[
https://issues.apache.org/jira/browse/HBASE-50?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12865853#action_12865853
]
Jonathan Gray commented on HBASE-50:
------------------------------------
Also, rather than creating a mock Master and mock RegionServer, it might be
more useful to just create classes that can be used by the real Master and RS
processes.
For example, the Master class you have could become SnapshotManager or
SnapshotMonitor. The RegionServer might become SnapshotExecutor. Just
suggestions for names but I would move to making patches against the svn
repository and building stuff in a way that will make it easier to integrate
after you get through early testing.
Does your current design have it so that all RegionServers wait for every other
RS to be ready, and then they concurrently perform their snapshots? It seems
like this could be dangerous behavior. We might want to stagger snapshots
across the cluster, otherwise this process could create too much load.
The flow chart is great. Could you add some different notation about the znode
names and what things will have more than one process performing it? For
example, each RS might make the znode: /snapshot/ready/RSNAME or some such.
As far as this thing working in failure scenarios, that could potentially be
very complex (or if designed in a region-centric view rather than
regionserver-centric view maybe not so bad). Let's try to flesh out the
requirements early/now. If we decide we won't work under failure, then let's
make that explicit and ensure that under failure we can roll everything back.
If we want it to work under failure (Master and/or RegionServer), let's talk
about it more now because the basic design could have big implications.
In your flow chart, once the snapshot starts it looks like an RS won't
communicate until it finishes. If that RS is serving 1000 regions of the table
being snapshotted, this could take a long time. Does the RS need to keep a
timestamp updated to let the master/client know it's still working (and maybe
it's progress?). Also, are splits and load balancing blocked during snapshot
time across the whole cluster?
I think it would be useful to generate a design document. This would include
your current flow chart as well as some of the stuff in comments on this jira.
I'd like to have something we can continue to iterate on rather than just
comments.
Great work so far!
> Snapshot of table
> -----------------
>
> Key: HBASE-50
> URL: https://issues.apache.org/jira/browse/HBASE-50
> Project: Hadoop HBase
> Issue Type: New Feature
> Reporter: Billy Pearson
> Assignee: Li Chongxin
> Priority: Minor
> Attachments: snapshot-flowchart.png, snapshot-src.zip
>
>
> Havening an option to take a snapshot of a table would be vary useful in
> production.
> What I would like to see this option do is do a merge of all the data into
> one or more files stored in the same folder on the dfs. This way we could
> save data in case of a software bug in hadoop or user code.
> The other advantage would be to be able to export a table to multi locations.
> Say I had a read_only table that must be online. I could take a snapshot of
> it when needed and export it to a separate data center and have it loaded
> there and then i would have it online at multi data centers for load
> balancing and failover.
> I understand that hadoop takes the need out of havening backup to protect
> from failed servers, but this does not protect use from software bugs that
> might delete or alter data in ways we did not plan. We should have a way we
> can roll back a dataset.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
