Negative sub-second durations lose their sign: Serial(-500ms) equals Serial(500ms)

[slachiewicz]

Any time.Duration strictly between -1s and 0 is serialised with its sign lost, so a negative interval silently becomes positive.

Serial(-500ms ) -> "INTERVAL '0.5' SECOND"
Serial(-999ms ) -> "INTERVAL '0.999' SECOND"
Serial(-100ms ) -> "INTERVAL '0.1' SECOND"
Serial( 500ms ) -> "INTERVAL '0.5' SECOND"    <- identical
Serial(-1.5s  ) -> "INTERVAL '-1.5' SECOND"   <- correct

Cause

serialMillisecondsInterval (serial.go:253):

seconds := int64(dur / time.Second)
millisInSecond := dur.Abs().Milliseconds() % 1000
intervalNr := strings.TrimRight(fmt.Sprintf("%d.%03d", seconds, millisInSecond), "0")

Integer division truncates toward zero, so for -500ms seconds is 0, not -0, and the sign disappears. The fractional part then comes from dur.Abs(), so nothing restores it. The sign only survives when the magnitude is at least one second, which is why -1.5s is fine.

Why it has not been noticed

Every duration case in serial_test.go is at whole-second magnitude or larger, so none of them enters the broken range, and there is no integration coverage for intervals.

This is on the default parameter path — Serial runs for every argument regardless of useExplicitPrepare (trino.go:1260) — so it is silent wrong data rather than a rejected query.

Fix

Take the sign from the duration rather than from the truncated seconds, and apply it once to the formatted value. A test at sub-second negative magnitudes would have caught this and should come with the fix.

_{This work was created with AI assistance.}