mck created CASSANDRA-8098:
------------------------------
Summary: Allow CqlInputFormat to be restricted to more than one
data-center
Key: CASSANDRA-8098
URL: https://issues.apache.org/jira/browse/CASSANDRA-8098
Project: Cassandra
Issue Type: Improvement
Components: Hadoop
Reporter: mck
Assignee: mck
Today, using CqlInputFormat, it's only possible to
- enforce data-locality to one specific data-center, or
- disable it by changing CL from LOCAL_ONE to ONE.
We need a way to enforce data-locality to specific *data-centers*, and would
like to contribute a solution.
Suggested ideas
- CqlInputFormat (gently) calls describeLocalRing against all the listed
connection addresses and merge the results into one masterRangeNodes list, or
- changing the signature of describeLocalRing(..) to describeRings(String
keyspace, String[] dc) and having the job specify which DCs it will be running
within.
*Long description*
A lot has changed since CASSANDRA-2388 that has made life a lot easier with
integrating c* and hadoop, for example: CqlInputFormat, CL.LOCAL_ONE,
LimitedLocalNodeFirstLocalBalancingPolicy, vnodes, and describe_local_ring.
When using CqlInputFormat, if you don't want to be stuck within
datacenter-locality you can for example change the consistency level from
LOCAL_ONE to ONE. That's great, but describe_local_ring + CL.LOCAL_ONE in its
current implementation isn't enough for us. We have multiple datacenters for
offline, multiple for online, because we still want the availability advantages
that come from aligning virtual datacenters to physical datacenters for the
offline stuff too.
Some of our jobs have their own InputFormat implementation that uses
describe_ring, LOCAL_ONE, and data with only replica in the offline
datacenters. Works very well, except the last point kinda sucks because we have
online clients that want to read this data and have to then do so through nodes
in the offline datacenters. Underlying performance improvements: eg
cross_node_timeout and speculative requests; have helped but there's still the
need to separate online and offline. If we wanted to push replica out on to the
online nodes, i think the best approach is for us is to have to filter out
those splits/locations in getRangeMap(..)
Back to this issue we also have jobs using CqlInputFormat. Specifying multiple
client input addresses doesn't help take advantage of the multiple offline
datacenters because the Cassandra.Client only makes one call to
describe_local_ring, and StorageService.describeLocalRing(..) only checks
against its own address. It would work to have either a) CqlInputFormat call
describeLocalRing against all the listed connection addresses and merge the
results into one masterRangeNodes list, or b) something along the lines of
changing the signature of describeLocalRing(..) to describeRings(String
keyspace, String[] dc) and having the job specify which DCs it will be running
within.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
