-
Notifications
You must be signed in to change notification settings - Fork 329
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Nullpointer exception for pekko-instrumentation in version 1.1.0-M1 onward #1352
Comments
Pekko 1.1.0-M1 is released. It is a milestone release but the jars are all up on Maven Central. |
There are some changes in the Pekko dispatch code in v1.1.0-M1. I wouldn't have thought them major. apache/pekko#485 is the biggest change |
@pjfanning sry for the late response. |
i can try if this would help with our problem but im not to optimistic that it will help. |
@ivantopo do you have any idea how to create tests for that? |
I created https://github.com/pjfanning/kamon-pekko-http-test based on the description here and this runs fine with no issues. |
Did you start it with the kanela java-agent? |
Thanks. I hadn't added kanela-agent. The issue happens with it enabled. I updated the sample project. I haven't used Kamon since its early days. I have my own fork that I use myself and it works with Pekko 1.1 (https://github.com/pjfanning/micrometer-pekko). |
I may have messed up the test run but when I tried Scala 3.3.3 and 2.12.19, it seemed ok. Maybe this only happens with Scala 2.13. |
I'm getting this issue now with Pekko 1.1.0. |
I think it's related to inlining. This commit still works: apache/pekko@d829637. This one doesn't: apache/pekko@0f1db53. And if I disable inlining with the |
Seems plausible. Do you see the same behaviour as me - in that Scala 3 and Scala 2.12 work fine but that the NPE pops up with Scala 2.13? The Pekko inlining does not affect the Scala 3 release of Pekko. |
2.12.19 and 3.5.0 didn't give a NPE. |
Is this a bug in Kamon, or a bug in Pekko ? |
Pekko 1.1 uses Scala 2 compiler inlining. Kamon doesn't seem to be able to find the inlined code in Pekko. |
I do have the same issue with 2.13.14 |
To be clear, the Apache Pekko 1.1 releases won't be modified to suit Kamon. We reserve the right in minor releases to make some changes. @mdedetrich just adding to you to the issue just in case you have any thoughts on this. If anyone has any thoughts on hooks that could be added to Pekko 1.2 or 2.0 that makes it easier for Kamon to instrument Pekko then feel free to raise issues or PRs in the Pekko repo(s). |
I created #1361 as a workaround. |
Users hitting this issue should also be able to disable dispatcher instrumentation. See https://github.com/kamon-io/Kamon/blob/124e223de4d7d806b47c76323f5d91edf626fa81/instrumentation/kamon-pekko/src/main/resources/reference.conf and check about excluding dispatchers. |
Just catching up on all this, been neglecting it over the summer (sorry about that!). The instrumentation is definitely failing because of the inlining (added some comments about this on a draft pr before I realised you guys had already spotted it 😅). If we can get some instrumentation hooks into pekko I'm sure that would mitigate the issue... I will have a poke about with what might work on the pekko side and see if we can find a place of agreement. Almost all the instrumentation should work for scala 2.12, all of it should work for scala 3, so it's really only 2.13 that suffers much as it currently stands; one fix is, thus, to upgrade to scala 3 (although that's clearly not always possible). From my experiments so far, I don't think it's possible to fix this purely on the Kamon side. |
I defintiely think that instrumentation hooks is the way to go. While inlining may have caused the issue, as @pjfanning pointed out Pekko reserves the right to change the structure of stack/methods as it pleases as long as the methods are not public and so the current implementation of kamon does appear to be brittle for this reason. If you have some ideas about instrumentation, PR's/design documents are welcome! Supporting this officially via an API is the way to go. |
I debugged a broken test in https://github.com/pjfanning/micrometer-pekko (uses AOP and has similar issues to Kamon) and in this once case that calls to ActorCell.stop were being inlined on one particular call path (Scala 2.13 only). The call path related to stopping a RobinRobinRouter instance that had 5 actor refs to round robin over. What I couldn't work out is why the ActorCell.stop got inlined because it is not marked with |
@mdedetrich is there any way to mark a function so it can't be inlined? Edit - apparently there is a |
I raised apache/pekko#1484 - no guarantees that the community will agree to make changes but it is open for discussion. |
Thanks @pjfanning ! I'll see if I can get a pekko build that passes the suite with a sprinkle of those annotations 🙏 |
when testing pekko with the version 1.1.0-M1 which will be released soon we've ran into a nullpointer error when running with Kamon + kanela-instrumentation.
to Reproduce:
build.sbt
Main.scala:
when starting with
-javaagent:javaagent/kanela-agent.jar
(version 1.0.18)i'm receiving the exception:
i've tried to analyze the error and got to the point that
kamon/instrumentation/pekko/instrumentations/DispatcherInstrumentation.scala:164
does not seem to get triggered with pekko 1.1.0-M1 anymore resulting in
dispatcherPrerequisites
being null inkamon/instrumentation/pekko/instrumentations/DispatcherInstrumentation.scala:125
but i cannot understand what change in pekko 1.0.3 <-> 1.1.0-M1 made that difference and how to fix it
The text was updated successfully, but these errors were encountered: