My takeaways: > 4.2 Recommendations > The core problem is that MySQL claims to implement Repeatable Read but actually provides something much weaker. We see t

My takeaways: > 4.2 Recommendations > The core problem is that MySQL claims to implement Repeatable Read but actually provides something much weaker. We see two avenues to resolve this problem. > The first is to keep MySQL’s behavior as it is, and to clearly document the consistency model “Repeatable Read” actually provides. There is precedent in other databases: PostgreSQL’s Repeatable Read is actually Snapshot Isolation, and exhibits behaviors which violate PL-2.99 Repeatable Read. However, PostgreSQL’s documentation eventually mentions that their Repeatable Read implementation is actually Snapshot Isolation. MySQL could similarly document that their “Repeatable Read” means “Read Committed, plus some sort of guarantees that hold until the transaction writes something, at which point mysteries occur.” A precise characterization of those mysteries would be most welcome. Calling what MySQL's "Repeatable Read" as "Read Committed plus..." would be more confusing as even the simplest repeated read without mutations wouldn't work as expected. The documentation should be more upfront about how MySQL "Repeatable Read" doesn't mean what might be expected. In the meantime keep the "MySQL consistent read documentation"[0] close by. > The second option is to treat these behaviors as bugs and fix them. Jepsen would be delighted if MySQL and other vendors were to commit to providing PL-2.99 Repeatable Read. However, even satisfying the incomplete, ambiguous ANSI definition of Repeatable Read would be an improvement over current affairs. I doubt this would be feasible with the extent of deployment. At scale bug-fixes are bugs in themselves. The best that could be done is to create distinct isolation levels for the existing MySQL-RR and the compliant RR isolation levels (somewhat akin to the utf8/utf8mb4 evolution). [0] https://dev.mysql.com/doc/refman/8.0/en/innodb-consistent-read.html https://dev.mysql.com/doc/refman/8.0/en/innodb-consistent-re...