distribution/registry/storage/notifications/sinks.go

338 lines
8.4 KiB
Go
Raw Normal View History

Implement notification endpoint webhook dispatch This changeset implements webhook notification endpoints for dispatching registry events. Repository instances can be decorated by a listener that converts calls into context-aware events, using a bridge. Events generated in the bridge are written to a sink. Implementations of sink include a broadcast and endpoint sink which can be used to configure event dispatch. Endpoints represent a webhook notification target, with queueing and retries built in. They can be added to a Broadcaster, which is a simple sink that writes a block of events to several sinks, to provide a complete dispatch mechanism. The main caveat to the current approach is that all unsent notifications are inmemory. Best effort is made to ensure that notifications are not dropped, to the point where queues may back up on faulty endpoints. If the endpoint is fixed, the events will be retried and all messages will go through. Internally, this functionality is all made up of Sink objects. The queuing functionality is implemented with an eventQueue sink and retries are implemented with retryingSink. Replacing the inmemory queuing with something persistent should be as simple as replacing broadcaster with a remote queue and that sets up the sinks to be local workers listening to that remote queue. Metrics are kept for each endpoint and exported via expvar. This may not be a permanent appraoch but should provide enough information for troubleshooting notification problems. Signed-off-by: Stephen J Day <stephen.day@docker.com>
2015-01-28 07:27:46 +00:00
package notifications
import (
"container/list"
"fmt"
"sync"
"time"
"github.com/Sirupsen/logrus"
)
// NOTE(stevvooe): This file contains definitions for several utility sinks.
// Typically, the broadcaster is the only sink that should be required
// externally, but others are suitable for export if the need arises. Albeit,
// the tight integration with endpoint metrics should be removed.
// Broadcaster sends events to multiple, reliable Sinks. The goal of this
// component is to dispatch events to configured endpoints. Reliability can be
// provided by wrapping incoming sinks.
type Broadcaster struct {
sinks []Sink
events chan []Event
closed chan chan struct{}
}
// NewBroadcaster ...
// Add appends one or more sinks to the list of sinks. The broadcaster
// behavior will be affected by the properties of the sink. Generally, the
// sink should accept all messages and deal with reliability on its own. Use
// of EventQueue and RetryingSink should be used here.
func NewBroadcaster(sinks ...Sink) *Broadcaster {
b := Broadcaster{
sinks: sinks,
events: make(chan []Event),
closed: make(chan chan struct{}),
}
// Start the broadcaster
go b.run()
return &b
}
// Write accepts a block of events to be dispatched to all sinks. This method
// will never fail and should never block (hopefully!). The caller cedes the
// slice memory to the broadcaster and should not modify it after calling
// write.
func (b *Broadcaster) Write(events ...Event) error {
select {
case b.events <- events:
case <-b.closed:
return ErrSinkClosed
}
return nil
}
// Close the broadcaster, ensuring that all messages are flushed to the
// underlying sink before returning.
func (b *Broadcaster) Close() error {
logrus.Infof("broadcaster: closing")
select {
case <-b.closed:
// already closed
return fmt.Errorf("broadcaster: already closed")
default:
// do a little chan handoff dance to synchronize closing
closed := make(chan struct{})
b.closed <- closed
close(b.closed)
<-closed
return nil
}
}
// run is the main broadcast loop, started when the broadcaster is created.
// Under normal conditions, it waits for events on the event channel. After
// Close is called, this goroutine will exit.
func (b *Broadcaster) run() {
for {
select {
case block := <-b.events:
for _, sink := range b.sinks {
if err := sink.Write(block...); err != nil {
logrus.Errorf("broadcaster: error writing events to %v, these events will be lost: %v", sink, err)
}
}
case closing := <-b.closed:
// close all the underlying sinks
for _, sink := range b.sinks {
if err := sink.Close(); err != nil {
logrus.Errorf("broadcaster: error closing sink %v: %v", sink, err)
}
}
closing <- struct{}{}
logrus.Debugf("broadcaster: closed")
return
}
}
}
// eventQueue accepts all messages into a queue for asynchronous consumption
// by a sink. It is unbounded and thread safe but the sink must be reliable or
// events will be dropped.
type eventQueue struct {
sink Sink
events *list.List
listeners []eventQueueListener
cond *sync.Cond
mu sync.Mutex
closed bool
}
// eventQueueListener is called when various events happen on the queue.
type eventQueueListener interface {
ingress(events ...Event)
egress(events ...Event)
}
// newEventQueue returns a queue to the provided sink. If the updater is non-
// nil, it will be called to update pending metrics on ingress and egress.
func newEventQueue(sink Sink, listeners ...eventQueueListener) *eventQueue {
eq := eventQueue{
sink: sink,
events: list.New(),
listeners: listeners,
}
eq.cond = sync.NewCond(&eq.mu)
go eq.run()
return &eq
}
// Write accepts the events into the queue, only failing if the queue has
// beend closed.
func (eq *eventQueue) Write(events ...Event) error {
eq.mu.Lock()
defer eq.mu.Unlock()
if eq.closed {
return ErrSinkClosed
}
for _, listener := range eq.listeners {
listener.ingress(events...)
}
eq.events.PushBack(events)
eq.cond.Signal() // signal waiters
return nil
}
// Close shutsdown the event queue, flushing
func (eq *eventQueue) Close() error {
eq.mu.Lock()
defer eq.mu.Unlock()
if eq.closed {
return fmt.Errorf("eventqueue: already closed")
}
// set closed flag
eq.closed = true
eq.cond.Signal() // signal flushes queue
eq.cond.Wait() // wait for signal from last flush
return eq.sink.Close()
}
// run is the main goroutine to flush events to the target sink.
func (eq *eventQueue) run() {
for {
block := eq.next()
if block == nil {
return // nil block means event queue is closed.
}
if err := eq.sink.Write(block...); err != nil {
logrus.Warnf("eventqueue: error writing events to %v, these events will be lost: %v", eq.sink, err)
}
for _, listener := range eq.listeners {
listener.egress(block...)
}
}
}
// next encompasses the critical section of the run loop. When the queue is
// empty, it will block on the condition. If new data arrives, it will wake
// and return a block. When closed, a nil slice will be returned.
func (eq *eventQueue) next() []Event {
eq.mu.Lock()
defer eq.mu.Unlock()
for eq.events.Len() < 1 {
if eq.closed {
eq.cond.Broadcast()
return nil
}
eq.cond.Wait()
}
front := eq.events.Front()
block := front.Value.([]Event)
eq.events.Remove(front)
return block
}
// retryingSink retries the write until success or an ErrSinkClosed is
// returned. Underlying sink must have p > 0 of succeeding or the sink will
// block. Internally, it is a circuit breaker retries to manage reset.
// Concurrent calls to a retrying sink are serialized through the sink,
// meaning that if one is in-flight, another will not proceed.
type retryingSink struct {
mu sync.Mutex
sink Sink
closed bool
// circuit breaker hueristics
failures struct {
threshold int
recent int
last time.Time
backoff time.Duration // time after which we retry after failure.
}
}
type retryingSinkListener interface {
active(events ...Event)
retry(events ...Event)
}
// TODO(stevvooe): We are using circuit break here, which actually doesn't
// make a whole lot of sense for this use case, since we always retry. Move
// this to use bounded exponential backoff.
// newRetryingSink returns a sink that will retry writes to a sink, backing
// off on failure. Parameters threshold and backoff adjust the behavior of the
// circuit breaker.
func newRetryingSink(sink Sink, threshold int, backoff time.Duration) *retryingSink {
rs := &retryingSink{
sink: sink,
}
rs.failures.threshold = threshold
rs.failures.backoff = backoff
return rs
}
// Write attempts to flush the events to the downstream sink until it succeeds
// or the sink is closed.
func (rs *retryingSink) Write(events ...Event) error {
rs.mu.Lock()
defer rs.mu.Unlock()
retry:
if rs.closed {
return ErrSinkClosed
}
if !rs.proceed() {
logrus.Warnf("%v encountered too many errors, backing off", rs.sink)
rs.wait(rs.failures.backoff)
goto retry
}
if err := rs.write(events...); err != nil {
if err == ErrSinkClosed {
// terminal!
return err
}
logrus.Errorf("retryingsink: error writing events: %v, retrying", err)
goto retry
}
return nil
}
// Close closes the sink and the underlying sink.
func (rs *retryingSink) Close() error {
rs.mu.Lock()
defer rs.mu.Unlock()
if rs.closed {
return fmt.Errorf("retryingsink: already closed")
}
rs.closed = true
return rs.sink.Close()
}
// write provides a helper that dispatches failure and success properly. Used
// by write as the single-flight write call.
func (rs *retryingSink) write(events ...Event) error {
if err := rs.sink.Write(events...); err != nil {
rs.failure()
return err
}
rs.reset()
return nil
}
// wait backoff time against the sink, unlocking so others can proceed. Should
// only be called by methods that currently have the mutex.
func (rs *retryingSink) wait(backoff time.Duration) {
rs.mu.Unlock()
defer rs.mu.Lock()
// backoff here
time.Sleep(backoff)
}
// reset marks a succesful call.
func (rs *retryingSink) reset() {
rs.failures.recent = 0
rs.failures.last = time.Time{}
}
// failure records a failure.
func (rs *retryingSink) failure() {
rs.failures.recent++
rs.failures.last = time.Now().UTC()
}
// proceed returns true if the call should proceed based on circuit breaker
// hueristics.
func (rs *retryingSink) proceed() bool {
return rs.failures.recent < rs.failures.threshold ||
time.Now().UTC().After(rs.failures.last.Add(rs.failures.backoff))
}