Robert Joseph Evans commented on ZOOKEEPER-2845:
Perhaps I don't understand the issue well enough which is totally possible
because I am not a frequent contributor and the path for all of the request
processors is kind of complex.
My understanding is that the SyncRequestProcessor handles writing out edits to
the edit log and snapshots, there are a few other places where this happens at
startup though. The SyncRequestProcessor writes out edits as they arrive and
will flush them to disk periodically in batches. It also takes snapshots
The in memory portion appears to be updated by the FinalRequestProcessor prior
to a quorum of acks being received.
So yes there is the possibility that something is written to the transaction
log that is not applied to memory. This means that when ZKDatabase.clear() is
called it should actually fast forward the in memory changes to match those in
the edit log + snapshot.
So you are saying that
1) proposals come in, are written to the transaction log, but the in memory
database is not updated yet.
2) the server does a soft restart for some reason and some transactions appear
to be lost (because the in memory DB was not fast forwarded).
3) more transactions come in (possibly conflicting with the first set of
4) before a snapshot can happen the leader or follower restarts and has to
reconstruct the in memory DB from edits + snapshot. This would then reapply the
edits that originally appeared to be lost.
This does look like it might happen, so I will look into that as well.
But the test in [https://github.com/apache/zookeeper/pull/310] didn't appear to
trigger this. I could be wrong because I concentrated most of my debugging on
the original leader and what was happening with it, instead of the followers
and what was happening with them. I also didn't understand how clearing the
leader's in memory database caused an edit to be lost, if the edits are being
written out to disk before the in memory DB is updated. What I saw was that
1) a bunch of edits and leaders/followers being restarted that didn't really do
much of anything.
2) the original leader lost a connection to the followers.
3a) A transaction was written to the leader in memory DB but it didn't get a
quorum of acks
3b) The followers restarted and formed a new quorum
4) The original leader timed out and joined the new quorum
5) As part of the sync when the old leader joined the new quorum it got a diff
(not a snap), but it had an edit that was not a part of the new leader so it
was off from the others.
I could see this second part happening even without my change so I don't really
understand how that clearing the database would prevent it. My thinking was
that it was a race condition where the edits in the edit log were not flushed
yet, and as such when we cleared the DB they were lost. But I didn't confirm
> Data inconsistency issue due to retain database in leader election
> Key: ZOOKEEPER-2845
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-2845
> Project: ZooKeeper
> Issue Type: Bug
> Components: quorum
> Affects Versions: 3.4.10, 3.5.3, 3.6.0
> Reporter: Fangmin Lv
> Assignee: Robert Joseph Evans
> Priority: Critical
> In ZOOKEEPER-2678, the ZKDatabase is retained to reduce the unavailable time
> during leader election. In ZooKeeper ensemble, it's possible that the
> snapshot is ahead of txn file (due to slow disk on the server, etc), or the
> txn file is ahead of snapshot due to no commit message being received yet.
> If snapshot is ahead of txn file, since the SyncRequestProcessor queue will
> be drained during shutdown, the snapshot and txn file will keep consistent
> before leader election happening, so this is not an issue.
> But if txn is ahead of snapshot, it's possible that the ensemble will have
> data inconsistent issue, here is the simplified scenario to show the issue:
> Let's say we have a 3 servers in the ensemble, server A and B are followers,
> and C is leader, and all the snapshot and txn are up to T0:
> 1. A new request reached to leader C to create Node N, and it's converted to
> txn T1
> 2. Txn T1 was synced to disk in C, but just before the proposal reaching out
> to the followers, A and B restarted, so the T1 didn't exist in A and B
> 3. A and B formed a new quorum after restart, let's say B is the leader
> 4. C changed to looking state due to no enough followers, it will sync with
> leader B with last Zxid T0, which will have an empty diff sync
> 5. Before C take snapshot it restarted, it replayed the txns on disk which
> includes T1, now it will have Node N, but A and B doesn't have it.
> Also I included the a test case to reproduce this issue consistently.
> We have a totally different RetainDB version which will avoid this issue by
> doing consensus between snapshot and txn files before leader election, will
> submit for review.
This message was sent by Atlassian JIRA