Adar Dembo has posted comments on this change. ( 
http://gerrit.cloudera.org:8080/8090 )

Change subject: KUDU-1125: issue one catalog write per tablet report
......................................................................


Patch Set 8:

(1 comment)

http://gerrit.cloudera.org:8080/#/c/8090/6//COMMIT_MSG
Commit Message:

http://gerrit.cloudera.org:8080/#/c/8090/6//COMMIT_MSG@10
PS6, Line 10: When the
            : master is configured to fsync WAL writes, this can add a lot of 
load during
            : election storms and when the master is restarted.
> That's fair. If it's going to take a lot of effort I don't think it's worth
On our internal flash cluster I ran the following experiment:
1. Reconfigure the master to use log_force_fsync_all=true.
2. Stop the master.
3. Restart a tserver.
4. Start the master.
5. Look at the TSHeartbeat samples in the master's /rpcz page.

In this experiment not only does the master receive a full heartbeat from every 
tserver, but it also has to update a bunch of persistent tablet state for the 
tablets whose leaders were on the restarted tserver but are now elsewhere.

- Without this change, I saw a "fat" heartbeat sample which updated 140 tablets 
and took 10304 ms.
- With this change, an equivalent sample updated 308 tablets and took 157 ms.

As for measuring when the number of tablets may run up against the maximum RPC 
size, I modified the master to LOG during a tablet report the number of tablets 
and the size of the overall report. I restarted the master and calculated an 
average across the full tablet reports sent by all of the tservers in the 
cluster. It worked out to about ~283 bytes per tablet. With our 50MB maximum 
RPC size, we'll be in danger of blowing that out at around 185,000 tablets. 
Note that this applies to single-master deployments too as the maximum RPC size 
applies to the TSHeartbeat itself.



--
To view, visit http://gerrit.cloudera.org:8080/8090
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings

Gerrit-Project: kudu
Gerrit-Branch: master
Gerrit-MessageType: comment
Gerrit-Change-Id: Ie6f5cf0e4b1cd1160b3b310d89c6dbf3dd62e43b
Gerrit-Change-Number: 8090
Gerrit-PatchSet: 8
Gerrit-Owner: Adar Dembo <a...@cloudera.com>
Gerrit-Reviewer: Adar Dembo <a...@cloudera.com>
Gerrit-Reviewer: Dan Burkert <danburk...@apache.org>
Gerrit-Reviewer: Mike Percy <mpe...@apache.org>
Gerrit-Reviewer: Todd Lipcon <t...@apache.org>
Gerrit-Comment-Date: Sat, 07 Oct 2017 01:23:27 +0000
Gerrit-HasComments: Yes

Reply via email to