replace PODArray with PODArrayWithStackMemory in AggregateFunctionWindowFunnelData

[ucasfl]

I hereby agree to the terms of the CLA available at: https://yandex.ru/legal/cla/?lang=en

Changelog category (leave one):

• Performance Improvement

Changelog entry (a user-readable short description of the changes that goes to CHANGELOG.md):

Replace PODArray with PODArrayWithStackMemory in AggregateFunctionWindowFunnelData to improvement windowFunnel function performance.

Detailed description / Documentation draft:

Funnel analysis is very useful in many scenarios.

In Tencent, we use windowFunnel aggregate function to analyze user events, and it's slow. we profile with perf and found that lots of time is used in memory allocation, CPU utilization is not very high.

We consider the implementation of this function, and analyze it with specific sql, then we found that usually, we always caculate windowFunnel with group by userid and lots of fileter condition, for example:

SELECT
    level,
    count()
FROM
(
    SELECT
        uid,
        windowFunnel(86400)(time, event = '浏览', event = '点击', event = '下单', event = '支付') AS level
    FROM action where ...// lots of filter condition
    GROUP BY uid
)
GROUP BY level

After filter, each user's event list is not very big, so we replace PODArray with PODArrayWithStackMemory in AggregateFunctionWindowFunnelData and gain a lot of performance improvement, maybe we should consider about the implementation here?

[alexey-milovidov]

Yes, this is very typical optimization for aggregate functions!
I did not dig into windowFunnel implementation in more detail - maybe there are more possible optimizations.

This PR is Ok to merge, but let's add performance test to prove and keep the improvement, see tests/performance.

[ucasfl]

Yes, this is very typical optimization for aggregate functions!
I did not dig into windowFunnel implementation in more detail - maybe there are more possible optimizations.

This PR is Ok to merge, but let's add performance test to prove and keep the improvement, see tests/performance.

Ok, currently, we still working to find more possible optimization.

[alexey-milovidov]

#18878 @alesapin
#18879 @tavplubix