A couple days ago, a mistake caused a key-value table to generate new keys for duplicate values when there should have been a unique constraint on key. This affected the stat gathering table (stats for one value are now associated with many keys) and I have to clean up the data -- while preserving the stats that have already been generated. In trying to do this, I think I found a bug in replication .... an UPDATE statement that aggregates the data produces different results on my replication slave. I've submitted a formal bug report along with SQL to duplicate the error here. I'm curious if anyone else has encountered this bug, or knows of a way to achieve the result that I want without breaking replication.

Here's a description of the test case:

Two tables store key-value pairs (in the example, `color` and `animal`), and `a_to_c` stores a relationship …