The GitHub Actions job "Benchmarks" on texera.git/feat/standalone-base-transforms has failed. Run started by GitHub user kz930 (triggered by kz930).
Head commit for run: 20e19e0bdfb06f06ed982e0e6dbccc727d0a8d3a / kary zheng <[email protected]> fix(operator): answer as the engine does when a column becomes text Five operators render a column as text before matching, splitting or storing it, and each read the value's own type to decide how. A hole costs the column that type on the way through a file: pandas reads a holed integer column as a float and a holed boolean one as 1.0 and 0.0, so 6 was matched as "6.0" and true as "1.0". The declared type is the only thing that still says which was meant, a real DOUBLE holding 6.0 looking exactly the same, so the five now take it and narrow first. Four more answers were wrong for their own reasons, each observed against the JVM rather than assumed: - Keyword Search ignored its own Case Sensitive flag, always matching without case. The base fixture is lower-case throughout, so sweeping the flag there decided nothing. - Type Casting emitted one line per unit, where `tupleCasting` takes a Map and reads each column's original value once: two units naming one column collapse to the last. - A cast reaches `parseField(force = true)`, whose numeric branch is NumberFormat, which truncates a decimal and drops a grouping comma where Python's int refuses both. INTEGER and LONG cannot share one narrowing either: Long.toInt keeps the low 32 bits while Double.toInt saturates. A LONG read as a timestamp is milliseconds, and renders in the JVM's zone rather than UTC. - CONCAT skipped a leading null but not a leading empty string. The accumulator earns its separator only once it holds something, so "", "a", "" is "a," and not ",a,". Co-Authored-By: Claude Opus 5 (1M context) <[email protected]> Report URL: https://github.com/apache/texera/actions/runs/34524106683 With regards, GitHub Actions via GitBox
