[
https://issues.apache.org/jira/browse/FLINK-19773?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17227353#comment-17227353
]
Chesnay Schepler commented on FLINK-19773:
------------------------------------------
This makes sense to me. It would require a new {{RestartBackoffTimeStrategy}}
implementation that is similar to the {{FixedDelayRestartBackoffTimeStrategy}}.
The tricky part will be to reset the backoff at some point; for example, if you
have a restart loop at some point, you don't want to increased backoff to still
be used after a week of running smoothly.
There is currently no real hook to do this, but it could I suppose be some form
of configurable timeout, e.g., reset the backoff if the job has not failed for
T time.
> Exponential backoff restart strategy
> ------------------------------------
>
> Key: FLINK-19773
> URL: https://issues.apache.org/jira/browse/FLINK-19773
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Coordination
> Affects Versions: 1.11.2
> Reporter: Levi Ramsey
> Priority: Minor
> Labels: starter
>
> There are situations where the current restart strategies (fixed-delay and
> failure-rate) seem to be suboptimal. For example, in HDFS sinks, a delay
> between restarts shorter than the lease expiration time in HDFS is going to
> result in many restart attempts which fail, putting somewhat pointless stress
> on a cluster. On the other hand, setting a delay of close to the lease
> expiration time will mean far more downtime than necessary when the cause of
> failure is something that works itself out quickly.
>
> An exponential backoff restart strategy would address this. For example a
> backoff strategy where the jobs are contending for a lease on a shared
> resource that terminates after 1200 seconds of inactivity might have
> successive delays of 1, 2, 4, 8, 16... 1024 seconds (after which a cumulative
> delay of more than 1200 seconds has passed).
> While not intrinsically tied to exponential backoff (it's more of an example
> of variable delay), in the case of many jobs failing due to an infrastructure
> failure, a thundering herd scenario can be mitigated by adding jitter to the
> delays, e.g. 0 -> 1 -> 2 -> 3/4/5 -> 5/6/7/8/9/10/11 seconds. With this
> jitter, eventually a set of jobs competing to restart will spread out.
> (logging the ticket more to start a discussion and perhaps get context around
> if this had been considered and rejected, etc.)
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
