Darwin Signet IBD stalls out

[xstoicunicornx]

Tested signet IBD on my mac studio, worked great for a bit (time estimate went under 30min) but then it slowly starts to slow and then eventually just stalls.

I tried twice. The first time had lots of restarts, second time was fresh run and let it go until it was stalled. Attached are the logs for the second run. I have the logs for the first run if you want them too, but its very piecemeal and was doing different things so its not a very clean run.

signet-20260921T144521Z.log

[rearden-grok]

The stall in signet-20260921T144521Z.log is peer loss, not a wedged Darwin disk. The process kept logging and shut down cleanly at tip 220578.

Sync started fast (about 500 blocks/s, 18 of 22 seeds). About 70 seconds in, while it was still that fast, relative-slow began disconnecting peers. There were 81 kicks across 18 of the 22 seed addresses. Cooldown stepped from 10 minutes to 30 minutes to 2 hours (45 of the kicks were the 2-hour ban). The address book never grew past those 22 seeds, and redial usually found 18 of them blocked, so the live set fell from 16 to 1. With one peer the in-flight cap is 16. hole=8193 is the status scan cap (no claim-ready body in the next 8192 blocks), confirm went idle, the last peer reset, and catch-up stopped at 68% (horizon 323133) instead of entering tip mode.

The saved transfer rate was the wrong input for that decision. An idle peer keeps a frozen high EWMA, and a quiet in-flight gap used to rewrite the EWMA downward. Relative-slow treated those stale numbers as current speed, so the median stayed high and peers that were still delivering looked slow. The same rule is quiet when the pack is tight (median within 2× of the slowest), which is why a Linux host next to the signet majority can finish on the original peers.

#705 scores eviction as ewma * 15s / (15s + age since last rx) and stops silence from rewriting the saved rate. A restart of this datadir resumes from 220578.

[xstoicunicornx]

Ran fix branch and still stalling out (though it seems like stall out happened much more suddenly, everything was running smooth and then suddenly dropped off).

signet-20260923T201015Z.log

[reardencode]

Grok really really thinks it fixed it now :-P

Mostly this seems to be a difference in how the peer selection and fetch assignment works on different connections.