Pick correct denominator to make quantile aggregate function's result deterministic

[onkar]

Environment

Snuba SaaS

Steps to Reproduce

Some of the customers are reporting a problem that user's p50, p75 and p95 values are non-deterministic. There are quantile and quantileDeterministic aggregate functions in ClickHouse that are used in this case. They both use reservoir sampling method to compute approximate quantiles. The only difference is that quantile uses a random number generator to pickup the samples, making the results non-deterministic, while quantileDeterministic needs caller to pass in a denominator and that denominator is used to pick out the samples.

• This issue is reproduced in Duration values in dashboards non-deterministic sentry#80067
• Upon refresh of the page, the values of p50, p75 and p95 fluctuate on this dashboard
• Similar behavior can be found on the transaction data
• https://github.com/user-attachments/assets/4c55b439-0cf6-4a14-8b87-8b834d45b593

The main task is to pick a correct denominator that the customer can pass in. The lack of determinism comes from the random seed and a non-random denominator should be able to give the result we want. This ClickHouse documentation can be useful in this context. Specifically, this section of the documentation is a bit unclear:

If the same determinator value occurs too often, the function works incorrectly.

Expected Result

The fluctuation in the p50, p75 and p95 values does not occur.

Actual Result

p50, p75 and p95 values are non-deterministic.