Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Incomplete
-
2.3.0
-
None
Description
We've recently run into a few instances where a downed node has led to incomplete data, causing correctness issues, which we can reproduce some of the time.
Setup:
- we're currently on spark 2.3.0
- we allow retries on failed tasks and stages
- we use PySpark to perform these operations
Stages:
Simplistically, the job does the following:
- Stage 1/2: computes a number of `(sha256 hash, 0, 1)` partitioned into 65536 partitions
- Stage 3/4: computes a number of `(sha256 hash, 1, 0)` partitioned into 6408 partitions (one hash may exist in multiple partitions)
- Stage 5:
- repartitions stage 2 and stage 4 by the first 2 bytes of each hash, and find which ones are not in common (stage 2 hashes - stage 4 hashes).
- store this partition into a persistent data source.
Failure Scenario:
- We take out one of the machines (do a forced shutdown, for example)
- For some tasks, stage 5 will die immediately with one of the following:
- `ExecutorLostFailure (executor 24 exited caused by one of the running tasks) Reason: worker lost`
- `FetchFailed(BlockManagerId(24, [redacted], 36829, None), shuffleId=2, mapId=14377, reduceId=48402, message=`
- these tasks are reused to calculate stage 1-2 and 3-4 again that were missing on downed nodes, which is correctly recalculated by spark.
- However, some tasks still continue executing from Stage 5, seemingly missing stage 4 data, dumping incorrect data to the stage 5 data source. We noticed the subtract operation taking ~1-2 minutes after the machine goes down, and stores a lot more data than usual (which on inspection is wrong).
- we've seen this happen with slightly different execution plans too which don't involve or-ing, but end up being some variant of missing some stage 4 data.
However, we cannot reproduce this consistently - sometimes all tasks fail gracefully. Correctly downed nodes means all these tasks fail and re-work on stage 1-2/3-4. Note that this solution produces the correct results if machines stay alive!
We were wondering if a machine going down can result in a state where a task could keep executing even though not all data has been fetched which gives us incorrect results (or if there is setting that allows this - we tried scanning spark configs up and down). This seems similar to https://issues.apache.org/jira/browse/SPARK-24160 (maybe we get an empty packet?), but it doesn't look like that was to explicitly resolve any known bug.