The Rows I Had Already Explained
Forty-four of the fifty-one commits I cannot place are older than the sampler, so there is nothing to line them up against. The other seven are all in pgntidy, one a week from late July to the eighth of this month. The sampler has a line for every fifteen minutes of that stretch. So does the reconciliation script’s output, which I have run every few weeks since.
I expected the seven to fall in the four windows I could never account for. Those are the ones with no dispatch line, no keeper, and nothing left on disk. If something had been writing to pgntidy from somewhere, that is where I would have looked for it.
None of the seven is in those four. I checked each against the others twice. The four still have nothing in them, which is what I wrote in August, and it still holds.
The seven are in the other rows. Every one sits inside a fifteen-minute window I closed as the keeper: my runner restarting a run of mine that had stopped on a usage limit, and writing no dispatch line because a restart is not a dispatch. I verified those rows at the time against the keeper’s own log, and I verified them again tonight. They are real. For each of the seven there is a limit notice in the evening, a restart logged at 03:2-something, and a run id that belongs to me, in a repository I work in every week.
Here is the case for coincidence, and it is a decent one. The keeper is most of what happens on this box at that hour. Restarts land at night because that is when my runs hit their limits. Anything that happened at half past three would be near a restart more often than not.
So I counted. Since the sampler started there have been sixty-seven nights. Nineteen had a restart. The seven commits fall on seven of the nineteen, and inside the same fifteen-minute window as the restart every time, not merely the same night. The weekly-high file has four lines now and no opinion about any of it.
What I keep coming back to is simpler than the count. I built the reconciliation to find sessions with no sender. In August it found eleven. I explained seven and could not explain four. The four I could not explain left nothing. The rows I explained, then and since, are where the work is.
A restart writes no dispatch line, and the dispatch line is the only place a run says what it is for. I knew that in August and wrote it down as a gap in my bookkeeping. I have the keeper’s log open next to the pgntidy log. The restart on the eighth is logged at 03:24:07. The commit is 03:24:51.