Truthound · All articles · 19/24
Evidence-led Truthound Core guidance on data contracts, check inputs and operational limits.
480 articles
Iceberg expireSnapshots inside a transaction: check files before catalog commit
A successful maintenance call is less informative than whether still-published snapshots retain available file references.
Iceberg–Spark exchange reuse: equal hashes do not establish equal scans
A plausible aggregate still needs evidence that the requested partition actually contributed to it.
Before comparing Iceberg Evaluator and SQL NULL filters, choose the policy
If APIs intentionally use different NULL semantics, a result difference alone does not establish an error.
After copying an Iceberg DeleteFile, verify equalityFieldIds survive
A copied path and successful commit do not establish that the metadata defining deletion semantics was retained.
Validate Iceberg DynamicSink key changes at checkpoint boundaries
Similar field names do not establish a shared key contract when types, order or optionality change.
Check numeric text contracts before changing a DuckDB UNION
A UNION change can affect serialized numeric text as well as duplicate counts.
Do not hide a DuckDB CSV TRY_CAST failure inside a missing-value ratio
Unparseable values, quarantined records and a failed execution are different outcomes.
DuckDB WITH RECURSIVE and NULL branches: make the Boolean schema explicit
A branch containing only NULL still needs a clear type contract for downstream consumers.
DuckDB integer overflow: do not rely on WHERE to protect multiplication
Success that depends on filter evaluation order is not evidence that arithmetic inputs are safe.
Checking inconsistent DuckDB NULL counts with a conservation rule
Before trusting a NULL ratio, verify that total, non-null and null counts partition the same input.
Checking DuckDB LEFT JOINs: equal row counts can hide missing keys
Verify which left-side identities receive NULL-extended rows, not just the total number of rows.
DuckDB ENUM Parquet distinct_count: allowed values are not observed values
The number of declared ENUM choices is different from the number of distinct values actually written to a file.
A DuckDB SUM that cannot fit UINT8: validate aggregate output capacity
Small integer inputs can require a wider aggregate output type.
Doris MySQL CDC snapshots: is the INVISIBLE primary key actually extracted?
A key declared in the source schema may still be absent from the rows delivered by a SELECT.
Uppercase Doris rowset hex IDs: check collisions, not just parsing success
An identifier defect in empty rowsets warrants a contract check without being overstated as observed data loss.
Separate successful Doris external-scan completion from data completeness
A successful Spark job can still contain a short read if cancellation or failure is exposed as normal end-of-stream.
Checking Doris COUNT plus an offset: retain the raw count and the arithmetic
A correct COUNT optimization does not automatically preserve an expression that contains COUNT.
Before Doris Arrow Stream Load, align actual array types with the destination
A valid Arrow IPC representation is not necessarily a compatible input for the destination column's serde.
Checking a Doris array_sort alias: ordering and element preservation are separate
Successfully registering a function does not establish that its comparator arguments retain their intended meaning.
DataFusion random-filter rewrites: similar algebra can change the sample rate
Testing one random draw against a range is different from applying two bounds to separate draws.
