Skip to content

Apache DataFusion SQL Query Engine

docs: Document the TableProvider evaluation order for filter, limit and projection (#21091)

Andrew Lambbec6714d8a02
diff --git a/datafusion/catalog/src/table.rs b/datafusion/catalog/src/table.rs
index c9b4e974c8..361589c5b6 100644
--- a/datafusion/catalog/src/table.rs
+++ b/datafusion/catalog/src/table.rs
@@ -84,10 +84,10 @@ pub trait TableProvider: Debug + Sync + Send {
None
}
- /// Create an [`ExecutionPlan`] for scanning the table with optionally
- /// specified `projection`, `filter` and `limit`, described below.
+ /// Create an [`ExecutionPlan`] for scanning the table with optional
+ /// `projection`, `filter`, and `limit`, described below.
///
- /// The `ExecutionPlan` is responsible scanning the datasource's
+ /// The returned `ExecutionPlan` is responsible for scanning the datasource's
/// partitions in a streaming, parallelized fashion.
///
/// # Projection
@@ -96,33 +96,30 @@ pub trait TableProvider: Debug + Sync + Send {
/// specified. The projection is a set of indexes of the fields in
/// [`Self::schema`].
///
- /// DataFusion provides the projection to scan only the columns actually
- /// used in the query to improve performance, an optimization called
- /// "Projection Pushdown". Some datasources, such as Parquet, can use this
- /// information to go significantly faster when only a subset of columns is
- /// required.
+ /// DataFusion provides the projection so the scan reads only the columns
+ /// actually used in the query, an optimization called "Projection
+ /// Pushdown". Some datasources, such as Parquet, can use this information
+ /// to go significantly faster when only a subset of columns is required.
///
/// # Filters
///
/// A list of boolean filter [`Expr`]s to evaluate *during* the scan, in the
/// manner specified by [`Self::supports_filters_pushdown`]. Only rows for
- /// which *all* of the `Expr`s evaluate to `true` must be returned (aka the
- /// expressions are `AND`ed together).
+ /// which *all* of the `Expr`s evaluate to `true` must be returned (that is,
+ /// the expressions are `AND`ed together).
///
- /// To enable filter pushdown you must override
- /// [`Self::supports_filters_pushdown`] as the default implementation does
- /// not and `filters` will be empty.
+ /// To enable filter pushdown, override
+ /// [`Self::supports_filters_pushdown`]. The default implementation does not
+ /// push down filters, and `filters` will be empty.
///
- /// DataFusion pushes filtering into the scans whenever possible
- /// ("Filter Pushdown"), and depending on the format and the
- /// implementation of the format, evaluating the predicate during the scan
- /// can increase performance significantly.
+ /// DataFusion pushes filters into scans whenever possible ("Filter
+ /// Pushdown"). Depending on the data format and implementation, evaluating
+ /// predicates during the scan can significantly improve performance.
///
/// ## Note: Some columns may appear *only* in Filters
///
- /// In certain cases, a query may only use a certain column in a Filter that
- /// has been completely pushed down to the scan. In this case, the
- /// projection will not contain all the columns found in the filter
+ /// In some cases, a query may use a column only in a filter and the
+ /// projection will not contain all columns referenced by the filter
/// expressions.
///
/// For example, given the query `SELECT t.a FROM t WHERE t.b > 5`,
@@ -154,15 +151,40 @@ pub trait TableProvider: Debug + Sync + Send {
///
/// # Limit
///
- /// If `limit` is specified, must only produce *at least* this many rows,
- /// (though it may return more). Like Projection Pushdown and Filter
- /// Pushdown, DataFusion pushes `LIMIT`s as far down in the plan as
- /// possible, called "Limit Pushdown" as some sources can use this
- /// information to improve their performance. Note that if there are any
- /// Inexact filters pushed down, the LIMIT cannot be pushed down. This is
- /// because inexact filters do not guarantee that every filtered row is
- /// removed, so applying the limit could lead to too few rows being available
- /// to return as a final result.
+ /// If `limit` is specified, the scan must produce *at least* this many
+ /// rows, though it may return more. Like Projection Pushdown and Filter
+ /// Pushdown, DataFusion pushes `LIMIT`s as far down in the plan as
+ /// possible. This is called "Limit Pushdown", and some sources can use the
+ /// information to improve performance.
+ ///
+ /// Note: If any pushed-down filters are `Inexact`, the `LIMIT` cannot be
+ /// pushed down. Inexact filters do not guarantee that every filtered row is
+ /// removed, so applying the limit could leave too few rows to return in the
+ /// final result.
+ ///
+ /// # Evaluation Order
+ ///
+ /// The logical evaluation order is `filters`, then `limit`, then
+ /// `projection`.
+ ///
+ /// Note that `limit` applies to the filtered result, not to the unfiltered
+ /// input, and `projection` affects only which columns are returned, not
+ /// which rows qualify.
+ ///
+ /// For example, if a scan receives:
+ ///
+ /// - `projection = [a]`
+ /// - `filters = [b > 5]`
+ /// - `limit = Some(3)`
+ ///
+ /// It must logically produce results equivalent to:
+ ///
+ /// ```text
+ /// PROJECTION a (LIMIT 3 (SCAN WHERE b > 5))
+ /// ```
+ ///
+ /// As noted above, columns referenced only by pushed-down filters may be
+ /// absent from `projection`.
async fn scan(
&self,
state: &dyn Session,