#3960
  • Created on
    2014-03-24 15:04
  • Reported by
    Patrik Nordwall
  • Status
    Accepted
  • Priority
    Normal (3)
  • Assigned to
    Patrik Nordwall
  • Milestone
    Current
  • Plan Level
    -
  • Due Date
    -

Attachments

Related Tickets

Add new subtask or bug

Optimize EndpointWriter unstashing

Reported here: https://groups.google.com/d/msg/akka-user/5vYs4jKMaQU/Unjo3TaFG8QJ
with reproducible sample.

Major time is consumed in EndpointWriter stashing and unstashing messages. That happens when the transport can't write.
I think it goes into an very inefficient spin of unstashing all messages, and quickly discover that it can't write and then everything is stashed again, and so on...

We had something similar in EventsourcedProcessor, and then Martin introduced unstash() to avoid unstashAll().

I tried to increase backoff time and send buffer size, but could not see a big difference.
User picture
  on 2014-03-24 15:14 *
By drewhk
Hm, the unstash() spin is a reasonable assumption. We should fix that.

Once we change to stream based transports this issue will be mitigated since the buffering logic becomes simpler. Until then we need a workaround.
User picture
  on 2014-03-24 15:26 *
By drewhk
We should simply get rid of Stash here. The main reason it was used instead of an internal buffer is to prevent dropping messages during the restart of EndpointWriter -- which is no longer needed in 2.3.
User picture
  on 2014-03-24 15:35 *
By drewhk
Should this be fixed for 2.3.1? The fix is probably not too hard, but this is a sensitive change.
User picture
  on 2014-04-03 09:32 *
By Patrik Nordwall
Assigned to set to Patrik Nordwall
Status changed from New to Accepted
User picture
  on 2014-04-03 09:33 *
By Patrik Nordwall
Milestone set to Current