Hi, I've seen a fix merged for this.

It's a good idea to try to apply some boundaries, but with the patch which was applied it will occasionally reset collected data about some "randomly" selected queries (as picked by the LRU algorhytm - clearly there's some logic in that but it might as well be random for end users), this will lead to inaccurate or even misleading data.

I wonder if that makes the whole feature pointless. Maybe when there are lots of different queries being generated one should rather disable stats collection?

Or, probably safest, we could impose a hard limit when there's more than N QueryStatisticsImpl already collected, and just stop adding more. That will also be a bit unpredictable, but at least people will know that the data they have is reliable: for some other queries there will be no data at all.

Most importantly, such choices need to be clearly documented.