(Filing this as a follow-up to discussions on Slack)
Before making JSON_HB histograms enabled by default (
MDEV-27062), it would be nice to make sure they will not cause performance regressions at the users' systems.
Note: we want to check for performance regressions, not histogram accuracy (Elena has run some tests for histogram accuracy already)
Note: it won't be possible to avoid performance regressions completely. One can construct realistic-looking cases where improved histogram precision will cause a worse query plan due to the "Optimizer was making two wrong choices, we've fixed one" effect.
(The most likely second wrong choice is not accounting for correlations between conditions).
The following was suggested: run some test that was not specifically designed with JSON_HB histograms in mind and check the results. The problem is finding a benchmark or workload where JSON_HB vs DOUBLE_PREC_HB difference will be apparent.