BatchRequest) (*roachpb.BatchResponse, *roachpb.Error) {
if len(replicas) == 0 {
return nil, roachpb.NewError(noNodeAddrsAvailError{})
}
// Build a slice of replica addresses.
addrs := make([]net.Addr, 0, len(replicas))
replicaMap := make(map[string]*roachpb.ReplicaDescriptor, len(replicas))
for i := range replicas {
addr := replicas[i].NodeDesc.Address
addrs = append(addrs, addr)
replicaMap[addr.String()] = &replicas[i].ReplicaDescriptor
}
// TODO(pmattis): This needs to be tested. If it isn't set we'll
// still route the request appropriately by key, but won't receive
// RangeNotFoundErrors.
ba.RangeID = rangeID
// Set RPC opts with stipulation that one of N RPCs must succeed.
rpcOpts := rpc.Options{
N: 1,
Ordering: order,
SendNextTimeout: defaultSendNextTimeout,
Timeout: defaultRPCTimeout,
Trace: trace,
}
// getArgs clones the arguments on demand for all but the first replica.
firstArgs := true
getArgs := func(addr net.Addr) proto.Message {
var a *roachpb.BatchRequest
// Use the supplied args proto if this is our first address.
if firstArgs {
firstArgs = false
a = &ba
} else {
// Otherwise, copy the args value and set the replica in the header.
a = proto.Clone(&ba).(*roachpb.BatchRequest)
}
if addr != nil {
// TODO(tschottdorf): see len(replicas) above.
a.Replica = *replicaMap[addr.String()]
}
return a
}
// RPCs are sent asynchronously and there is no synchronized access to
// the reply object, so we don't pass itself to RPCSend.
// Otherwise there maybe a race case:
// If the RPC call times out using our original reply object,
// we must not use it any more; the rpc call might still return
// and just write to it at any time.
// args.CreateReply() should be cheaper than proto.Clone which use reflect.
getReply := func() proto.Message {
return &roachpb.BatchResponse{}
}
const method = "Node.Batch"
replies, err := ds.rpcSend(rpcOpts, method, addrs, getArgs, getReply, ds.rpcContext)
if err != nil {
return nil, roachpb.NewError(err)
}
return replies[0].(*roachpb.BatchResponse), nil
}
// process iterates through all keys in a replica's range, calling the garbage
// collector for each key and associated set of values. GC'd keys are batched
// into GC calls. Extant intents are resolved if intents are older than
// intentAgeThreshold. The transaction and sequence cache records are also
// scanned and old entries evicted. During normal operation, both of these
// records are cleaned up when their respective transaction finishes, so the
// amount of work done here is expected to be small.
//
// Some care needs to be taken to avoid cyclic recreation of entries during GC:
// * a Push initiated due to an intent may recreate a transaction entry
// * resolving an intent may write a new sequence cache entry
// * obtaining the transaction for a sequence cache entry requires a Push
//
// The following order is taken below:
// 1) collect all intents with sufficiently old txn record
// 2) collect these intents' transactions
// 3) scan the transaction table, collecting abandoned or completed txns
// 4) push all of these transactions (possibly recreating entries)
// 5) resolve all intents (unless the txn is still PENDING), which will recreate
// sequence cache entries (but with the txn timestamp; i.e. likely gc'able)
// 6) scan the sequence table for old entries
// 7) push these transactions (again, recreating txn entries).
// 8) send a GCRequest.
func (gcq *gcQueue) process(now roachpb.Timestamp, repl *Replica,
sysCfg config.SystemConfig) error {
snap := repl.store.Engine().NewSnapshot()
desc := repl.Desc()
iter := newReplicaDataIterator(desc, snap, true /* replicatedOnly */)
defer iter.Close()
defer snap.Close()
// Lookup the GC policy for the zone containing this key range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return util.Errorf("could not find zone config for range %s: %s", repl, err)
}
gc := engine.NewGarbageCollector(now, zone.GC)
// Compute intent expiration (intent age at which we attempt to resolve).
intentExp := now
intentExp.WallTime -= intentAgeThreshold.Nanoseconds()
txnExp := now
txnExp.WallTime -= txnCleanupThreshold.Nanoseconds()
gcArgs := &roachpb.GCRequest{}
// TODO(tschottdorf): This is one of these instances in which we want
// to be more careful that the request ends up on the correct Replica,
// and we might have to worry about mixing range-local and global keys
// in a batch which might end up spanning Ranges by the time it executes.
gcArgs.Key = desc.StartKey.AsRawKey()
gcArgs.EndKey = desc.EndKey.AsRawKey()
var expBaseKey roachpb.Key
var keys []engine.MVCCKey
var vals [][]byte
// Maps from txn ID to txn and intent key slice.
txnMap := map[uuid.UUID]*roachpb.Transaction{}
intentSpanMap := map[uuid.UUID][]roachpb.Span{}
// processKeysAndValues is invoked with each key and its set of
// values. Intents older than the intent age threshold are sent for
// resolution and values after the MVCC metadata, and possible
// intent, are sent for garbage collection.
var intentCount int
processKeysAndValues := func() {
// If there's more than a single value for the key, possibly send for GC.
if len(keys) > 1 {
meta := &engine.MVCCMetadata{}
if err := proto.Unmarshal(vals[0], meta); err != nil {
log.Errorf("unable to unmarshal MVCC metadata for key %q: %s", keys[0], err)
} else {
// In the event that there's an active intent, send for
// intent resolution if older than the threshold.
startIdx := 1
if meta.Txn != nil {
// Keep track of intent to resolve if older than the intent
// expiration threshold.
if meta.Timestamp.Less(intentExp) {
txnID := *meta.Txn.ID
txn := &roachpb.Transaction{
TxnMeta: *meta.Txn,
}
txnMap[txnID] = txn
intentCount++
intentSpanMap[txnID] = append(intentSpanMap[txnID], roachpb.Span{Key: expBaseKey})
}
// With an active intent, GC ignores MVCC metadata & intent value.
startIdx = 2
}
// See if any values may be GC'd.
if gcTS := gc.Filter(keys[startIdx:], vals[startIdx:]); !gcTS.Equal(roachpb.ZeroTimestamp) {
// TODO(spencer): need to split the requests up into
// multiple requests in the event that more than X keys
// are added to the request.
gcArgs.Keys = append(gcArgs.Keys, roachpb.GCRequest_GCKey{Key: expBaseKey, Timestamp: gcTS})
}
}
}
}
// Iterate through the keys and values of this replica's range.
for ; iter.Valid(); iter.Next() {
iterKey := iter.Key()
if !iterKey.IsValue() || !iterKey.Key.Equal(expBaseKey) {
// Moving to the next key (& values).
processKeysAndValues()
expBaseKey = iterKey.Key
if !iterKey.IsValue() {
keys = []engine.MVCCKey{iter.Key()}
vals = [][]byte{iter.Value()}
continue
}
// An implicit metadata.
keys = []engine.MVCCKey{engine.MakeMVCCMetadataKey(iterKey.Key)}
// A nil value for the encoded MVCCMetadata. This will unmarshal to an
// empty MVCCMetadata which is sufficient for processKeysAndValues to
// determine that there is no intent.
vals = [][]byte{nil}
}
keys = append(keys, iter.Key())
vals = append(vals, iter.Value())
}
if iter.Error() != nil {
return iter.Error()
}
// Handle last collected set of keys/vals.
processKeysAndValues()
gcq.eventLog.Infof(true, "assembled %d transactions from %d old intents; found %d gc'able keys", len(txnMap), intentCount, len(gcArgs.Keys))
txnKeys, err := gcq.processTransactionTable(repl, txnMap, txnExp)
if err != nil {
return err
}
// From now on, all newly added keys are range-local.
// TODO(tschottdorf): Might need to use two requests at some point since we
// hard-coded the full non-local key range in the header, but that does
// not take into account the range-local keys. It will be OK as long as
// we send directly to the Replica, though.
gcArgs.Keys = append(gcArgs.Keys, txnKeys...)
// Process push transactions in parallel.
var wg sync.WaitGroup
gcq.eventLog.Infof(true, "pushing %d txns", len(txnMap))
for _, txn := range txnMap {
if txn.Status != roachpb.PENDING {
continue
}
wg.Add(1)
go gcq.pushTxn(repl, now, txn, roachpb.PUSH_ABORT, &wg)
}
wg.Wait()
// Resolve all intents.
var intents []roachpb.Intent
for txnID, txn := range txnMap {
if txn.Status != roachpb.PENDING {
for _, intent := range intentSpanMap[txnID] {
intents = append(intents, roachpb.Intent{Span: intent, Status: txn.Status, Txn: txn.TxnMeta})
}
}
}
gcq.eventLog.Infof(true, "resolving %d intents", len(intents))
if pErr := repl.store.intentResolver.resolveIntents(repl.context(), repl, intents,
true /* wait */, false /* !poison */); pErr != nil {
return pErr.GoError()
}
// Deal with any leftover sequence cache keys. There shouldn't be many of
// them.
leftoverSeqCacheKeys := gcq.processSequenceCache(repl, now, txnExp, txnMap)
gcq.eventLog.Infof(true, "collected %d leftover sequence cache keys", len(leftoverSeqCacheKeys))
gcArgs.Keys = append(gcArgs.Keys, leftoverSeqCacheKeys...)
gcq.eventLog.Infof(true, "sending gc request for %d keys", len(gcArgs.Keys))
var ba roachpb.BatchRequest
// Technically not needed since we're talking directly to the Range.
ba.RangeID = desc.RangeID
ba.Timestamp = now
ba.Add(gcArgs)
if _, pErr := repl.Send(repl.context(), ba); pErr != nil {
return pErr.GoError()
}
return nil
}
// process iterates through all keys in a replica's range, calling the garbage
// collector for each key and associated set of values. GC'd keys are batched
// into GC calls. Extant intents are resolved if intents are older than
// intentAgeThreshold.
func (gcq *gcQueue) process(now roachpb.Timestamp, repl *Replica,
sysCfg *config.SystemConfig) error {
snap := repl.store.Engine().NewSnapshot()
desc := repl.Desc()
iter := newReplicaDataIterator(desc, snap)
defer iter.Close()
defer snap.Close()
// Lookup the GC policy for the zone containing this key range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return fmt.Errorf("could not find GC policy for range %s: %s", repl, err)
}
policy := zone.GC
gcMeta := roachpb.NewGCMetadata(now.WallTime)
gc := engine.NewGarbageCollector(now, *policy)
// Compute intent expiration (intent age at which we attempt to resolve).
intentExp := now
intentExp.WallTime -= intentAgeThreshold.Nanoseconds()
txnExp := now
txnExp.WallTime -= txnCleanupThreshold.Nanoseconds()
gcArgs := &roachpb.GCRequest{}
// TODO(tschottdorf): This is one of these instances in which we want
// to be more careful that the request ends up on the correct Replica,
// and we might have to worry about mixing range-local and global keys
// in a batch which might end up spanning Ranges by the time it executes.
gcArgs.Key = desc.StartKey.AsRawKey()
gcArgs.EndKey = desc.EndKey.AsRawKey()
var expBaseKey roachpb.Key
var keys []engine.MVCCKey
var vals [][]byte
// Maps from txn ID to txn and intent key slice.
txnMap := map[string]*roachpb.Transaction{}
intentSpanMap := map[string][]roachpb.Span{}
// processKeysAndValues is invoked with each key and its set of
// values. Intents older than the intent age threshold are sent for
// resolution and values after the MVCC metadata, and possible
// intent, are sent for garbage collection.
processKeysAndValues := func() {
// If there's more than a single value for the key, possibly send for GC.
if len(keys) > 1 {
meta := &engine.MVCCMetadata{}
if err := proto.Unmarshal(vals[0], meta); err != nil {
log.Errorf("unable to unmarshal MVCC metadata for key %q: %s", keys[0], err)
} else {
// In the event that there's an active intent, send for
// intent resolution if older than the threshold.
startIdx := 1
if meta.Txn != nil {
// Keep track of intent to resolve if older than the intent
// expiration threshold.
if meta.Timestamp.Less(intentExp) {
id := string(meta.Txn.ID)
txnMap[id] = meta.Txn
intentSpanMap[id] = append(intentSpanMap[id], roachpb.Span{Key: expBaseKey})
}
// With an active intent, GC ignores MVCC metadata & intent value.
startIdx = 2
}
// See if any values may be GC'd.
if gcTS := gc.Filter(keys[startIdx:], vals[startIdx:]); !gcTS.Equal(roachpb.ZeroTimestamp) {
// TODO(spencer): need to split the requests up into
// multiple requests in the event that more than X keys
// are added to the request.
gcArgs.Keys = append(gcArgs.Keys, roachpb.GCRequest_GCKey{Key: expBaseKey, Timestamp: gcTS})
}
}
}
}
// Iterate through the keys and values of this replica's range.
for ; iter.Valid(); iter.Next() {
baseKey, ts, isValue, err := engine.MVCCDecodeKey(iter.Key())
if err != nil {
log.Errorf("unable to decode MVCC key: %q: %v", iter.Key(), err)
continue
}
if !isValue {
// Moving to the next key (& values).
processKeysAndValues()
expBaseKey = baseKey
keys = []engine.MVCCKey{iter.Key()}
vals = [][]byte{iter.Value()}
} else {
if !baseKey.Equal(expBaseKey) {
log.Errorf("unexpectedly found a value for %q with ts=%s; expected key %q", baseKey, ts, expBaseKey)
continue
}
keys = append(keys, iter.Key())
vals = append(vals, iter.Value())
}
}
if iter.Error() != nil {
return iter.Error()
}
// Handle last collected set of keys/vals.
processKeysAndValues()
txnKeys, err := processTransactionTable(repl, txnMap, txnExp)
if err != nil {
return err
}
// From now on, all newly added keys are range-local.
// TODO(tschottdorf): Might need to use two requests at some point since we
// hard-coded the full non-local key range in the header, but that does
// not take into account the range-local keys. It will be OK as long as
// we send directly to the Replica, though.
gcArgs.Keys = append(gcArgs.Keys, txnKeys...)
// Process push transactions in parallel.
var wg sync.WaitGroup
for _, txn := range txnMap {
if txn.Status != roachpb.PENDING {
continue
}
wg.Add(1)
go pushTxn(repl, now, txn, roachpb.ABORT_TXN, &wg)
}
wg.Wait()
// Resolve all intents.
var intents []roachpb.Intent
for id, txn := range txnMap {
if txn.Status != roachpb.PENDING {
for _, intent := range intentSpanMap[id] {
intents = append(intents, roachpb.Intent{Span: intent, Txn: *txn})
}
}
}
if err := repl.resolveIntents(repl.context(), intents, true /* wait */, false /* !poison */); err != nil {
return err
}
// Deal with any leftover sequence cache keys. There shouldn't be many of
// them.
gcArgs.Keys = append(gcArgs.Keys, processSequenceCache(repl, now, txnExp, txnMap)...)
// Send GC request through range.
gcArgs.GCMeta = *gcMeta
var ba roachpb.BatchRequest
// Technically not needed since we're talking directly to the Range.
ba.RangeID = desc.RangeID
ba.Timestamp = now
ba.Add(gcArgs)
if _, pErr := repl.Send(repl.context(), ba); pErr != nil {
return pErr.GoError()
}
// Store current timestamp as last verification for this replica, as
// we've just successfully scanned.
if err := repl.SetLastVerificationTimestamp(now); err != nil {
log.Errorf("failed to set last verification timestamp for replica %s: %s", repl, err)
}
return nil
}
// InitOrJoinRequest executes a RequestLease command asynchronously and returns a
// channel on which the result will be posted. If there's already a request in
// progress, we join in waiting for the results of that request.
// It is an error to call InitOrJoinRequest() while a request is in progress
// naming another replica as lease holder.
//
// replica is used to schedule and execute async work (proposing a RequestLease
// command). replica.mu is locked when delivering results, so calls from the
// replica happen either before or after a result for a pending request has
// happened.
//
// transfer needs to be set if the request represents a lease transfer (as
// opposed to an extension, or acquiring the lease when none is held).
//
// Note: Once this function gets a context to be used for cancellation, instead
// of replica.store.Stopper().ShouldQuiesce(), care will be needed for cancelling
// the Raft command, similar to replica.addWriteCmd.
func (p *pendingLeaseRequest) InitOrJoinRequest(
replica *Replica,
nextLeaseHolder roachpb.ReplicaDescriptor,
timestamp hlc.Timestamp,
startKey roachpb.Key,
transfer bool,
) <-chan *roachpb.Error {
if nextLease := p.RequestPending(); nextLease != nil {
if nextLease.Replica.ReplicaID == nextLeaseHolder.ReplicaID {
// Join a pending request asking for the same replica to become lease
// holder.
return p.JoinRequest()
}
llChan := make(chan *roachpb.Error, 1)
// We can't join the request in progress.
llChan <- roachpb.NewErrorf("request for different replica in progress "+
"(requesting: %+v, in progress: %+v)",
nextLeaseHolder.ReplicaID, nextLease.Replica.ReplicaID)
return llChan
}
llChan := make(chan *roachpb.Error, 1)
// No request in progress. Let's propose a Lease command asynchronously.
// TODO(tschottdorf): get duration from configuration, either as a
// config flag or, later, dynamically adjusted.
startStasis := timestamp.Add(int64(replica.store.ctx.rangeLeaseActiveDuration), 0)
expiration := startStasis.Add(int64(replica.store.Clock().MaxOffset()), 0)
reqSpan := roachpb.Span{
Key: startKey,
}
var leaseReq roachpb.Request
reqLease := roachpb.Lease{
Start: timestamp,
StartStasis: startStasis,
Expiration: expiration,
Replica: nextLeaseHolder,
}
if transfer {
leaseReq = &roachpb.TransferLeaseRequest{
Span: reqSpan,
Lease: reqLease,
}
} else {
leaseReq = &roachpb.RequestLeaseRequest{
Span: reqSpan,
Lease: reqLease,
}
}
if replica.store.Stopper().RunAsyncTask(func() {
// Propose a RequestLease command and wait for it to apply.
var execPErr *roachpb.Error
ba := roachpb.BatchRequest{}
ba.Timestamp = replica.store.Clock().Now()
ba.RangeID = replica.RangeID
ba.Add(leaseReq)
// Send lease request directly to raft in order to skip unnecessary
// checks from normal request machinery, (e.g. the command queue).
// Note that the command itself isn't traced, but usually the caller
// waiting for the result has an active Trace.
ch, _, err := replica.proposeRaftCommand(
replica.context(context.Background()), ba)
if err != nil {
execPErr = roachpb.NewError(err)
} else {
// If the command was committed, wait for the range to apply it.
select {
case c := <-ch:
if c.Err != nil {
if log.V(1) {
log.Infof("failed to acquire lease for replica %s: %s", replica.store, c.Err)
}
execPErr = c.Err
}
case <-replica.store.Stopper().ShouldQuiesce():
execPErr = roachpb.NewError(
replica.newNotLeaseHolderError(nil, replica.store.StoreID(), replica.Desc()))
}
}
// Send result of lease to all waiter channels.
replica.mu.Lock()
defer replica.mu.Unlock()
for i, llChan := range p.llChans {
// Don't send the same pErr object twice; this can lead to races. We could
// clone every time but it's more efficient to send pErr itself to one of
// the channels (the last one; if we send it earlier the race can still
// happen).
if i == len(p.llChans)-1 {
llChan <- execPErr
} else {
llChan <- protoutil.Clone(execPErr).(*roachpb.Error) // works with `nil`
}
}
p.llChans = p.llChans[:0]
p.nextLease = roachpb.Lease{}
}) != nil {
// We failed to start the asynchronous task. Send a blank NotLeaseHolderError
// back to indicate that we have no idea who the range lease holder might
// be; we've withdrawn from active duty.
llChan <- roachpb.NewError(
replica.newNotLeaseHolderError(nil, replica.store.StoreID(), replica.mu.state.Desc))
return llChan
}
p.llChans = append(p.llChans, llChan)
p.nextLease = reqLease
return llChan
}
// process iterates through all keys in a replica's range, calling the garbage
// collector for each key and associated set of values. GC'd keys are batched
// into GC calls. Extant intents are resolved if intents are older than
// intentAgeThreshold.
func (gcq *gcQueue) process(now roachpb.Timestamp, repl *Replica,
sysCfg *config.SystemConfig) error {
snap := repl.store.Engine().NewSnapshot()
desc := repl.Desc()
iter := newReplicaDataIterator(desc, snap)
defer iter.Close()
defer snap.Close()
// Lookup the GC policy for the zone containing this key range.
zone, err := sysCfg.GetZoneConfigForKey(desc.StartKey)
if err != nil {
return fmt.Errorf("could not find GC policy for range %s: %s", repl, err)
}
policy := zone.GC
gcMeta := roachpb.NewGCMetadata(now.WallTime)
gc := engine.NewGarbageCollector(now, *policy)
// Compute intent expiration (intent age at which we attempt to resolve).
intentExp := now
intentExp.WallTime -= intentAgeThreshold.Nanoseconds()
// TODO(tschottdorf): execution will use a leader-assigned local
// timestamp to compute intent age. While this should be fine, could
// consider adding a Now timestamp to GCRequest which would be used
// instead.
gcArgs := &roachpb.GCRequest{}
var mu sync.Mutex
var oldestIntentNanos int64 = math.MaxInt64
var expBaseKey roachpb.Key
var keys []engine.MVCCKey
var vals [][]byte
// Maps from txn ID to txn and intent key slice.
txnMap := map[string]*roachpb.Transaction{}
intentSpanMap := map[string][]roachpb.Span{}
// updateOldestIntent atomically updates the oldest intent.
updateOldestIntent := func(intentNanos int64) {
mu.Lock()
defer mu.Unlock()
if intentNanos < oldestIntentNanos {
oldestIntentNanos = intentNanos
}
}
// processKeysAndValues is invoked with each key and its set of
// values. Intents older than the intent age threshold are sent for
// resolution and values after the MVCC metadata, and possible
// intent, are sent for garbage collection.
processKeysAndValues := func() {
// If there's more than a single value for the key, possibly send for GC.
if len(keys) > 1 {
meta := &engine.MVCCMetadata{}
if err := proto.Unmarshal(vals[0], meta); err != nil {
log.Errorf("unable to unmarshal MVCC metadata for key %q: %s", keys[0], err)
} else {
// In the event that there's an active intent, send for
// intent resolution if older than the threshold.
startIdx := 1
if meta.Txn != nil {
// Keep track of intent to resolve if older than the intent
// expiration threshold.
if meta.Timestamp.Less(intentExp) {
id := string(meta.Txn.ID)
txnMap[id] = meta.Txn
intentSpanMap[id] = append(intentSpanMap[id], roachpb.Span{Key: expBaseKey})
} else {
updateOldestIntent(meta.Txn.OrigTimestamp.WallTime)
}
// With an active intent, GC ignores MVCC metadata & intent value.
startIdx = 2
}
// See if any values may be GC'd.
if gcTS := gc.Filter(keys[startIdx:], vals[startIdx:]); !gcTS.Equal(roachpb.ZeroTimestamp) {
// TODO(spencer): need to split the requests up into
// multiple requests in the event that more than X keys
// are added to the request.
gcArgs.Keys = append(gcArgs.Keys, roachpb.GCRequest_GCKey{Key: expBaseKey, Timestamp: gcTS})
}
}
}
}
// Iterate through the keys and values of this replica's range.
for ; iter.Valid(); iter.Next() {
baseKey, ts, isValue, err := engine.MVCCDecodeKey(iter.Key())
if err != nil {
log.Errorf("unable to decode MVCC key: %q: %v", iter.Key(), err)
continue
}
if !isValue {
// Moving to the next key (& values).
processKeysAndValues()
expBaseKey = baseKey
keys = []engine.MVCCKey{iter.Key()}
vals = [][]byte{iter.Value()}
} else {
if !baseKey.Equal(expBaseKey) {
log.Errorf("unexpectedly found a value for %q with ts=%s; expected key %q", baseKey, ts, expBaseKey)
continue
}
keys = append(keys, iter.Key())
vals = append(vals, iter.Value())
}
}
if iter.Error() != nil {
return iter.Error()
}
// Handle last collected set of keys/vals.
processKeysAndValues()
// Process push transactions in parallel.
var wg sync.WaitGroup
for _, txn := range txnMap {
wg.Add(1)
go gcq.pushTxn(repl, now, txn, updateOldestIntent, &wg)
}
wg.Wait()
// Resolve all intents.
var intents []roachpb.Intent
for id, txn := range txnMap {
if txn.Status != roachpb.PENDING {
for _, intent := range intentSpanMap[id] {
intents = append(intents, roachpb.Intent{Span: intent, Txn: *txn})
}
}
}
done := true
if len(intents) > 0 {
done = false
repl.resolveIntents(repl.context(), intents)
}
// Set start and end keys.
if len(gcArgs.Keys) > 0 {
done = false
gcArgs.Key = gcArgs.Keys[0].Key
gcArgs.EndKey = gcArgs.Keys[len(gcArgs.Keys)-1].Key.Next()
}
if done {
return nil
}
// Send GC request through range.
gcMeta.OldestIntentNanos = proto.Int64(oldestIntentNanos)
gcArgs.GCMeta = *gcMeta
var ba roachpb.BatchRequest
// Technically not needed since we're talking directly to the Range.
ba.RangeID = desc.RangeID
ba.Add(gcArgs)
if _, pErr := repl.Send(repl.context(), ba); pErr != nil {
return pErr.GoError()
}
// Store current timestamp as last verification for this replica, as
// we've just successfully scanned.
if err := repl.SetLastVerificationTimestamp(now); err != nil {
log.Errorf("failed to set last verification timestamp for replica %s: %s", repl, err)
}
return nil
}