[GitHub] activemq-artemis pull request #1119: ARTEMIS-1025 OutOfDirectMemoryError rai...

[franz1981 <gi...@git.apache.org>]

Github user franz1981 commented on a diff in the pull request:

    https://github.com/apache/activemq-artemis/pull/1119#discussion_r109278189
  
    --- Diff: artemis-core-client/src/main/java/org/apache/activemq/artemis/core/remoting/impl/netty/NettyConnection.java ---
    @@ -92,69 +86,143 @@ public NettyConnection(final Map<String, Object> configuration,
     
           this.listener = listener;
     
    +      this.directDeliver = directDeliver;
    +
           this.batchingEnabled = batchingEnabled;
     
    -      this.directDeliver = directDeliver;
    +      this.writeBufferHighWaterMark = this.channel.config().getWriteBufferHighWaterMark();
    +
    +      this.batchLimit = batchingEnabled ? Math.min(this.writeBufferHighWaterMark, DEFAULT_BATCH_BYTES) : 0;
    +   }
    +
    +   private static void waitFor(ChannelPromise promise, long millis) {
    +      try {
    +         final boolean completed = promise.await(millis);
    +         if (!completed) {
    +            ActiveMQClientLogger.LOGGER.timeoutFlushingPacket();
    +         }
    +      } catch (InterruptedException e) {
    +         throw new ActiveMQInterruptedException(e);
    +      }
    +   }
    +
    +   /**
    +    * Returns an estimation of the current size of the write buffer in the channel.
    +    * To obtain a more precise value is necessary to use the unsafe API of the channel to
    +    * call the {@link io.netty.channel.ChannelOutboundBuffer#totalPendingWriteBytes()}.
    +    * Anyway, both these values are subject to concurrent modifications.
    +    */
    +   private static int batchBufferSize(Channel channel, int writeBufferHighWaterMark) {
    +      //Channel::bytesBeforeUnwritable is performing a volatile load
    +      //this is the reason why writeBufferHighWaterMark is passed as an argument
    +      final int bytesBeforeUnwritable = (int) channel.bytesBeforeUnwritable();
    +      assert bytesBeforeUnwritable >= 0;
    +      final int writtenBytes = writeBufferHighWaterMark - bytesBeforeUnwritable;
    +      assert writtenBytes >= 0;
    +      return writtenBytes;
    +   }
    +
    +   /**
    +    * When batching is not enabled, it tries to back-pressure the caller thread.
    +    * The back-pressure provided is not before the writeAndFlush request, buf after it: too many threads that are not
    +    * using {@link Channel#isWritable} to know when push unbatched data will risk to cause OOM due to the enqueue of each own {@link Channel#writeAndFlush} requests.
    +    * Trying to provide back-pressure before the {@link Channel#writeAndFlush} request could work, but in certain scenarios it will block {@link Channel#isWritable} to be true.
    +    */
    +   private static ChannelFuture backPressuredWriteAndFlush(final ByteBuf bytes,
    +                                                           final int readableBytes,
    +                                                           final Channel channel,
    +                                                           final ChannelPromise promise) {
    +      final ChannelFuture future;
    +      if (!channel.isWritable()) {
    +         final ChannelPromise channelPromise = promise.isVoid() ? channel.newPromise() : promise;
    +         future = channel.writeAndFlush(bytes, channelPromise);
    +         //is the channel is not writable wait the current request to be flushed, providing backpressuring on the caller thread
    +         if (!channel.isWritable() && !future.awaitUninterruptibly(DEFAULT_BACK_PRESSURE_WAIT_MILLIS)) {
    +            if (ActiveMQClientLogger.LOGGER.isTraceEnabled()) {
    --- End diff --
    
    Agree with the warn and about reaction strategy: a sensible default and customization via connection properties would be welcome.
    I'm personally a fan of "kill the slowest connections", but I'm worried it could lead to a storm of users complaining that while sending 1 [EB](https://it.wikipedia.org/wiki/Exabyte) of message strangely the broker has died while screaming: ``"WARN: killed connection X - too slow; required 10 days to send -9223372036854775808 bytes"``  :)
    I'm joking, but effectively I'm just worried to do not break everything adding here new behaviours that would break (too much) the expectations of any piece of code that will request to write something on the network: maybe an external job (running in the even loop) that will monitor the connections and detecting the slowest ones (using the new methods I've added to monitor the buffering state) applying the configured countermeasures could do the job, leaving the ``NettyConnection`` simpler and with less responsabilities, wdyt?
    Just thinking loud



---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastructure@apache.org or file a JIRA ticket
with INFRA.
---