Comment 2 for bug 504291

I missed a vital point in my explanation above: what this all boils down to is that AFAICS a connection gets stuck in this bad state if it's broken (e.g. by a database restart) just when it's in the middle of a commit.

The bad connection will stay in the store and get reused by requests, all of which fail with this error. And because they fail quickly, the bad connection becomes available again quickly to carry another request to its doom.