> Actually, if we wanted to improve things in this area, we should have a
> set of queries that don't chose optimal plans we can test with. We used
> to see them a lot before we had extended statistics, but I don't
> remember seeing many recently, let alone a collection of them. I guess
> that is good.
In the VLDB paper they actually created their own "Join Order Benchmark", which is publicly available under
https://github.com/gregrahn/join-order-benchmark
It would probably be more suited for this kind of testing than, e.g. the TPC benchmarks.
If there is interest, I could also compile a set of relevant cases based on the message history of the performance
mailinglist.