Gotcha. That's definitely the biggest factor that I was aware of. I
wasn't sure if you knew more than I did by now ;). I can respect the
implications of too much bookkeeping going on. That might really start
pounding the metadata and replication tables.
Happy to do lunch, also happy to just have a video call too if that's
more convenient.
Adam J. Shook wrote:
Thanks, Josh. I think the main pain-point is that replication doesn't
occur until the WAL is closed. We've made some aggressive configuration
changes to Accumulo to reduce the WAL time rollover and minor compaction
frequency to force replication to go faster. It is down to around 20
minutes or so on our production clusters, but we are kind of at our
limit -- Accumulo is spending a lot more time doing bookkeeping tasks
and it is starting to affect our query performance.
My initial thoughts are to increase the replication parallelism and
start replicating the WAL before it is closed (I see a few JIRAs open
already that mention these things), but I haven't done enough digging in
the code base to see what is really available.
Are you free for a bit in the near future to meet up for a bit and talk
replication? I'll buy lunch!
Cheers,
--Adam
On Wed, Feb 15, 2017 at 2:52 PM, Josh Elser <josh.el...@gmail.com
<mailto:josh.el...@gmail.com>> wrote:
Hi Adam,
I'm not presently working on anything (too many irons in other
fires), but I'd be happy to help work through a design doc for
improvements.
Do you have a list of pain-points which are the primary causes of
latency? That would help in identifying the changes to make and how
best to implement them.
- Josh
Adam J. Shook wrote:
I'm currently scoping what it would take to improve the latency
in the
replication feature of Accumulo. I'm interested in knowing what
work,
if any, is being done to improve replication latency? If work
is being
done, would there be some interest in collaborating on that effort?
If nothing is currently being planned, I'd be interested in
design ideas
and pointers from the community for improvements to the existing
implementation. We're looking to get replication down to less
than five
minutes and are willing to put in the effort to implement the
improvements.
Thank you for your time!
Cheers,
--Adam
