Notes from Backlog Refinement:

1. Figure out if the relevant data (number of requests and request execution time) can be found in Hadoop logs already and whether it is sufficient for Wikidata's tracking needs
2. If not, define what service interface wikidata wants to use: Statsd/Graphite? Prometheus?

For the general understanding of REST API usage the following data could be worth tracking:

• response codes by endpoint and method
• number of PUT (maybe PATCH?) requests that do not have any effect on the data (see Idempotence)

The Wikidata team ticket for visualizing the metrics is now in T326724.

With confirmed by querying Wikimedia Hadoop that request data is actually being recorded for any visualization or analysis that'd be intended.

Update: I have now created a dashboard prototype in Superset based on 1/128 sampled data for the launch. However, using an approach based on Grafana and Graphite seems more feasible long term.

For privacy reasons, the relevant data in Hive/Hadoop is limited to the last 30 days. Also, unsampled data is available only in a hard-to-use form. That means we would need to create a pipeline aggregating the data first and saving it in a new table to allow for long-term monitoring. Another downside of this approach is that the available visualization tools cannot create public dashboards yet (we would need to set up our own tools if we wanted to enable this).

This is why it seems more feasible to create a dashboard using Grafana. The relevant API data appears to be already sent to Graphite thanks to a patch by Daniel Kinzler (the data can be found under Graphite Browser, "Metrics" > "MediaWiki" > "rest_api_*").

See also: T327154: Create a basic monitoring of usage of Wikibase REST API on wikidata

@Manuel as a note, it really depends what one wants to find out.
Data which Mediawiki sends to Graphite only records request that made it to mediawiki - will not show that a certain request are made many times more but cached etc.
To know ALL request Hadoop is the way to go IMO. Preprocessed (no need private data for most of monitoring i'd guess etc) and stored in aggregated form somewhere as you say but if one wanted to see all usage on Wikidata that'd be the source of truth.

Wikibase Product Platform's job is done here though, so I will be closing this task.

To know ALL requests Hadoop is the way to go IMO.

Yes, that's true, we will have to keep that in mind! We will be able to measure the gap by comparing the results based on the two sources to see how urgent this is.

Also, the Grafana dashboard seems to not separate between test.wikidata.org and www.wikidata.org. We might want to address this before we run our next tests.