Yes, I think this approach would be more efficient, but Ian's point about failed runs is well taken. It is still a problem with this approach. I may have to introduce a scheme where Hadoop's output is written to a new column family, and then some sort of pointer is updated to point to this column family at the end of a successful run. That way, if a run failed, the pointer would still point to the old data, and we could just rerun the update task. (Older and failed versions of the data would have to be periodically purged as well.)
Ian: I think that as get_range_slice gets faster, the approach that Mark was heading toward may be considerably more efficient than reading the old value in the OutputFormat.
Mark: Reading all of the data you want to update out of Cassandra using the InputFormat, merging it with (tagged) new data, and then performing deletes/inserts (without reads) in a Cassandra OutputFormat makes a lot of sense to me. It gives you much better data locality: all reads happen in a streaming fashion (mostly: this is where we need improvement for get_range_slices) at the beginning of the job.
From: "Ian Kallen" <firstname.lastname@example.org>
Sent: Thursday, May 6, 2010 5:14pm
Subject: Re: Updating (as opposed to just setting) Cassandra data via Hadoop
I have inputs that are text logs and I wrote a Cassandra OutputFormat, the
reducers read the old values from their respective column families,
increment the counts and write back the new values. Since all of the writes
are done by the hadoop jobs and we're not running multiple jobs
concurrently, the writes aren't clobbering any values except the ones from
the prior hadoop run. If/when atomic increments are available, we'd be able
to run concurrent log processing jobs for but for now, this seems to work. I
think the biggest risk is that a reduce task fails, hadoop restarts it and
the replacement task re-increments the values. We're going to wait and see
to determine how much of a problem this is in practice.
On Tue, May 4, 2010 at 8:53 PM, Mark Schnitzius
> I have a situation where I need to accumulate values in Cassandra on an
> ongoing basis. Atomic increments are still in the works apparently (see
> https://issues.apache.org/jira/browse/CASSANDRA-721) so for the time being
> I'll be using Hadoop, and attempting to feed in both the existing values and
> the new values to a M/R process where they can be combined together and
> written back out to Cassandra.
> The approach I'm taking is to use Hadoop's CombineFileInputFormat to blend
> the existing data (using Cassandra's ColumnFamilyInputFormat) with the newly
> incoming data (using something like Hadoop's SequenceFileInputFormat).
> I was just wondering, has anyone here tried this, and were there issues?
> I'm worried because the CombineFileInputFormat has restrictions around
> splits being from different pools so I don't know how this will play out
> with data from both Cassandra and HDFS. The other option, I suppose, is to
> use a separate M/R process to replicate the data onto HDFS first, but I'd
> rather avoid the extra step and duplication of storage.
> Also, if you've tackled a similar situation in the past using a different
> approach, I'd be keen to hear about it...