There are all good points, and we certainly do need better tooling for individual developers.
There are a lot of things a developer can do on just a laptop in terms of profiling code, that if done consistently, could go a long way, even without it looking anything like production. Things like understanding if algorithms or queries are O(n) or O(2^n), etc. and thinking about the potential size of the relevant production data set might be more useful at that stage than raw numbers. When it comes to gathering numbers in such an environment, it would be helpful if either the mediawiki profiler could gain an easy visualization interface appropriate for such environments, or if we standardized around something like xdebug.
The beta cluster has some potential as a performance test bed if only it could gain a guarantee that the compute nodes it runs on aren't oversubscribed or that the beta virts were otherwise consistently resourced. By running a set of performance benchmarks against beta and production, we may be able to gain insight on how new features are likely to perform.
Beyond due diligence while architecting and implementing a feature, I'm actually a proponent of testing in production, albeit in limited ways. Not as with test.wikipedia.org which ran on the production cluster, but by deploying a feature to 5% of enwiki users, or 10% of pages, or 20% of editors. Once something is deployed like that, we do indeed have tooling available to gather hard performance metrics of the sort I proposed, though they can always be improved upon.
It became apparent that ArticleFeedbackV5 had severe scaling issues after being enabled on 10% of the articles on enwiki. For that example, I think it could have been caught in an architecture review or in local testing by the developers that issuing 17 database write statements per submission of an anonymous text box that would go at the bottom of every wikipedia article was a bad idea. But it's really great that it was incrementally deployed and we could halt its progress before the resulting issues got too serious.
That rollout methodology should be considered a great success. If it can become the norm, perhaps it won't be difficult to get to the point where we can have actionable performance standards for new features, via a process that actually encourages getting features in production instead of being a complicated roadblock.
On Fri, Mar 22, 2013 at 1:20 PM, Arthur Richards arichards@wikimedia.orgwrote:
Right now, I think many of us profile locally or in VMs, which can be useful for relative metrics or quickly identifying bottlenecks, but doesn't really get us the kind of information you're talking about from any sort of real-world setting, or in any way that would be consistent from engineer to engineer, or even necessarily from day to day. From network topology to article counts/sizes/etc and everything in between, there's a lot we can't really replicate or accurately profile against. Are there plans to put together and support infrastructure for this? It seems to me that this proposal is contingent upon a consistent environment accessible by engineers for performance testing.
On Thu, Mar 21, 2013 at 10:55 PM, Yuri Astrakhan yastrakhan@wikimedia.orgwrote:
API is fairly complex to meassure and performance target. If a bot
requests
5000 pages in one call, together with all links & categories, it might
take
a very long time (seconds if not tens of seconds). Comparing that to another api request that gets an HTML section of a page, which takes a fraction of a second (especially when comming from cache) is not very useful.
On Fri, Mar 22, 2013 at 1:32 AM, Peter Gehres lists@pgehres.com wrote:
From where would you propose measuring these data points? Obviously network latency will have a great impact on some of the metrics and a consistent location would help to define the pass/fail of each test. I
do
think another benchmark Ops "features" would be a set of latency-to-datacenter values, but I know that is a much harder taks.
Thanks
for putting this together.
On Thu, Mar 21, 2013 at 6:40 PM, Asher Feldman <afeldman@wikimedia.org
wrote:
I'd like to push for a codified set of minimum performance standards
that
new mediawiki features must meet before they can be deployed to
larger
wikimedia sites such as English Wikipedia, or be considered complete.
These would look like (numbers pulled out of a hat, not actual suggestions):
- p999 (long tail) full page request latency of 2000ms
- p99 page request latency of 800ms
- p90 page request latency of 150ms
- p99 banner request latency of 80ms
- p90 banner request latency of 40ms
- p99 db query latency of 250ms
- p90 db query latency of 50ms
- 1000 write requests/sec (if applicable; writes operations must be
free
from concurrency issues)
- guidelines about degrading gracefully
- specific limits on total resource consumption across the stack per
request
- etc..
Right now, varying amounts of effort are made to highlight potential performance bottlenecks in code review, and engineers are encouraged
to
profile and optimize their own code. But beyond "is the site still
up
for
everyone / are users complaining on the village pump / am I ranting
in
irc", we've offered no guidelines as to what sort of request latency
is
reasonable or acceptable. If a new feature (like aftv5, or flow)
turns
out
not to meet perf standards after deployment, that would be a high
priority
bug and the feature may be disabled depending on the impact, or if
not
addressed in a reasonable time frame. Obviously standards like this
can't
be applied to certain existing parts of mediawiki, but systems other
than
the parser or preprocessor that don't meet new standards should at
least
be
prioritized for improvement.
Thoughts?
Asher _______________________________________________ Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l
Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l
Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l
-- Arthur Richards Software Engineer, Mobile [[User:Awjrichards]] IRC: awjr +1-415-839-6885 x6687 _______________________________________________ Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l