The Analytics Query Service (AQS) is a public facing API that serves analytics data from both a Cassandra and a Druid backend.
- Analytics/AQS/Unique Devices
- Analytics/AQS/Wikistats 2
More and up to date info in: https://wikimedia.org/api/rest_v1/?doc#/
Scaling: Settings, Failover and Capacity Projections
- Cassandra: https://phabricator.wikimedia.org/T116590#1754386
- Druid: https://grafana.wikimedia.org/dashboard/db/prometheus-druid
- PageViews: https://grafana.wikimedia.org/dashboard/db/pageviews
- ElasticSearch: https://logstash.wikimedia.org/#/dashboard/elasticsearch/restbase => Query: analytics.wikimedia.org
Sum up:Throttling is enforced at the restbase/AQS layer thus requests that are served by varnish are not throttled. This is an important point. It means that the throughput of the API in the top endpoints is real high as the same data is requested over and over as on those endpoints we mostly serve"daily top" data. Throttling is done per (IP/endpoint/second) and if a client breaks throttling limits it will receive a 429 response code to its http request.
At the time of this writing throttling is set to trigger at reqs per (IP/endpoint/second) and thus far we are only logging when limits are breached, we are not enforcing throttling quite yet. Why? Cause if we get more that 30 concurrent requests in cassandra at any one time cassandra lookups time out. This, likely, will not be true after we finish our work in scaling the storage layer of the API.
Ticket in which we discussed throttling: 
Throttling limits breached are logged in to: https://logstash.wikimedia.org/#/dashboard/temp/AVTsUtpi_LTxu7wlBfI-
Config for throttling is at: https://github.com/wikimedia/restbase/blob/master/v1/metrics.yaml
Bumping up throttling limits after scaling work to 400 requests per second, after our load tests. See: Analytics/AQS/Scaling#Load_testing
This step-by-step serves for deploying to both staging (beta) and production. Watch out for specific differences between beta and prod in each step of this section.
Step 0: Testing AQS locally
Testing your change in our staging environment (beta) requires either having a stable patch merged in AQS and deployed through scap, or a lot of git black magic and messing around that you shouldn't do. A good solution for quick testing is setting up your own mini AQS in your local machine where you can make changes to the APIs instantly, update dependencies, load data... without switching between machines or sending gerrit patches.
- Install Zookeeper (
brew install zookeeperon mac).
- Install Cassandra (
brew install email@example.com mac. Be aware that without the @2.2, brew will install version 3, which we don't have yet in production).
- Make sure you're using the right Java version (8). Cassandra will complain a lot about Java 9 and 10, so make sure that your JAVA_HOME environment variable points to your Java 8 installation (
/usr/libexec/java_home -Vwill show the versions currently installed). To do that, set
export JAVA_HOME=`/usr/libexec/java_home -v 1.8`
- Once the Cassandra service is running, start AQS by running the server with the default config provided on the repo:
./server.js -c config.example.wikimedia.yaml
- To load data or make changes in Cassandra, run
The easiest to test AQS druid integration is to use the production druid cluster (AQS can only query druid, no data loss is possible). This is doable by following those steps:
- Start an SSH tunnel between your machine and the druid-public broker :
ssh -N druid1004.eqiad.wmnet -L 8082:druid-public-broker.svc.eqiad.wmnet:8082
- Start AQS locally with the appropriate configuration, as suggested in this gerrit patch (WARNING: Update the
datasources: mediawiki_historyto the correct value for testing).
- You should be able to query your local AQS for druid-oriented queries (for instance: http://localhost:7231/analytics.wikimedia.org/v1/edits/aggregate/all-projects/all-editor-types/all-page-types/monthly/20180101/20190101)
Step 1: Update the AQS deploy repository
Note: Be aware that this process requires having Docker installed as an instantiation of docker is done when building.
Note: Even if you're deploying to staging (beta), the code you want to deploy should be merged to master. Otherwise, the whole deployment process won't work.
- If it's the first time you deploy:
- Get the deploy repository: git clone ssh://$USER@gerrit.wikimedia.org:29418/analytics/aqs/deploy .
- Make sure AQS source git repo has the deploy.dir config variable set (see Services/FirstDeployment#Local Git).
npm installin the source repository and make sure that no error is returned. Do also the same thing with
- Are you deploying a new endpoint? You need to add a bit of code to the fake data script that matches the x-amples definition in AQS's v1 yaml. Otherwise endpoint checks will fail on deployment.
- Then (regardless if first time or not):
- Make sure both aqs-deploy and aqs repositories are on master, have latest, are clean, including submodules updated
- Follow Services/Deployment#Preparing_the_Deploy_Repository (basically, run
./server.js build --deploy-repo --force --review -c config.test.yamlin the source folder).
- Check that src's sha1 in the review corresponds to the code you want to deploy).
- Merge the newly created change to aqs deploy repo to master.
Issues with "src" path
Remove src path from deploy repo. (We're not sure why this was added to the docs, we should discuss and explain or remove.)
Issues with git review
It uses git review only if you pass it the --review param, omit it and it will not try to submit patch, it will commit it but it will not be pushed. Sometimes the build hangs. In this case, check the sync-repo branch of the deploy repository. It should have the commit in there and that can be pushed to gerrit. It's ok to kill the build if it's been hanging for a while.
Whenever possible, it is convenient to run
npm audit and make sure that no dependencies pose a threat to the service. Most vulnerabilities will be solved by upgrading packages, but in some cases they will correspond to a second or third-level dependency that can only be upgraded by forcing versions in
package-lock.json . Forcing versions can be avoided if you are certain that the code carrying the vulnerability will not be run by AQS (task T207945 is an example of this). If this is not the case, you can enforce the new version by editing
package-lock.json and making sure that the version change doesn't break tests.
See note about hoek npm vulnerability here: https://phabricator.wikimedia.org/T206474
NPM has more information about dealing with vulnerabilities here.
Step 2: Deploy using scap
- Tell the
#wikimedia-operationsIRC channels that you are deploying (use
- Ssh into the deployment machine that suits your needs:
- For staging (beta) use:
- For production use:
- For staging (beta) use:
- Execute scap:
git submodule update --init
scap deploy "YOUR DEPLOYMENT MESSAGE"
- [optional] To see more detailed error logs during deployment, run
/srv/deployment/analytics/aqs/deploywhile you deploy.
Note: after T156049 scap will deploy only to aqs1004 (or deployment-aqs01 in case of beta) as first step (canary) and it will ask for confirmation before proceeding to the rest of the cluster. After that, it will deploy to one host at the time serially. You can force scap to ask for confirmation after each host or not, but telling him to proceed to all the other hosts (after the canary) will not cause a deployment to all of them at the same time, since the previously mentioned constraint will hold. Each host will be de-pooled from the load-balancer before the aqs restart, and re-pooled after that.
Step 3: Test
Beta thus far just has a modest dataset with pageviews to Barack Obama page in 2016 from es.wikipedia, en.wikipedia and de.wikipedia
You can run some queries like the following to see that aqs is running well:
Should return daily records
Should return monthly records
Should return aggreggate data for en.wikipedia, if any
Should return monthly aggreggate data for en.wikipedia
From (one of) the deployed machine, run
Issues with deployment to labs deploy
SSH_AUTH_SOCK=/run/keyholder/proxy.sock ssh -l deploy-service deployment-aqs01.deployment-prep.eqiad.wmflabs
Issues with scap
- Depool machine
- Delete deployment directory
- Run puppet
- Try to deploy again.
Check deploy logs:
scap deploy-log -v
Check AQS logs:
sudo journalctl -u aqs
Journalctl might not have a lot of information since by default Restbase is configured to push logs to logstash. So in order to disable this behavior, remove the following from the AQS configuration file under /etc:
logging: name: aqs level: warn streams: - # XXX: Use gelf-stream -> logstash - - type: gelf - host: localhost - port: 12201
Manual AQS restart:
sudo systemctl restart aqs
Cqlsh is a python-based CLI for executing Cassandra Query Language commands. To start cqlsh in beta (password is public, this is labs)
cqlsh deployment-aqs01.deployment-prep.eqiad.wmflabs -u cassandra -p cassandra
ssh deployment-aqs01.deployment-prep.eqiad.wmflabs cqlsh -u cassandra -p cassandra deployment-aqs01
Load data into cassandra in beta
- Generate a CSV with the data you want to load. You have basically 2 options:
- Generate it via a query to production cassandra:
cqlsh -u user -p pwd aqs1004-a -e "select <<see below for column list>> from \"local_group_default_T_pageviews_per_article_flat\".data where article='Barack_Obama' and timestamp >'2016010100' and timestamp <'2017012000' and project in ('en.wikipedia', 'de.wikipedia' , 'jp.wikipedia', 'es.wikipedia') and granularity='daily' and \"_domain\"='analytics.wikimedia.org' " > out.csv
- This will generate a file separated by | with a header and footer, so you have to adjust it manually.
- Generate it yourself. If you do this, take the following into account:
- You have to include the underscore-prefixed columns of your table, i.e. _domain or _tid. Look at the table description to get them.
- The _tid column needs to have valid timeuuid values. You can grab an existing _tid value from the data that is already loaded in another table. It's OK for testing purposes to give the same _tid value to all rows.
- The column _del (that exists in all tables) must be left out. It should not be populated, otherwise the table will interpret the record as deleted.
- Be careful with the values you insert. The COPY command checks for data types, but not for value correctness (as an insert statement would). So if you insert values that do not match the possible options for that column, your queries may not find the data. Example: If a column accepts a string among (a, b, c), and you give it d, the COPY command will not complain, but you'll not be able to find any data with your queries.
- The CSV should have no header.
- Generate it via a query to production cassandra:
- move data into beta (deployment-aqs01.deployment-prep.eqiad.wmflabs)
- load data into beta using cqlsh COPY command:
- (NOTE: you may need to use the latest version of cqlsh if we haven't upgraded it yet, see this if you get the error "'NoneType' object has no attribute 'datacenter'" https://stackoverflow.com/questions/39308666/cassandra-error-nonetype-object-has-no-attribute-datacenter-while-importing/39358147#39358147)
cqlsh -u cassandra -p cassandra deployment-aqs01
- Generate INSERT statements like the following:
insert into "local_group_default_T_pageviews_per_article_flat".data ("_domain", project, article, granularity, timestamp, "_tid", aa, ab, as, au, da, db, ds, du, maa, mab, mas, mau, mwa, mwb, mws, mwu) VALUES ('analytics.wikimedia.org','de.wikipedia','Barack_Obama','daily','2016010200', 13814000-1dd2-11b2-8080-808080808080,null,3527,null,28,3499,1398,null,22,1376,145,null,null,145,1984,null,6,1978);
- NOTE: The COPY command doesn't seem to work anymore, it gets really picky about any null values you exported, and as of this writing I couldn't figure out how to fix it:
COPY "local_group_default_T_pageviews_per_article_flat".data ("_domain", "project", "article", "granularity", "timestamp", "_tid", "aa", "ab", "as", "au", "da", "db", "ds", "du", "maa", "mab", "mas", "mau", "mwa", "mwb", "mws", "mwu") from '/home/your_user/cassandra_test_input.csv';
On the host to check live requests:
elukey@aqs1003:~$ sudo httpry -i eth0 tcp
Check Restbase status:
elukey@aqs1003:~$ systemctl status aqs ● aqs.service - "aqs service" Loaded: loaded (/lib/systemd/system/aqs.service; enabled) Active: active (running) since Tue 2016-05-17 15:45:58 UTC; 1 day 21h ago Main PID: 25226 (firejail) CGroup: /system.slice/aqs.service ├─25226 /usr/bin/firejail --blacklist=root --blacklist=/home/* --tmpfs=/tmp --caps --seccomp /usr/bin/nodejs src/server.js -c /etc/aqs/config.yaml ├─25227 /usr/bin/nodejs src/server.js -c /etc/aqs/config.yaml ├─25254 /usr/bin/nodejs /srv/deployment/analytics/aqs/deploy-cache/revs/a38e4d78718b072a70514477c3b268baaf8e1d29/src/server.js -c /etc/aqs/config.yaml [...] ├─25493 /usr/bin/nodejs /srv/deployment/analytics/aqs/deploy-cache/revs/a38e4d78718b072a70514477c3b268baaf8e1d29/src/server.js -c /etc/aqs/config.yaml └─25504 /usr/bin/nodejs /srv/deployment/analytics/aqs/deploy-cache/revs/a38e4d78718b072a70514477c3b268baaf8e1d29/src/server.js -c /etc/aqs/config.yaml
Check Cassandra cluster status (UN == Up Normal):
elukey@aqs1001:~$ nodetool status Datacenter: eqiad ================= Status=Up/Down |/ State=Normal/Leaving/Joining/Moving -- Address Load Tokens Owns Host ID Rack UN 10.64.48.117 1.5 TB 256 ? eb6d11b8-7a7e-4820-b56b-23869d2b79da rack1 UN 10.64.0.123 1.49 TB 256 ? db9cd8c1-910a-49af-9605-38af3a064788 rack1 UN 10.64.32.175 1.49 TB 256 ? 434ae715-b2a9-459b-91e9-4f29764939fd rack1 elukey@aqs1001:~$ nodetool info ID : db9cd8c1-910a-49af-9605-38af3a064788 Gossip active : true Thrift active : false Native Transport active: true Load : 1.49 TB Generation No : 1459781811 Uptime (seconds) : 912 Heap Memory (MB) : 8146.42 / 16384.00 Off Heap Memory (MB) : 3492.32 Data Center : eqiad Rack : rack1 Exceptions : 0 Key Cache : entries 1977574, size 364.84 MB, capacity 400 MB, 4097 hits, 78494 requests, 0.052 recent hit rate, 14400 save period in seconds Row Cache : entries 0, size 0 bytes, capacity 200 MB, 0 hits, 0 requests, NaN recent hit rate, 0 save period in seconds Counter Cache : entries 0, size 0 bytes, capacity 50 MB, 0 hits, 0 requests, NaN recent hit rate, 7200 save period in seconds Token : (invoke with -T/--tokens to see all 256 tokens)
Please note: aqs100 are running two cassandra instance per node so you'll need to use nodetool-a or nodetool-b:
elukey@aqs1004:~$ nodetool-a status Datacenter: eqiad ================= Status=Up/Down |/ State=Normal/Leaving/Joining/Moving -- Address Load Tokens Owns Host ID Rack UN 10.64.48.148 495.72 KB 256 ? da129795-421f-439b-bd29-6a4cd9f18813 rack1 UN 10.64.48.149 552.43 KB 256 ? e28f73cd-93c6-47e6-b046-8bbf801389f6 rack1 UN 10.64.32.189 578.18 KB 256 ? f05db2ca-61c4-4324-8f9a-d11d3cf66e95 rack1 UN 10.64.0.126 624.22 KB 256 ? af353a9f-0dd4-41f1-8a08-b1c7e57b2c68 rack1 UN 10.64.32.190 524.25 KB 256 ? 571af44e-23c3-4140-b59c-66fbdc16af6a rack1 UN 10.64.0.127 610.39 KB 256 ? 06dc704b-b39b-4d2a-8d9e-81368163221f rack1 Note: Non-system keyspaces don't have the same replication settings, effective ownership information is meaningless
Most useful one is /var/log/cassandra/system.log, that becomes -a and -b on aqs100 since we have two cassandra instances running:
elukey@aqs1004:/var/log/cassandra$ ls gc-a.log.0.current gc-b.log.0.current system-a.log system-b.log system.log
The AQS IPs are deployed in the Production network, meanwhile the Hadoop IPs are running in the Analytics network. The traffic flow is guarded by ACLs on switches/routers that needs to be updated if you need to connect new AQS IPs to the Analytics network. For example, this is the error that we were getting from analytics1* hosts while trying to upload data to the aqs1004-a.eqiad.wmnet Cassandra instance:
Caused by: com.datastax.driver.core.exceptions.NoHostAvailableException: All host(s) tried for query failed (tried: aqs1004-a.eqiad.wmnet/10.64.0.126:9042 (com.datastax.driver.core.TransportException: [aqs1004-a.eqiad.wmnet/10.64.0.126:9042] Cannot connect))
To solve the issue ops extended the existing ACL for aqs100.eqiad.wmnet to allow all the Cassandra Instances IPs too.
Deploy new History snapshot for Wikistats Backend
As of Q4 2018 every snapshot of mediawiki history we load into druid is a new datasource named after the snapshot. For example: "mediawiki-2018-05" AQS will not serve this data until told to do so (this is so we can actually rollback to a prior snapshot easily). In order to enable a new snapshot you need to change the hiera config for AQS that points to the active snapshot. See patch for example: https://gerrit.wikimedia.org/r/#/c/operations/puppet/+/440148
A quick note on caching. After deploying a new snapshot, you can check the data by hitting AQS directly with curl:
Real people out in the world won't see the new data until they also clear cache or it expires (86400 seconds / 1 day).
See table schema:
cassandra@cqlsh> describe table "local_group_default_T_pageviews_per_article_flat".data
Add fake data to Cassandra after wiping the cluster
cqlsh -u cassandra -p cassandra aqs1004-a -f /srv/deployment/analytics/aqs/deploy/scripts/insert_monitoring_fake_data.cql
This commands will ensure that no AQS related alarm will fire.