I noticed the de-optimization bits when the compiler handles corner cases. How often did Vitess notice that the VM had to switch back to the old AST implementation?
Very very seldomly! Obviously we need to be correct in 100% of the cases, which is why the de-optimizer is there, but in practice it just never triggers. Right now the things that trigger a deopt are very limited — malformed time literals, oferflowing integer negations and very little else. The performance impact is essentially zero.
