I'm looking into this guys. The largest cohort we have so far has over
2500 editors in it, so 700 is definitely not too big. I'm not sure how
large something would have to be before it broke the system, but I'm
guessing tens of thousands if not hundreds like Dario says.
The 504 timeout should not be happening. Could one of you email me the
cohort so I can test?
Dan
On Fri, Oct 11, 2013 at 1:58 PM, Dario Taraborelli <
dtaraborelli(a)wikimedia.org> wrote:
On Fri, Oct 11, 2013 at 10:42 AM, Dario Taraborelli
<
dtaraborelli(a)wikimedia.org> wrote:
I think we should set the right expectations
about working with large
cohorts.
I would not call 700 editors a large cohort :) this should just work fine.
We don't know yet whether this was related to labs maintenance or not and I
asked Steven to share the cohort with us to verify that it works.
I tried uploading a CSV with 500 user_ids and got the same 504 Gateway
Time-out error as Steven
Quoting Dan's response from September 17:
I just wanted to correct one small misunderstanding.
Running large cohorts does *not* work in wikimetrics at this time for two
reasons:
1. You'll have a problem uploading them as Dario mentioned (because it
validates each user individually against the database, as Dario guessed).
The best solution for this is to create a temp table of all the users we
are trying to upload and verify them in one query. This would be very fast
and not too hard to implement.
Uploading a cohort this should work work but it's a blocking operation
which
is not very user friendly, Mingle card 818 addresses this issue.
Thanks, bookmarked :)
2. A large cohort will not fit in the
"IN" clause of a SQL query. This
is a known limitation and we have to fix it by creating a temporary table
from the cohort. We can then join to the temp table for any metrics. The
reason I've delayed this is because the same mechanism could be used to
implement dynamic cohorts, boolean cohort combinations, and project
level cohorts. We should prioritize these technically related features and
then I can come up with a plan to do the minimally viable thing without
shooting ourselves in the foot.
I did some calculations and it seems that this is only an issue with
cohorts
larger than 200k editors.
We will need to discuss and benchmark performance for these jobs, this is
one of the issues that I'd personally like to see prioritized over UX as
it's something the entire Product team would benefit from.
Dario
_______________________________________________
Analytics mailing list
Analytics(a)lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics