[Serve] Scale down old terminal replicas

[cblmemo]

Previous implementation will keep old terminated (i.e. failed) replicas forever. This PR changes the behaviour to scale down old terminal replicas as well when an update is initiated.

Tested (run the relevant ones):

• Code formatting: bash format.sh
• Any manual or new tests for this PR (please specify below)

$ sky serve up tests/skyserve/failures/initial_delay.yaml -n fail-2
$ sky serve status
Services
NAME    VERSION  UPTIME  STATUS      REPLICAS  ENDPOINT                                                                  
fail-2  -        -       NO_REPLICA  0/0       http://localhost:30072/skypilot/sky-serve-controller-402b1bba-402b/30001  

Service Replicas
SERVICE_NAME  ID  VERSION  ENDPOINT  LAUNCHED  RESOURCES  STATUS                REGION  
fail-2        1   1        -         -         -          FAILED_INITIAL_DELAY  -       
$ sky serve update fail-2 tests/skyserve/failures/initial_delay.yaml
$ sky serve status
Services
NAME    VERSION  UPTIME  STATUS      REPLICAS  ENDPOINT                                                                  
fail-2  -        -       NO_REPLICA  0/2       http://localhost:30072/skypilot/sky-serve-controller-402b1bba-402b/30001  

Service Replicas
SERVICE_NAME  ID  VERSION  ENDPOINT  LAUNCHED     RESOURCES              STATUS         REGION      
fail-2        1   1        -         -            -                      SHUTTING_DOWN  -           
fail-2        2   2        -         30 secs ago  1x Kubernetes(vCPU=2)  STARTING       Kubernetes 

• All smoke tests: pytest tests/test_smoke.py
• Relevant individual smoke tests: pytest tests/test_smoke.py::test_fill_in_the_name
• Backward compatibility tests: conda deactivate; bash -i tests/backward_compatibility_tests.sh

[Michaelvll]

Thanks for submitting the PR @cblmemo! Left several questions.

[Michaelvll]

This refactoring seems not necessary.

Suggested change

	all_replica_ids_to_scale_down.extend(
	self._select_outdated_nonterminal_replicas_to_scale_down(
	replica_infos))
	return all_replica_ids_to_scale_down
	def _select_outdated_nonterminal_replicas_to_scale_down(
	self,
	replica_infos: List['replica_managers.ReplicaInfo']) -> List[int]:
	"""Select outdated nonterminal replicas to scale down."""

[Michaelvll]

IIRC the original behavior will terminate the failed replicas already. Is this only for removing the FAILED replica entries in the serve status table? In that case, should we make the sky serve down command able to remove a FAILED entry in the table instead?

I feel keeping the FAILED replicas for the previous version might be still useful.

[cblmemo]

Yes, it is only for removing those failed records from a user's request. Agree that adding the ability to remove a record in sky serve down is a better idea. It is implemented in #3179 and I'll take a look soon 🫡

[github-actions]

This PR is stale because it has been open 120 days with no activity. Remove stale label or comment or this will be closed in 10 days.