Add minimal APIs / hooks for granular statistics collection in TableProvider implementations#22300
Conversation
|
Thank you for opening this pull request! Reviewer note: cargo-semver-checks reported the current version number is not SemVer-compatible with the changes in this pull request (compared against the base branch). Details |
0ef815a to
e2838e8
Compare
|
cc @asolimando |
070700d to
ae549d8
Compare
9c6bbc9 to
d20c5aa
Compare
`TableScan::try_new` takes five positional arguments and bare
`TableScan { .. }` struct literals are scattered across the codebase,
making both fragile to field additions.
Introduce `TableScanBuilder` (with `From<TableScan>`, so an existing
scan can be decomposed, tweaked, and rebuilt) and move the
schema-derivation logic into `build()`. `TableScan::try_new` is now
deprecated and delegates to the builder; all in-tree callers are
migrated to the builder. Pure refactor, no behavior change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add `StatisticsRequest` to `datafusion-expr-common::statistics` — a small vocabulary for query-aware statistics: a caller can ask a provider for a specific statistic (Min/Max/NullCount/DistinctCount/Sum/ByteSize per column, plus RowCount and TotalByteSize) instead of for a dense `Statistics` covering every column. It derives `Ord` so requests can be held in a deduplicating, deterministically-ordered `BTreeSet` (see the next commit). It is intentionally just a vocabulary; nothing in DataFusion populates or consumes it yet. Re-exported via `datafusion_expr::statistics`. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add an advisory `statistics_requests: BTreeSet<StatisticsRequest>` field to `TableScan`. A custom optimizer rule can record the statistics the surrounding plan shape would benefit from (e.g. Min/Max for sort keys); the physical planner threads them into the table provider (next commit). A `BTreeSet` rather than a `Vec`: the optimizer runs rules to a fixpoint, so the request collection must be idempotent under re-derivation. With a set a rule simply `insert`s its requests — re-running is a no-op and composes with other rules — instead of every rule having to dedupe a `Vec` itself. `BTree` ordering also keeps plans deterministic. The field is empty by default and DataFusion's own rules never populate it. It can be set when building a scan via `TableScanBuilder::with_statistics_requests`, or mutated directly (it is `pub`, like `TableScan`'s other fields). `Debug`/`PartialEq`/`Eq`/`Hash`/ `PartialOrd` for `TableScan` are left unchanged — it is advisory metadata, not part of plan identity. `map_expressions` in `tree_node.rs` is rewritten to rebuild `TableScan` via `..scan` instead of an exhaustive destructure, so it carries this (and any future) field through untouched. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a `statistics_requests` field to `ScanArgs` (with `with_statistics_requests` / `statistics_requests` accessors) and have the physical planner thread `TableScan::statistics_requests` into it. This completes the request-side path: a custom optimizer rule annotates `TableScan`, and the request reaches a custom `TableProvider` in `scan_with_args`. DataFusion's own providers ignore the field; the default `ScanArgs` value is an empty slice. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a self-contained integration test that plays both external roles: a custom `OptimizerRule` annotates each `TableScan` with `StatisticsRequest`s, and a custom `TableProvider` records the `ScanArgs::statistics_requests` it receives in `scan_with_args`. This demonstrates the request-side hooks are sufficient to build the feature entirely outside of DataFusion. A second test confirms that without such a rule the provider sees an empty request list. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
d20c5aa to
cd2202e
Compare
| /// Set the statistics the caller would like the provider to answer for | ||
| /// this scan, if it can do so cheaply. | ||
| /// | ||
| /// Providers read these via [`Self::statistics_requests()`]; anything a |
There was a problem hiding this comment.
What i the Provider here? Do you mean "Table Provider"?
| } | ||
| } | ||
|
|
||
| // --------------------------------------------------------------------------- |
There was a problem hiding this comment.
I think this comment should be made into a real /// comment and folded into the doc strings of StatisticsRequests
By being here it is likely invisible / harder to find I think
| #[derive(Debug, Clone, PartialEq, Eq, PartialOrd, Ord)] | ||
| pub enum StatisticsRequest { | ||
| /// Smallest non-null value of `column`. | ||
| Min(Column), |
There was a problem hiding this comment.
A column has a bunch of owned strings: https://github.com/alamb/datafusion/blob/0181c7906b47f9b17fbde92bf30b2d939a80230c/datafusion/common/src/column.rs#L30-L37
What do you think about making this cheaper by making it take &Column or Arc<Column> rather than Column?
| ) -> Result<Self> { | ||
| let table_scan = | ||
| TableScan::try_new(table_name, table_source, projection, filters, fetch)?; | ||
| let table_scan = TableScanBuilder::new(table_name, table_source) |
| } | ||
| } | ||
|
|
||
| /// Builder for [`TableScan`]. |
| projected_schema, | ||
| filters, | ||
| fetch, | ||
| LogicalPlan::TableScan(mut scan) => { |
There was a problem hiding this comment.
I feel like the older code was easier to verify that it didn't carry along any other fields with expressions -- why change it here?
Specificially, I worry if we ever add another field to Filter then we would be more likely to miss this callsite
There was a problem hiding this comment.
Okay will revert it
| //! `TableScan` with `StatisticsRequest`s and have them reach a *custom* | ||
| //! `TableProvider`'s `scan_with_args`. | ||
| //! | ||
| //! DataFusion ships no rule that populates `TableScan::statistics_requests` |
Co-authored-by: Andrew Lamb <andrew@nerdnetworks.org>
Restores the explicit field destructuring for the LogicalPlan::TableScan arm of map_expressions instead of the `mut scan` + `..scan` shorthand. Explicit destructuring forces a compile error if a new field carrying expressions is added to TableScan, so this callsite cannot be silently missed. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Resolves an ambiguity raised in review: the doc referred to "the
provider"/"Providers" without saying which provider. Use `TableProvider`
consistently, and fix a typo ("for to allow") and trailing whitespace.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The "Query-aware statistics requests" explanation lived in a `//` block comment above the enum, where it is invisible in rustdoc. Move it into the enum's `///` doc comment so it is discoverable. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A Column owns its relation/name strings, so cloning a StatisticsRequest cloned those strings. The requests live in a BTreeSet<StatisticsRequest> on TableScan that is cloned with the plan throughout optimization, so wrap the per-column payload in Arc<Column> to make those clones cheap. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Leftover trailing whitespace from a previously applied review suggestion; removed by cargo fmt. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
|
Thanks for the review @alamb! |
Which issue does this PR close?
datafusion.execution.collect_statisticson wide tables #21624This is not a replacement for #21996 — it is a minimal subset of it, carved out so the feature can be discussed/merged in smaller pieces.
Rationale for this change
#21996 ("Query-aware statistics requests via ScanArgs / ScanResult") is a full vertical slice: new statistics types, request threading optimizer → planner → provider, a built-in
RequestStatisticsoptimizer rule, and a consumer integration (FilePruner/ListingTable).This PR extracts only the framework hooks — just enough that the rest can be implemented entirely outside of DataFusion. A third party can write their own optimizer rule to derive statistics requests, and their own
TableProviderto consume them, without DataFusion shipping any rule or consumer of its own.In stock DataFusion nothing observable changes: no rule populates the new field, and the built-in providers ignore it.
What changes are included in this PR?
Five small, independently-reviewable commits:
refactor: add TableScanBuilder, deprecate TableScan::try_new—TableScan::try_newtakes five positional args and bareTableScan { .. }literals are fragile to field additions. IntroduceTableScanBuilder(withFrom<TableScan>), move schema derivation intobuild(), deprecatetry_new(delegates to the builder), migrate all in-tree callers. Pure refactor.feat: add StatisticsRequest type— a new publicStatisticsRequestenum indatafusion-expr-common::statistics(Min/Max/NullCount/DistinctCount/Sum/ByteSize per column, plus RowCount/TotalByteSize). Nothing consumes it yet.feat: add TableScan::statistics_requests field— an advisoryBTreeSet<StatisticsRequest>onTableScan. A set so request-deriving optimizer rules stay idempotent under fixpoint iteration (a ruleinserts its requests; re-running is a no-op and composes with other rules — no per-rule dedup). Empty by default; DataFusion's own rules never populate it.feat: thread statistics requests into ScanArgs—ScanArgsgainsstatistics_requests; the physical planner threadsTableScan::statistics_requestsinto it so the request reachesTableProvider::scan_with_args.test: e2e statistics-request flow via a custom optimizer rule— an integration test playing both external roles.Deliberately left out vs #21996, since this PR is request-side only: the built-in
RequestStatisticsoptimizer rule, theFilePruner/ListingTableconsumer integration, thePartitionedFile::satisfied_statsper-file response field, and the response-side types (StatisticsValue/SatisfiedStatistics). Those belong with whatever actually wires the response side.Are these changes tested?
Yes:
datafusion/core/tests/user_defined/statistics_requests.rs: an end-to-end integration test where a customOptimizerRuleannotatesTableScanand a customTableProviderasserts the requests reachscan_with_args— plus a test that without such a rule the provider sees an empty request list.datafusion-expr/datafusion-optimizer/datafusion-prototests pass against theTableScanBuilderrefactor.Are there any user-facing changes?
Yes — this needs the
api changelabel:StatisticsRequest(re-exported viadatafusion_expr::statistics).TableScanBuilder;TableScan::try_newis deprecated (still works, delegates to the builder).TableScangains a new public fieldstatistics_requests— this breaks exhaustiveTableScan { .. }struct literals downstream (the recommended fix isTableScanBuilder).ScanArgsgainswith_statistics_requests/statistics_requests.🤖 Generated with Claude Code