Truthound · Query correctness · 1/3
Compare aggregates, joins and filters against independent expectations.
49 articles
An OpenSearch date_nanos bucket returns 0: inspect range-rewrite units before changing source data
Separate search results from the aggregation fast path.
Check reserved-name readiness in Trino metadata queries
Exercise each consumer-facing metadata query for a table with a reserved-word column. Success on another metadata path, or an abbreviated SQL description, cannot stand in for the required query parsing successfully.
Read final Pinot metadata before accepting an empty Trino split
Exclude a Pinot split from successful aggregation when final metadata reports an error, even if its data blocks are empty. Keep missing metadata distinct from a valid empty result so a low count cannot conceal a missing segment.
Test Trino bloom filters with the opposite sign of zero
Compare keys at one snapshot to check whether a bloom filter drops an oppositely signed zero that SQL treats as equal. Matching counts can conceal changed identities, so reconcile key sets with pruning enabled and disabled.
Do not assume local dates are monotone across a midnight clock fallback
Local dates need not increase with UTC through a midnight fallback. Fix the historical timezone interval and expression type when comparing pruned key sets, and do not treat disabling skipping as a cure for every rewrite.
Match Hudi incremental pruning statistics to historical file coverage
A current snapshot can conceal pruning omissions in a historical incremental read. Fix the requested timeline range and reconcile its keys with a control read covering that same range.
Bind Hudi expression indexes to functions as well as columns
An expression index must match the query function as well as its source column. Reconcile filters using a different function against an unindexed control at the same snapshot to detect incorrect pruning.
Keep empty Xarray selections distinct from backend execution failures
An empty axis can be a valid result. Keep execution errors separate from the resulting axis structure.
Before inheriting UTC sortedness for local calendar dates
A calendar reversal after timezone conversion differs from UTC order. Derive the minimum from actual values rather than trusting the flag.
A small experiment before deduplicating NULL structs and NULL children
Define whether an absent parent differs from a present parent with a NULL child. Check overall distinct counts and row-level decisions separately.
When a filter moved before shift changes transition statistics
Derive previous states from the original sequence. A test expecting only row counts is insufficient to establish operation order.
Check whether another parent’s element moves into an empty list
Equal total element counts can conceal changed parent ownership. Retain an empty parent in the small fixture.
Why unknown NaN-page statistics cannot be omitted from row-group bounds
An unknown range is different from an empty page. Check that the advertised lower bound does not exclude original valid values.
Polars mixed-key full join: two rows still require key and payload reconciliation
Equal row counts do not establish preservation of keys or payloads after a full join.
Check observation presence when a singleton GPU datetime mean is NULL
Use the identity that one valid instant has that same mean. Distinguish operation support from CPU-fallback requirements.
Do not read truncated Delta timestamp statistics as exact upper bounds
Preserve the precision difference between millisecond statistics and microsecond values. Read duplicate reporting and the follow-up fix together.
Inspect execution when aggregation CSE changes a scalar into a list
Expression reuse can change an aggregate’s role and dtype. Admit the plan and the executed result separately.
Dask array conversion after Parquet projection: align partition and length-vector units
Projection can combine partitions in the execution graph, so align declared partition metadata with the length vector’s actual ownership unit.
171 versus 13 in Dask loc: do not force an index mask into positional alignment
Check partition correspondence of the data and index mask at the execution boundary.
Do not interpret a nonfinite singleton variance as stable zero
Keep the population-variance result of a non-finite singleton separate from a finite singleton's no-variation state. Retain input classes and result tags so indeterminate arithmetic is not published as a stability metric.
