reader: ScanIterator.compareValues swallows ClassCastException, silently disabling zone-map pruning on type-mismatched filters

[dfa1]

Summary

ScanIterator.compareValues(Object a, Object b) catches ClassCastException and returns 0. Because the canPruneChunk arms treat compareValues(...) < 0 / > 0 as the prune condition, a 0 means "cannot prune" — so a filter whose value type does not match the stored zone-map stat type silently prunes nothing and the whole column is decoded. No error, no log: a correctness-safe but performance-silent failure.

Where

reader/src/main/java/io/github/dfa1/vortex/reader/ScanIterator.java

private static int compareValues(Object a, Object b) {
    try {
        return ((Comparable<Object>) a).compareTo(b);
    } catch (ClassCastException _) {
        return 0;   // <-- mismatch => 0 => canPruneChunk yields false => no pruning
    }
}

Integer zone stats decode as Long (ArrayStats.decodeScalar returns int64_value()), and floats as Double. So a caller that builds a RowFilter with an Integer (or Float) value against an integer (or f32) column gets zero pruning even though the predicate is perfectly valid — the comparator just throws ClassCastException internally and is swallowed.

How it surfaced

Building the Calcite adapter's filter push-down (feat/vortex-calcite-demo), a WHERE date BETWEEN ... on an I32 column pruned 0 of 100 chunks until the literal was coerced from Integer to Long. The filter, expansion, and canPruneChunk path were all correct — only the boxed type differed, and the swallow hid it.

Impact

• Silent performance cliff: a valid, selective filter degrades to a full scan with no signal.
• Easy to hit from any caller that boxes filter values at the column's "natural" width (Integer for I32, Float for F32) rather than the stat's storage width (Long / Double).

Suggested fix (pick one)

1. Normalise numeric comparison in compareValues — if both args are Number but different boxed types, compare via Long/Double (or BigDecimal) instead of returning 0. Most robust; makes pruning width-agnostic.
2. Coerce on RowFilter construction — normalise integer values to Long and float values to Double when a RowFilter is built, matching the stat storage types.
3. At minimum, do not return 0 on mismatch — a genuine type mismatch (e.g. comparing a string filter to a numeric column) should be a clear error or a logged "pruning skipped", not a silent no-op.

Option 1 is preferred: it fixes every caller, not just the careful ones.

Notes

The Calcite adapter worked around this by coercing integer literals to Long (commit on feat/vortex-calcite-demo), so it is unblocked — but the underlying reader trap should be fixed so the next caller does not rediscover it.