You are browsing a read-only backup copy of Wikitech. The live site can be found at wikitech.wikimedia.org

Analytics/Systems/Presto/Administration

From Wikitech-static
Jump to navigation Jump to search

View the Presto UI

File:Screenshot of Presto UI.png
The Presto home page shows Presto cluster metrics and recent queries.

The Presto coordinator hosts an admin interface that can be accessed via browser.

Read the upstream documentation here.

To access the presto UI for the production cluster, first open an SSH tunnel:

ssh -NL 8280:an-coord1001.eqiad.wmnet:8280 an-coord1001.eqiad.wmnet

Then open http://localhost:8280/ in your browser.

For the test cluster, use an-test-cood1001.eqiad.wmnet instead.

⚠️ Both clusters say the environment is "TEST"; ideally this would be "PRODUCTION" for the production cluster instead.

Roll restart the Presto cluster

There is a Spicerack cookbook for it!

Check metrics about Presto

The first thing to check is the Grafana dashboard: https://grafana.wikimedia.org/d/pMd25ruZz/presto

Presto is currently deployed in the following hosts:

  • an-coord1001- Presto Server (acting as coordinator)
  • an-presto100[1-5] - Presto Server (acting as worker)

All the queries are managed by the Presto Coordinator, that uses the workers to fetch data from HDFS when needed.

There is a test cluster deployed to the following hosts:

  • an-test-coord1001 - Presto Server (acting as coordinator)
  • an-test-presto1001 - Presto Server (acting as worker)

At the moment there is no working failover mechanism for the presto-coordinator, so an-coord1001 is a SPOF. However work to improve this is being done under T280905

Who/What uses Presto?

Presto is a new service that the Analytics team is experimenting with to provide a faster and better read-access to HDFS data compared to other systems like Hive. It is used mostly by Analysts/Researchers via CLI or via Superset.

Presto server down

There is a nagios check to control that the Presto server daemon is up and running on every Presto node. As described above, if one of the daemons is reported down please check metrics and contact the Analytics team.