planner, statistics: maintain histogram for inner join #8097
Conversation
winoros
added
type/enhancement
winoros
force-pushed
the
inner-join-hist
branch
from
232fe9b to
2adb8dd
Compare
winoros
force-pushed
the
inner-join-hist
branch
from
72667ab to
ea60d70
Compare
winoros
force-pushed
the
inner-join-hist
branch
from
63cf9bb to
8a9ac4f
Compare
|
I'll add more tests to cover the case that the count of |
Codecov Report
@@ Coverage Diff @@
## master #8097 +/- ##
==========================================
+ Coverage 67.16% 67.23% +0.06%
==========================================
Files 371 371
Lines 76393 76543 +150
==========================================
+ Hits 51311 51461 +150
- Misses 20494 20496 +2
+ Partials 4588 4586 -2
Continue to review full report at Codecov.
|
planner/core/stats.go
Outdated
| @@ -281,6 +281,9 @@ func (p *LogicalJoin) DeriveStats(childStats []*property.StatsInfo) (*property.S | |||
| leftKeys = append(leftKeys, eqCond.GetArgs()[0].(*expression.Column)) | |||
| rightKeys = append(rightKeys, eqCond.GetArgs()[1].(*expression.Column)) | |||
| } | |||
| if p.JoinType == InnerJoin && p.ctx.GetSessionVars().OptimizerSelectivityLevel >= 1 { | |||
| return p.deriveInnerJoinStatsWithHist(leftKeys, rightKeys) | |||
There was a problem hiding this comment.
Shall we pass the childStats parameter down to deriveInnerJoinStatsWithHist? so this DeriveStats function can be used by cascades planner as well.
There was a problem hiding this comment.
Oh, using childStats in new commits. But it uses the schema of join's children inside this method. Seems that i need some way to not rely on it.
winoros
force-pushed
the
inner-join-hist
branch
from
cc8ea08 to
15485aa
Compare
| leftCol := &statistics.Column{Info: leftHist.Info, Histogram: *newHist} | ||
| rightCol := &statistics.Column{Info: rightHist.Info, Histogram: *newHist} | ||
| lIncreaseFactor := leftHist.GetIncreaseFactor(leftProfile.HistColl.Count) | ||
| // The factor is used to scale the NDV. When it's higher than one. NDV doesn't need to be changed. |
There was a problem hiding this comment.
Why don't we change NDV if this factor is larger than one?
| keyNdv = float64(newHist.NDV) * lIncreaseFactor * rIncreaseFactor | ||
| lPosNew := p.schema.ColumnIndex(leftKeys[i]) | ||
| rPosNew := p.schema.ColumnIndex(rightKeys[i]) | ||
| cardinality[lPosNew] = float64(newHist.NDV) |
There was a problem hiding this comment.
Why not multiply this cardinality by lIncreaseFactor like keyNdv.
|
closed for a while. will reopen it in the future. |
What problem does this PR solve?
Initial commit for maintaining histogram for inner join.
Don't consider use index to calculate join's
StatsInfocurrently.What is changed and how it works?
Following the description in #7605
Check List
Tests
Not enough. I'll add more, but maybe not in this pr.
Code changes
Side effects
This change is