Styxis
Company
Knowledge

Trino decimal-bucket join omissions: compare parallel runs on one snapshot

When fixed inputs produce unstable results, inspect key-level task assignment and match presence rather than only value ranges.

Repeated runs are part of the evidence

Equal input row counts do not establish identical key-to-task placement or join completeness. Giving Truthound one successful output can miss nondeterministic omissions. Keep separate results from repeated runs against the same snapshot. [1, 2, 3]

A suspected shared decimal-hash buffer

The public Trino report observes different results on the same Iceberg snapshots and points to a possible race in a shared decimal hash buffer. It includes an operational comparison with an earlier version but explicitly does not claim an executed reduced reproduction. Keep the observed scope separate from the strength of the causal explanation. [1]

Establish which keys should match

Collect keys, buckets and match states for repeated runs on a fixed snapshot. Compare parallel omissions with a serial control and hold the transition when they disagree. Preserve independent evidence that reference-side keys exist. Runs against a changing source are not automatically the same experiment. [1, 3]

An illustrative missing-match condition

In a synthetic input where every inspected key has an approved reference match, unexpected unmatched keys must total 0. An unexpected-unmatched count of 0 preserves reference matching in that run. A count of 2 identifies business keys whose existing reference matches were not joined. Legitimate outer-join rows without reference keys are outside this scope. A single result of 0 does not prove the absence of future races. [1, 3]

Check conditional completeness with Truthound

Truthound Core's ConditionalExpressionValidator can require match_present when a reference key exists. Join independent reference evidence and observed output by the same business identity. Truthound does not repair a threading race. Evaluate a control without bucket optimization and repeat the original parallel conditions after correction. [1, 2, 3]

Application conditions

Collect run_id, business_key, bucket_id and match_present across repeated executions on a fixed snapshot, independently establishing reference-key presence.

Alternatives

Repeat serial or non-bucket-optimized controls on the fixed snapshot.

Limitations

  • Do not claim that the issue’s reduced reproduction was executed. Truthound neither repairs threading races nor guarantees absence of nondeterminism from one check.

Evidence

Event checked
Subject checked
Modified