Re: Need help to review a serializability implementation for MySQL Cluster

[Alex Ou]

A review of TiDB is up.

In Section 5, we've generalized Serializable Snapshot Isolation to a distributed database system based on the Clock Condition. This review complements Section 5 in the sense that it demonstrates how to apply the method developed there to a distributed Snapshot Isolation implementation in the case that its distributed clock doesn't satisfy the Clock Condition. This gives us the second serializability implementation for a distributed database system at 2-tier. The nice thing about it is that its SQL interface is also MySQL and we may again use conflict_parser to process and analyze its applications.

Unfortunately, right now I can't mount a serializability implementation to TiDB's pessimistic isolation levels – Read-Committed and Repeatable-Read – as I did for NDB Cluster. A tiny piece of the puzzle is missing and it is explained in this subsection. I'll file a feature request to see if there is any luck.

More details about applying this work to MySQL InnoDB will be available in a few weeks.