...
When running the query against the read replica, the expected index is being used. Removing `author_id_1` should not have affected this because the compound index `author_id_1_course_id_1` would be used instead. The median remains the same and the 95th percentile did not spike like the 99th. It seems like the proper indexes are in place, otherwise, there would be a regression for the median as well. One possible explanation could be the increase of index size explained in this wiki below.
The threads endpoint is experiencing a similar situation as the users endpoint.
Index size:
The indexes that we added takes up more RAM than the ones removed. To add to that, the `delete_spam` index that was removed had a size of 0 which does not really improve performance at all. Overall it seems like another 1GB of indexes was added.
...