Hey Mathias Stearn,
Would appreciate your input here. As a consequence of making the changes we discussed around returning CommandNotSupported by default, there are a large number of test failures for rocksdb.
Looking over those failures, they mostly seem to come from places where we are using fsyncLock to block writes, rather than guarantee filesystem consistency (replication tests where we block writes to secondary's are a common example). There are about 10~ test suites which would need to be modified in order to stop these errors.
Wanted your input on how best to proceed, currently I see the following as options:
- Leave the tests broken until we fix the issue in mongo-rocks
- Modify all of theses tests to skip the relevant part when fsync returns CommandNotSupported