fix: validate inter-file ordering in eq_properties() (#20329)

## Summary

Discovered this bug while working on #19724.

TLDR: just because the files themselves are sorted doesn't mean the
partition streams are sorted.

- **`eq_properties()` in `FileScanConfig` blindly trusted
`output_ordering`** (set from Parquet `sorting_columns` metadata)
without verifying that files within a group are in the correct
inter-file order
- `EnforceSorting` then removed `SortExec` based on this unvalidated
ordering, producing **wrong results** when filesystem order didn't match
data order
- Added `validated_output_ordering()` that filters orderings using
`MinMaxStatistics::new_from_files()` + `is_sorted()` to verify
inter-file sort order before reporting them to the optimizer

## Changes

### `datafusion/datasource/src/file_scan_config.rs`
- Added `validated_output_ordering()` method on `FileScanConfig` that
validates each output ordering against actual file group statistics
- Changed `eq_properties()` to call `self.validated_output_ordering()`
instead of `self.output_ordering.clone()`

### `datafusion/sqllogictest/test_files/sort_pushdown.slt`
Added 8 new regression tests (Tests 4-11):

| Test | Scenario | Key assertion |
|------|----------|---------------|
| **4** | Reversed filesystem order (inferred ordering) | SortExec
retained — wrong inter-file order detected |
| **5** | Overlapping file ranges (inferred ordering) | SortExec
retained — overlapping ranges detected |
| **6** | `WITH ORDER` + reversed filesystem order | SortExec retained
despite explicit ordering |
| **7** | Correctly ordered multi-file group (positive) | SortExec
eliminated — validation passes |
| **8** | DESC ordering with wrong inter-file DESC order | SortExec
retained for DESC direction |
| **9** | Multi-column sort key (overlapping vs non-overlapping) |
Conservative rejection with overlapping stats; passes with clean
boundaries |
| **10** | Correctly ordered + `WITH ORDER` (positive) | SortExec
eliminated — both ordering and stats agree |
| **11** | Multiple partitions (one file per group) |
`SortPreservingMergeExec` merges; no per-partition sort needed |

## Test plan

- [x] `cargo test --test sqllogictests -- sort_pushdown` — all new +
existing tests pass
- [x] `cargo test -p datafusion-datasource` — 97 unit tests + 6 doc
tests pass
- [x] Existing Test 1 (single-file sort pushdown with `WITH ORDER`)
still eliminates SortExec (no regression)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: adriangb

datafusion/core/tests/physical_optimizer/partition_statistics.rs

@@ -826,7 +826,7 @@ mod test {
         let plan_string = get_plan_string(&aggregate_exec_partial).swap_remove(0);
         assert_snapshot!(
             plan_string,
-            @"AggregateExec: mode=Partial, gby=[id@0 as id, 1 + id@0 as expr], aggr=[COUNT(c)], ordering_mode=Sorted"
+            @"AggregateExec: mode=Partial, gby=[id@0 as id, 1 + id@0 as expr], aggr=[COUNT(c)]"
         );
 
         let p0_statistics = aggregate_exec_partial.partition_statistics(Some(0))?;

datafusion/datasource/src/file_scan_config.rs

@@ -736,7 +736,7 @@ impl DataSource for FileScanConfig {
         let schema = self.file_source.table_schema().table_schema();
         let mut eq_properties = EquivalenceProperties::new_with_orderings(
             Arc::clone(schema),
-            self.output_ordering.clone(),
+            self.validated_output_ordering(),
         )
         .with_constraints(self.constraints.clone());
 
@@ -926,6 +926,40 @@ impl DataSource for FileScanConfig {
 }
 
 impl FileScanConfig {
+    /// Returns only the output orderings that are validated against actual
+    /// file group statistics.
+    ///
+    /// For example, individual files may be ordered by `col1 ASC`,
+    /// but if we have files with these min/max statistics in a single partition / file group:
+    ///
+    /// - file1: min(col1) = 10, max(col1) = 20
+    /// - file2: min(col1) = 5, max(col1) = 15
+    ///
+    /// Because reading file1 followed by file2 would produce out-of-order output (there is overlap
+    /// in the ranges), we cannot retain `col1 ASC` as a valid output ordering.
+    ///
+    /// Similarly this would not be a valid order (non-overlapping ranges but not ordered):
+    ///
+    /// - file1: min(col1) = 20, max(col1) = 30
+    /// - file2: min(col1) = 10, max(col1) = 15
+    ///
+    /// On the other hand if we had:
+    ///
+    /// - file1: min(col1) = 5, max(col1) = 15
+    /// - file2: min(col1) = 16, max(col1) = 25
+    ///
+    /// Then we know that reading file1 followed by file2 will produce ordered output,
+    /// so `col1 ASC` would be retained.
+    ///
+    /// Note that we are checking for ordering *within* *each* file group / partition,
+    /// files in different partitions are read independently and do not affect each other's ordering.
+    /// Merging of the multiple partition streams into a single ordered stream is handled
+    /// upstream e.g. by `SortPreservingMergeExec`.
+    fn validated_output_ordering(&self) -> Vec<LexOrdering> {
+        let schema = self.file_source.table_schema().table_schema();
+        validate_orderings(&self.output_ordering, schema, &self.file_groups, None)
+    }
+
     /// Get the file schema (schema of the files without partition columns)
     pub fn file_schema(&self) -> &SchemaRef {
         self.file_source.table_schema().file_schema()
@@ -1300,6 +1334,51 @@ fn ordered_column_indices_from_projection(
         .collect::<Option<Vec<usize>>>()
 }
 
+/// Check whether a given ordering is valid for all file groups by verifying
+/// that files within each group are sorted according to their min/max statistics.
+///
+/// For single-file (or empty) groups, the ordering is trivially valid.
+/// For multi-file groups, we check that the min/max statistics for the sort
+/// columns are in order and non-overlapping (or touching at boundaries).
+///
+/// `projection` maps projected column indices back to table-schema indices
+/// when validating after projection; pass `None` when validating at
+/// table-schema level.
+fn is_ordering_valid_for_file_groups(
+    file_groups: &[FileGroup],
+    ordering: &LexOrdering,
+    schema: &SchemaRef,
+    projection: Option<&[usize]>,
+) -> bool {
+    file_groups.iter().all(|group| {
+        if group.len() <= 1 {
+            return true; // single-file groups are trivially sorted
+        }
+        match MinMaxStatistics::new_from_files(ordering, schema, projection, group.iter())
+        {
+            Ok(stats) => stats.is_sorted(),
+            Err(_) => false, // can't prove sorted → reject
+        }
+    })
+}
+
+/// Filters orderings to retain only those valid for all file groups,
+/// verified via min/max statistics.
+fn validate_orderings(
+    orderings: &[LexOrdering],
+    schema: &SchemaRef,
+    file_groups: &[FileGroup],
+    projection: Option<&[usize]>,
+) -> Vec<LexOrdering> {
+    orderings
+        .iter()
+        .filter(|ordering| {
+            is_ordering_valid_for_file_groups(file_groups, ordering, schema, projection)
+        })
+        .cloned()
+        .collect()
+}
+
 /// The various listing tables does not attempt to read all files
 /// concurrently, instead they will read files in sequence within a
 /// partition.  This is an important property as it allows plans to
@@ -1366,52 +1445,47 @@ fn get_projected_output_ordering(
     let projected_orderings =
         project_orderings(&base_config.output_ordering, projected_schema);
 
-    let mut all_orderings = vec![];
-    for new_ordering in projected_orderings {
-        // Check if any file groups are not sorted
-        if base_config.file_groups.iter().any(|group| {
-            if group.len() <= 1 {
-                // File groups with <= 1 files are always sorted
-                return false;
-            }
-
-            let Some(indices) = base_config
-                .file_source
-                .projection()
-                .as_ref()
-                .map(|p| ordered_column_indices_from_projection(p))
-            else {
-                // Can't determine if ordered without a simple projection
-                return true;
-            };
-
-            let statistics = match MinMaxStatistics::new_from_files(
-                &new_ordering,
+    let indices = base_config
+        .file_source
+        .projection()
+        .as_ref()
+        .map(|p| ordered_column_indices_from_projection(p));
+
+    match indices {
+        Some(Some(indices)) => {
+            // Simple column projection — validate with statistics
+            validate_orderings(
+                &projected_orderings,
                 projected_schema,
-                indices.as_deref(),
-                group.iter(),
-            ) {
-                Ok(statistics) => statistics,
-                Err(e) => {
-                    log::trace!("Error fetching statistics for file group: {e}");
-                    // we can't prove that it's ordered, so we have to reject it
-                    return true;
-                }
-            };
-
-            !statistics.is_sorted()
-        }) {
-            debug!(
-                "Skipping specified output ordering {:?}. \
-                Some file groups couldn't be determined to be sorted: {:?}",
-                base_config.output_ordering[0], base_config.file_groups
-            );
-            continue;
+                &base_config.file_groups,
+                Some(indices.as_slice()),
+            )
+        }
+        None => {
+            // No projection — validate with statistics (no remapping needed)
+            validate_orderings(
+                &projected_orderings,
+                projected_schema,
+                &base_config.file_groups,
+                None,
+            )
+        }
+        Some(None) => {
+            // Complex projection (expressions, not simple columns) — can't
+            // determine column indices for statistics. Still valid if all
+            // file groups have at most one file.
+            if base_config.file_groups.iter().all(|g| g.len() <= 1) {
+                projected_orderings
+            } else {
+                debug!(
+                    "Skipping specified output orderings. \
+                     Some file groups couldn't be determined to be sorted: {:?}",
+                    base_config.file_groups
+                );
+                vec![]
+            }
         }
-
-        all_orderings.push(new_ordering);
     }
-    all_orderings
 }
 
 /// Convert type to a type suitable for use as a `ListingTable`

datafusion/datasource/src/statistics.rs

@@ -266,11 +266,12 @@ impl MinMaxStatistics {
     }
 
     /// Check if the min/max statistics are in order and non-overlapping
+    /// (or touching at boundaries)
     pub fn is_sorted(&self) -> bool {
         self.max_by_sort_order
             .iter()
             .zip(self.min_by_sort_order.iter().skip(1))
-            .all(|(max, next_min)| max < next_min)
+            .all(|(max, next_min)| max <= next_min)
     }
 }
 

datafusion/sqllogictest/test_files/parquet_sorted_statistics.slt

@@ -274,4 +274,4 @@ logical_plan
 02)--TableScan: test_table projection=[constant_col]
 physical_plan
 01)SortPreservingMergeExec: [constant_col@0 ASC NULLS LAST]
-02)--DataSourceExec: file_groups={2 groups: [[WORKSPACE_ROOT/datafusion/sqllogictest/test_files/scratch/parquet_sorted_statistics/test_table/partition_col=A/0.parquet, WORKSPACE_ROOT/datafusion/sqllogictest/test_files/scratch/parquet_sorted_statistics/test_table/partition_col=B/1.parquet], [WORKSPACE_ROOT/datafusion/sqllogictest/test_files/scratch/parquet_sorted_statistics/test_table/partition_col=C/2.parquet]]}, projection=[constant_col], file_type=parquet
+02)--DataSourceExec: file_groups={2 groups: [[WORKSPACE_ROOT/datafusion/sqllogictest/test_files/scratch/parquet_sorted_statistics/test_table/partition_col=A/0.parquet, WORKSPACE_ROOT/datafusion/sqllogictest/test_files/scratch/parquet_sorted_statistics/test_table/partition_col=B/1.parquet], [WORKSPACE_ROOT/datafusion/sqllogictest/test_files/scratch/parquet_sorted_statistics/test_table/partition_col=C/2.parquet]]}, projection=[constant_col], output_ordering=[constant_col@0 ASC NULLS LAST], file_type=parquet