restic/vendor/cloud.google.com/go/pubsub/iterator.go
2017-08-06 21:47:56 +02:00

520 lines
14 KiB
Go

// Copyright 2016 Google Inc. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
package pubsub
import (
"log"
"sync"
"time"
"golang.org/x/net/context"
"google.golang.org/api/iterator"
"google.golang.org/api/support/bundler"
pb "google.golang.org/genproto/googleapis/pubsub/v1"
"google.golang.org/grpc"
"google.golang.org/grpc/codes"
)
type messageIterator struct {
impl interface {
next() (*Message, error)
stop()
}
}
type pollingMessageIterator struct {
// kaTicker controls how often we send an ack deadline extension request.
kaTicker *time.Ticker
// ackTicker controls how often we acknowledge a batch of messages.
ackTicker *time.Ticker
ka *keepAlive
acker *acker
nacker *bundler.Bundler
puller *puller
// mu ensures that cleanup only happens once, and concurrent Stop
// invocations block until cleanup completes.
mu sync.Mutex
// closed is used to signal that Stop has been called.
closed chan struct{}
}
var useStreamingPull = false
// newMessageIterator starts a new messageIterator. Stop must be called on the messageIterator
// when it is no longer needed.
// subName is the full name of the subscription to pull messages from.
// ctx is the context to use for acking messages and extending message deadlines.
func newMessageIterator(ctx context.Context, s service, subName string, po *pullOptions) *messageIterator {
if !useStreamingPull {
return &messageIterator{
impl: newPollingMessageIterator(ctx, s, subName, po),
}
}
sp := s.newStreamingPuller(ctx, subName, int32(po.ackDeadline.Seconds()))
err := sp.open()
if grpc.Code(err) == codes.Unimplemented {
log.Println("pubsub: streaming pull unimplemented; falling back to legacy pull")
return &messageIterator{
impl: newPollingMessageIterator(ctx, s, subName, po),
}
}
// TODO(jba): handle other non-nil error?
log.Println("using streaming pull")
return &messageIterator{
impl: newStreamingMessageIterator(ctx, sp, po),
}
}
func newPollingMessageIterator(ctx context.Context, s service, subName string, po *pullOptions) *pollingMessageIterator {
// TODO: make kaTicker frequency more configurable.
// (ackDeadline - 5s) is a reasonable default for now, because the minimum ack period is 10s. This gives us 5s grace.
keepAlivePeriod := po.ackDeadline - 5*time.Second
kaTicker := time.NewTicker(keepAlivePeriod) // Stopped in it.Stop
// Ack promptly so users don't lose work if client crashes.
ackTicker := time.NewTicker(100 * time.Millisecond) // Stopped in it.Stop
ka := &keepAlive{
s: s,
Ctx: ctx,
Sub: subName,
ExtensionTick: kaTicker.C,
Deadline: po.ackDeadline,
MaxExtension: po.maxExtension,
}
ack := &acker{
s: s,
Ctx: ctx,
Sub: subName,
AckTick: ackTicker.C,
Notify: ka.Remove,
}
nacker := bundler.NewBundler("", func(ackIDs interface{}) {
// NACK by setting the ack deadline to zero, to make the message
// immediately available for redelivery.
//
// If the RPC fails, nothing we can do about it. In the worst case, the
// deadline for these messages will expire and they will still get
// redelivered.
_ = s.modifyAckDeadline(ctx, subName, 0, ackIDs.([]string))
})
nacker.DelayThreshold = 100 * time.Millisecond // nack promptly
nacker.BundleCountThreshold = 10
pull := newPuller(s, subName, ctx, po.maxPrefetch, ka.Add, ka.Remove)
ka.Start()
ack.Start()
return &pollingMessageIterator{
kaTicker: kaTicker,
ackTicker: ackTicker,
ka: ka,
acker: ack,
nacker: nacker,
puller: pull,
closed: make(chan struct{}),
}
}
// Next returns the next Message to be processed. The caller must call
// Message.Done when finished with it.
// Once Stop has been called, calls to Next will return iterator.Done.
func (it *messageIterator) Next() (*Message, error) {
return it.impl.next()
}
func (it *pollingMessageIterator) next() (*Message, error) {
m, err := it.puller.Next()
if err == nil {
m.doneFunc = it.done
return m, nil
}
select {
// If Stop has been called, we return Done regardless the value of err.
case <-it.closed:
return nil, iterator.Done
default:
return nil, err
}
}
// Client code must call Stop on a messageIterator when finished with it.
// Stop will block until Done has been called on all Messages that have been
// returned by Next, or until the context with which the messageIterator was created
// is cancelled or exceeds its deadline.
// Stop need only be called once, but may be called multiple times from
// multiple goroutines.
func (it *messageIterator) Stop() {
it.impl.stop()
}
func (it *pollingMessageIterator) stop() {
it.mu.Lock()
defer it.mu.Unlock()
select {
case <-it.closed:
// Cleanup has already been performed.
return
default:
}
// We close this channel before calling it.puller.Stop to ensure that we
// reliably return iterator.Done from Next.
close(it.closed)
// Stop the puller. Once this completes, no more messages will be added
// to it.ka.
it.puller.Stop()
// Start acking messages as they arrive, ignoring ackTicker. This will
// result in it.ka.Stop, below, returning as soon as possible.
it.acker.FastMode()
// This will block until
// (a) it.ka.Ctx is done, or
// (b) all messages have been removed from keepAlive.
// (b) will happen once all outstanding messages have been either ACKed or NACKed.
it.ka.Stop()
// There are no more live messages, so kill off the acker.
it.acker.Stop()
it.nacker.Flush()
it.kaTicker.Stop()
it.ackTicker.Stop()
}
func (it *pollingMessageIterator) done(ackID string, ack bool) {
if ack {
it.acker.Ack(ackID)
// There's no need to call it.ka.Remove here, as acker will
// call it via its Notify function.
} else {
it.ka.Remove(ackID)
_ = it.nacker.Add(ackID, len(ackID)) // ignore error; this is just an optimization
}
}
type streamingMessageIterator struct {
ctx context.Context
po *pullOptions
sp *streamingPuller
kaTicker *time.Ticker // keep-alive (deadline extensions)
ackTicker *time.Ticker // message acks
nackTicker *time.Ticker // message nacks (more frequent than acks)
failed chan struct{} // closed on stream error
stopped chan struct{} // closed when Stop is called
drained chan struct{} // closed when stopped && no more pending messages
msgc chan *Message
wg sync.WaitGroup
mu sync.Mutex
keepAliveDeadlines map[string]time.Time
pendingReq *pb.StreamingPullRequest
err error // error from stream failure
}
func newStreamingMessageIterator(ctx context.Context, sp *streamingPuller, po *pullOptions) *streamingMessageIterator {
// TODO: make kaTicker frequency more configurable. (ackDeadline - 5s) is a
// reasonable default for now, because the minimum ack period is 10s. This
// gives us 5s grace.
keepAlivePeriod := po.ackDeadline - 5*time.Second
kaTicker := time.NewTicker(keepAlivePeriod)
// Ack promptly so users don't lose work if client crashes.
ackTicker := time.NewTicker(100 * time.Millisecond)
nackTicker := time.NewTicker(100 * time.Millisecond)
it := &streamingMessageIterator{
ctx: ctx,
sp: sp,
po: po,
kaTicker: kaTicker,
ackTicker: ackTicker,
nackTicker: nackTicker,
failed: make(chan struct{}),
stopped: make(chan struct{}),
drained: make(chan struct{}),
// use maxPrefetch as the channel's buffer size.
msgc: make(chan *Message, po.maxPrefetch),
keepAliveDeadlines: map[string]time.Time{},
pendingReq: &pb.StreamingPullRequest{},
}
it.wg.Add(2)
go it.receiver()
go it.sender()
return it
}
func (it *streamingMessageIterator) next() (*Message, error) {
// If ctx has been cancelled or the iterator is done, return straight
// away (even if there are buffered messages available).
select {
case <-it.ctx.Done():
return nil, it.ctx.Err()
case <-it.failed:
break
case <-it.stopped:
break
default:
// Wait for a message, but also for one of the above conditions.
select {
case msg := <-it.msgc:
// Since active select cases are chosen at random, this can return
// nil (from the channel close) even if it.failed or it.stopped is
// closed.
if msg == nil {
break
}
msg.doneFunc = it.done
return msg, nil
case <-it.ctx.Done():
return nil, it.ctx.Err()
case <-it.failed:
break
case <-it.stopped:
break
}
}
// Here if the iterator is done.
it.mu.Lock()
defer it.mu.Unlock()
return nil, it.err
}
func (it *streamingMessageIterator) stop() {
it.mu.Lock()
select {
case <-it.stopped:
it.mu.Unlock()
it.wg.Wait()
return
default:
close(it.stopped)
}
if it.err == nil {
it.err = iterator.Done
}
// Before reading from the channel, see if we're already drained.
it.checkDrained()
it.mu.Unlock()
// Nack all the pending messages.
// Grab the lock separately for each message to allow the receiver
// and sender goroutines to make progress.
// Why this will eventually terminate:
// - If the receiver is not blocked on a stream Recv, then
// it will write all the messages it has received to the channel,
// then exit, closing the channel.
// - If the receiver is blocked, then this loop will eventually
// nack all the messages in the channel. Once done is called
// on the remaining messages, the iterator will be marked as drained,
// which will trigger the sender to terminate. When it does, it
// performs a CloseSend on the stream, which will result in the blocked
// stream Recv returning.
for m := range it.msgc {
it.mu.Lock()
delete(it.keepAliveDeadlines, m.ackID)
it.addDeadlineMod(m.ackID, 0)
it.checkDrained()
it.mu.Unlock()
}
it.wg.Wait()
}
// checkDrained closes the drained channel if the iterator has been stopped and all
// pending messages have either been n/acked or expired.
//
// Called with the lock held.
func (it *streamingMessageIterator) checkDrained() {
select {
case <-it.drained:
return
default:
}
select {
case <-it.stopped:
if len(it.keepAliveDeadlines) == 0 {
close(it.drained)
}
default:
}
}
// Called when a message is acked/nacked.
func (it *streamingMessageIterator) done(ackID string, ack bool) {
it.mu.Lock()
defer it.mu.Unlock()
delete(it.keepAliveDeadlines, ackID)
if ack {
it.pendingReq.AckIds = append(it.pendingReq.AckIds, ackID)
} else {
it.addDeadlineMod(ackID, 0) // Nack indicated by modifying the deadline to zero.
}
it.checkDrained()
}
// addDeadlineMod adds the ack ID to the pending request with the given deadline.
//
// Called with the lock held.
func (it *streamingMessageIterator) addDeadlineMod(ackID string, deadlineSecs int32) {
pr := it.pendingReq
pr.ModifyDeadlineAckIds = append(pr.ModifyDeadlineAckIds, ackID)
pr.ModifyDeadlineSeconds = append(pr.ModifyDeadlineSeconds, deadlineSecs)
}
// fail is called when a stream method returns a permanent error.
func (it *streamingMessageIterator) fail(err error) {
it.mu.Lock()
if it.err == nil {
it.err = err
close(it.failed)
}
it.mu.Unlock()
}
// receiver runs in a goroutine and handles all receives from the stream.
func (it *streamingMessageIterator) receiver() {
defer it.wg.Done()
defer close(it.msgc)
for {
// Stop retrieving messages if the context is done, the stream
// failed, or the iterator's Stop method was called.
select {
case <-it.ctx.Done():
return
case <-it.failed:
return
case <-it.stopped:
return
default:
}
// Receive messages from stream. This may block indefinitely.
msgs, err := it.sp.fetchMessages()
// The streamingPuller handles retries, so any error here
// is fatal to the iterator.
if err != nil {
it.fail(err)
return
}
// We received some messages. Remember them so we can
// keep them alive.
deadline := time.Now().Add(it.po.maxExtension)
it.mu.Lock()
for _, m := range msgs {
it.keepAliveDeadlines[m.ackID] = deadline
}
it.mu.Unlock()
// Deliver the messages to the channel.
for _, m := range msgs {
select {
case <-it.ctx.Done():
return
case <-it.failed:
return
// Don't return if stopped. We want to send the remaining
// messages on the channel, where they will be nacked.
case it.msgc <- m:
}
}
}
}
// sender runs in a goroutine and handles all sends to the stream.
func (it *streamingMessageIterator) sender() {
defer it.wg.Done()
defer it.kaTicker.Stop()
defer it.ackTicker.Stop()
defer it.nackTicker.Stop()
defer it.sp.closeSend()
done := false
for !done {
send := false
select {
case <-it.ctx.Done():
// Context canceled or timed out: stop immediately, without
// another RPC.
return
case <-it.failed:
// Stream failed: nothing to do, so stop immediately.
return
case <-it.drained:
// All outstanding messages have been marked done:
// nothing left to do except send the final request.
it.mu.Lock()
send = (len(it.pendingReq.AckIds) > 0 || len(it.pendingReq.ModifyDeadlineAckIds) > 0)
done = true
case <-it.kaTicker.C:
it.mu.Lock()
send = it.handleKeepAlives()
case <-it.nackTicker.C:
it.mu.Lock()
send = (len(it.pendingReq.ModifyDeadlineAckIds) > 0)
case <-it.ackTicker.C:
it.mu.Lock()
send = (len(it.pendingReq.AckIds) > 0)
}
// Lock is held here.
if send {
req := it.pendingReq
it.pendingReq = &pb.StreamingPullRequest{}
it.mu.Unlock()
err := it.sp.send(req)
if err != nil {
// The streamingPuller handles retries, so any error here
// is fatal to the iterator.
it.fail(err)
return
}
} else {
it.mu.Unlock()
}
}
}
// handleKeepAlives modifies the pending request to include deadline extensions
// for live messages. It also purges expired messages. It reports whether
// there were any live messages.
//
// Called with the lock held.
func (it *streamingMessageIterator) handleKeepAlives() bool {
live, expired := getKeepAliveAckIDs(it.keepAliveDeadlines)
for _, e := range expired {
delete(it.keepAliveDeadlines, e)
}
dl := trunc32(int64(it.po.ackDeadline.Seconds()))
for _, m := range live {
it.addDeadlineMod(m, dl)
}
it.checkDrained()
return len(live) > 0
}