Make cost model more precisely

[hanfei1991]

There're several physical executions that need to be selected during cbo. Currently we only estimate the cost roughly. We need more experiments to prove the correctness of estimation.

• For DataSource, we need to compare the cost between TableReader, IndexReader, Index Look up.
• For Aggregation, we need to compare the cost between HashAggregation and StreamAggregation.
• For Join, we need to compare the cost between HashJoin, IndexJoin and MergeJoin.

Besides these, we still need to test the composition of two executors. Say

• DataSource + (Selection) + Sort (asc/desc) , we need to compare the keep order and memory sort.
• DataSource + (Selection) + TopN, we need to compare the keep order + limit and TopN.
• DataSource + Aggregation, we need to check push down or not, if the cardinality of the keys is huge.
• Multi Join
• Join + Aggregation, -> Index Join + Stream, Index Join + Hash, Hash Join + Hash, Merge Join + Stream
• And so on

The result of expriments will be posted here.

[morgo]

@zz-jason can you help re-triage?

[zz-jason]

OK. We still need to update the cost formular for these physical operators, it's a long term issue, I think we can leave it open.

/cc @eurekaka

[zz-jason]

hope we can finish it in release 3.0