The formula of calculating net cost for TiFlash plans is not aligned with TiKV plans'

[qw4990]

Enhancement

In copTask.convertToRootTaskImpl, we can see the formula of calculating net cost for TiKV plans is est-row * row-size * net-factor, while in mppTask.convertToRootTaskImpl the formula for TiFlash plans is est-row * net-factor where the row-size is lost.
That would let the cost of TiFlash plans be always relatively lower than TiKV plans', which can cause some wrong plans.

[zhangjinpeng87]

It should be a bug? @qw4990

[qw4990]

It should be a bug? @qw4990

Yep, we can regard this issue as a bug also, but it may not be appropriate.
To fix it, we have to design some experiments and test cases to evaluate the new formula to ensure there is no regression, which is what we are doing in Sprint6.
It seems that we cannot fix it like a bug quickly, so I just labeled it with enhancement and planned to solve it in Sprint6 by the project Cardinality&Cost Enhancement.

[yudongusa]

This is part of cost model enhancement planned in Sprint 6 in which this and other related costing enhancement and issues will be examined and improved. For this particular issue, a simple addition of row_size may not work as TiFlash only scans and transmits certain columns, so rather a fine tuning of the TiFlash cost formula would be needed based on some experiments.