[multistage]partition assignment refactor by walterddr · Pull Request #12079 · apache/pinot

walterddr · 2023-12-01T00:50:17Z

step 2 in #12015

this PR:

loosen the criteria for "co-assigning workers to the same partition to avoid shuffling" to non-leaf children
reshape the rule of partition assignment & mailbox rule.
- At logical plan time:
  - ExchangeNode now knows whether an exchange is "pre-partitioned" --> e.g. the input partition is the same desired by the exchange (exchange might still be needed if the keys are the same but functions are different)
- At worker assignment time:
  - For intermediate-stage, ExchangeNode with "pre-partitioned" is required for assigning the same set of workers from the 1st child;
  - For leaf-stage, worker assignment is only controlled by tableOptions - e.g. assigned as table partition
- At mailbox assignment time:
  - For all mailboxes: to determine whether a "direct" exchange can apply.
    - check partition_parallelism
      - if both sender/receiver are leaf (semi-join, partition_parallelism shouldn't apply; num_worker_sender == num_worker_receiver = num_partition_of_table (both))
      - if both sender/receiver are intermediate stage, partition_parallelism shouldn't apply num_worker_sender == num_worker_receiver)
      - if sender is leaf and receiver is intermediate: num_worker_received == num_worker_sender * partition_parallelism; num_worker_sender == num_partition_of_table)
    - check partition_function
      - sender stage plan metadata partition function == receiving stage plan metadata partition function
  - NOTE:
    - Exchange "pre-partitioned" is necessary and sufficient to assign direct exchange for 1st child and intermediate stage;
    - for partition function check it is also necessary to check if the mailbox send is using the same partition function as the receiving stage metadata indicates (tracked in TODO)

Work List

redefine partition info in DispatchablePlanMetadata and clear boundary between server/worker assignment and partitioned/non-partitioned optimization;
add partition function info as part of the validation step during worker assignment

TODO:

allow select partition function via config for all exchanges; nowadays it is hardcoded to "table hash function" for leaf; and "abs_hash_order_invariant" hash for non-leaf stages (if not using direct exchange);
- once allowed, the condition for prepartition assignment can be loosen further
throw when "colocated hint is given but prepartition assignment condition check fails"
allow utilizing the same hash-function to recompute hashes when doing parallelism fan-out

codecov-commenter · 2023-12-01T01:26:05Z

Codecov Report

Attention: 40 lines in your changes are missing coverage. Please review.

Comparison is base (3474db4) 61.68% compared to head (154f6ef) 61.48%.
Report is 5 commits behind head on master.

Files	Patch %	Lines
.../org/apache/pinot/query/routing/WorkerManager.java	75.88%	24 Missing and 10 partials ⚠️
...ery/planner/physical/MailboxAssignmentVisitor.java	58.33%	0 Missing and 5 partials ⚠️
.../query/planner/logical/RelToPlanNodeConverter.java	50.00%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@             Coverage Diff              @@
##             master   #12079      +/-   ##
============================================
- Coverage     61.68%   61.48%   -0.20%     
- Complexity     1151     1153       +2     
============================================
  Files          2391     2406      +15     
  Lines        130211   130851     +640     
  Branches      20141    20218      +77     
============================================
+ Hits          80317    80457     +140     
- Misses        44046    44513     +467     
- Partials       5848     5881      +33

Flag	Coverage Δ
custom-integration1	`?`
integration	`<0.01% <0.00%> (-0.01%)`	⬇️
integration1	`<0.01% <0.00%> (-0.01%)`	⬇️
integration2	`0.00% <0.00%> (ø)`
java-11	`27.63% <0.00%> (-34.02%)`	⬇️
java-21	`34.85% <76.19%> (-0.11%)`	⬇️
skip-bytebuffers-false	`61.46% <76.19%> (-0.22%)`	⬇️
skip-bytebuffers-true	`34.83% <76.19%> (?)`
temurin	`61.48% <76.19%> (-0.20%)`	⬇️
unittests	`61.48% <76.19%> (-0.20%)`	⬇️
unittests1	`46.56% <76.19%> (-0.39%)`	⬇️
unittests2	`27.63% <0.00%> (+0.03%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

walterddr

CC @Jackie-Jiang to take a look

walterddr · 2023-12-04T23:50:53Z

pinot-query-planner/src/main/java/org/apache/pinot/query/routing/WorkerManager.java

not sure if this is the best way to go. i am also ok with having to ask this pass in as mandatory

walterddr · 2023-12-04T23:51:40Z

pinot-query-planner/src/main/java/org/apache/pinot/query/routing/WorkerManager.java

technically the class should be named "PartitionTableInfo". this doesn't indicate "co-locate" at all

Either change it or add a TODO

walterddr · 2023-12-04T23:52:23Z

pinot-query-planner/src/test/resources/queries/PinotHintablePlans.json

TODO: fix this test, it should always check the nested reason b/c that's always wrapped in parser throw or planner throw. doesn't make sense to check just the top level msg

walterddr · 2023-12-04T23:53:05Z

pinot-query-runtime/src/test/resources/queries/QueryHints.json

previously we decided to allow this but i guess we should not. putting an ignore here first but i think we should not allow this and should throw exception

note: allowed here but it will revert back to generic assignment

this field is to indicate whether the exchange (or mailbox) requires to reshuffle the data isPartition semantically means if the data is already partitioned, which can be true, just not the same partition as the mailbox/exchange expects

fix several rebase error

1. we don't support rehash by pre-hash method (exchange doesn't support the same with table) so check all pre-partition instead of check single child 2. set default for exchange and default for table separately 3. fix tests

...-planner/src/main/java/org/apache/pinot/query/planner/physical/DispatchablePlanMetadata.java

Jackie-Jiang · 2023-12-07T17:37:14Z

pinot-query-planner/src/main/java/org/apache/pinot/query/routing/WorkerManager.java

Either change it or add a TODO

Jackie-Jiang · 2023-12-07T17:39:07Z

pinot-query-planner/src/main/java/org/apache/pinot/query/routing/WorkerManager.java

+    Preconditions.checkState(numPartitions > 0, "'%s' must be positive, got: %s",
+        PinotHintOptions.TableHintOptions.PARTITION_SIZE, numPartitions);
+
+    String partitionFunction = tableOptions.getOrDefault(PinotHintOptions.TableHintOptions.PARTITION_FUNCTION,


Don't add default. In most cases this is murmur but we should not assume it

this would be backward-incompatible, all existing queries with table hints without hash function will fail

Understood. Currently we assume all table options using the same partition function. Putting Murmur as the default might be safer

Jackie-Jiang · 2023-12-07T17:39:44Z

...-planner/src/main/java/org/apache/pinot/query/planner/physical/MailboxAssignmentVisitor.java

+      // leaf-to-intermediate condition
+      return numSenders * sender.getPartitionParallelism() == numReceivers
+          && sender.getPartitionFunction() != null
+          && sender.getPartitionFunction().equals(receiver.getPartitionFunction());


We should compare ignore cases

walterddr mentioned this pull request Dec 1, 2023

[multistage][feature] RelDistribution-based optimization #12015

Open

3 tasks

walterddr changed the title ~~[multistage][draft] partition assignment refactor~~ [multistage]partition assignment refactor Dec 4, 2023

walterddr marked this pull request as ready for review December 4, 2023 23:31

walterddr force-pushed the pr_partition_assignment_refactor branch 2 times, most recently from ad29058 to 78f9d7a Compare December 4, 2023 23:36

walterddr commented Dec 4, 2023

View reviewed changes

Jackie-Jiang added refactor multi-stage Related to the multi-stage query engine labels Dec 5, 2023

Rong Rong added 6 commits December 6, 2023 15:21

[draft] allow partition info to pull all the way through above

190388a

[test] fix tests

c6d5733

adding hash-function as part of the physical plan info

4c81606

add tests for partition functions and

471deea

fix several rebase error

further clean up

325c1b2

walterddr force-pushed the pr_partition_assignment_refactor branch from a63a686 to 325c1b2 Compare December 6, 2023 23:22

Rong Rong added 2 commits December 6, 2023 15:53

fixing sevearl bug

04f98ac

1. we don't support rehash by pre-hash method (exchange doesn't support the same with table) so check all pre-partition instead of check single child 2. set default for exchange and default for table separately 3. fix tests

fix regex escape

c70dda0

Jackie-Jiang approved these changes Dec 7, 2023

View reviewed changes

Rong Rong added 3 commits December 7, 2023 12:34

rename colocated to partition table info

1c92ca5

compare partition function in case insensitive manner

afd4291

changing default to murmur and modify all tests

154f6ef

walterddr merged commit e207844 into apache:master Dec 7, 2023

Conversation

walterddr commented Dec 1, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Work List

Uh oh!

codecov-commenter commented Dec 1, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

walterddr left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

walterddr commented Dec 1, 2023 •

edited

Loading

codecov-commenter commented Dec 1, 2023 •

edited

Loading