### What changes were proposed in this pull request?
The DecommissionWorkerSuite started becoming flaky and it revealed a real regression. Recently closed#29211 necessitates remembering the decommissioning shortly beyond the removal of the executor.
In addition to fixing this issue, ensure that DecommissionWorkerSuite continues to pass when executors haven't had a chance to exit eagery. That is the old behavior before #29211 also still works.
Added some more tests to TaskSchedulerImpl to ensure that the decommissioning information is indeed purged after a timeout.
Hardened the test DecommissionWorkerSuite to make it wait for successful job completion.
### Why are the changes needed?
First, let me describe the intended behavior of decommissioning: If a fetch failure happens where the source executor was decommissioned, we want to treat that as an eager signal to clear all shuffle state associated with that executor. In addition if we know that the host was decommissioned, we want to forget about all map statuses from all other executors on that decommissioned host. This is what the test "decommission workers ensure that fetch failures lead to rerun" is trying to test. This invariant is important to ensure that decommissioning a host does not lead to multiple fetch failures that might fail the job. This fetch failure can happen before the executor is truly marked "lost" because of heartbeat delays.
- However, #29211 eagerly exits the executors when they are done decommissioning. This removal of the executor was racing with the fetch failure. By the time the fetch failure is triggered the executor is already removed and thus has forgotten its decommissioning information. (I tested this by delaying the decommissioning). The fix is to keep the decommissioning information around for some time after removal with some extra logic to finally purge it after a timeout.
- In addition the executor loss can also bump up `shuffleFileLostEpoch` (added in #28848). This happens because when the executor is lost, it forgets the shuffle state about just that executor and increments the `shuffleFileLostEpoch`. This incrementing precludes the clearing of state of the entire host when the fetch failure happens because the failed task is still reusing the old epoch. The fix here is also simple: Ignore the `shuffleFileLostEpoch` when the shuffle status is being cleared due to a fetch failure resulting from host decommission.
I am strategically making both of these fixes be very local to decommissioning to avoid other regressions. Especially the version stuff is tricky (it hasn't been fundamentally changed since it was first introduced in 2013).
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
Manually ran DecommissionWorkerSuite several times using a script and ensured it all passed.
### (Internal) Configs added
I added two configs, one of which is sort of meant for testing only:
- `spark.test.executor.decommission.initial.sleep.millis`: Initial delay by the decommissioner shutdown thread. Default is same as before of 1 second. This is used for testing only. This one is kept "hidden" (ie not added as a constant to avoid config bloat)
- `spark.executor.decommission.removed.infoCacheTTL`: Number of seconds to keep the removed executors decom entries around. It defaults to 5 minutes. It should be around the average time it takes for all of the shuffle data to be fetched from the mapper to the reducer, but I think that can take a while since the reducers also do a multistep sort.
Closes#29422 from agrawaldevesh/decom_fixes.
Authored-by: Devesh Agrawal <devesh.agrawal@gmail.com>
Signed-off-by: Wenchen Fan <wenchen@databricks.com>
1ac23dea52