For correctness testing it's useful to be able to capture production traffic so that it can be replayed against both the old and new versions of Cassandra while comparing the results.
Implementing this functionality once inside the database is high performance and presents less operational complexity.
In this patch there is an implementation of a full query log that logs uses chronicle-queue (apache licensed, the maven artifacts are labeled incorrectly in some cases, dependencies are also apache licensed) to implement a rotating log of queries.
- Single thread asynchronously writes log entries to disk to reduce impact on query latency
- Heap memory usage bounded by a weighted queue with configurable maximum weight sitting in front of logging thread
- If the weighted queue is full producers can be blocked or samples can be dropped
- Disk utilization is bounded by deleting old log segments once a configurable size is reached
- The on disk serialization uses a flexible schema binary format (chronicle-wire) making it easy to skip unrecognized fields, add new ones, and omit old ones.
- Can be enabled and configured via JMX, disabled, and reset (delete on disk data), logging path is configurable via both JMX and YAML
- Introduce new fqltool in /bin that currently implements Dump which can dump in a human readable format full query logs as well as follow active full query logs
Follow up work:
- Introduce new fqltool command Replay which can replay N full query logs to two different clusters and compare the result and check for inconsistencies. <- Actively working on getting this done
- Log not just queries but their results to facilitate a comparison between the original query result and the replayed result. <- Really just don't have specific use case at the moment
- "Consistent" query logging allowing replay to fully replicate the original order of execution and completion even in the face of races (including CAS). <- This is more speculative