[jira] [Commented] (AURORA-197) aurora client fails in rollback if rolled-back instances already dead

["Bill Farner (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/AURORA-197?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14008016#comment-14008016 ] 

Bill Farner commented on AURORA-197:
------------------------------------

There's a good chance this was inadvertently fixed in https://reviews.apache.org/r/20732/.  [~wickman], what do you think?

> aurora client fails in rollback if rolled-back instances already dead
> ---------------------------------------------------------------------
>
>                 Key: AURORA-197
>                 URL: https://issues.apache.org/jira/browse/AURORA-197
>             Project: Aurora
>          Issue Type: Bug
>          Components: Client
>            Reporter: brian wickman
>
> User reported:
> {noformat}
> I was trying to test our deployment failure and rollback today, and this happened:
>  INFO] Detected RUNNING instance 0
>  INFO] Instance 0 is unhealthy
> ERROR] Failed instances: set([0])
>  WARN] Not restarting failed instances [0], which exceeded maximum allowed instance failure limit of 0
> ERROR] 1 failed instances observed, maximum allowed is 0
> ERROR] 1 instance failures for instance 0, maximum allowed is 0
> ERROR] 1 failed instances observed, maximum allowed is 0
> ERROR] 1 instance failures for instance 0, maximum allowed is 0
>  INFO] Reverting update for [0]
>  INFO] Examining instances: [0]
>  INFO] Killing instances: [0]
>  INFO] Response from scheduler: ERROR (message: Aborting update without rollback! Fatal error: No jobs to kill)
> $
> So instance 0 failed (as expected; I was testing this), then the aurora client tries to kill the failed instance. That fails because the failed instance is already dead (maybe a race condition?), so it gives up on rolling back.
> Did I interpret that correctly? If so, is that expected? If so, is there a way to get it to gracefully roll back in this case (maybe via update configuration)? I can paste more details if it's helpful, but here's my update config:
> UpdateConfig(restart_threshold = 120, watch_secs = 60, batch_size = 1)
> The failure also left a lock on the deployment, which is a bit of a pain.
> Thanks for the help,
> {noformat}



--
This message was sent by Atlassian JIRA
(v6.2#6252)