jackrabbit-oak-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Andrei Dulceanu (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (OAK-6888) Flushing the FileStore might return before data is persisted
Date Thu, 02 Nov 2017 12:12:01 GMT

    [ https://issues.apache.org/jira/browse/OAK-6888?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16235631#comment-16235631
] 

Andrei Dulceanu commented on OAK-6888:
--------------------------------------

[~frm], I was wondering how the current version of the patch solves the general case, since
currently the {{TarMK flush}} thread still uses {{tryFlush}}. In {{DataStoreTestBase#testSync}}
and {{DataStoreTestBase#testSyncBigBlob}} there are calls to {{primary.flush()}} before the
sync happens. This forces the flush on the primary, before that single client sync run. 

My question is: how are we protecting ourselves from scenarios like the one you already described,
if we still use {{FileStore#tryFlush}} in the {{TarMK flush}} thread? Doesn't this defeat
the whole purpose of this fix? OTOH, waiting for every flush to succeed (w, w/o a cold standby
attached) will have a tremendous impact on performance, right?

/cc [~mduerig]

> Flushing the FileStore might return before data is persisted
> ------------------------------------------------------------
>
>                 Key: OAK-6888
>                 URL: https://issues.apache.org/jira/browse/OAK-6888
>             Project: Jackrabbit Oak
>          Issue Type: Bug
>          Components: segment-tar
>            Reporter: Francesco Mari
>            Assignee: Francesco Mari
>            Priority: Major
>             Fix For: 1.8, 1.7.11
>
>         Attachments: failure.txt
>
>
> The implementation of {{FileStore#flush}} might return before all the expected data is
persisted on disk. 
> The root cause of this behaviour is the implementation of {{TarRevisions#flush}}, which
is too lenient when acquiring the lock for the journal file. If a background flush operation
is in progress and a user calls {{FileStore#flush}}, that method will immediately return because
the lock of the journal file is already owned by the background flush operation. The caller
doesn't have the guarantee that everything committed before {{FileStore#flush}} is persisted
to disk when the method returns. 
> A fix for this problem might be to create an additional implementation of flush. The
current implementation, needed for the background flush thread, will not be exposed to the
users of {{FileStore}}. The new implementation of {{TarRevisions#flush}} should have stricter
semantics and always guarantee that the persisted head contains everything visible to the
user of {{FileStore}} before the flush operation was started.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Mime
View raw message