fix Kafka flaky test #105
fix Kafka flaky test #105
Conversation
|
After four times CI checks(see: https://github.com/apache/incubator-pekko-projection/actions/runs/7831923269), this kind of test has been successfully passed, and we should be able to assume that this kind of flaky has been fixed. |
|
@raboof Could you please do a code review? Thanks. |
There was a problem hiding this comment.
TBH I'm not convinced this change fixes the root cause of the instability: because of the assertion records.count(_.partition() == tp0.partition()) shouldBe 5 AFAICS it should be guaranteed that tp0TestCount would always be 5, so this change should not have an effect other than timing differences - or did I miss something?
The producer produces a total of 20 elements, and each partition has 10 evenly. Then Sink consumes 10 elements from two partitions. There is no assertion that each partition consumes 5 evenly. If it is not consumed evenly, it will lead to problems with subsequent assertions, because partition 0 may consume 6 records in the previous step. This PR does not force the assertion that Sink consumes evenly from all partitions. If Kafka Consumer is not balanced consume from partitions, I don't think this is the responsibility of this test or should be added a new test to assert. |
Isn't that asserted by https://github.com/apache/incubator-pekko-projection/pull/105/files#diff-fdbaabc373547fac2368cc881beaa945f8d0225553d3e964dbe09eaf792592adR108-R112 ?
That makes sense to me. |
| @@ -109,6 +110,12 @@ class KafkaSourceProviderImplSpec extends ScalaTestWithActorTestKit with LogCapt | |||
| records.count(_.partition() == tp1.partition()) shouldBe 5 | |||
| } | |||
|
|
|||
| // because source push to handle(probe) before sinkProbe request pull, it made probe cache random one record | |||
| val eagerMessage = probe.receiveMessage() | |||
There was a problem hiding this comment.
It's kind of like akka/akka-projection#462, i won't say this is best solution.
There was a problem hiding this comment.
@He-Pin Is there any way to make the source and sink reach a consensus on the number of elements?
|
Would you like to take a look? @raboof This may not the best solution, but it does reduce flaky. |
Trying to solve #103, I am very doubtful that KafakConsumer will not be consumed evenly from two partitions.