// UpdateDeadlineMaybe sets the transactions deadline to the lower of the
// current one (if any) and the passed value.
func (txn *Txn) UpdateDeadlineMaybe(deadline hlc.Timestamp) bool {
if txn.deadline == nil || deadline.Less(*txn.deadline) {
txn.deadline = &deadline
return true
}
return false
}
// UpdateObservedTimestamp stores a timestamp off a node's clock for future
// operations in the transaction. When multiple calls are made for a single
// nodeID, the lowest timestamp prevails.
func (t *Transaction) UpdateObservedTimestamp(nodeID NodeID, maxTS hlc.Timestamp) {
if t.ObservedTimestamps == nil {
t.ObservedTimestamps = make(map[NodeID]hlc.Timestamp)
}
if ts, ok := t.ObservedTimestamps[nodeID]; !ok || maxTS.Less(ts) {
t.ObservedTimestamps[nodeID] = maxTS
}
}
// shouldQueueAgain is a helper function to determine whether the
// replica should be queued according to the current time, the last
// time the replica was processed, and the minimum interval between
// successive processing. Specifying minInterval=0 queues all replicas.
// Returns a bool for whether to queue as well as a priority based
// on how long it's been since last processed.
func shouldQueueAgain(now, last hlc.Timestamp, minInterval time.Duration) (bool, float64) {
if minInterval == 0 || last == hlc.ZeroTimestamp {
return true, 0
}
if diff := now.GoTime().Sub(last.GoTime()); diff >= minInterval {
priority := float64(1)
// If there's a non-zero last processed timestamp, adjust the
// priority by a multiple of how long it's been since the last
// time this replica was processed.
if last != hlc.ZeroTimestamp {
priority = float64(diff.Nanoseconds()) / float64(minInterval.Nanoseconds())
}
return true, priority
}
return false, 0
}
func replicaGCShouldQueueImpl(
now, lastCheck, lastActivity hlc.Timestamp, isCandidate bool,
) (bool, float64) {
timeout := ReplicaGCQueueInactivityThreshold
priority := replicaGCPriorityDefault
if isCandidate {
// If the range is a candidate (which happens if its former replica set
// ignores it), let it expire much earlier.
timeout = ReplicaGCQueueCandidateTimeout
priority = replicaGCPriorityCandidate
} else if now.Less(lastCheck.Add(ReplicaGCQueueInactivityThreshold.Nanoseconds(), 0)) {
// Return false immediately if the previous check was less than the
// check interval in the past. Note that we don't do this if the
// replica is in candidate state, in which case we want to be more
// aggressive - a failed rebalance attempt could have checked this
// range, and candidate state suggests that a retry succeeded. See
// #7489.
return false, 0
}
shouldQ := lastActivity.Add(timeout.Nanoseconds(), 0).Less(now)
if !shouldQ {
return false, 0
}
return shouldQ, priority
}
// NewTransaction creates a new transaction. The transaction key is
// composed using the specified baseKey (for locality with data
// affected by the transaction) and a random ID to guarantee
// uniqueness. The specified user-level priority is combined with a
// randomly chosen value to yield a final priority, used to settle
// write conflicts in a way that avoids starvation of long-running
// transactions (see Replica.PushTxn).
func NewTransaction(
name string,
baseKey Key,
userPriority UserPriority,
isolation enginepb.IsolationType,
now hlc.Timestamp,
maxOffset int64,
) *Transaction {
u := uuid.MakeV4()
return &Transaction{
TxnMeta: enginepb.TxnMeta{
Key: baseKey,
ID: &u,
Isolation: isolation,
Timestamp: now,
Priority: MakePriority(userPriority),
Sequence: 1,
},
Name: name,
OrigTimestamp: now,
MaxTimestamp: now.Add(maxOffset, 0),
}
}
// leaseStatus returns lease status. If the lease is epoch-based,
// the liveness field will be set to the liveness used to compute
// its state, unless state == leaseError.
//
// - The lease is considered valid if the timestamp is covered by the
// supplied lease. This is determined differently depending on the
// lease properties. For expiration-based leases, the timestamp is
// covered if it's less than the expiration (minus the maximum
// clock offset). For epoch-based "node liveness" leases, the lease
// epoch must match the owner node's liveness epoch -AND- the
// timestamp must be within the node's liveness expiration (also
// minus the maximum clock offset).
//
// To be valid, a lease which contains a valid ProposedTS must have
// a proposed timestamp greater than the minimum proposed timestamp,
// which prevents a restarted process from serving commands, since
// the command queue has been wiped through the restart.
//
// - The lease is considered in stasis if the timestamp is within the
// maximum clock offset window of the lease expiration.
//
// - The lease is considered expired in all other cases.
//
// The maximum clock offset must always be taken into consideration to
// avoid a failure of linearizability on a single register during
// lease changes. Without that stasis period, the following could
// occur:
//
// * a range lease gets committed on the new lease holder (but not the old).
// * client proposes and commits a write on new lease holder (with a
// timestamp just greater than the expiration of the old lease).
// * client tries to read what it wrote, but hits a slow coordinator
// (which assigns a timestamp covered by the old lease).
// * the read is served by the old lease holder (which has not
// processed the change in lease holdership).
// * the client fails to read their own write.
func (r *Replica) leaseStatus(
lease *roachpb.Lease, timestamp, minProposedTS hlc.Timestamp,
) LeaseStatus {
status := LeaseStatus{timestamp: timestamp, lease: lease}
if lease == nil {
status.state = leaseExpired
return status
}
var expiration hlc.Timestamp
if lease.Type() == roachpb.LeaseExpiration {
expiration = lease.Expiration
} else {
var err error
status.liveness, err = r.store.cfg.NodeLiveness.GetLiveness(lease.Replica.NodeID)
if err != nil || status.liveness.Epoch < *lease.Epoch {
// If lease validity can't be determined (e.g. gossip is down
// and liveness info isn't available for owner), we can neither
// use the lease nor do we want to attempt to acquire it.
status.state = leaseError
return status
}
if status.liveness.Epoch > *lease.Epoch {
status.state = leaseExpired
return status
}
expiration = status.liveness.Expiration
}
stasis := expiration.Add(-int64(r.store.Clock().MaxOffset()), 0)
if timestamp.Less(stasis) {
status.state = leaseValid
// If the replica owns the lease, additional verify that the lease's
// proposed timestamp is not earlier than the min proposed timestamp.
if lease.Replica.StoreID == r.store.StoreID() &&
lease.ProposedTS != nil && lease.ProposedTS.Less(minProposedTS) {
status.state = leaseProscribed
}
} else if timestamp.Less(expiration) {
status.state = leaseStasis
} else {
status.state = leaseExpired
}
return status
}
// 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, ok := p.RequestPending(); ok {
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.cfg.RangeLeaseActiveDuration), 0)
expiration := startStasis.Add(int64(replica.store.Clock().MaxOffset()), 0)
reqSpan := roachpb.Span{
Key: startKey,
}
var leaseReq roachpb.Request
now := replica.store.Clock().Now()
reqLease := roachpb.Lease{
Start: timestamp,
StartStasis: startStasis,
Expiration: expiration,
Replica: nextLeaseHolder,
ProposedTS: &now,
}
if transfer {
leaseReq = &roachpb.TransferLeaseRequest{
Span: reqSpan,
Lease: reqLease,
}
} else {
leaseReq = &roachpb.RequestLeaseRequest{
Span: reqSpan,
Lease: reqLease,
}
}
if replica.store.Stopper().RunAsyncTask(context.TODO(), func(ctx context.Context) {
ctx = replica.AnnotateCtx(ctx)
// Propose a RequestLease command and wait for it to apply.
ba := roachpb.BatchRequest{}
ba.Timestamp = replica.store.Clock().Now()
ba.RangeID = replica.RangeID
ba.Add(leaseReq)
if log.V(2) {
log.Infof(ctx, "sending lease request %v", leaseReq)
}
_, pErr := replica.Send(ctx, ba)
// We reset our state below regardless of whether we've gotten an error or
// not, but note that an error is ambiguous - there's no guarantee that the
// transfer will not still apply. That's OK, however, as the "in transfer"
// state maintained by the pendingLeaseRequest is not relied on for
// correctness (see replica.mu.minLeaseProposedTS), and resetting the state
// is beneficial as it'll allow the replica to attempt to transfer again or
// extend the existing lease in the future.
// Send result of lease to all waiter channels.
replica.mu.Lock()
defer replica.mu.Unlock()
for _, llChan := range p.llChans {
// Don't send the same transaction object twice; this can lead to races.
if pErr != nil {
pErrClone := *pErr
pErrClone.SetTxn(pErr.GetTxn())
llChan <- &pErrClone
} else {
llChan <- 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(
newNotLeaseHolderError(nil, replica.store.StoreID(), replica.mu.state.Desc))
return llChan
}
p.llChans = append(p.llChans, llChan)
p.nextLease = reqLease
return llChan
}
// 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, ok := p.RequestPending(); ok {
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.cfg.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(context.TODO(), func(ctx context.Context) {
ctx = replica.AnnotateCtx(ctx)
// Propose a RequestLease command and wait for it to apply.
ba := roachpb.BatchRequest{}
ba.Timestamp = replica.store.Clock().Now()
ba.RangeID = replica.RangeID
ba.Add(leaseReq)
if log.V(2) {
log.Infof(ctx, "sending lease request %v", leaseReq)
}
_, pErr := replica.Send(ctx, ba)
// 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 <- pErr
} else {
llChan <- protoutil.Clone(pErr).(*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(
newNotLeaseHolderError(nil, replica.store.StoreID(), replica.mu.state.Desc))
return llChan
}
p.llChans = append(p.llChans, llChan)
p.nextLease = reqLease
return llChan
}
// Covers returns true if the given timestamp can be served by the Lease.
// This is the case if the timestamp precedes the Lease's stasis period.
// Note that the fact that a lease covers a timestamp is not enough for the
// holder of the lease to be able to serve a read with that timestamp;
// pendingLeaderLeaseRequest.TransferInProgress() should also be consulted to
// account for possible lease transfers.
func (l Lease) Covers(timestamp hlc.Timestamp) bool {
return timestamp.Less(l.StartStasis)
}
// RunGC runs garbage collection for the specified descriptor on the provided
// Engine (which is not mutated). It uses the provided functions pushTxnFn and
// resolveIntentsFn to clarify the true status of and clean up after encountered
// transactions. It returns a slice of gc'able keys from the data, transaction,
// and abort spans.
func RunGC(
ctx context.Context,
desc *roachpb.RangeDescriptor,
snap engine.Reader,
now hlc.Timestamp,
policy config.GCPolicy,
pushTxnFn pushFunc,
resolveIntentsFn resolveFunc,
) ([]roachpb.GCRequest_GCKey, GCInfo, error) {
iter := NewReplicaDataIterator(desc, snap, true /* replicatedOnly */)
defer iter.Close()
var infoMu = lockableGCInfo{}
infoMu.Policy = policy
infoMu.Now = now
{
realResolveIntentsFn := resolveIntentsFn
resolveIntentsFn = func(intents []roachpb.Intent, poison bool, wait bool) (err error) {
defer func() {
infoMu.Lock()
infoMu.ResolveTotal += len(intents)
if err == nil {
infoMu.ResolveSuccess += len(intents)
}
infoMu.Unlock()
}()
return realResolveIntentsFn(intents, poison, wait)
}
realPushTxnFn := pushTxnFn
pushTxnFn = func(ts hlc.Timestamp, txn *roachpb.Transaction, typ roachpb.PushTxnType) {
infoMu.Lock()
infoMu.PushTxn++
infoMu.Unlock()
realPushTxnFn(ts, txn, typ)
}
}
// Compute intent expiration (intent age at which we attempt to resolve).
intentExp := now
intentExp.WallTime -= intentAgeThreshold.Nanoseconds()
txnExp := now
txnExp.WallTime -= txnCleanupThreshold.Nanoseconds()
abortSpanGCThreshold := now.Add(-int64(abortCacheAgeThreshold), 0)
gc := engine.MakeGarbageCollector(now, policy)
infoMu.Threshold = gc.Threshold
infoMu.TxnSpanGCThreshold = txnExp
var gcKeys []roachpb.GCRequest_GCKey
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.
processKeysAndValues := func() {
// If there's more than a single value for the key, possibly send for GC.
if len(keys) > 1 {
meta := &enginepb.MVCCMetadata{}
if err := proto.Unmarshal(vals[0], meta); err != nil {
log.Errorf(ctx, "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
infoMu.IntentsConsidered++
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(hlc.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.
gcKeys = append(gcKeys, 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 nil, GCInfo{}, iter.Error()
}
// Handle last collected set of keys/vals.
processKeysAndValues()
infoMu.IntentTxns = len(txnMap)
infoMu.NumKeysAffected = len(gcKeys)
txnKeys, err := processTransactionTable(ctx, snap, desc, txnMap, txnExp, &infoMu, resolveIntentsFn)
if err != nil {
return nil, GCInfo{}, 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.
gcKeys = append(gcKeys, txnKeys...)
// Process push transactions in parallel.
var wg sync.WaitGroup
sem := make(chan struct{}, gcTaskLimit)
for _, txn := range txnMap {
if txn.Status != roachpb.PENDING {
continue
}
wg.Add(1)
sem <- struct{}{}
// Avoid passing loop variable into closure.
txnCopy := txn
go func() {
defer func() {
<-sem
wg.Done()
}()
pushTxnFn(now, txnCopy, roachpb.PUSH_ABORT)
}()
}
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})
}
}
}
if err := resolveIntentsFn(intents, true /* wait */, false /* !poison */); err != nil {
return nil, GCInfo{}, err
}
// Clean up the abort cache.
gcKeys = append(gcKeys, processAbortCache(
ctx, snap, desc.RangeID, abortSpanGCThreshold, &infoMu, pushTxnFn)...)
return gcKeys, infoMu.GCInfo, nil
}
// selectEventTimestamp selects a timestamp for this log message. If the
// transaction this event is being written in has a non-zero timestamp, then that
// timestamp should be used; otherwise, the store's physical clock is used.
// This helps with testing; in normal usage, the logging of an event will never
// be the first action in the transaction, and thus the transaction will have an
// assigned database timestamp. However, in the case of our tests log events
// *are* the first action in a transaction, and we must elect to use the store's
// physical time instead.
func (ev EventLogger) selectEventTimestamp(input hlc.Timestamp) time.Time {
if input == hlc.ZeroTimestamp {
return ev.LeaseManager.clock.PhysicalTime()
}
return input.GoTime()
}
// TestTxnCoordSenderHeartbeat verifies periodic heartbeat of the
// transaction record.
func TestTxnCoordSenderHeartbeat(t *testing.T) {
defer leaktest.AfterTest(t)()
s, sender := createTestDB(t)
defer s.Stop()
defer teardownHeartbeats(sender)
// Set heartbeat interval to 1ms for testing.
sender.heartbeatInterval = 1 * time.Millisecond
initialTxn := client.NewTxn(context.Background(), *s.DB)
if err := initialTxn.Put(roachpb.Key("a"), []byte("value")); err != nil {
t.Fatal(err)
}
// Verify 3 heartbeats.
var heartbeatTS hlc.Timestamp
for i := 0; i < 3; i++ {
util.SucceedsSoon(t, func() error {
txn, pErr := getTxn(sender, &initialTxn.Proto)
if pErr != nil {
t.Fatal(pErr)
}
// Advance clock by 1ns.
// Locking the TxnCoordSender to prevent a data race.
sender.Lock()
s.Manual.Increment(1)
sender.Unlock()
if txn.LastHeartbeat != nil && heartbeatTS.Less(*txn.LastHeartbeat) {
heartbeatTS = *txn.LastHeartbeat
return nil
}
return errors.Errorf("expected heartbeat")
})
}
// Sneakily send an ABORT right to DistSender (bypassing TxnCoordSender).
{
var ba roachpb.BatchRequest
ba.Add(&roachpb.EndTransactionRequest{
Commit: false,
Span: roachpb.Span{Key: initialTxn.Proto.Key},
})
ba.Txn = &initialTxn.Proto
if _, pErr := sender.wrapped.Send(context.Background(), ba); pErr != nil {
t.Fatal(pErr)
}
}
util.SucceedsSoon(t, func() error {
sender.Lock()
defer sender.Unlock()
if txnMeta, ok := sender.txns[*initialTxn.Proto.ID]; !ok {
t.Fatal("transaction unregistered prematurely")
} else if txnMeta.txn.Status != roachpb.ABORTED {
return fmt.Errorf("transaction is not aborted")
}
return nil
})
// Trying to do something else should give us a TransactionAbortedError.
_, err := initialTxn.Get("a")
assertTransactionAbortedError(t, err)
}
// isAsOf analyzes a select statement to bypass the logic in newPlan(),
// since that requires the transaction to be started already. If the returned
// timestamp is not nil, it is the timestamp to which a transaction should
// be set.
//
// max is a lower bound on what the transaction's timestamp will be. Used to
// check that the user didn't specify a timestamp in the future.
func isAsOf(planMaker *planner, stmt parser.Statement, max hlc.Timestamp) (*hlc.Timestamp, error) {
s, ok := stmt.(*parser.Select)
if !ok {
return nil, nil
}
sc, ok := s.Select.(*parser.SelectClause)
if !ok {
return nil, nil
}
if sc.From == nil || sc.From.AsOf.Expr == nil {
return nil, nil
}
te, err := sc.From.AsOf.Expr.TypeCheck(nil, parser.TypeString)
if err != nil {
return nil, err
}
d, err := te.Eval(&planMaker.evalCtx)
if err != nil {
return nil, err
}
var ts hlc.Timestamp
switch d := d.(type) {
case *parser.DString:
// Allow nanosecond precision because the timestamp is only used by the
// system and won't be returned to the user over pgwire.
dt, err := parser.ParseDTimestamp(string(*d), time.Nanosecond)
if err != nil {
return nil, err
}
ts.WallTime = dt.Time.UnixNano()
case *parser.DInt:
ts.WallTime = int64(*d)
case *parser.DDecimal:
// Format the decimal into a string and split on `.` to extract the nanosecond
// walltime and logical tick parts.
s := d.String()
parts := strings.SplitN(s, ".", 2)
nanos, err := strconv.ParseInt(parts[0], 10, 64)
if err != nil {
return nil, errors.Wrap(err, "parse AS OF SYSTEM TIME argument")
}
var logical int64
if len(parts) > 1 {
// logicalLength is the number of decimal digits expected in the
// logical part to the right of the decimal. See the implementation of
// cluster_logical_timestamp().
const logicalLength = 10
p := parts[1]
if lp := len(p); lp > logicalLength {
return nil, errors.Errorf("bad AS OF SYSTEM TIME argument: logical part has too many digits")
} else if lp < logicalLength {
p += strings.Repeat("0", logicalLength-lp)
}
logical, err = strconv.ParseInt(p, 10, 32)
if err != nil {
return nil, errors.Wrap(err, "parse AS OF SYSTEM TIME argument")
}
}
ts.WallTime = nanos
ts.Logical = int32(logical)
default:
return nil, fmt.Errorf("unexpected AS OF SYSTEM TIME argument: %s (%T)", d.ResolvedType(), d)
}
if max.Less(ts) {
return nil, fmt.Errorf("cannot specify timestamp in the future")
}
return &ts, nil
}
func TestBatchBuilderStress(t *testing.T) {
defer leaktest.AfterTest(t)()
stopper := stop.NewStopper()
defer stopper.Stop()
e := NewInMem(roachpb.Attributes{}, 1<<20)
stopper.AddCloser(e)
rng, _ := randutil.NewPseudoRand()
for i := 0; i < 1000; i++ {
count := 1 + rng.Intn(1000)
func() {
batch := e.NewBatch().(*rocksDBBatch)
// Ensure that, even though we reach into the batch's internals with
// dbPut etc, asking for the batch's Repr will get data from C++ and
// not its unused builder.
batch.flushes++
defer batch.Close()
builder := &RocksDBBatchBuilder{}
for j := 0; j < count; j++ {
var ts hlc.Timestamp
if rng.Float32() <= 0.9 {
// Give 90% of keys timestamps.
ts.WallTime = rng.Int63()
if rng.Float32() <= 0.1 {
// Give 10% of timestamps a non-zero logical component.
ts.Logical = rng.Int31()
}
}
key := MVCCKey{
Key: []byte(fmt.Sprintf("%d", rng.Intn(10000))),
Timestamp: ts,
}
// Generate a random mixture of puts, deletes and merges.
switch rng.Intn(3) {
case 0:
if err := dbPut(batch.batch, key, []byte("value")); err != nil {
t.Fatal(err)
}
builder.Put(key, []byte("value"))
case 1:
if err := dbClear(batch.batch, key); err != nil {
t.Fatal(err)
}
builder.Clear(key)
case 2:
if err := dbMerge(batch.batch, key, appender("bar")); err != nil {
t.Fatal(err)
}
builder.Merge(key, appender("bar"))
}
}
batchRepr := batch.Repr()
builderRepr := builder.Finish()
if !bytes.Equal(batchRepr, builderRepr) {
t.Fatalf("expected [% x], but got [% x]", batchRepr, builderRepr)
}
}()
}
}
func (l *Liveness) isLive(now hlc.Timestamp, maxOffset time.Duration) bool {
expiration := l.Expiration.Add(-maxOffset.Nanoseconds(), 0)
return now.Less(expiration)
}
// add the specified timestamp to the cache as covering the range of
// keys from start to end. If end is nil, the range covers the start
// key only. txnID is nil for no transaction. readTSCache specifies
// whether the command adding this timestamp should update the read
// timestamp; false to update the write timestamp cache.
func (tc *timestampCache) add(
start, end roachpb.Key, timestamp hlc.Timestamp, txnID *uuid.UUID, readTSCache bool,
) {
// This gives us a memory-efficient end key if end is empty.
if len(end) == 0 {
end = start.Next()
start = end[:len(start)]
}
tc.latest.Forward(timestamp)
// Only add to the cache if the timestamp is more recent than the
// low water mark.
if tc.lowWater.Less(timestamp) {
tcache := tc.wCache
if readTSCache {
tcache = tc.rCache
}
addRange := func(r interval.Range) {
value := cacheValue{timestamp: timestamp, txnID: txnID}
key := tcache.MakeKey(r.Start, r.End)
entry := makeCacheEntry(key, value)
tcache.AddEntry(entry)
}
r := interval.Range{
Start: interval.Comparable(start),
End: interval.Comparable(end),
}
// Check existing, overlapping entries and truncate/split/remove if
// superseded and in the past. If existing entries are in the future,
// subtract from the range/ranges that need to be added to cache.
for _, entry := range tcache.GetOverlaps(r.Start, r.End) {
cv := entry.Value.(*cacheValue)
key := entry.Key.(*cache.IntervalKey)
sCmp := r.Start.Compare(key.Start)
eCmp := r.End.Compare(key.End)
if cv.timestamp.Less(timestamp) {
// The existing interval has a timestamp less than the new
// interval. Compare interval ranges to determine how to
// modify existing interval.
switch {
case sCmp == 0 && eCmp == 0:
// New and old are equal; replace old with new and avoid the need to insert new.
//
// New: ------------
// Old: ------------
//
// New: ------------
// Old:
*cv = cacheValue{timestamp: timestamp, txnID: txnID}
tcache.MoveToEnd(entry)
return
case sCmp <= 0 && eCmp >= 0:
// New contains or is equal to old; delete old.
//
// New: ------------ ------------ ------------
// Old: -------- or ---------- or ----------
//
// New: ------------ ------------ ------------
// Old:
tcache.DelEntry(entry)
case sCmp > 0 && eCmp < 0:
// Old contains new; split up old into two.
//
// New: ----
// Old: ------------
//
// New: ----
// Old: ---- ----
oldEnd := key.End
key.End = r.Start
newKey := tcache.MakeKey(r.End, oldEnd)
newEntry := makeCacheEntry(newKey, *cv)
tcache.AddEntryAfter(newEntry, entry)
case eCmp >= 0:
// Left partial overlap; truncate old end.
//
// New: -------- --------
// Old: -------- or ------------
//
// New: -------- --------
// Old: ---- ----
key.End = r.Start
case sCmp <= 0:
// Right partial overlap; truncate old start.
//
// New: -------- --------
// Old: -------- or ------------
//
// New: -------- --------
// Old: ---- ----
key.Start = r.End
default:
panic(fmt.Sprintf("no overlap between %v and %v", key.Range, r))
}
} else if timestamp.Less(cv.timestamp) {
// The existing interval has a timestamp greater than the new interval.
// Compare interval ranges to determine how to modify new interval before
// adding it to the timestamp cache.
switch {
case sCmp >= 0 && eCmp <= 0:
// Old contains or is equal to new; no need to add.
//
// Old: ----------- ----------- ----------- -----------
// New: ----- or ----------- or -------- or --------
//
// Old: ----------- ----------- ----------- -----------
// New:
return
case sCmp < 0 && eCmp > 0:
// New contains old; split up old into two. We can add the left piece
// immediately because it is guaranteed to be before the rest of the
// overlaps.
//
// Old: ------
// New: ------------
//
// Old: ------
// New: --- ---
lr := interval.Range{Start: r.Start, End: key.Start}
addRange(lr)
r.Start = key.End
case eCmp > 0:
// Left partial overlap; truncate new start.
//
// Old: -------- --------
// New: -------- or ------------
//
// Old: -------- --------
// New: ---- ----
r.Start = key.End
case sCmp < 0:
// Right partial overlap; truncate new end.
//
// Old: -------- --------
// New: -------- or ------------
//
// Old: -------- --------
// New: ---- ----
r.End = key.Start
default:
panic(fmt.Sprintf("no overlap between %v and %v", key.Range, r))
}
} else if (cv.txnID == nil && txnID == nil) ||
(cv.txnID != nil && txnID != nil && *cv.txnID == *txnID) {
// The existing interval has a timestamp equal to the new
// interval, and the same transaction ID.
switch {
case sCmp >= 0 && eCmp <= 0:
// Old contains or is equal to new; no need to add.
//
// New: ----- or ----------- or -------- or --------
// Old: ----------- ----------- ----------- -----------
//
// New:
// Old: ----------- ----------- ----------- -----------
return
case sCmp <= 0 && eCmp >= 0:
// New contains old; delete old.
//
// New: ------------ ------------ ------------
// Old: -------- or ---------- or ----------
//
// New: ------------ ------------ ------------
// Old:
tcache.DelEntry(entry)
case eCmp >= 0:
// Left partial overlap; truncate old end.
//
// New: -------- --------
// Old: -------- or ------------
//
// New: -------- --------
// Old: ---- ----
key.End = r.Start
case sCmp <= 0:
// Right partial overlap; truncate old start.
//
// New: -------- --------
// Old: -------- or ------------
//
// New: -------- --------
// Old: ---- ----
key.Start = r.End
default:
panic(fmt.Sprintf("no overlap between %v and %v", key.Range, r))
}
} else {
// The existing interval has a timestamp equal to the new
// interval and a different transaction ID.
switch {
case sCmp == 0 && eCmp == 0:
// New and old are equal. Segment is no longer owned by any
// transaction.
//
// New: ------------
// Old: ------------
//
// New:
// Nil: ============
// Old:
cv.txnID = nil
return
case sCmp == 0 && eCmp > 0:
// New contains old, left-aligned. Clear ownership of the
// existing segment and truncate new.
//
// New: ------------
// Old: ----------
//
// New: --
// Nil: ==========
// Old:
cv.txnID = nil
r.Start = key.End
case sCmp < 0 && eCmp == 0:
// New contains old, right-aligned. Clear ownership of the
// existing segment and truncate new.
//
// New: ------------
// Old: ----------
//
// New: --
// Nil: ==========
// Old:
cv.txnID = nil
r.End = key.Start
case sCmp < 0 && eCmp > 0:
// New contains old; split into three segments with the
// overlap owned by no txn.
//
// New: ------------
// Old: --------
//
// New: -- --
// Nil: ========
// Old:
cv.txnID = nil
newKey := tcache.MakeKey(r.Start, key.Start)
newEntry := makeCacheEntry(newKey, cacheValue{timestamp: timestamp, txnID: txnID})
tcache.AddEntryAfter(newEntry, entry)
r.Start = key.End
case sCmp > 0 && eCmp < 0:
// Old contains new; split up old into two. New segment is
// owned by no txn.
//
// New: ----
// Old: ------------
//
// New:
// Nil: ====
// Old: ---- ----
txnID = nil
oldEnd := key.End
key.End = r.Start
newKey := tcache.MakeKey(r.End, oldEnd)
newEntry := makeCacheEntry(newKey, *cv)
tcache.AddEntryAfter(newEntry, entry)
case eCmp == 0:
// Old contains new, right-aligned; truncate old end and clear
// ownership of new segment.
//
// New: --------
// Old: ------------
//
// New:
// Nil: ========
// Old: ----
txnID = nil
key.End = r.Start
case sCmp == 0:
// Old contains new, left-aligned; truncate old start and
// clear ownership of new segment.
// New: --------
// Old: ------------
//
// New:
// Nil: ========
// Old: ----
txnID = nil
key.Start = r.End
case eCmp > 0:
// Left partial overlap; truncate old end and split new into
// segments owned by no txn (the overlap) and the new txn.
//
// New: --------
// Old: --------
//
// New: ----
// Nil: ====
// Old: ----
key.End, r.Start = r.Start, key.End
newKey := tcache.MakeKey(key.End, r.Start)
newCV := cacheValue{timestamp: cv.timestamp, txnID: nil}
newEntry := makeCacheEntry(newKey, newCV)
tcache.AddEntryAfter(newEntry, entry)
case sCmp < 0:
// Right partial overlap; truncate old start and split new into
// segments owned by no txn (the overlap) and the new txn.
//
// New: --------
// Old: --------
//
// New: ----
// Nil: ====
// Old: ----
key.Start, r.End = r.End, key.Start
newKey := tcache.MakeKey(r.End, key.Start)
newCV := cacheValue{timestamp: cv.timestamp, txnID: nil}
newEntry := makeCacheEntry(newKey, newCV)
tcache.AddEntryAfter(newEntry, entry)
default:
panic(fmt.Sprintf("no overlap between %v and %v", key.Range, r))
}
}
}
addRange(r)
}
}
// selectEventTimestamp selects a timestamp for this log message. If the
// transaction this event is being written in has a non-zero timestamp, then that
// timestamp should be used; otherwise, the store's physical clock is used.
// This helps with testing; in normal usage, the logging of an event will never
// be the first action in the transaction, and thus the transaction will have an
// assigned database timestamp. However, in the case of our tests log events
// *are* the first action in a transaction, and we must elect to use the store's
// physical time instead.
func selectEventTimestamp(s *Store, input hlc.Timestamp) time.Time {
if input == hlc.ZeroTimestamp {
return s.Clock().PhysicalTime()
}
return input.GoTime()
}
