Holistic Repartition Optimisation #1731
Labels
enhancement
New feature or request
Comments
tustvold
added a commit
to tustvold/arrow-datafusion
that referenced
this issue
Feb 2, 2022
tustvold
added a commit
to tustvold/arrow-datafusion
that referenced
this issue
Feb 2, 2022
tustvold
added a commit
to tustvold/arrow-datafusion
that referenced
this issue
Feb 2, 2022
alamb
added a commit
that referenced
this issue
Feb 15, 2022
* 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the register_table test to use the correct result set for comparison * Consolidate group/agg exprs * Format * Remove outdated comment Co-authored-by: Andrew Lamb 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>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Is your feature request related to a problem or challenge? Please describe what you are trying to do.
The
Repartitionoptimizer recurses into plans and looks for operators with an output partition count less than the target partition count, and on finding such a partition adds aRepartitionExecnode.Unfortunately this does not take into account how this might impact the concurrency of the plan as a whole.
For example, in IOx it is fairly common to have a
UnionExecwith a list of manyIOxReadFilterNodeas children. Each of theseIOxReadFilterNodeis a single partition, and therefore theRepartitionoptimizer repartitions the output of each of theseIOxReadFilterNode.This leads to a partition explosion, with the
UnionExecgoing from potentially havingxpartitions, ton * xwherenis the target parallelism. In some of our plans this is leading to thousands of partitions, which is not ideal 😁A similar quirk can occur when there is an operator further up the chain that imposes some partitioning constraint, e.g. a sort, but isn't the direct parent. I'm fairly certain that I've seen plans generated that
RepartitionExecinto aLimitExecinto aCoalescePartitionsExec.Describe the solution you'd like
I don't really know how to solve this, but I wanted to flag up that the current greedy approach can be pretty sub-optimal, and that perhaps a more holistic optimizer that can take into account the plan as a whole might be beneficial
The text was updated successfully, but these errors were encountered: