[
https://issues.apache.org/jira/browse/RATIS-1804?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Tsz-wo Sze updated RATIS-1804:
------------------------------
Description:
The following is an observation from [~William Song]:
{quote}In some of our scenarios, this problem happens when a new leader is
elected and start sending AppendEntries to a slow follower. Unfortunately, this
slow follower is going through a GC pause, so these AppendEntries just
accumulates at gRPC buffer. When the slow follower comes around and starts
dealing with these piled RPCs, it immediately finds the first AppendEntries
nextIndex is inconsistent. So it replies INCONSISTENT to all of these
accumulated AE RPCs, all with the same nextIndex.
{quote}
See [https://github.com/apache/ratis/pull/832#issuecomment-1451186915]
was:
The following an observation from [~William Song]:
bq. In some of our scenarios, this problem happens when a new leader is elected
and start sending AppendEntries to a slow follower. Unfortunately, this slow
follower is going through a GC pause, so these AppendEntries just accumulates
at gRPC buffer. When the slow follower comes around and starts dealing with
these piled RPCs, it immediately finds the first AppendEntries nextIndex is
inconsistent. So it replies INCONSISTENT to all of these accumulated AE RPCs,
all with the same nextIndex.
See https://github.com/apache/ratis/pull/832#issuecomment-1451186915
> Limit the number of outstanding append entires
> ----------------------------------------------
>
> Key: RATIS-1804
> URL: https://issues.apache.org/jira/browse/RATIS-1804
> Project: Ratis
> Issue Type: Improvement
> Components: server
> Reporter: Tsz-wo Sze
> Priority: Major
>
> The following is an observation from [~William Song]:
> {quote}In some of our scenarios, this problem happens when a new leader is
> elected and start sending AppendEntries to a slow follower. Unfortunately,
> this slow follower is going through a GC pause, so these AppendEntries just
> accumulates at gRPC buffer. When the slow follower comes around and starts
> dealing with these piled RPCs, it immediately finds the first AppendEntries
> nextIndex is inconsistent. So it replies INCONSISTENT to all of these
> accumulated AE RPCs, all with the same nextIndex.
> {quote}
> See [https://github.com/apache/ratis/pull/832#issuecomment-1451186915]
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
