Tianyi Wang has posted comments on this change. ( 
http://gerrit.cloudera.org:8080/10277 )

Change subject: IMPALA-5737: Tighten minicluster memory limit
......................................................................


Patch Set 1:

> Thanks for working on this.
 >
 > From a flaky test perspective, I worry that this kind of approach
 > is too clever. For the in-repo tests, we want to assert that they
 > have some requirements in terms of memory, and, if those
 > requirements are met, they always work. To me, the easiest way to
 > make that happen is always have consistent memory limits, and run
 > within the same constraints (as much as possible) even if we're
 > running on a machine with a ton of memory.
 >
 > In that sense, I think the existing "70% of system memory" for
 > impalad memlimit is a bit of an anti-pattern for running tests. (I
 > think it's actually fine for "start-impala-cluster", which is used
 > not just for tests, but I wish that run-all-tests pinned that
 > explicitly to X GB, and we always ran with that.) I had a lot of
 > trouble tuning that when I worked on the test-with-docker stuff,
 > and I still see failures (both inside of Docker and out) with
 > random-seeming mem-limit-exceeded errors.
 >
 > One other interesting bit is that the amount of memory we use
 > actually also changes with number of cores, both because "parallel"
 > tests scale that way and because there are parallel things going on
 > inside of Impala.
 >
 > Perhaps a simpler approach would be to simply have variables for
 > all the daemons, and have a "profile" of sorts that tunes them in a
 > small number of ways. i.e., instead of supporting all possible
 > memory sizes, we support "I see you have 20GB of RAM" and "I see
 > you have over 40GB" of RAM, and that reduces the potential for
 > noise.
 >
 > If you do go with this approach, I'd put all the math in one place.
 > Leave all the scripts to have simple variables like HS2_LIMIT_MB
 > (i.e, not just Hive, pin it to the specific daemon), and do all the
 > math centrally. I think it'll be easier to follow.

The profile idea sounds good. I think I'll go with that.


--
To view, visit http://gerrit.cloudera.org:8080/10277
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings

Gerrit-Project: Impala-ASF
Gerrit-Branch: master
Gerrit-MessageType: comment
Gerrit-Change-Id: If8240551e726c6da546a926a1ce3444f41ef87fe
Gerrit-Change-Number: 10277
Gerrit-PatchSet: 1
Gerrit-Owner: Tianyi Wang <[email protected]>
Gerrit-Reviewer: Philip Zeyliger <[email protected]>
Gerrit-Reviewer: Tianyi Wang <[email protected]>
Gerrit-Comment-Date: Wed, 02 May 2018 03:28:57 +0000
Gerrit-HasComments: No

Reply via email to