Another approach we discussed back in the day was setting up a canary script to send known good messages whose delivery is monitored. This might be a bit easier to set up.

It's been effective on other systems I've worked on; also a good way to measure delivery latency. 

-Toby

On Friday, July 8, 2016, Jeff Green <jgreen@wikimedia.org> wrote:
On Fri, 8 Jul 2016, Andrew Otto wrote:

We’ll, you won’t be able to do it exactly how we do, since we are loading the data into Hadoop and then checking it there, so we use Hadoop tools.  Here’s what we got:

https://github.com/wikimedia/analytics-refinery/blob/master/oozie/webrequest/load/generate_sequence_statistics.hql

https://github.com/wikimedia/analytics-refinery/blob/master/oozie/webrequest/load/generate_sequence_statistics_hourly.hql

This old udp2log tool did a similar thing, so it is worth knowing about: https://github.com/wikimedia/analytics-udplog/blob/master/srcmisc/packet-loss.cpp However, it only worked with TSV udp2logs, and I think it won’t work with a multi-partition kafka topic, since seqs could be out of order based on partition read order.

You guys do some kind of 15 (10?) minute roll ups, right?  You could probably do some very rough guesses on data loss in each 15 minute bucket.  You’d have to be careful though, since the order of the data is not guaranteed.  We have the luxury of being over to query over our hourly buckets and assuming that all (most, really) of the data belongs in that hour bucket.  But, we use Camus to read from Kafka, which handles the time bucket sorting for us.

Yep, the pipeline is kafkatee->udp2log->files rotated on a 15 min interval, and parser-script->mysql which runs on a separate system.

Since the log files are stored one option would be to have a script that runs merges several files for a longer period sample, and sort and check for sequence gaps. Another option would be to modify the parse-to-mysql script to do the same thing.

But the part I don't get yet is how a script looking at output logs would identify a problematic gap in sequence numbers. We have two collectors, one is 1:1 and the other sampled 1:10, and both filter on the GET string. So if my understanding of the sequence numbers is correct (they're per-proxy right?) we should see only a small sample of sequence numbers, and how that sample relates to overall traffic will vary greatly depending on fundraising campaign and what else is going on on the site.

jg


Happy to chat more here or IRC. :)

On Fri, Jul 8, 2016 at 9:17 AM, Jeff Green <jgreen@wikimedia.org> wrote:
      Hi Nuria, thanks for raising the issue. Could you point me to the script you're using for sequence checks? I'm definitely
      interested in looking at how we might integrate that into fundraising monitoring.

      On Thu, 7 Jul 2016, Nuria Ruiz wrote:

            (cc-ing analytics public list)
            Fundraising folks:

            We were talking about the problems we have had with clickstream data and kafka as of late and how to prevent
            issues like this one going forward:
            (https://phabricator.wikimedia.org/T132500)

            We think you guys could benefit from setting up the same set of alarms on data integrity that we have on the
            webrequest end and we ill be happy
            to help with that at your convenience. 

            An example of how these alarms could work (simplified version): every message that comes from kafka has a
            sequence Id, if sorted those sequence
            Ids should be more or less contiguous, a gap in sequence ids indicates an issue with data loss at the kafka
            source. A script checks for sequence
            ids and number of records and triggers an alarm if those two do not match.

            Let us know if you want to proceed with this work.

            Thanks,

            Nuria


_______________________________________________
Analytics mailing list
Analytics@lists.wikimedia.org
https://lists.wikimedia.org/mailman/listinfo/analytics