Details
-
Bug
-
Status: Resolved
-
P3
-
Resolution: Fixed
-
2.35.0
-
None
Description
Discussed here: https://lists.apache.org/thread/rcto0yggqygfk0ljnd8mg69h83rhvvs5
—
We (Spotify) are experiencing flaky tests on beam 2.35.0 as a result of NullPointerExceptions during pipeline construction.
Stacktrace [1]. Root cause appears to be multiple threads accessing DeserializationContext from ValueProvider.Deserializer.deserialize [2] which, according to the javadocs "is guaranteed to only be used from single-threaded context" [3]
In the context internals, the mutable _currentType is set [4] then accessed a few lines later [5], but in a multithreaded situation the writes clobber each other and a call to next() can fail with a NPE.
I think synchronizing on ctxt in [2] should be sufficient to avoid this issue, but I'm not sure if the NPEs are due to beam misuse or something on the jackson side that should be changed.
Failing (scala) test here, scalatest runs tests in parallel:
https://gist.github.com/kellen/124185463c16a66167a7fa704147c510#file-fail-scala
Test stacktrace:
https://gist.github.com/kellen/124185463c16a66167a7fa704147c510#file-test_stacktrace-txt
[1] https://gist.github.com/kellen/124185463c16a66167a7fa704147c510#file-stacktrace-txt
[2] https://github.com/apache/beam/blob/master/sdks/java/core/src/main/java/org/apache/beam/sdk/options/ValueProvider.java#L359
[3] https://github.com/joansmith/jackson-databind/blob/master/src/main/java/com/fasterxml/jackson/databind/DeserializationContext.java#L36
[4] https://github.com/joansmith/jackson-databind/blob/master/src/main/java/com/fasterxml/jackson/databind/DeserializationContext.java#L652
[5] https://github.com/joansmith/jackson-databind/blob/master/src/main/java/com/fasterxml/jackson/databind/DeserializationContext.java#L656
Attachments
Issue Links
- links to
