CAPACITY NOTES
Deploy a poller without missing a price
A poller restart does not have to mean a gap. The cursor handoff, the brief overlap, and the budget check that make deploys boring.
Every deploy of a polling service risks the same two failures: a gap in your price history while nothing polls, or a burst of doubled requests while two versions overlap. Both are avoidable with a cursor handoff and a short, planned overlap that stays inside your rate budget.
The handoff is the cursor
The deploy is safe when the new process resumes exactly where the old one stopped. That is what the stored cursor is for: the integration guide has the poller persist the cursor with the merged state and resume from it after a restart, checked 2026-10-06. The deploy procedure is then three steps: start the new poller, let it take over the cursor and make its first delta read, then stop the old one.
Stopping the old first invites a gap; starting the new first invites a brief overlap. The overlap is the right choice, because the budget absorbs it.
Budget the overlap window
During the handoff minute, two pollers run the same boards. That doubles your request rate for the overlap's length, which is fine if your steady state sits near 70 percent of the ceiling for seconds at a time: the rate limits guide's headroom rule exists for exactly these bursts. If your steady state already hugs the ceiling, slow the interval during the deploy instead of doubling through it.
Watch the 429 count during the window. A clean deploy shows a flat error rate; a spike means the overlap ran long or the interval needs slack.
What the storage does with the overlap
The doubled reads do not double your history. The storage guide's insert rule, write a row only when a selection's price differs from the latest you hold, absorbs the overlap automatically: two pollers writing the same unchanged prices produce one set of rows.
Prove it once, rehearse it forever
Run the deploy against a quiet board first and confirm the sequence in your logs: new process up, first delta read from the stored cursor, old process down, zero gaps in snapshot age. The monitoring guide's snapshot-age metric is the witness; if it never blips during the deploy, the handoff worked.
Cursor first, overlap briefly, watch the age metric. Then a deploy is a non-event, and the price history never notices.