After meeting with the team: - Encoding issue was due to locale wrongly set on some machines (but we don't know why) - We will find a way to enforce file.encoding, first looking for a java-global way, if not feasible, a process-local way. - We will NOT spend computing resource on a job trying to detect this issue (too costly for occurrence probability, particularly if we force file.encoding) Cheers Joseph
On Wed, Mar 2, 2016 at 11:24 AM, Joseph Allemandou < jallemandou@wikimedia.org> wrote:
@Ori: Needs to be discussed with the team - My 2 cents
- Detection: possible to implement as part of one of the oozie jobs.
we will compute number of pages having different page_title for the same uri_path (if high, not good).
- Prevention: 2 things possible
think, possibly related to weird state after upgrade) and ensure we don't fall into that state again
- Try to understand WHY this thing happened (very difficult I
(probably easier but not really easy not to forget anything)
- Force JVM file.encoding for every java process of the cluster
I'd love to have your thoughts / ideas and discuss them with the team. Thanks
On Wed, Mar 2, 2016 at 10:53 AM, Ori Livneh ori@wikimedia.org wrote:
So: what is the planning for making sure this doesn't happen the next time around? :)
On Tue, Mar 1, 2016 at 5:26 AM, Joseph Allemandou < jallemandou@wikimedia.org> wrote:
Hi,
*TL,DR: Please don't use hive / spark / hadoop before next week.*
Last week the Analytics Team performed an upgrade to the Hadoop Cluster. It went reasonably well except for many of the hadoop processes were launched with a special option to NOT use utf-8 as default encoding. This issue caused trouble particularly in page title extraction and was detected last sunday (many kudos to the people having filled bugs on Analytics API about encoding :) We found the bug and fixed it yesterday, and backfill starts today, with the cluster recomputing every dataset starting 2016-02-23 onward. This means you shouldn't query last week data during this week, first because it is incorrect, and second because you'll curse the cluster for being too slow :)
We are sorry for the inconvenience. Don't hesitate to contact us if you have any question
-- *Joseph Allemandou* Data Engineer @ Wikimedia Foundation IRC: joal
Engineering mailing list Engineering@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/engineering
-- *Joseph Allemandou* Data Engineer @ Wikimedia Foundation IRC: joal