Zach Amsden has posted comments on this change.

Change subject: IMPALA-5164: Fix flaky benchmarks
......................................................................


Patch Set 2:

(5 comments)

http://gerrit.cloudera.org:8080/#/c/6935/2/be/src/util/benchmark.cc
File be/src/util/benchmark.cc:

Line 35: double Benchmark::Measure(BenchmarkFunction function, void* args,
> Can we change the behaviour so that benchmarks can opt in or out from the c
Definitely.  I noticed a few tests with more complicated behavior that 
definitely don't want this (unfortunately they crash instead of running).


Line 62:       throw std::runtime_error("Benchmark failed to complete due to 
context switching.");
> We usually use LOG(FATAL) for fatal errors or Status for non-fatal errors. 
That is true, but having an error message and call graph doesn't help here.  
Instead, I want to throw an error which can be caught and logged at the level 
where we have context on which benchmark was being run (see the sample error 
message in the summary).


Line 69:   }
> Maybe we should reset 'sw' and 'lost_time' after doing the tuning step so w
I thought so as well, but there is no reset method currently, so this actually 
seems easiest.


PS2, Line 94: 10
> Where does the 10 come from?
Arbitrary flake factor.  If we lose 90% of measurements, we bail.  This 
parameter just made the arithmetic easy.


Line 95:     throw std::runtime_error("Benchmark failed to complete due to 
noisy measurements.");
> Same comment here about using Status/LOG(FATAL).
Same reason - we want context on which benchmark was failing, which we don't 
have here.


-- 
To view, visit http://gerrit.cloudera.org:8080/6935
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings

Gerrit-MessageType: comment
Gerrit-Change-Id: I7e0ddfaa9bd1340990ae2b6ff946d72bbd26c274
Gerrit-PatchSet: 2
Gerrit-Project: Impala-ASF
Gerrit-Branch: master
Gerrit-Owner: Zach Amsden <[email protected]>
Gerrit-Reviewer: Matthew Jacobs <[email protected]>
Gerrit-Reviewer: Tim Armstrong <[email protected]>
Gerrit-Reviewer: Zach Amsden <[email protected]>
Gerrit-HasComments: Yes

Reply via email to