[Serve] Fail early for user app failure and expose failure reasons #3411
Conversation
…il-early-for-user-error
…arly-for-user-error
sky/serve/autoscalers.py
Outdated
| if info.is_ready: | ||
| self.latest_version_ever_ready = self.latest_version | ||
| elif (info.status_property.unrecoverable_failure() and | ||
| self.latest_version_ever_ready < self.latest_version): | ||
| # Stop scaling if one of replica of the latest version | ||
| # failed, it is likely that a fatal error happens to the | ||
| # user application and may lead to a infinte termination | ||
| # and restart. | ||
| return [] |
There was a problem hiding this comment.
Should we do the elif in a new for loop? If replica 1 is unrecoverable and replica 2 is ready, and the order in replica_infos is [replica_1, replica_2], iiuc here we will first execute the loop body for replica 1 and stop scaling, which is not expected as the replica 2 is ready. instead, should we loop for all replicas and update the self.latest_version_every_ready first?
There was a problem hiding this comment.
Good point! When unrecoverable_failure() happens, it is likely all the replicas will fail, but changing it to two loops should be safer.
There was a problem hiding this comment.
just wondering some edge case like the initial delay seconds is about to miss and dependent on network speed 🤔
There was a problem hiding this comment.
Good point! I suppose the latest two loops should guard the concerns : )
| @@ -333,17 +340,17 @@ def to_replica_status(self) -> serve_state.ReplicaStatus: | |||
| return serve_state.ReplicaStatus.FAILED | |||
There was a problem hiding this comment.
| return serve_state.ReplicaStatus.FAILED | |
| return serve_state.ReplicaStatus.FAILED_USER_APP |
should we update this to a status w/ more information as well?
There was a problem hiding this comment.
I am thinking to use FAILED to always mean FAILED_USER_APP, as FAILED represents user app error in managed spot job. Wdyt? cc'ing @concretevitamin @romilbhardwaj for some input here as well : )
There was a problem hiding this comment.
Oh if it is the case, agreed that we keep align w/ spot jobs ;)
| @@ -333,17 +340,17 @@ def to_replica_status(self) -> serve_state.ReplicaStatus: | |||
| return serve_state.ReplicaStatus.FAILED | |||
There was a problem hiding this comment.
Oh if it is the case, agreed that we keep align w/ spot jobs ;)
Co-authored-by: Tian Xia <cblmemo@gmail.com>
Co-authored-by: Tian Xia <cblmemo@gmail.com>
Co-authored-by: Tian Xia <cblmemo@gmail.com>
…arly-for-user-error
…kypilot into fail-early-for-user-error
Tested (run the relevant ones):
bash format.shpytest tests/test_smoke.pypytest tests/test_smoke.py --servepytest tests/test_smoke.py::test_fill_in_the_namepytest tests/test_smoke.py::test_skyserve_user_bug_restartpytest tests/test_smoke.py::test_skyserve_failurespytest tests/test_smoke.py --servebash tests/backward_comaptibility_tests.sh