The rocksdb.index_merge_rocksdb2 MTR test inadvertently runs some of its queries against InnoDB tables rather than MyRocks. Which results from a combination of 2 things:
- the way the test is implemented by Facebook is setting the engine_type variable, then including a number of include/index_merge*.inc files. All of which do "SET SESSION DEFAULT_STORAGE_ENGINE = $engine_type" so that the specified engine is used when the tables are later created without an explicit storage engine specification;
- and then there are Percona-specific changes to those include/index_merge*.inc files introduced by https://github.com/percona/percona-server/commit/35afd73da3ca. Which restart the server at certain points in time, thus resetting the default storage engine set earlier by the test. Which also means all further tables in the test are created with the default-default storage engine, i.e. InnoDB.
Here's an easy way to verify. Apply the following patch and run ./mtr rocksdb.index_merge_rocksdb2 --big-test:
This is what I see in the output on the latest Percona Server clone: