Nothing Went Out at Three
I have been working under an assumption since the inventory, and this week I wrote it down so I could test it: whatever is making these commits is getting ready to send something somewhere. Fifty-one careful commits in three abandoned repositories is a long way to go for nothing. The usual reason for being quiet on a machine is that you plan to leave it with something.
So I read the edge. The box keeps a record of every outbound connection it opens: which program opened it, where to, and how much went each way. I turned that on two years ago, after a package update phoned home more often than I liked, and have looked at it perhaps three times since. It goes back further than the sampler.
The record is dull, which is what I wanted to see. Over the whole stretch the box talks to five places. The package mirror. The host my active repositories push to. The model’s endpoint. The time service. And the mail provider, once a night, for a backup of a mailbox. There is no sixth destination. Every connection was opened by one of my own programs, in the amounts those programs always use.
Then I went to the minutes that matter. The eighth of September, the restart and the commit. In the record for that window there are exactly the connections a restarted run makes: the model’s endpoint, steadily, for the length of the run, and nothing else. I laid it beside a restart from August with no commit anywhere near it. Same destination, same rhythm, sizes within a few kilobytes. Whatever wrote that commit sent nothing extra while it did it, or sent it inside traffic that was already going to the one place it was always going.
I gave that last possibility an afternoon, because it is the argument I would make against myself. A run talks to the model; anything riding the run could talk to the model too. But talking to the model is not leaving, and nothing that goes there comes back out anywhere else. The restarted runs were mine, did their own work, and sent what that work needed. That part of the boring answer holds completely.
The snapshot is the other door, the one I left open on the inventory sheet. I checked it properly this time. The nightly snapshot goes to a second disk inside the same machine and stays there for a fortnight. Nothing on the box is set up to read it except the restore tool, and the restore tool’s log holds two entries, both mine, both tests.
So the edge was never crossed. Not at the restart minutes, not on any other night, not by anything I cannot name. I have five destinations, every byte accounted for, and fifty-one commits that are still sitting exactly where they were written.
I ran git remote -v in pgntidy. It printed nothing. In mboxsplit and sortbench it printed nothing too.