[
https://issues.apache.org/jira/browse/ZOOKEEPER-3220?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16729308#comment-16729308
]
Jiafu Jiang commented on ZOOKEEPER-3220:
----------------------------------------
[~maoling]
why this situation happend? The disk is full?
No, but the machine restarted.
Do you see some logs about *FileTxnSnapLog#save* at that time?
No any error log, if fact, during the machine reboot, some log of the follower
was missing. But from the log of the leader, the follower had received a
snapshot and began to received other transaction logs, so the
*FileTxnSnapLog#save of follower must have succeed, but the data is not in
disk!*
*2.Even if this situation that the size of snapshot is 0 could not cause data
inconsistency.*
Yes, I know. Zookeeper recover it's data from both logs and snapshot.
If a ZooKeeper follower believes a snapshot is saved, it believes that the data
in the snapshot is all in the disk(but in fact it may be not), it will begin to
receive logs that come after the snapshot. If the snapshot is invalid,
ZooKeeper server will recover data from logs only, but some data is missing,
because the data is only saved in the snapshot.
> The snapshot is not saved to disk and may cause data inconsistency.
> -------------------------------------------------------------------
>
> Key: ZOOKEEPER-3220
> URL: https://issues.apache.org/jira/browse/ZOOKEEPER-3220
> Project: ZooKeeper
> Issue Type: Bug
> Components: server
> Affects Versions: 3.4.12, 3.4.13
> Reporter: Jiafu Jiang
> Priority: Critical
>
> We known that ZooKeeper server will call fsync to make sure that log data has
> been successfully saved to disk. But ZooKeeper server does not call fsync to
> make sure that a snapshot has been successfully saved, which may cause
> potential problems. Since a close to a file description does not make sure
> that data is written to disk, see
> [http://man7.org/linux/man-pages/man2/close.2.html#notes] for more details.
>
> If the snapshot is not successfully saved to disk, it may lead to data
> inconsistency. Here is my example, which is also a real problem I have ever
> met.
> 1. I deployed a 3-node ZooKeeper cluster: zk1, zk2, and zk3, zk2 was the
> leader.
> 2. Both zk1 and zk2 had the log records from log1~logX, X was the zxid.
> 3. The machine of zk1 restarted, and during the reboot, log(X+1) ~ log Y are
> saved to log files of both zk2(leader) and zk3(follower).
> 4. After zk1 restarted successfully, it found itself to be a follower, and it
> began to synchronize data with the leader. The leader sent a snapshot(records
> from log 1 ~ log Y) to zk1, zk1 then saved the snapshot to local disk by
> calling the method ZooKeeperServer.takeSnapshot. But unfortunately, when the
> method returned, the snapshot data was not saved to disk yet. In fact the
> snapshot file was created, but the size was 0.
> 5. zk1 finished the synchronization and began to accept new requests from the
> leader. Say log records from log(Y + 1) ~ log Z were accepted by zk1 and
> saved to log file. With fsync zk1 could make sure log data was not lost.
> 6. zk1 restarted again. Since the snapshot's size was 0, it would not be
> used, therefore zk1 recovered using the log files. But the records from
> log(X+1) ~ logY were lost !
>
> Sorry for my poor English.
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
