Publication counter seemly wrong

[callmerockett]

TL;DR: Should pub-pos be greater than pub-lmt?

Hi, I was testing a combination of tethered and untethered IPC subscriptions, with the following options:

 ...
      -Dagrona.disable.bounds.checks=false
      -Daeron.mtu.length=1024
      -Daeron.ipc.mtu.length=1024
      -Daeron.socket.so_sndbuf=2m
      -Daeron.socket.so_rcvbuf=2m
      -Daeron.rcv.initial.window.length=2m
      -Daeron.term.buffer.length=2m
      -Daeron.ipc.term.buffer.length=2m
      -Daeron.publication.term.window.length=128k
      -Daeron.ipc.publication.term.window.length=128k
...

And using AeronStat got this results:

 56:           13,768,608 - pub-pos (sampled): 22 -1336219639 2000 aeron:ipc
 57:           13,767,392 - pub-lmt: 22 -1336219639 2000 aeron:ipc
 58:    1,731,101,471,743 - client-heartbeat: id=21
 59:    1,731,101,471,750 - client-heartbeat: id=24
 60:           13,768,608 - sub-pos: 25 -1336219639 2000 aeron:ipc?tether=false @0
 64:    1,731,101,471,535 - client-heartbeat: id=50
 66:           13,636,320 - sub-pos: 51 -1336219639 2000 aeron:ipc @9303456

In line 56 pub-pos is greater than pub-lmt, is this right?

Some context:

• Aeron client version: 1.35.1
• Aeron driver version: 1.46.7

[mjpt777]

It is possible for the publication limit to be clamped back to the max consumer position when subscribers go away and this could be less than the producer position, i.e. the publication position.

BTW we would not recommend such extreme version range between client and driver. We recommend no more than 1 year of releases between versions.

Also those term window and buffer lengths are in ratios that I would not find useful. It seems like your system could benefit from some consulting on how to tune it.

[callmerockett]

Hi @mjpt777 , thanks for the reply.

Still in this issue:

It is possible for the publication limit to be clamped back to the max consumer position when subscribers go away and this could be less than the producer position, i.e. the publication position.

This could happen even if the subscriptions are still alive? With the above sample the untethered subscription has the max consumer position while the publication is hold by the slowest tethered consumer, both still running. To provide more context below are the configured timeouts for the test

      -Daeron.threading.mode=SHARED
      -Daeron.spies.simulate.connection=true
      -Daeron.publication.connection.timeout=20s
      -Daeron.publication.unblock.timeout=150s
      -Daeron.publication.linger.timeout=15s
      -Daeron.client.liveness.timeout=30s

These parameters configurations were set intentionally to analyse our system under back pressure scenarios. I could reproduce this behavior with driver at version 1.39.0 as well.

Thanks for the recommendations.

[mjpt777]

Please provide a full example as a test if you wish to have it further investigated.