Prevent repartitioning of certain operator's direct children (#1731) #1732

tustvold · 2022-02-02T15:51:26Z

Which issue does this PR close?

Closes #1731.

Rationale for this change

Not all operators benefit from additional concurrency, in particular those that are not CPU-intensive (e.g. UnionExec, ProjectionExec) or are not embarrassingly parallel (e.g. LimitExec). Introducing additional partitioning is unlikely to yield a performance benefit for these operators, and may in fact make execution slower. It also makes the plans harder to follow.

What changes are included in this PR?

Adds a new trait method ExecutionPlan::should_repartition_children that allows operators to indicate that their direct children should not be repartitioned.

Are there any user-facing changes?

There is a new optional ExecutionPlan method

…1731)

tustvold · 2022-02-02T17:21:28Z

datafusion/tests/sql/explain_analyze.rs

@@ -700,15 +700,15 @@ async fn test_physical_plan_display_indent_multi_children() {
        "    HashJoinExec: mode=Partitioned, join_type=Inner, on=[(Column { name: \"c1\", index: 0 }, Column { name: \"c2\", index: 0 })]",
        "      CoalesceBatchesExec: target_batch_size=4096",
        "        RepartitionExec: partitioning=Hash([Column { name: \"c1\", index: 0 }], 3)",
-        "          ProjectionExec: expr=[c1@0 as c1]",
+        "          RepartitionExec: partitioning=RoundRobinBatch(3)",


This plan is a little bit strange, but not a strangeness introduced by this PR.

Perhaps RepartitionExec should itself be should_repartition_children 🤔

Edit: I did this in the end

Dandandan

We

datafusion/src/physical_plan/projection.rs

datafusion/src/physical_plan/repartition.rs

Dandandan · 2022-02-02T18:26:14Z

I put some comments on ProjectionExec and RepartitionExec which might in some cases benefit from being repartitioned (maybe not in many cases, but local round-robin repartitioning is AFAIK very cheap compared to e.g. hashing string columns or parsing numbers etc.)

tustvold · 2022-02-02T18:38:54Z

@Dandandan I've reverted the changes to those, the main operator I'm interested in excluding is UnionExec as it is showing up in IOx scan plans and yielding 3000 odd partitions as it repartitions every parquet file 28 times. 😆

For context one of the reasons I'm interested in reducing the concurrency is somewhat stupid, the plans and the resulting metrics and tracing become HUGE, but I also can't help but feel that always pursuing the maximum concurrency is liable to hit a wall somewhere - probably memory 😁

alamb · 2022-02-02T21:40:38Z

For context one of the reasons I'm interested in reducing the concurrency is somewhat stupid, the plans and the resulting metrics and tracing become HUGE, but I also can't help but feel that always pursuing the maximum concurrency is liable to hit a wall somewhere - probably memory 😁

I think pursuing the maximum concurrency specified in target_partitions is a good strategy, but it needs to be complimented by a more sophisticated scheduler, that breaks the plans down into smaller pieces, perhaps as described / explained in #64

alamb · 2022-02-02T21:43:08Z

datafusion/src/physical_optimizer/repartition.rs

        )?;

+        let plan = displayable(optimized.as_ref()).indent().to_string();
+
+        let expected = &[


I think this is a good improvement to the tests 👍

datafusion/src/physical_plan/mod.rs

datafusion/src/physical_plan/limit.rs

datafusion/src/physical_optimizer/repartition.rs

Dandandan · 2022-02-03T09:31:25Z

For context one of the reasons I'm interested in reducing the concurrency is somewhat stupid, the plans and the resulting metrics and tracing become HUGE, but I also can't help but feel that always pursuing the maximum concurrency is liable to hit a wall somewhere - probably memory grin

I think pursuing the maximum concurrency specified in target_partitions is a good strategy, but it needs to be complimented by a more sophisticated scheduler, that breaks the plans down into smaller pieces, perhaps as described / explained in #64

Agreed with this approach 💯

alamb

LGTM -- any final thoughts @Dandandan ?

Dandandan

Looks good!

Dandandan · 2022-02-03T16:35:26Z

Thanks @tustvold !

alamb · 2022-02-04T16:15:55Z

datafusion/src/physical_plan/limit.rs

+
+    fn should_repartition_children(&self) -> bool {
+        // No reason to repartition children as this node is just limiting each input partition.
+        false


This also fixed #423 I believe

As currently expressed this only prevent repartitioning of direct children, I think you need something more than that for sortedness as discussed on #424

alamb · 2022-02-10T20:36:02Z

Removing API change label as it was showing up prominently on the changelog and is not a breaking API change (really)

* feat: add join type for logical plan display (#1674) * (minor) Reduce memory manager and disk manager logs from `info!` to `debug!` (#1689) * Move `information_schema` tests out of execution/context.rs to `sql_integration` tests (#1684) * Move tests from context.rs to information_schema.rs * Fix up tests to compile * Move timestamp related tests out of context.rs and into sql integration test (#1696) * Move some tests out of context.rs and into sql * Move support test out of context.rs and into sql tests * Fixup tests and make them compile * Add `MemTrackingMetrics` to ease memory tracking for non-limited memory consumers (#1691) * Memory manager no longer track consumers, update aggregatedMetricsSet * Easy memory tracking with metrics * use tracking metrics in SPMS * tests * fix * doc * Update datafusion/src/physical_plan/sorts/sort.rs Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> * make tracker AtomicUsize Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> * Implement TableProvider for DataFrameImpl (#1699) * Add TableProvider impl for DataFrameImpl * Add physical plan in * Clean up plan construction and names construction * Remove duplicate comments * Remove unused parameter * Add test * Remove duplicate limit comment * Use cloned instead of individual clone * Reduce the amount of code to get a schema Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> * Add comments to test * Fix plan comparison * Compare only the results of execution * Remove println * Refer to df_impl instead of table in test Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> * Fix the register_table test to use the correct result set for comparison * Consolidate group/agg exprs * Format * Remove outdated comment Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> * refine test in repartition.rs & coalesce_batches.rs (#1707) * Fuzz test for spillable sort (#1706) * Lazy TempDir creation in DiskManager (#1695) * Incorporate dyn scalar kernels (#1685) * Rebase * impl ToNumeric for ScalarValue * Update macro to be based on * Add floats * Cleanup * Newline * add annotation for select_to_plan (#1714) * Support `create_physical_expr` and `ExecutionContextState` or `DefaultPhysicalPlanner` for faster speed (#1700) * Change physical_expr creation API * Refactor API usage to avoid creating ExecutionContextState * Fixup ballista * clippy! * Fix can not load parquet table form spark in datafusion-cli. (#1665) * fix can not load parquet table form spark * add Invalid file in log. * fix fmt * add upper bound for pub fn (#1713) Signed-off-by: remzi <13716567376yh@gmail.com> * Create SchemaAdapter trait to map table schema to file schemas (#1709) * Create SchemaAdapter trait to map table schema to file schemas * Linting fix * Remove commented code * approx_quantile() aggregation function (#1539) * feat: implement TDigest for approx quantile Adds a [TDigest] implementation providing approximate quantile estimations of large inputs using a small amount of (bounded) memory. A TDigest is most accurate near either "end" of the quantile range (that is, 0.1, 0.9, 0.95, etc) due to the use of a scalaing function that increases resolution at the tails. The paper claims single digit part per million errors for q ≤ 0.001 or q ≥ 0.999 using 100 centroids, and in practice I have found accuracy to be more than acceptable for an apprixmate function across the entire quantile range. The implementation is a modified copy of https://github.com/MnO2/t-digest, itself a Rust port of [Facebook's C++ implementation]. Both Facebook's implementation, and Mn02's Rust port are Apache 2.0 licensed. [TDigest]: https://arxiv.org/abs/1902.04023 [Facebook's C++ implementation]: https://github.com/facebook/folly/blob/main/folly/stats/TDigest.h * feat: approx_quantile aggregation Adds the ApproxQuantile physical expression, plumbing & test cases. The function signature is: approx_quantile(column, quantile) Where column can be any numeric type (that can be cast to a float64) and quantile is a float64 literal between 0 and 1. * feat: approx_quantile dataframe function Adds the approx_quantile() dataframe function, and exports it in the prelude. * refactor: bastilla approx_quantile support Adds bastilla wire encoding for approx_quantile. Adding support for this required modifying the AggregateExprNode proto message to support propigating multiple LogicalExprNode aggregate arguments - all the existing aggregations take a single argument, so this wasn't needed before. This commit adds "repeated" to the expr field, which I believe is backwards compatible as described here: https://developers.google.com/protocol-buffers/docs/proto3#updating Specifically, adding "repeated" to an existing message field: "For ... message fields, optional is compatible with repeated" No existing tests needed fixing, and a new roundtrip test is included that covers the change to allow multiple expr. * refactor: use input type as return type Casts the calculated quantile value to the same type as the input data. * fixup! refactor: bastilla approx_quantile support * refactor: rebase onto main * refactor: validate quantile value Ensures the quantile values is between 0 and 1, emitting a plan error if not. * refactor: rename to approx_percentile_cont * refactor: clippy lints * suppport bitwise and as an example (#1653) * suppport bitwise and as an example * Use $OP in macro rather than `&` * fix: change signature to &dyn Array * fmt Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> * fix: substr - correct behaivour with negative start pos (#1660) * minor: fix cargo run --release error (#1723) * Convert boolean case expressions to boolean logic (#1719) * Convert boolean case expressions to boolean logic * Review feedback * substitute `parking_lot::Mutex` for `std::sync::Mutex` (#1720) * Substitute parking_lot::Mutex for std::sync::Mutex * enable parking_lot feature in tokio * Add Expression Simplification API (#1717) * Add Expression Simplification API * fmt * Add tests and CI for optional pyarrow module (#1711) * Implement other side of conversion * Add test workflow * Add (failing) tests * Get unit tests passing * Use python -m pip * Debug LD_LIBRARY_PATH * Set LIBRARY_PATH * Update help with better info * Update parking_lot requirement from 0.11 to 0.12 (#1735) Updates the requirements on [parking_lot](https://github.com/Amanieu/parking_lot) to permit the latest version. - [Release notes](https://github.com/Amanieu/parking_lot/releases) - [Changelog](https://github.com/Amanieu/parking_lot/blob/master/CHANGELOG.md) - [Commits](Amanieu/parking_lot@0.11.0...0.12.0) --- updated-dependencies: - dependency-name: parking_lot dependency-type: direct:production ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * Prevent repartitioning of certain operator's direct children (#1731) (#1732) * Prevent repartitioning of certain operator's direct children (#1731) * Update ballista tests * Don't repartition children of RepartitionExec * Revert partition restriction on Repartition and Projection * Review feedback * Lint * API to get Expr's type and nullability without a `DFSchema` (#1726) * API to get Expr type and nullability without a `DFSchema` * Add test * publically export * Improve docs * Fix typos in crate documentation (#1739) * add `cargo check --release` to ci (#1737) * remote test * Update .github/workflows/rust.yml Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> * Move optimize test out of context.rs (#1742) * Move optimize test out of context.rs * Update * use clap 3 style args parsing for datafusion cli (#1749) * use clap 3 style args parsing for datafusion cli * upgrade cli version * Add partitioned_csv setup code to sql_integration test (#1743) * use ordered-float 2.10 (#1756) Signed-off-by: Andy Grove <agrove@apache.org> * #1768 Support TimeUnit::Second in hasher (#1769) * Support TimeUnit::Second in hasher * fix linter * format (#1745) * Create built-in scalar functions programmatically (#1734) * create build-in scalar functions programatically Signed-off-by: remzi <13716567376yh@gmail.com> * solve conflict Signed-off-by: remzi <13716567376yh@gmail.com> * fix spelling mistake Signed-off-by: remzi <13716567376yh@gmail.com> * rename to call_fn Signed-off-by: remzi <13716567376yh@gmail.com> * [split/1] split datafusion-common module (#1751) * split datafusion-common module * pyarrow * Update datafusion-common/README.md Co-authored-by: Andy Grove <agrove@apache.org> * Update datafusion/Cargo.toml * include publishing Co-authored-by: Andy Grove <agrove@apache.org> * fix: Case insensitive unquoted identifiers (#1747) * move dfschema and column (#1758) * add datafusion-expr module (#1759) * move column, dfschema, etc. to common module (#1760) * include window frames and operator into datafusion-expr (#1761) * move signature, type signature, and volatility to split module (#1763) * [split/10] split up expr for rewriting, visiting, and simplification traits (#1774) * split up expr for rewriting, visiting, and simplification * add docs * move built-in scalar functions (#1764) * split expr type and null info to be expr-schemable (#1784) * rewrite predicates before pushing to union inputs (#1781) * move accumulator and columnar value (#1765) * move accumulator and columnar value (#1762) * fix bad data type in test_try_cast_decimal_to_decimal * added projections for avro columns Co-authored-by: xudong.w <wxd963996380@gmail.com> Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org> Co-authored-by: Yijie Shen <henry.yijieshen@gmail.com> Co-authored-by: Phillip Cloud <417981+cpcloud@users.noreply.github.com> Co-authored-by: Matthew Turner <matthew.m.turner@outlook.com> Co-authored-by: Yang <37145547+Ted-Jiang@users.noreply.github.com> Co-authored-by: Remzi Yang <59198230+HaoYang670@users.noreply.github.com> Co-authored-by: Dan Harris <1327726+thinkharderdev@users.noreply.github.com> Co-authored-by: Dom <dom@itsallbroken.com> Co-authored-by: Kun Liu <liukun@apache.org> Co-authored-by: Dmitry Patsura <talk@dmtry.me> Co-authored-by: Raphael Taylor-Davies <1781103+tustvold@users.noreply.github.com> Co-authored-by: Will Jones <willjones127@gmail.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: r.4ntix <r.4ntix@gmail.com> Co-authored-by: Jiayu Liu <Jimexist@users.noreply.github.com> Co-authored-by: Andy Grove <agrove@apache.org> Co-authored-by: Rich <jychen7@users.noreply.github.com> Co-authored-by: Marko Mikulicic <mmikulicic@gmail.com> Co-authored-by: Eduard Karacharov <13005055+korowa@users.noreply.github.com>

github-actions bot added the datafusion Changes in the datafusion crate label Feb 2, 2022

tustvold force-pushed the constrain-repartitioning branch from 60ad5af to 1fa17bc Compare February 2, 2022 16:19

Prevent repartitioning of certain operator's direct children (apache#…

8c58b48

…1731)

tustvold force-pushed the constrain-repartitioning branch from 1fa17bc to 8c58b48 Compare February 2, 2022 16:46

Update ballista tests

9cfe2ff

tustvold commented Feb 2, 2022

View reviewed changes

Don't repartition children of RepartitionExec

abb1e9d

Dandandan reviewed Feb 2, 2022

View reviewed changes

datafusion/src/physical_plan/projection.rs Outdated Show resolved Hide resolved

datafusion/src/physical_plan/repartition.rs Outdated Show resolved Hide resolved

Revert partition restriction on Repartition and Projection

c4c0a1c

alamb reviewed Feb 2, 2022

View reviewed changes

houqp added the api change Changes the API exposed to users of the crate label Feb 3, 2022

houqp requested a review from Dandandan February 3, 2022 06:07

tustvold added 2 commits February 3, 2022 09:35

Review feedback

e6e7b13

Lint

87da45d

alamb approved these changes Feb 3, 2022

View reviewed changes

Dandandan approved these changes Feb 3, 2022

View reviewed changes

Dandandan merged commit 78c30b6 into apache:master Feb 3, 2022

alamb mentioned this pull request Feb 4, 2022

Repartition Optimisation Pass Breaks Sorting (wrong answer with limit) #423

Closed

alamb reviewed Feb 4, 2022

View reviewed changes

This was referenced Feb 4, 2022

Do not repartition sorted inputs SortPreservingMerge #1748

Closed

Update ExecutionPlan to know about sortedness and repartitioning optimizer pass respect the invariants #1776

Merged

alamb removed the api change Changes the API exposed to users of the crate label Feb 10, 2022

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Prevent repartitioning of certain operator's direct children (#1731) #1732

Prevent repartitioning of certain operator's direct children (#1731) #1732

tustvold commented Feb 2, 2022

tustvold Feb 2, 2022 •

edited

Loading

Dandandan left a comment

Dandandan commented Feb 2, 2022

tustvold commented Feb 2, 2022

alamb commented Feb 2, 2022

alamb Feb 2, 2022

Dandandan commented Feb 3, 2022

alamb left a comment

Dandandan left a comment

Dandandan commented Feb 3, 2022

alamb Feb 4, 2022

tustvold Feb 4, 2022

alamb commented Feb 10, 2022

Prevent repartitioning of certain operator's direct children (#1731) #1732

Prevent repartitioning of certain operator's direct children (#1731) #1732

Conversation

tustvold commented Feb 2, 2022

Which issue does this PR close?

Rationale for this change

What changes are included in this PR?

Are there any user-facing changes?

tustvold Feb 2, 2022 • edited Loading

Choose a reason for hiding this comment

Dandandan left a comment

Choose a reason for hiding this comment

Dandandan commented Feb 2, 2022

tustvold commented Feb 2, 2022

alamb commented Feb 2, 2022

alamb Feb 2, 2022

Choose a reason for hiding this comment

Dandandan commented Feb 3, 2022

alamb left a comment

Choose a reason for hiding this comment

Dandandan left a comment

Choose a reason for hiding this comment

Dandandan commented Feb 3, 2022

alamb Feb 4, 2022

Choose a reason for hiding this comment

tustvold Feb 4, 2022

Choose a reason for hiding this comment

alamb commented Feb 10, 2022

tustvold Feb 2, 2022 •

edited

Loading