Fair enough, just a thought.

On Wednesday, April 15, 2015, Nuria Ruiz <[email protected]> wrote:

> Given that batching code is been deployed since earlier (March16th)  than
> the 1st event listed by Marcel (April 9th) and since then we have swapped
> the EL box (April 3rd/4th) we probably want to look at system issues.
>
> On my opinion it is probably easier to see with tcpdump whether inserts
> are being sent, rather than trying to reply traffic to repro the problem.
>
> Thanks,
>
> Nuria
>
>
> On Wed, Apr 15, 2015 at 5:14 PM, Dan Andreescu <[email protected]
> <javascript:_e(%7B%7D,'cvml','[email protected]');>> wrote:
>
>> >This sounds like the fixes we did last quarter to the batch insertion
>>> basically hid the problem instead of making it go away.
>>> I think we are mixing things here, when we had issues with batching code
>>> we never saw a pattern of "no-events-whatsoever-in-any-table for an hour".
>>> We saw events dropped in bursts here and there but certainly not  an "hour
>>> long blackout".
>>>
>>> Also, there were no events dropped when we did the major backfilling in
>>> early march where the db sustained quite a bit of load as we had to insert
>>> those one by one.
>>>
>>> So (while I am not saying we could not uncover a code issue in our end)
>>> we have not seen this particular error pattern before.
>>>
>>
>> I didn't mean to suggest we saw this error before.  I was trying to say
>> that intuitively the error seems very similar.  That is, over time, the lag
>> grows and at some point it's so big that we lose a bunch of data all at
>> once.  I was just saying that because the first place I'd look is at that
>> change.  For example, I'd try replicating by simulating the same traffic
>> and then I'd revert to the original logic before batch inserts and try
>> that.  We've all looked at this code but we must be missing something big.
>>
>> _______________________________________________
>> Analytics mailing list
>> [email protected]
>> <javascript:_e(%7B%7D,'cvml','[email protected]');>
>> https://lists.wikimedia.org/mailman/listinfo/analytics
>>
>>
>
_______________________________________________
Analytics mailing list
[email protected]
https://lists.wikimedia.org/mailman/listinfo/analytics

Reply via email to