Use bounds to prove a region cannot match

Suppose a region's nonnull finite values lie within inclusive bounds L and U under the same order used by the query. For the predicate x at least a and strictly below b, no value can match when U is below a or L is at least b. If the intervals overlap, the region merely remains a candidate: bounds alone cannot establish which intermediate values occur. This asymmetry makes a false positive acceptable as extra work, while a false negative is a query correctness failure.

Parquet can supply column-chunk statistics and an optional page index for finer navigation. Page-index bounds can be conservative rather than actual values present, allowing truncated summaries of long strings. Readers must interpret the specified column order and metadata contract. Bytewise string order, signed numbers, decimal interpretation, NaNs, and application collations cannot be casually interchanged. If a cast or user-defined predicate does not have a justified relationship to the stored bounds, retain the region or use another sound rule instead of guessing that a familiar-looking range test still applies.

SQL comparison predicates do not evaluate to true for a null input, so an entirely null region cannot contribute to this ordinary range filter. But an IS NULL query asks a different question and needs validity information. Parquet statistics fields are optional; an absent null count does not mean zero nulls. A missing or unusable min/max also does not mean an empty region. Distinguish known all-null, known no-null, and unknown states explicitly in the pruning logic, then test each with actual query semantics.

Composing predicates needs the same care. A region can be rejected for an AND when one conjunct is impossible; rejecting an OR generally requires proving every alternative impossible. A predicate that involves several columns may leave many false positives because independent per-column bounds do not preserve their correlations. For instance, a region can contain low x only with high y and high x only with low y, while separate ranges suggest a combination that never occurs. Retain candidates and evaluate rows rather than pretending a zone map is a multidimensional exact index.

A sorted or clustered range key often produces narrower region bounds than uniformly mixed values, improving the opportunity to reject groups. Row selectivity and region selectivity are different: one percent of matching rows spread across every page can require reading almost every page. Smaller regions can sharpen summaries but add metadata and requests. Measure regions considered, regions rejected, ranges fetched, and rows evaluated. Compare query answers with pruning disabled so an optimization cannot hide a mistaken type or null interpretation behind a faster runtime.

For a probability illustration, Reset the lab, set both p=0.2 and n=20, predict four relevant regions on average, then Play or Step. Interpret p as a hypothetical independent probability that a whole region is relevant, not as row selectivity. Playback highlights exact binomial bars rather than sampling files. Reset and re-enter both controls before another case because edits retain elapsed time. The model computes neither min/max summaries nor actual pruning decisions; those require the logical conditions and metadata checks above.