DROP ALERT NOTES
Your reconnect count is a health metric
Every reconnect writes a control frame into your log. Counted per day, they are the earliest record of whether your stream setup is stable.
Buried in your relay's log is a number nobody charts: how often the stream reconnects. Every reconnection writes a control frame, and the connected messages your log has been collecting are a week-by-week record of how stable your setup really is. That record is worth reading.
Every reconnect leaves a mark
The stream opens with a control object announcing the connection, as the setup guide documents the frame, checked 2026-11-12. A clean week has a small number of them: one per restart you meant, plus a few for the network being a network. A rough week has dozens, and the count climbing week over week is the stream telling you something before your alerts do.
The frames are free data. You already log them or can start today, and the only work is counting them with a date attached.
What the count separates
Reconnects have two families of causes, and the count tells them apart by when they happen. Reconnects that cluster around your deploys are self-inflicted: every push replaces the process, and each account holds one connection per stream, so a new process politely closes the old one. Those are yours to schedule and yours to forgive.
Reconnects at random hours, with nothing of yours changing, point at the path: a flaky network, a proxy idling you out, a host with a tired NIC. A rising baseline of those is the early warning for the silence problems nobody enjoys debugging on a Saturday.
Give the count a threshold
The number means little without a line. Count your connected frames per day for two normal weeks, and set a soft alarm at double the median. A normal week stays flat; a creeping count is the trend you want to see before the market's busiest night finds it for you.
Pair it with the silence triage you already have: a high reconnect count plus long quiet spells is a different diagnosis than a quiet market, and the dedupe guide's state store depends on the stream being steady enough to keep one, checked 2026-11-12.
What a healthy count buys you
A flat reconnect baseline is the boring evidence that your reader is stable, your network path is clean, and your deploys are the only interruption. That evidence pays the day something else breaks, because the first question in any alert outage is whether the stream itself was steady, and the count answers it without a search.
Count the hellos, not just the alerts. The stream's handshake is the earliest signal of its health, and it has been writing itself into your logs all along.