cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jonathan Ellis (JIRA)" <>
Subject [jira] Updated: (CASSANDRA-1602) Commit Log archivation and rolling forward utility (AKA Retaining commit logs)
Date Wed, 13 Oct 2010 01:12:33 GMT


Jonathan Ellis updated CASSANDRA-1602:

    Attachment: 1602-v2.txt

v2 attached.

The main difference is it cuts down on the special cases in CommitLog by creating a hard link
to the new segment as soon as it is created.  Then the rest of the logic can proceed oblivious
to archive mode.

> Commit Log archivation and rolling forward utility (AKA Retaining commit logs)
> ------------------------------------------------------------------------------
>                 Key: CASSANDRA-1602
>                 URL:
>             Project: Cassandra
>          Issue Type: New Feature
>          Components: Core, Tools
>            Reporter: Oleg Anastasyev
>             Fix For: 0.6.7
>         Attachments: 1602-0.6.4.txt, 1602-cassandra0.6.txt, 1602-v2.txt
> As couple of people from mailing list suggested me to share this patch (see discussion
at ) to retain (archive) commit
logs generated by cassandra and  restore data by rolling forward of commit logs to previously
backed up or snapshotted data files.
> Here is an instruction of how to use it, which i extracted from out internal wiki:
> We rely on cassandra replication factor for disaster recovery.
> But there is another problem: there are bugs in data manipulation logic, which can lead
to data destruction on the whole cluster. But the freshest backup to restore from is last
snapshot, which can be up to 24h as old.
> To fight with it, we decided to implement snapshot + log archive&nbsp; backup strategy
- i.e. we collect commit logs and snapshotted data files. On event of data loss, either due
hardware failure or logical bug, we restore last snapshot and roll forward all logs, collected
since last snapshot time to it.
> Originally cassandra does not support log archive , so I implemented it by myself.
> The idea is simple:
> # As soon as commit log file is not needed anymore by cassadra, a hardlink (unix command
"ln $1 $2") is created from just closed commit log file to commit log archive directory. Both
commit log and commit log archive are on the same volume, of course.
> # Some script (which authoring i left to admin) then takes files from commit log archive
dir and copies them over net to a backup location. As soon as file is transferred, it can
be safely deleted from commit log archive dir.
> # Dont forget there must be some script (also authored by admins), which do data snapshots
from time to time using "nodetool snapshot" command, available from standard cassandra distribution
and copies snapshot files to backup location.
> ## Creating a snapshot is very light operation for cassandra - under the hood it is just
hardlinking currently existing files to "snapshot/<timestamp-millis>" directory. So
frequence is up to our ability to pull snapshotted data files over network.
> # To restore data, admin must:
> ## stop cassandra instance
> ## remove all corrupted data files from /data directory. Leave only commit logs you want
to roll forward in /commitlog dir
> ## copy to /data last snapshot data files
> ## copy to /commitlog all archived commit log files (only files not older than snapshot
data files could be copied. copying too old will do no harm, but will extend processing time)
> ## *not starting* cassandra instance, run roll forward utility (bin/logreplay) with option
-forced or -forcedcompaction and wait for its completion
> ## then start cassandra node instance as usual.
> Log archive&nbsp; logic is activated using CommitLogArchive directive in storage-conf.xml:
> {code:xml}
> <CommitLogDirectory>/commitlog</CommitLogDirectory>
> <CommitLogArchive>true</CommitLogArchive>
> <DataFileDirectories>
> <DataFileDirectory>/data</DataFileDirectory>
> </DataFileDirectories>
> {code}Log files will be archived to <commit-dir>/.archive directory, i.e. as shown
in example above, to /commitlog/.archive directory.
> Archived logs replay process is launched by running
with option "-forced" locally on node (use option "-forcedcompact" to do major compaction
right after log roll forward process completion). I also made a script named <cassandra-dir>/bin/logreplay.

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message