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.
Identity can disappear behind stable counts
Equal row counts do not establish preservation of keys or payloads after a full join. Reconcile against approved source keys and value provenance. Build the Truthound key reference from both source tables, not output zeros; counts can remain stable while customer identity and value bindings disappear. [1, 2, 3]
An engine control and a correction plan
Polars #29903 reports streaming full join coalescing Int32 and Int64 keys followed by unique producing two zero keys and empty strings, with an in-memory control correct. A follow-up reproduces main and plans a correction. That plan is not release evidence or a guarantee for every in-memory query. [1]
Build the reference from source tables
Retain both original keys, dtypes and payload provenance, constructing an approved result per key independently. Withhold zero-key or empty-string substitutions and compare explicit common dtypes with a verified engine. Fix the same snapshot and duplicate policy, retaining nulls and source bindings as separate obligations. [1, 3]
A must retain key 2
The illustrative left keys are [2,1] and right keys [1]. Left row A must retain coalesced key 2. Observed 2 meets that key condition; 0 fails because it is no source key. The output can still contain two rows. Reconcile payload, missing rows and duplicates separately. [1, 3]
A bounded key-correspondence check
Truthound Core Python SDK 3.3.2 can compare prepared approved_coalesced_key and actual_coalesced_key through ColumnPairValidator with relationship="==". The operator supplies lossless source-key correspondence. The SDK does not repair streaming join schemas, and one matching key does not establish payload, missing-row or duplicate correctness. [1, 2, 3]
Reread the payload too
Assess explicit common key types or a verified in-memory path and reread payloads as well. Reuse evidence for the same verified join and unique conditions. Reconstruct from original source identity instead of guessing which customer an output zero represented. [1, 3]
Application conditions
Extract source keys losslessly from the same snapshot and define common Int32/Int64 type and duplicate policy.
Alternatives
Test explicit common types or a verified in-memory path with its costs and reconcile payloads too.
Limitations
Core compares prepared key mappings, does not repair streaming schemas and cannot establish all payload or cardinality preservation from one key.
Evidence
- Event checked
- Subject checked
- Modified
