Breaking that `group_concat()` example certainly looks like an unacceptable change - I wouldn't be at all surprised to see real-world code that relies on that. In my particular case queries that I was running extremely frequently (often several times for a single HTTP page load) got nearly a 10x speed-up - but I've now changed my code to remove the inner `order by` (see <https://github.com/simonw/datasette/issues/1394>) so my project would no longer benefit from this optimization.