Just a quick update on the same two demo.peer.observer addrmans (hal / len).
| new table (hal / len) |
pre-flood | during flood | 19 Jul | 22 Jul – 5 Aug | 15 Aug |
|---|---|---|---|---|---|
| 7d timestamp | 71–76% | 95–97% | 78.4% / 72.4% | 42–46% | 83.1% / 81.7% |
| daily turnover | 1–1.5% | 0.3–0.5% | 0.4–0.5% | 0.4–0.5% | 4.2% / 4.4% |
The thing I wanted to check from 19 July was addrman’s 30-day horizon. After 15 July the 7-day share fell to 42–46% for a couple of weeks. The July bulge aged out of that window, but the same endpoints stayed in the table, so turnover stayed low. Then both lines jump in mid-August. Freshness going back up looks like a new flood, but the turnover spike is the 30-day recovery. Those July addresses can finally be overwritten.
During the flood, freshness was high and turnover was low: same addresses getting new timestamps. This time turnover peaked around 12%/day on 13–14 August, and the rows that left hal almost all had gossip timestamps from 12–14 July. That is 30 days later, which is addrman’s horizon, and it lines up with the freeze timed around 14 July in #164. The replacements are current gossip, so the 7-day share goes through the old 71–76% range and is 83.1% / 81.7% on 15 August.
I’d expect this to settle back toward the old range soon. If it doesn’t, I’ll post another update. A good new discussion of what this did to the network is in Address Relay under stress.