[jira] [Commented] (CASSANDRA-3690) Streaming CommitLog backup

["Jonathan Ellis (Commented) (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CASSANDRA-3690?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13247729#comment-13247729 ] 

Jonathan Ellis commented on CASSANDRA-3690:
-------------------------------------------

I've made a bunch of minor changes and pushed to https://github.com/jbellis/cassandra/branches/3690-v3.

I noticed that we need to wait for the archive to finish whether we end up recycling or not.  Seems to me it would be simpler to continue to always recycle, but (as we have here) wait for the archive first.  So archive can copy off to s3 or whatever directly, instead of ln somewhere else as an intermediate step.  Total i/o will be lower and commitlog will create extra segments if needed in the meantime.

Maybe we should also have a restore_list_segments command as well, so we can query s3 (again for instance) directly and have restore_command pull from there, rather than requiring a local directory?
                
> Streaming CommitLog backup
> --------------------------
>
>                 Key: CASSANDRA-3690
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-3690
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Tools
>            Reporter: Vijay
>            Assignee: Vijay
>            Priority: Minor
>             Fix For: 1.1.1
>
>         Attachments: 0001-CASSANDRA-3690-v2.patch, 0001-Make-commitlog-recycle-configurable.patch, 0002-support-commit-log-listener.patch, 0003-helper-jmx-methods.patch, 0004-external-commitlog-with-sockets.patch, 0005-cmmiting-comments-to-yaml.patch
>
>
> Problems with the current SST backups
> 1) The current backup doesn't allow us to restore point in time (within a SST)
> 2) Current SST implementation needs the backup to read from the filesystem and hence additional IO during the normal operational Disks
> 3) in 1.0 we have removed the flush interval and size when the flush will be triggered per CF, 
>           For some use cases where there is less writes it becomes increasingly difficult to time it right.
> 4) Use cases which needs BI which are external (Non cassandra), needs the data in regular intervals than waiting for longer or unpredictable intervals.
> Disadvantages of the new solution
> 1) Over head in processing the mutations during the recover phase.
> 2) More complicated solution than just copying the file to the archive.
> Additional advantages:
> Online and offline restore.
> Close to live incremental backup.
> Note: If the listener agent gets restarted, it is the agents responsibility to Stream the files missed or incomplete.
> There are 3 Options in the initial implementation:
> 1) Backup -> Once a socket is connected we will switch the commit log and send new updates via the socket.
> 2) Stream -> will take the absolute path of the file and will read the file and send the updates via the socket.
> 3) Restore -> this will get the serialized bytes and apply's the mutation.
> Side NOTE: (Not related to this patch as such) The agent which will take incremental backup is planned to be open sourced soon (Name: Priam).

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira