planner: split avg to count and sum for TableReader cop task #11926
Conversation
Codecov Report
@@ Coverage Diff @@
## master #11926 +/- ##
===========================================
Coverage 81.4033% 81.4033%
===========================================
Files 444 444
Lines 95560 95560
===========================================
Hits 77789 77789
Misses 12291 12291
Partials 5480 5480 |
|
/run-all-tests |
|
/run-integration-common-test |
1 similar comment
|
/run-integration-common-test |
|
@lzmhhh123 Does it affect the execution logic on TiKV Coprocessor? |
I'm checking this situation. |
|
/run-all-tests |
| @@ -864,7 +864,7 @@ func (s *testPlanSuite) TestDAGPlanBuilderAgg(c *C) { | |||
| // Test agg + table. | |||
| { | |||
| sql: "select sum(a), avg(b + c) from t group by d", | |||
| best: "TableReader(Table(t)->HashAgg)->HashAgg", | |||
| best: "TableReader(Table(t))->Projection->HashAgg", | |||
There was a problem hiding this comment.
I felt confused too. So I'm investigating the reason now.
There was a problem hiding this comment.
I guess it should be caused by the fact that the number of AggFuncs increases, so the cost of Aggregation is changed, and planner decides not to push this Agg down to coprocessor.
There was a problem hiding this comment.
According to #11926 (comment), seems we should modify the cost model to decrease the impact of number of aggregate functions to make this aggregation can be pushed to tikv?
There was a problem hiding this comment.
We can temporarily not to modify the cost model. That's another task of cost model improvement with the consideration of aggregation type.
There was a problem hiding this comment.
OK, How about filing an issue about this improvement?
It passed the integration-common-test at least. |
Yeah, does it mean that the execution instruction in TiKV is not changed? I'm afraid that this pr would impact the end-to-end performance of a SQL. |
Only a little. Because TiKV executes |
| @@ -864,7 +864,7 @@ func (s *testPlanSuite) TestDAGPlanBuilderAgg(c *C) { | |||
| // Test agg + table. | |||
| { | |||
| sql: "select sum(a), avg(b + c) from t group by d", | |||
| best: "TableReader(Table(t)->HashAgg)->HashAgg", | |||
| best: "TableReader(Table(t))->Projection->HashAgg", | |||
There was a problem hiding this comment.
I guess it should be caused by the fact that the number of AggFuncs increases, so the cost of Aggregation is changed, and planner decides not to push this Agg down to coprocessor.
zz-jason
added
status/LGT2
|
/run-all-tests |
What problem does this PR solve?
Because the calculation of
avgin the coprocessor is split assumandcount, and it also returns two columns. We can splitavgin the planner. Now it's only effective for cop agg inTableReader.A sub PR for support the coprocessor in tiflash.
What is changed and how it works?
Check List
Tests
Side effects
Related changes