Hi Gary,
Thanks for answering my question. I thought that could be a tough
question as Cassandra is quickly evolving recently, but it was great
to know that you guys are always trying to make rolling upgrade
possible for minor upgrades.
My friend is evaluating Cassandra as an HA alterative of enterprise
service bus. He found Cassandra can be ideal for receiveing real-time
feeds from hundreds of systems in the company and passing the feeds to
Hadoop Map Reduce jobs. I thought it's very helpful for him if he can
run rolling upgrade for minor bug fixes.
Thanks,
Tatsuya
On May 11, 2010, at 10:05 PM, Gary Dusbabek <[email protected]> wrote:
On Mon, May 10, 2010 at 17:01, Tatsuya Kawano
<[email protected]> wrote:
Hi,
Does Cassandra support rolling restart recipe between minor version
upgrade? I mean rolling restart is a way to upgrade Cassandra version
or change configuration **without** bringing down the whole cluster.
The recipe will be something like killing a couple of nodes at a time
and starting them again with newer version (or newer configuration.)
We aim for this, but fell short on the 0.6 -> 0.6.1 upgrade. A change
made it in that broker inter-node communication across versions.
Sorry.
To make this possible, I think at least the RPC and SSTable format
should be compatible between older and newer versions. Do you have
any
plan to change them in the future? Also, there might be some
exceptions and certain combinations of versions can't coexist in a
cluster (e.g. RPC or SSTable bug in the older version.) It will be
great if such incompatibilities are documented in the release note.
RPC has versioning built in, so clients should be able to communicate
with servers that share the same major version. Note that this is not
the same as the inter-node message format that nodes use to
communicate with each other.
The SSTable format has been stable since 0.4 (compatible with 0.5 and
0.6). There are some breaking changes scheduled for 0.7, but none of
them have made it into trunk yet.
The commit log disk format has changed and will continue to change
from version to version. You should flush it prior to upgrading a
node. 0.6.1 added 'drain' to the nodetool command to make this
process easier.
Gary.
Thanks,
Tatsuya