Cost model of aggregation should consider the agg type.

[lzmhhh123]

Feature Request

Is your feature request related to a problem? Please describe:

In pull request #11926. We found the coprocessor of tikv execute avg very similar with sum and count. But the cost model always regards its a more expensive plan. The reason is the cost model of aggregation only take the number of agg functions into consideration. So we should improve the cost model by taking care of the type of agg functions.

Describe the feature you'd like:

Improve the cost model of aggregation by taking care of the type of agg functions.

Describe alternatives you've considered:

None.

[eurekaka]

Besides from Avg, is there any other kind of aggregation that would be split during execution?

[lzmhhh123]

That's another problem. We can support split agg functions one by one, and resolve it step by step. Then add a rule to merge the same agg functions for a logicalAgg plan.

[jingyugao]

How about the cost of sum(a)+sum(b) compared to sum(a+b)

[lzmhhh123]

How about the cost of sum(a)+sum(b) compared to sum(a+b)

When thinking about the null column situation, this split may cause wrong results.

[breezewish]

Does the cost model needs to be updated systematically since performance characters are very different in TiKV batch execution model compared to the old execution model?

[lzmhhh123]

Does the cost model needs to be updated systematically since performance characters are very different in TiKV batch execution model compared to the old execution model?

That's necessary. We should take it into consideration in future.