[SPARK-34922][SQL][3.0] Use a relative cost comparison function in the CBO #32076
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Kubernetes integration test starting |
|
Kubernetes integration test status failure |
|
Refer to this link for build results (access rights to CI server needed): |
|
Test build #137007 has finished for PR 32076 at commit
|
|
Refer to this link for build results (access rights to CI server needed): |
Contributor
Author
maropu
pushed a commit
that referenced
this pull request
Apr 8, 2021
…e CBO ### What changes were proposed in this pull request? Changed the cost comparison function of the CBO to use the ratios of row counts and sizes in bytes. ### Why are the changes needed? In #30965 we changed to CBO cost comparison function so it would be "symetric": `A.betterThan(B)` now implies, that `!B.betterThan(A)`. With that we caused a performance regressions in some queries - TPCDS q19 for example. The original cost comparison function used the ratios `relativeRows = A.rowCount / B.rowCount` and `relativeSize = A.size / B.size`. The changed function compared "absolute" cost values `costA = w*A.rowCount + (1-w)*A.size` and `costB = w*B.rowCount + (1-w)*B.size`. Given the input from wzhfy we decided to go back to the relative values, because otherwise one (size) may overwhelm the other (rowCount). But this time we avoid adding up the ratios. Originally `A.betterThan(B) => w*relativeRows + (1-w)*relativeSize < 1` was used. Besides being "non-symteric", this also can exhibit one overwhelming other. For `w=0.5` If `A` size (bytes) is at least 2x larger than `B`, then no matter how many times more rows does the `B` plan have, `B` will allways be considered to be better - `0.5*2 + 0.5*0.00000000000001 > 1`. When working with ratios, then it would be better to multiply them. The proposed cost comparison function is: `A.betterThan(B) => relativeRows^w * relativeSize^(1-w) < 1`. ### Does this PR introduce _any_ user-facing change? Comparison of the changed TPCDS v1.4 query execution times at sf=10: | absolute | multiplicative | | additive | -- | -- | -- | -- | -- | -- q12 | 145 | 137 | -5.52% | 141 | -2.76% q13 | 264 | 271 | 2.65% | 271 | 2.65% q17 | 4521 | 4243 | -6.15% | 4348 | -3.83% q18 | 758 | 466 | -38.52% | 480 | -36.68% q19 | 38503 | 2167 | -94.37% | 2176 | -94.35% q20 | 119 | 120 | 0.84% | 126 | 5.88% q24a | 16429 | 16838 | 2.49% | 17103 | 4.10% q24b | 16592 | 16999 | 2.45% | 17268 | 4.07% q25 | 3558 | 3556 | -0.06% | 3675 | 3.29% q33 | 362 | 361 | -0.28% | 380 | 4.97% q52 | 1020 | 1032 | 1.18% | 1052 | 3.14% q55 | 927 | 938 | 1.19% | 961 | 3.67% q72 | 24169 | 13377 | -44.65% | 24306 | 0.57% q81 | 1285 | 1185 | -7.78% | 1168 | -9.11% q91 | 324 | 336 | 3.70% | 337 | 4.01% q98 | 126 | 129 | 2.38% | 131 | 3.97% All times are in ms, the change is compared to the situation in the master branch (absolute). The proposed cost function (multiplicative) significantlly improves the performance on q18, q19 and q72. The original cost function (additive) has similar improvements at q18 and q19. All other chagnes are within the error bars and I would ignore them - perhaps q81 has also improved. ### How was this patch tested? PlanStabilitySuite Closes #32076 from tanelk/SPARK-34922_cbo_better_cost_function_3.0. Lead-authored-by: Tanel Kiis <[email protected]> Co-authored-by: [email protected] <[email protected]> Signed-off-by: Takeshi Yamamuro <[email protected]>
Member
|
Thanks! Merged to branch-3.0. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What changes were proposed in this pull request?
Changed the cost comparison function of the CBO to use the ratios of row counts and sizes in bytes.
Why are the changes needed?
In #30965 we changed to CBO cost comparison function so it would be "symetric":
A.betterThan(B)now implies, that!B.betterThan(A).With that we caused a performance regressions in some queries - TPCDS q19 for example.
The original cost comparison function used the ratios
relativeRows = A.rowCount / B.rowCountandrelativeSize = A.size / B.size. The changed function compared "absolute" cost valuescostA = w*A.rowCount + (1-w)*A.sizeandcostB = w*B.rowCount + (1-w)*B.size.Given the input from @wzhfy we decided to go back to the relative values, because otherwise one (size) may overwhelm the other (rowCount). But this time we avoid adding up the ratios.
Originally
A.betterThan(B) => w*relativeRows + (1-w)*relativeSize < 1was used. Besides being "non-symteric", this also can exhibit one overwhelming other.For
w=0.5IfAsize (bytes) is at least 2x larger thanB, then no matter how many times more rows does theBplan have,Bwill allways be considered to be better -0.5*2 + 0.5*0.00000000000001 > 1.When working with ratios, then it would be better to multiply them.
The proposed cost comparison function is:
A.betterThan(B) => relativeRows^w * relativeSize^(1-w) < 1.Does this PR introduce any user-facing change?
Comparison of the changed TPCDS v1.4 query execution times at sf=10:
All times are in ms, the change is compared to the situation in the master branch (absolute).
The proposed cost function (multiplicative) significantlly improves the performance on q18, q19 and q72. The original cost function (additive) has similar improvements at q18 and q19. All other chagnes are within the error bars and I would ignore them - perhaps q81 has also improved.
How was this patch tested?
PlanStabilitySuite