The order of aggregation is not guaranteed to be deterministic, see note here. My understanding is that HLL sketches will have minor variations from run to run for this reason. The results should still fall within the accuracy level expected.
a
Andrei Harbunou
04/03/2023, 4:50 PM
Is it only because of underlying HLL lib behavior? I ran Druid 25 with old HLL lib (1.1.0) and results are consistent.
s
Sergio Ferragut
04/03/2023, 5:02 PM
I'm not sure. Perhaps something changed in the order of aggregation in between 0.19 and 25.0. What difference do you see?
a
Andrei Harbunou
04/03/2023, 5:13 PM
In general, just inconsistent result for the same request. I haven't found big difference in HLL accuracy.
0.18.1 groupby result is always deterministic for HLL.
It is a big shift for customers as even though they know that value is approximate people used to rely on this because of stability.
s
Sergio Ferragut
04/03/2023, 5:17 PM
I'll ask around for more clarity on why this change occurred. I think that as long as the accuracy is consistent, it is working as expected.