[
https://issues.apache.org/jira/browse/FLINK-17871?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17238030#comment-17238030
]
fanxin edited comment on FLINK-17871 at 11/24/20, 10:44 AM:
------------------------------------------------------------
hi, [~azagrebin] [~chesnay]
Sorry for the late reply. We have a yarn cluster with about 200 nodes(128G, 32
cores). Based on our experience, it should greater than 30s. A single Flink on
yarn distribute it’s jars and launch the AppMaster will cost at least 10s, if
we don't put these jars in HDFS in advance.
was (Author: fanxiin):
hi, [~azagrebin] [~chesnay]
Sorry for the late reply. We have a yarn cluster with about 200 nodes(128G, 32
cores and hhd). Based on our experience, it should greater than 30s. A single
Flink on yarn distribute it’s jars and launch the AppMaster will cost at least
10s, if we don't put these jars in HDFS in advance.
> Make the default value of attemptFailuresValidityInterval more reasonable
> -------------------------------------------------------------------------
>
> Key: FLINK-17871
> URL: https://issues.apache.org/jira/browse/FLINK-17871
> Project: Flink
> Issue Type: Improvement
> Components: Deployment / YARN
> Reporter: fanxin
> Priority: Minor
>
> Default value of `yarn.application-attempt-failures-validity-interval` is
> `10000` milliseconds at present. Usually preparing the context alone can take
> seconds, which means that default value of 10000 is too small to even prepare
> the runtime context. With a default config, a flink on yarn job in will
> hardly meet the condition of ”fail 2 times in 10s“. If the job has some
> internal problems, unfortunately, it can easily get bogged down in endless
> retries.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
