- 1 Overview
- 2 Important notes
- 3 Dumping data via sqoop from eventlogging to hdfs
- 4 EventLogging Routine Maintenance for the oncall
- 5 OnCall Setup
- 6 How Tos
- 6.1 Restart all EventLogging processes
- 6.2 Start/Stop/Restart individual EventLogging processes
- 6.3 Backfill schemas
- 6.4 Check logs
- 6.5 Get a first feeling of end-to-end issues on the whole pipeline of a schema
- 6.6 Troubleshoot events coming in in real time
- 6.7 Troubleshoot insufficient permission
- 6.8 Deploy EventLogging
- 6.9 Blacklist a schema
- 6.10 Resolve lagging replication between MySQL Master and Slaves
The following diagram should be a companion of the excellent explanation in https://github.com/wikimedia/eventlogging
The diagram has been created with https://www.draw.io/. If you want the source code, please check https://gist.github.com/elukey/975fab2bcf2ea6398fe1
statsv is running on hafnium.eqiad,wmnet and it is a daemon responsible to aggregate performance data before sending it to Statsd. There are metrics in the Even Logging dashboard like https://grafana.wikimedia.org/dashboard/db/eventlogging?panelId=11&fullscreen that are counting on this service to work properly. If you observe datapoint loss in the metric please check the status of the statsv service on hafnium and restart it if needed.
As depicted in the picture we are monitoring:
- Lag between messages landing in Kafka topics and message consumption rate from EventLogging's processes using Burrow. The alarm will be triggered as email sent to analytics-alert@.
- Insertion rate to Mysql master from the Consumer processes using Graphite/Icinga (config file). You will see alerts in the wikimedia-analytics IRC channel.
- [IN PROGRESS] Replication lag between for MySQL slaves (https://phabricator.wikimedia.org/T124306)
Consumption Lag "Alarms": Burrow
Alarms for burrow report numbers like
eventlogging-valid-mixed:2 (1454622892611, 199315032, 17) -> (1454622901672, 199315032, 17)
(timestamp, offset, #of messages behind (lag))
- Burrow evaluates lag in length window of 10 offsets by default. We are committing offsets every second, this frequency would make us evaluate lag every 10 secs which seems too frequent so we have changed the length of lag window window to 100 secs.
- Lag is evaluated aprox every couple minutes. (we are fine tuning this)
- docs about burrow lag - https://github.com/linkedin/Burrow/wiki/Consumer-Lag-Evaluation-Rules
An interesting alert use case is the following one:
Cluster: eqiad Group: eventlogging-00 Status: ERR Complete: true Errors: 1 partitions have problems Format: Topic:Partition (timestamp, start-offset, start-lag) -> (timestamp, end-offset, end-lag) eventlogging-client-side:11 (1455812030875, 188196575, 0) -> (1455813022220, 188205237, 0)
As you can see the end offset is greater than the start one (meanwhile the lag is zero) and the partition status is ERROR. This is due to the current Burrow's rule 4 in https://github.com/linkedin/Burrow/wiki/Consumer-Lag-Evaluation-Rules, that should be read in this way:
Burrow will alert you when a consumer is so slow that the time elapsed between the last offset committed and now is bigger than the time taken to commit all the offsets belonging the last window.
This can happen when a Kafka broker goes offline and it is a partition leader. EventLogging will need a bit of time to recognize the problem and request new metadata from Kafka, and hence the related consumer status according to Burrow will look like it is stalled or completely blocked even though it is only a temporary stop.
Reseting burrow consumer group monitoring
If you happen to change the topic assignment for a consumer group, burrow will continue to think that that group should consume from its previous assigned topics, and report lag if it stops doing so. To reset what topics burrow should monitor for a given consumer group, you should delete the consumer group monitoring from burrow. After deletion, consumer group monitoring will be automatically recreated for new topic partitions that get offset commits.
curl -X DELETE localhost:8000/v2/kafka/<cluster>/consumer/<consumer_group>
curl -X DELETE localhost:8000/v2/kafka/eqiad/consumer/eventlogging_consumer_mysql_00
MySQL Master/Slave replication
Eventlogging databases (m4 shard):
- db1107 (m4-master)
- db1108 (analytics-slave) use a custom replication mechanism:
- db1046 used to be the m4-master (replaced with db1107 in T156844).
- db1047 used to be the analytics-slave (replaced with db1108 in T156844).
- dbstore1002 used to replicate the log database (changed with T156844).
Dumping data via sqoop from eventlogging to hdfs
We will be archiving large tables that do not need immediate data access at " /wmf/data/archive/eventlogging/Table_name" the archival is a plain sqoop dump of table, nothing else. Avro schemas for tables (autogenerated by sqoop upon import) can be found at /wmf/data/archive/eventlogging/avro-schemas.
Hive tables on top of data are generated on archive database;
Try that you can connect via scoop using a command with no side effects, for example list tables:
sudo -u hdfs sqoop list-tables --password-file '/user/hdfs/mysql-analytics-research-client-pw.txt' --username research --connect jdbc:mysql://analytics-store.eqiad.wmnet/log
A sample import (to /tmp/PageContentSaveCompleteAvro) to avro file format
time sudo -u hdfs sqoop import --as-avrodatafile --password-file '/user/hdfs/mysql-analytics-research-client-pw.txt' --username research --connect jdbc:mysql://analytics-store.eqiad.wmnet/log --table PageContentSaveComplete_5588433 --columns id,uuid,timestamp,webHost,wiki,event_isAPI,event_isMobile,event_revisionId --target-dir /tmp/PageContentSaveCompleteAvro
Another example (note column names with "." need to be quoted):
time sudo -u hdfs sqoop import --as-avrodatafile --username research --password-file '/user/hdfs/mysql-analytics-research-client-pw.txt' --connect jdbc:mysql://analytics-store.eqiad.wmnet/log --quer y 'select convert(uuid using utf8) uuid,convert(timestamp using utf8) timestamp, convert(wiki using utf8) wiki, convert(webHost using utf8) webHost, convert(event_action using utf8) event_action,conve rt(`event_action.abort.mechanism` using utf8) `event_action.abort.mechanism`, convert(`event_action.abort.timing` using utf8) `event_action.abort.timing`, convert(`event_action.abort.type` using utf8) `event_action.abort.type`, convert(`event_action.init.mechanism` using utf8) `event_action.init.mechanism`, convert(`event_action.init.timing` using utf8) `event_action.init.timing` , convert(`event_ action.init.type` using utf8) `event_action.init.type`, convert(`event_action.ready.timing` using utf8) `event_action.ready.timing`, convert(`event_action.saveAttempt.timing` using utf8) `event_action. saveAttempt.timing`, convert(`event_action.saveFailure.message` using utf8) `event_action.saveFailure.message`, convert(`event_action.saveFailure.timing` using utf8) `event_action.saveFailure.timing` , convert(`event_action.saveFailure.type` using utf8) `event_action.saveFailure.type` , convert(`event_action.saveIntent.timing` using utf8) `event_action.saveIntent.timing`, convert(`event_action.sa veSuccess.timing` using utf8) `event_action.saveSuccess.timing`, convert(`event_editingSessionId` using utf8) `event_editingSessionId`, convert(event_editor using utf8) event_editor, convert(event_int egration using utf8) event_integration, convert(`event_mediawiki.version` using utf8) `event_mediawiki.version`, convert(`event_page.id` using utf8) `event_page.id`, convert(`event_page.ns` using utf8) `event_page.ns`, convert(`event_page.revid` using utf8) `event_page.revid`, convert(`event_page.title` using utf8) `event_page.title`, convert(event_platform using utf8) event_platform, convert(`event _user.class` using utf8) `event_user.class`, convert(`event_user.editCount` using utf8) `event_user.editCount`, convert(`event_user.id` using utf8) `event_user.id`, convert(event_version using utf8) event_version from Edit_13457736_15423246 where $CONDITIONS ' --target-dir /wmf/data/archive/eventlogging/Edit_13457736_15423246 --split-by uuid
Remapping columns and using a custom query:
time sudo -u hdfs sqoop import --as-avrodatafile --password-file '/user/hdfs/mysql-analytics-research-client-pw.txt' --username research --connect jdbc:mysql://analytics-store.eqiad.wmnet/log --query 'select id,uuid,convert(timestamp using utf8) timestamp,convert(webHost using utf8) webhost,wiki,cast(event_isAPI as Integer) event_isAPI,cast(event_isMobile as Integer) event_isMobile,cast(event_revisionId as Integer) event_revisionId from PageContentSaveComplete_5588433 where $CONDITIONS' --map-column-java event_isAPI=Integer,event_isMobile=Integer,event_revisionId=Integer --target-dir /tmp/PageContentSaveCompleteAvro --split-by id
An example of how to set up a table on top of avro files can be found here: https://github.com/wikimedia/analytics-refinery/blob/master/hive/mediawiki/history/create_mediawiki_page_table.hql
I (Nuria) could not get the direct mapping on hive to work, rather I had to use a syntax that included the schema to create the table: https://gist.github.com/nuria/acd67dc1d237c59a2dda9799e82da4c3#file-create-avro-table-with-schema-hql
EventLogging Routine Maintenance for the oncall
- Check grafana for problems with raw vs. validated events, or other apparent problems
- Check database for any gaps. A few different scripts exist, Milimetric's gist for example.
- Decide on whether we need to deploy on that week, avoid Friday deployments
- Remember to log all actions to SAL log (!log <something> on ops channel)
- Report outages as part of wikimedia's incident reports so there is a reference
- Follow up on any alarms that might be raised
- Ensure you can access graphite (wikitech credentials).
- Get Access to eventlog1002.eqiad.wmnet with root access.
- Ask Andrew (ottomata on IRC) to setup the system to send you alarms.
- In case of DB errors, the person to contact is Jaime Crespo (jcrespo on IRC), usually on #wikimedia-operations channel.
Restart all EventLogging processes
sudo eventloggingctl status
sudo eventloggingctl restart
sudo eventloggingctl stop
The config applied to create logs and such is at:
Start/Stop/Restart individual EventLogging processes
EventLogging processes are managed by systemd. Each config file in /etc/eventlogging.d/*/* corresponds to a single eventlogging process. Let's call the pieces of this hierarchy /etc/eventlogging.d/$service/$name.
To stop one of them, you can do something like:
systemctl stop eventlogging-$service@$name
elukey@eventlog1002:~$ sudo systemctl status email@example.com ● firstname.lastname@example.org - Eventlogging Consumer mysql-m4-master-00 Loaded: loaded (/email@example.com; enabled; vendor preset: enabled) Active: active (running) since Thu 2018-03-08 13:32:20 UTC; 5 days ago Main PID: 7882 (python) Tasks: 2 (limit: 4915) CGroup: /system.slice/system-eventlogging\firstname.lastname@example.org └─7882 python /srv/deployment/eventlogging/analytics/bin/eventlogging-consumer @/etc/eventlogging.d/consumers/mysql-m4-master-00 elukey@eventlog1002:~$ sudo systemctl restart email@example.com
You can stop, start, and restart any individual EventLogging process using variations of this command.
Raw logs are at:
Process logs are at:
Get a first feeling of end-to-end issues on the whole pipeline of a schema
If you're for example interested in NavigationTiming 10374055 run
mysql --defaults-extra-file=/etc/mysql/conf.d/research-client.cnf --host analytics-slave.eqiad.wmnet -e "select left(timestamp,10) ts , COUNT(*) from log.NavigationTiming_10374055 where left(timestamp,8) >= '20150101' group by ts order by ts;" >out && tail -n 100 out
on stat1006 (You need to be in the researchers group to access the
research-client.cnf). That will dump recent hourly totals to the screen, and if you prefer graphs, more data is stored in
out and only waiting to get plotted.
If you need different granularity, just change the
10 in the query to the granularity you need (like 8 => per day, 12 => per minute, 14 => per second).
If the numbers you get indicate issues, you can go to Graphite to sanity check the early parts of the pipeline. A first indicator is typically the
overall counts. Like comparing
eventlogging.overall.valid.rate. Then one can bisect into
eventlogging.server_side_events.*, directly drill down into per schema counts, like looking at the graph for
If you're good at staring at graphs, go right to the all in one graph.
If the numbers you get indicate issues, you can also repeat the database query to the
m2 master database directly (credentials are in
vanadium. That allows to exclude replication issues.
Troubleshoot events coming in in real time
- Incoming events counts are logged to graphite, both the count of validating and non validating events per schema are available
using those users can get a sense of change, graphite is populated real-time and if all of a sudden events for an schema do not validate it is clearly visible.
- EvenLogging slave database (for users with access to 'research' user) is also populated real-time.
Lastly, Event Logging events coming on real time are written to files that are sync-ed to stat1005 and stat1006 once a day, these files can be found here:
If you detect an issue or suspicious change , please notify analytics@ and escalate with analytics devs.
Troubleshoot insufficient permission
"error: insufficient permission for adding an object to repository database .git/objects"
List > groups to see if you are on wikidev group, if so likely some files on .git directory are not writable by "wikidev" group. Make them so.
EventLogging is deployed using scap3. The scap deployment configuration for various EventLogging deployments can be found in specific scap repos in gerrit: eventlogging/scap/<deployment-name>. The EventLogging Analytics deployment scap configs are at at eventlogging/scap/analytics. Deployment on deployment.eqiad.wmnet using
# ssh to production deploy server ssh deployment.eqiad.wmnet # cd to the EventLogging Analytics instance deploy source cd /srv/deployment/eventlogging/analytics # Checkout the revision you want to deploy git pull # Update the submodules git submodule update --init # Run scap3 deployment deploy
ssh eventlog1002.eqiad.wmnet (or wherever eventlogging is deployed.)
Go to /srv/deployment/eventlogging/analytics
See that checkout is there from what you just pulled in from tin (via git log).
Restart EL on target host (eventlog1002.eqiad.wmnet)
eventloggingctl stop eventloggingctl start
Check various logs in /var/log/upstart/eventlogging_* to see that things are running as they should.
Check that /srv/log/eventlogging/all-events.log has data flowing in.
Hop in the Ops IRC channel and !log that you upgraded & restarted EventLogging and add the commit hash that you deployed.
!log Redeployed eventlogging with revert to batch/queue size change - https://gerrit.wikimedia.org/r/#/c/258384/
Now please deploy latest code to Beta Cluster to keep things in sync: EventLogging/Testing/BetaLabs#How_to_deploy_code
Blacklist a schema
Resolve lagging replication between MySQL Master and Slaves
If replication is lagging we should open a ticket with the info we have and tag it with phabricator DBA task so DBA gets ping-ed. An example of a prior problem: 
Have in mind that while monitoring for database is here: , lag reported does not apply as replication on EL doesn't go through regular channels.
Ad hoc replication script is here: