Hi Team! Thank you for your hard work on 4.0 release. Iceberg tables query performance looks great! I have a question regarding hll functions that were added in 4.0

ds_hll_combine

and

ds_hll_estimate

. We run StarRocks 3.5.7 right now so we have to use custom Java UDFs for HLL. I'm testing the performance of new functions to replace existed custom UDFs. It works good with mat views. I see overall 10-30% improvements in query speed. BUT I see 2x-3x performance degradation for Iceberg tables i.e. my Java UDF works faster. Java UDFs we have are simple wrapper around DS HLL java lib without any extra tweaks. I did

explain analyze

and found that:

Plan 1: AggComputeTime: 1s177ms - my UDF
Plan 2: AggComputeTime: 3s365ms - ds_hll_estimate

What can be the possible reason here so ds_hll* functions works faster for mat views but not for Iceberg tables? Is there any setting I need to update? Let me know if you need more details.