Why declare clock models instead of inferring offsets, and what breaks if the anchor is wrong?

[aLima89]

The clock alignment module takes explicit offset/skew/anchor per source rather than inferring them from the data. I understand the honesty argument, but mechanically: how much does a wrong anchor actually cost, and when does anchor choice stop mattering?

[michellsilva]

The anchor matters exactly in proportion to skew - which is the cleanest way to reason about it.

1. The map is affine: reference = raw + offset + skew * (raw - anchor). Read it as: at the anchor instant, the correction is exactly offset; away from the anchor, skew adds a term that grows with the distance from the anchor. So the anchor is the point where your stated offset is pure.

2. Wrong anchor, zero skew: harmless. If skew = 0 the anchor term vanishes and only the constant offset shifts the timeline - you can move the anchor anywhere and nothing changes. That is why anchor choice only shows up in sources with measurable rate error.

3. Wrong anchor, nonzero skew: a tilt. Mis-stating the anchor by Δ applies an error of skew * Δ at a given raw timestamp relative to the intended model - the timeline is rotated around the wrong point. For small skew (a few ms per minute) the error is small; for a clock that drifts seconds per hour it is exactly the kind of alignment bug that makes two sources look ordered when they are not.

4. Why declare, not infer. Inferring would need a shared reference point anyway (you cannot estimate skew or offset without assuming something is true), and the tool refuses to smuggle that assumption in. Declaring the anchor makes the assumption auditable: an NTP sample, a deploy marker, a known log line. The math stays a deterministic affine transform, so the same declarations always produce the same timeline.

The practical rule: put the anchor at the most trustworthy instant you actually have for that source (a sync point), not at the start of the window. A sync point late in the window keeps both the offset and the inferred drift small near the events you care about.