Description
Summary:
- I ran into this issue using AWS EMR.
- Frequency of the issue varies. On one EMR cluster, I typically see ~10-20% chance of hitting the issue. But on another EMR cluster, the chance is ~1%. I suspect the chance depends on how busy AWS hardware actually was (my EMR likely share hardware resources with other AWS tenants).
- I believe that I have identify the root cause in Livy source code (refer to a later section).
How to reproduce:
- An EMR with Spark, Yarn and Livy configured.
- Use the attached livy_batch.py to trigger a Livy batch by using livy python client (0.8.0). See attached livy_client.py.
- Repeat the testing and you should see when the issue happens, even though the spark program errors out, Livy still reports the batch as SUCCEED.
Livy log for a good case when Livy returns batch as DEAD (expected behavior):
22/10/14 02:46:22 INFO BatchSessionManager: Registered new session 1
22/10/14 02:46:42 DEBUG BatchSession: BatchSession 1 state changed from STARTING to RUNNING
22/10/14 02:46:43 WARN BatchSession$: spark-submit exited with code 1
22/10/14 02:46:47 DEBUG BatchSession: BatchSession 1 state changed from RUNNING to FINISHED
22/10/14 02:46:47 DEBUG BatchSession: BatchSession 1 state changed from FINISHED to FAILED
Livy log for bad case when Livy returns batch as SUCCEED (bug):
22/10/14 02:47:40 INFO BatchSessionManager: Registered new session 3
22/10/14 02:48:00 DEBUG BatchSession: BatchSession 3 state changed from STARTING to FINISHED
22/10/14 02:48:01 WARN BatchSession$: spark-submit exited with code 1
Root cause analysis:
- I think the bug is in the YarnAppMonitorThread of Livy server.
- The bug is in this section: https://github.com/apache/incubator-livy/blob/v0.7.1-incubating-rc1/server/src/main/scala/org/apache/livy/utils/SparkYarnApp.scala#L283-L299
- When the right timing happens,
- Yarn sees the application completed
- But spark-submit process is still running
- So line 286 returns app report saying the application completes successfully (Yarn has not context of the Spark application succeeds or not).
- Line 288 set the Livy session as succeed.
- Line 293 returns false as the spark-submit is still running.
- So we hit the bug.
Even without hitting the timing condition, the code logic itself is still incorrect.
If you take a look at the log from a "good" case, the session state was updated twice: FINISHED, then FAILED. If a client query arrives on the perfect timing, the livy server could can still return a wrong state.
22/10/14 02:46:47 DEBUG BatchSession: BatchSession 1 state changed from RUNNING to FINISHED 22/10/14 02:46:47 DEBUG BatchSession: BatchSession 1 state changed from FINISHED to FAILED
I hope we can work together to have the issue addressed ASAP as the bug hit our production code pretty bad. I think the right code logic should be:
- read the spark-submit process's state, if still running, do nothing
- If the spark-submit process finishes, read Yarn report, and determines the actual application finish state in a single shot.
- Update the session state in a single step.
At the same time, I will see if I can create a PR with suggested fix soon. The challenge on my side is that it's almost impossible for me to swap a few jars from open-source code base on AWS EMR (not compatible with EMR runtime).
Thank you, Livy team!
Regards,
Jeff Xu, a Workday engineer