In most respects, PostgreSQL behaved as expected: both read uncommitted and read committed prevent write skew and aborted reads. We observed no internal consistency violations. However, we have two surprising results to report. The first is that PostgreSQL’s “repeatable read” is weaker than repeatable read, at least as defined by Berenson, Adya, Bailis, et al. This is not necessarily wrong: the ANSI SQL standard is ambiguous. The second result, which is definitely wrong, is that PostgreSQL’s “serializable” isolation level isn’t serializable: it allows G2-item during normal operation. "
