[jira] [Updated] (RATIS-1707) Fix corner case when getPrevious in LogAppenderBase

["Tsz-wo Sze (Jira)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/RATIS-1707?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Tsz-wo Sze updated RATIS-1707:
------------------------------
    Fix Version/s: 2.4.0

> Fix corner case when getPrevious in LogAppenderBase
> ---------------------------------------------------
>
>                 Key: RATIS-1707
>                 URL: https://issues.apache.org/jira/browse/RATIS-1707
>             Project: Ratis
>          Issue Type: Bug
>          Components: server
>            Reporter: Xu Shao Hong
>            Assignee: Xu Shao Hong
>            Priority: Major
>             Fix For: 2.4.0, 3.0.0
>
>         Attachments: 745_review.patch, ratis.png
>
>          Time Spent: 3h 20m
>  Remaining Estimate: 0h
>
> In the cluster, we found that there are many exceptions on {*}Failed appendEntries{*}.
> _Unexpected Index: previous is null but entries[0].getIndex()=3402406. So the follower cannot append an entry due to some limit._
> The scenario:
> Follower B restarted, leader A sent the entry and B can not find the previous log entry. A sent the notifyInstallSnapshot request to B, and B found its next index is larger than the leader's firstAvailableLogIndex(the index to install snapshot). A updated B's index according to the reply and sent the entries to B.  A will find the previous entry TermIndex through ``getPrevious(long nextIndex)``, if nextIndex of raft log of B is exactly the same as startIndex of leader A (B needs the entries since A's firstAvailableLogIndex), A has purged its raft log and will check the snapshot Index through +server.getStateMachine().getLatestSnapshot()+ whether it equals to nextIndex - 1, if not then returns null. 
> The problem is due to the uncertainty of purging raft log. If A has also been stopped, and thus triggered the takeSnapshot, the raft log may not be purged up to the snapshot index. The latest snapshot index from SM is not equal to the raft log's first available index, which leads to this corner case.
> We could add a case check when getPrevious.
>  



--
This message was sent by Atlassian Jira
(v8.20.10#820010)