Wrong results: filter marked as pushed down at plan time can be silently dropped per-file when schema evolution makes the runtime pushdown checker reject the adapted predicate

[zhuqi-lucas]

Describe the bug

With pushdown_filters = true, a filter conjunct that plan time marks as pushed down (removing it from the parent FilterExec) can be silently dropped at runtime when the per-file RowFilter is rebuilt against the physical file schema — the query then returns rows that violate the WHERE clause. No error, no metric, only a debug! line.

The safety invariant is "the plan-time pushdown checker must be strictly more conservative than runtime row-filter candidate building". Adapter-inserted casts validated by arrow's permissive can_cast_types break it:

1. Plan time says Yes: try_pushdown_filters checks can_expr_be_pushed_down_with_schemas against the table schema (datafusion/datasource-parquet/src/source.rs). With table schema a: Utf8, a = 'keep' is a primitive-column predicate → PushedDown::Yes → the conjunct is removed from FilterExec.
2. Runtime adapter inserts a valid cast: for a file whose physical schema stores a as List<Int32> (schema evolution), the physical expr rewriter validates with arrow's can_cast_types, which allows List → Utf8 → a CastExpr is inserted with no error (schema_rewriter.rs → nested_struct.rs::validate_data_type_compatibility, _ fallback arm).
3. Runtime pushdown checker rejects the shape: prebuild_row_filter_candidates → PushdownChecker::check_single_column sees the physical List type; = is not in supports_list_predicates (only IsNull/IsNotNull/array_has*) → the candidate comes back Ok(None).
4. Silent drop: the None vanishes in .flatten() in build_row_filter/prebuild_row_filter_candidates (row_filter.rs, the doc comment even states "Conjuncts that cannot be evaluated as an ArrowPredicate are ignored"). No RowFilter is installed for that file; row-group/page pruning only proves "may contain matches" and never filters rows → the predicate is applied nowhere for that file.

The cast itself is perfectly evaluable — the pushdown_filters=false path proves it (the retained FilterExec evaluates the adapted expression correctly). The runtime checker is just more conservative than the adapter's validator.

A second reachable instance of the same shape: physical FixedSizeList(T, 1) vs logical primitive T (arrow allows FSL(1) → T, the checker rejects FixedSizeList for non-list predicates).

There is also a latent adjacent hazard: if prebuild_row_filter_candidates returns Err, the entire RowFilter is dropped with only a debug! (push_decoder.rs, Err(e) => { debug!(...); None }). I could not construct a reachable input for that leg, but it silently discards ALL pushed conjuncts if ever hit.

To Reproduce

Failing test (appended to datafusion/core/tests/parquet/schema_coercion.rs; two parquet files in one listing table — a: Utf8 in f1, a: List<Int32> in f2, table schema declares a Utf8):

Result with pushdown_filters = true (physical plan is a bare DataSourceExec ... predicate=a@1 = keep, no FilterExec):

| id | a      |
|----|--------|
| 1  | keep   |
| 3  | [1, 2] |   <- violates WHERE a = 'keep'
| 4  | [3]    |   <- violates WHERE a = 'keep'

Same query with pushdown_filters = false correctly returns only id = 1.

Full repro test

#[tokio::test]
async fn pushdown_filter_dropped_conjunct_returns_wrong_rows() {
    use arrow::array::{AsArray, Int32Array, ListArray};
    use arrow::buffer::OffsetBuffer;
    use datafusion::datasource::file_format::parquet::ParquetFormat;
    use datafusion::datasource::listing::ListingOptions;
    use datafusion::prelude::SessionConfig;

    let dir = tempfile::tempdir().unwrap();

    // file 1: a is Utf8 (matches table schema)
    let batch1 = RecordBatch::try_from_iter(vec![
        (
            "id",
            Arc::new(Int64Array::from(vec![1_i64, 2])) as ArrayRef,
        ),
        (
            "a",
            Arc::new(StringArray::from(vec!["keep", "drop"])) as ArrayRef,
        ),
    ])
    .unwrap();

    // file 2: a is physically List<Int32> (schema evolution)
    let list = ListArray::new(
        Arc::new(Field::new("item", DataType::Int32, true)),
        OffsetBuffer::from_lengths([2, 1]),
        Arc::new(Int32Array::from(vec![1, 2, 3])),
        None,
    );
    let batch2 = RecordBatch::try_from_iter(vec![
        (
            "id",
            Arc::new(Int64Array::from(vec![3_i64, 4])) as ArrayRef,
        ),
        ("a", Arc::new(list) as ArrayRef),
    ])
    .unwrap();

    for (name, batch) in [("f1.parquet", batch1), ("f2.parquet", batch2)] {
        let file = std::fs::File::create(dir.path().join(name)).unwrap();
        let mut w = ArrowWriter::try_new(file, batch.schema(), None).unwrap();
        w.write(&batch).unwrap();
        w.close().unwrap();
    }

    let mut cfg = SessionConfig::new();
    cfg.options_mut().execution.parquet.pushdown_filters = true;
    let ctx = SessionContext::new_with_config(cfg);

    let table_schema = Arc::new(Schema::new(vec![
        Field::new("id", DataType::Int64, true),
        Field::new("a", DataType::Utf8, true),
    ]));
    let opts = ListingOptions::new(Arc::new(ParquetFormat::default()));
    ctx.register_listing_table(
        "t",
        dir.path().to_str().unwrap(),
        opts,
        Some(table_schema),
        None,
    )
    .await
    .unwrap();

    // Show the physical plan (the FilterExec should have been removed /
    // the predicate reported as handled by the scan)
    let plan = ctx
        .sql("EXPLAIN SELECT id, a FROM t WHERE a = 'keep'")
        .await
        .unwrap()
        .collect()
        .await
        .unwrap();
    println!(
        "{}",
        datafusion_common::test_util::batches_to_string(&plan)
    );

    let results = ctx
        .sql("SELECT id, a FROM t WHERE a = 'keep' ORDER BY id")
        .await
        .unwrap()
        .collect()
        .await
        .unwrap();

    println!(
        "{}",
        datafusion_common::test_util::batches_to_string(&results)
    );

    // Sanity: with pushdown_filters=false the same query is answered correctly
    let ctx2 = SessionContext::new();
    let opts2 = ListingOptions::new(Arc::new(ParquetFormat::default()));
    ctx2.register_listing_table(
        "t",
        dir.path().to_str().unwrap(),
        opts2,
        Some(Arc::new(Schema::new(vec![
            Field::new("id", DataType::Int64, true),
            Field::new("a", DataType::Utf8, true),
        ]))),
        None,
    )
    .await
    .unwrap();
    let no_pushdown = ctx2
        .sql("SELECT id, a FROM t WHERE a = 'keep' ORDER BY id")
        .await
        .unwrap()
        .collect()
        .await
        .unwrap();
    println!(
        "no-pushdown result:\n{}",
        datafusion_common::test_util::batches_to_string(&no_pushdown)
    );
    let no_pushdown_rows: usize = no_pushdown.iter().map(|b| b.num_rows()).sum();
    assert_eq!(no_pushdown_rows, 1, "without pushdown only id=1 matches");

}

Expected behavior

Either the query returns correct results (runtime accepts the evaluable adapted cast), or the scan fails loudly. Never silent extra rows.

Fix directions (not mutually exclusive):

1. Make the drop loud: when a conjunct that plan time reported as PushedDown::Yes produces no candidate at runtime, return an error (or fall back to evaluating the residual conjuncts post-decode) instead of ignoring it; at minimum bump a metric.
2. Close the checker/adapter gap: let the runtime PushdownChecker accept cast-over-nested-column shapes the adapter can produce and evaluate.

Additional context

• Verified present on current main (load-bearing files row_filter.rs, push_decoder.rs, projection_read_plan.rs byte-identical to tip at time of writing).
• Related history: Parquet filter pushdown silently drops a get_field predicate when the file needs schema adaptation (wrong results) #24109 was the same bug class for struct columns and was fixed by try_narrow_struct_cast; missing columns, primitive type evolution, and explicit user struct casts are all guarded correctly (I checked each).
• This directly affects the pushdown-by-default work (WIP: DO NOT REVIEW: Benchmark enabling parquet filter pushdown by default with one-shot row group I/O #24426): the blast radius of the silent drop grows if pushdown_filters becomes the default.
• Found during an aggregation/scan audit; repro + investigation notes available. I plan to follow up with a fix PR.