*/
/**
* Backend aliveness timer. This is reset whenever we receive a response from the backend and kept armed whenever
- * we have an outstanding request. If when this time expires, we tear down this connection and attept to reconnect
+ * we have an outstanding request. If when this time expires, we tear down this connection and attempt to reconnect
* it.
*/
- @VisibleForTesting
- static final long BACKEND_ALIVE_TIMEOUT_NANOS = TimeUnit.SECONDS.toNanos(30);
+ public static final long DEFAULT_BACKEND_ALIVE_TIMEOUT_NANOS = TimeUnit.SECONDS.toNanos(30);
/**
* Request timeout. If the request fails to complete within this time since it was originally enqueued, we time
* the request out.
*/
- @VisibleForTesting
- static final long REQUEST_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(2);
+ public static final long DEFAULT_REQUEST_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(2);
/**
* No progress timeout. A client fails to make any forward progress in this time, it will terminate itself.
*/
- @VisibleForTesting
- static final long NO_PROGRESS_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(15);
+ public static final long DEFAULT_NO_PROGRESS_TIMEOUT_NANOS = TimeUnit.MINUTES.toNanos(15);
// Emit a debug entry if we sleep for more that this amount
private static final long DEBUG_DELAY_NANOS = TimeUnit.MILLISECONDS.toNanos(100);
private volatile RequestException poisoned;
+ // Private constructor to avoid code duplication.
+ private AbstractClientConnection(final AbstractClientConnection<T> oldConn, final TransmitQueue newQueue) {
+ this.context = Preconditions.checkNotNull(oldConn.context);
+ this.cookie = Preconditions.checkNotNull(oldConn.cookie);
+ this.queue = Preconditions.checkNotNull(newQueue);
+ // Will be updated in finishReplay if needed.
+ this.lastReceivedTicks = oldConn.lastReceivedTicks;
+ }
+
+ // This constructor is only to be called by ConnectingClientConnection constructor.
// Do not allow subclassing outside of this package
- AbstractClientConnection(final ClientActorContext context, final Long cookie,
- final TransmitQueue queue) {
+ AbstractClientConnection(final ClientActorContext context, final Long cookie, final int queueDepth) {
this.context = Preconditions.checkNotNull(context);
this.cookie = Preconditions.checkNotNull(cookie);
- this.queue = Preconditions.checkNotNull(queue);
+ this.queue = new TransmitQueue.Halted(queueDepth);
this.lastReceivedTicks = currentTime();
}
+ // This constructor is only to be called (indirectly) by ReconnectingClientConnection constructor.
+ // Do not allow subclassing outside of this package
+ AbstractClientConnection(final AbstractClientConnection<T> oldConn) {
+ this(oldConn, new TransmitQueue.Halted(oldConn.queue, oldConn.currentTime()));
+ }
+
+ // This constructor is only to be called (indirectly) by ConnectedClientConnection constructor.
// Do not allow subclassing outside of this package
- AbstractClientConnection(final AbstractClientConnection<T> oldConnection, final int targetQueueSize) {
- this.context = oldConnection.context;
- this.cookie = oldConnection.cookie;
- this.queue = new TransmitQueue.Halted(targetQueueSize);
- this.lastReceivedTicks = oldConnection.lastReceivedTicks;
+ AbstractClientConnection(final AbstractClientConnection<T> oldConn, final T newBackend, final int queueDepth) {
+ this(oldConn, new TransmitQueue.Transmitting(oldConn.queue, queueDepth, newBackend, oldConn.currentTime(),
+ Preconditions.checkNotNull(oldConn.context).messageSlicer()));
}
public final ClientActorContext context() {
enqueueEntry(new ConnectionEntry(request, callback, enqueuedTicks), currentTime());
}
- public final long enqueueEntry(final ConnectionEntry entry, final long now) {
+ private long enqueueOrForward(final ConnectionEntry entry, final long now) {
lock.lock();
try {
- final RequestException maybePoison = poisoned;
- if (maybePoison != null) {
- throw new IllegalStateException("Connection " + this + " has been poisoned", maybePoison);
- }
+ commonEnqueue(entry, now);
+ return queue.enqueueOrForward(entry, now);
+ } finally {
+ lock.unlock();
+ }
+ }
- if (queue.isEmpty()) {
- // The queue is becoming non-empty, schedule a timer.
- scheduleTimer(entry.getEnqueuedTicks() + REQUEST_TIMEOUT_NANOS - now);
- }
- return queue.enqueue(entry, now);
+ /**
+ * Enqueue an entry, possibly also transmitting it.
+ */
+ public final void enqueueEntry(final ConnectionEntry entry, final long now) {
+ lock.lock();
+ try {
+ commonEnqueue(entry, now);
+ queue.enqueueOrReplay(entry, now);
} finally {
lock.unlock();
}
}
+ @GuardedBy("lock")
+ private void commonEnqueue(final ConnectionEntry entry, final long now) {
+ final RequestException maybePoison = poisoned;
+ if (maybePoison != null) {
+ throw new IllegalStateException("Connection " + this + " has been poisoned", maybePoison);
+ }
+
+ if (queue.isEmpty()) {
+ // The queue is becoming non-empty, schedule a timer.
+ scheduleTimer(entry.getEnqueuedTicks() + context.config().getRequestTimeout() - now);
+ }
+ }
+
+ // To be called from ClientActorBehavior on ConnectedClientConnection after entries are replayed.
+ final void cancelDebt() {
+ queue.cancelDebt(currentTime());
+ }
+
public abstract Optional<T> getBackendInfo();
final Collection<ConnectionEntry> startReplay() {
RequestException runtimeRequestException);
final void sendEntry(final ConnectionEntry entry, final long now) {
- long delay = enqueueEntry(entry, now);
+ long delay = enqueueOrForward(entry, now);
try {
if (delay >= DEBUG_DELAY_NANOS) {
if (delay > MAX_DELAY_NANOS) {
LOG.info("Capping {} throttle delay from {} to {} seconds", this,
- TimeUnit.NANOSECONDS.toSeconds(delay), MAX_DELAY_SECONDS);
+ TimeUnit.NANOSECONDS.toSeconds(delay), MAX_DELAY_SECONDS, new Throwable());
delay = MAX_DELAY_NANOS;
}
if (LOG.isDebugEnabled()) {
TimeUnit.NANOSECONDS.sleep(delay);
} catch (InterruptedException e) {
Thread.currentThread().interrupt();
- LOG.debug("Interrupted after sleeping {}ns", e, currentTime() - now);
+ LOG.debug("Interrupted after sleeping {}ns", currentTime() - now, e);
}
}
// If the delay is negative, we need to schedule an action immediately. While the caller could have checked
// for that condition and take appropriate action, but this is more convenient and less error-prone.
- final long normalized = delay <= 0 ? 0 : Math.min(delay, BACKEND_ALIVE_TIMEOUT_NANOS);
+ final long normalized = delay <= 0 ? 0 : Math.min(delay, context.config().getBackendAlivenessTimerInterval());
final FiniteDuration dur = FiniteDuration.fromNanos(normalized);
LOG.debug("{}: connection {} scheduling timeout in {}", context.persistenceId(), this, dur);
/**
* Check this queue for timeout and initiate reconnection if that happened. If the queue has not made progress
- * in {@link #NO_PROGRESS_TIMEOUT_NANOS} nanoseconds, it will be aborted.
+ * in {@link #DEFAULT_NO_PROGRESS_TIMEOUT_NANOS} nanoseconds, it will be aborted.
*
* @param current Current behavior
* @return Next behavior to use
// The following line is only reliable when queue is not forwarding, but such state should not last long.
// FIXME: BUG-8422: this may not be accurate w.r.t. replayed entries
final long ticksSinceProgress = queue.ticksStalling(now);
- if (ticksSinceProgress >= NO_PROGRESS_TIMEOUT_NANOS) {
+ if (ticksSinceProgress >= context.config().getNoProgressTimeout()) {
LOG.error("Queue {} has not seen progress in {} seconds, failing all requests", this,
TimeUnit.NANOSECONDS.toSeconds(ticksSinceProgress));
}
final long backendSilentTicks = backendSilentTicks(now);
- if (backendSilentTicks >= BACKEND_ALIVE_TIMEOUT_NANOS) {
+ if (backendSilentTicks >= context.config().getBackendAlivenessTimerInterval()) {
LOG.debug("{}: Connection {} has not seen activity from backend for {} nanoseconds, timing out",
context.persistenceId(), this, backendSilentTicks);
return null;
int tasksTimedOut = 0;
for (ConnectionEntry head = queue.peek(); head != null; head = queue.peek()) {
final long beenOpen = now - head.getEnqueuedTicks();
- if (beenOpen < REQUEST_TIMEOUT_NANOS) {
- return Optional.of(REQUEST_TIMEOUT_NANOS - beenOpen);
+ final long requestTimeout = context.config().getRequestTimeout();
+ if (beenOpen < requestTimeout) {
+ return Optional.of(requestTimeout - beenOpen);
}
tasksTimedOut++;
queue.remove(now);
LOG.debug("{}: Connection {} timed out entry {}", context.persistenceId(), this, head);
- head.complete(head.getRequest().toRequestFailure(
- new RequestTimeoutException("Timed out after " + beenOpen + "ns")));
+
+ timeoutEntry(head, beenOpen);
}
LOG.debug("Connection {} timed out {} tasks", this, tasksTimedOut);
return Optional.empty();
}
+ private void timeoutEntry(final ConnectionEntry entry, final long beenOpen) {
+ // Timeouts needs to be re-scheduled on actor thread because we are holding the lock on the current queue,
+ // which may be the tail of a successor chain. This is a problem if the callback attempts to send a request
+ // because that will attempt to lock the chain from the start, potentially causing a deadlock if there is
+ // a concurrent attempt to transmit.
+ context.executeInActor(current -> {
+ final double time = beenOpen * 1.0 / 1_000_000_000;
+ entry.complete(entry.getRequest().toRequestFailure(
+ new RequestTimeoutException("Timed out after " + time + "seconds")));
+ return current;
+ });
+ }
+
final void poison(final RequestException cause) {
lock.lock();
try {
return poisoned;
}
- final void receiveResponse(final ResponseEnvelope<?> envelope) {
+ void receiveResponse(final ResponseEnvelope<?> envelope) {
final long now = currentTime();
lastReceivedTicks = now;