To answer a variety of data-related questions, ideally we are able to capture all query events that are issued to a database without sampling.
In practice, this can be prohibitively expensive. This issue tracks enhancements made to our current query event sampling approach which will allow us to track as much telemetry data as possible without impacting cluster performance. There are a few approaches we can consider.
Avoid sampling console queries (option: do not filter on a specific application name)
Only start sampling for clusters showing QPS of over a certain threshold
If we must sample due to the QPS threshold, the sample only DML statements where volume is over a certain threshold (threshold TBD based on perf testing)
As part of this issue, we should avoid sampling console queries as well. We should collect telemetry data and observability information of our own console to measure engagement and database load from the console.
To answer a variety of data-related questions, ideally we are able to capture all query events that are issued to a database without sampling.
In practice, this can be prohibitively expensive. This issue tracks enhancements made to our current query event sampling approach which will allow us to track as much telemetry data as possible without impacting cluster performance. There are a few approaches we can consider.
This issue builds on top of the following MVP: https://github.com/cockroachdb/cockroach/issues/65991.
Epic: CRDB-8032
Jira issue: CRDB-8094