[
https://issues.apache.org/jira/browse/HADOOP-11398?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14627032#comment-14627032
]
Li Lu commented on HADOOP-11398:
--------------------------------
bq. We cannot assume t1=t2 because t2 becomes larger than t1 when there are
underlying retries.
If this is the case, then only allowing retries to t1+delta_t is wrong because
we may accidentally but completely disable retry when t1 and t2 are
significantly different. What we want is to "retry for time t after an error
occurred" but not to "retry for time t after the retry object is initialized".
> RetryUpToMaximumTimeWithFixedSleep needs to behave more accurately
> ------------------------------------------------------------------
>
> Key: HADOOP-11398
> URL: https://issues.apache.org/jira/browse/HADOOP-11398
> Project: Hadoop Common
> Issue Type: Bug
> Reporter: Li Lu
> Assignee: Li Lu
> Attachments: HADOOP-11398-121114.patch, HADOOP-11398.002.patch
>
>
> RetryUpToMaximumTimeWithFixedSleep now inherits
> RetryUpToMaximumCountWithFixedSleep and just acts as a wrapper to decide
> maxRetries. The current implementation uses (maxTime / sleepTime) as the
> number of maxRetries. This is fine if the actual for each retry is
> significantly less than the sleep time, but it becomes less accurate if each
> retry takes comparable amount of time as the sleep time. The problem gets
> worse when there are underlying retries.
> We may want to use timers inside RetryUpToMaximumTimeWithFixedSleep to
> perform accurate timing.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
