CAPACITY NOTES
When a poll cycle outruns its interval
A slow cycle that queues behind itself pays for stale data forever. Skip the late tick, log the skip, and let the cursor cover the gap.
A 30-second interval assumes a cycle that takes seconds. Then a big Saturday arrives, the boards get heavier, the parser slows, and one cycle starts taking 40 seconds. Your loop is now behind by ten seconds every cycle, forever, unless you decided in advance what overlapping cycles should do. That decision is the whole post, and it is cheaper than it looks.
Measure the cycle before it matters
Log the duration of every poll cycle: when it started, when it finished, how many boards it covered. The number you care about is the p95 cycle time against your interval, and the pipeline monitoring guide is where such loop metrics belong, checked 2026-11-06. A cycle that averages 5 seconds against a 30-second interval is invisible until the day it is not.
Cycles grow for their own reasons: bigger slates mean bigger deltas, and a slow network or a cold parser adds seconds without asking. The interval stays flat while the work inflates, so the day the two cross always arrives unannounced.
Skip, do not queue
When a cycle runs long, the next one is already late, and the wrong instinct is to let them stack. A queue of overdue cycles is a queue of stale work: each one pays for data that is older than the next cycle's, and the backlog drains at the speed of the problem that caused it.
The clean rule is skip-and-log. If a cycle is still running when the next tick arrives, mark the tick as skipped, let the running cycle finish, and resume at the following tick. The cursor makes this safe: the next successful read covers everything the skipped one would have, because the since cursor only advances when a read completes.
When slow cycles become the normal
An occasional overrun on the biggest slate of the week is a scheduling footnote. A cycle that outruns its interval every day is the design telling you the interval no longer fits the work: widen the interval, split the boards across more workers, or move the heavy part out of the cycle, as the capacity planning guide frames the interval-cost trade, checked 2026-11-06.
The distinction matters because the fixes do not mix. A rare spike needs the skip rule. A chronic overrun needs a bigger interval or a smaller scope, and no amount of skip logic makes a 40-second job fit a 30-second budget.
The five-line guard
The implementation is small: a re-entrancy flag on the cycle, a counter of skipped ticks, and a log line when the counter moves. The counter over a week is the health metric; a nonzero trend is the earliest warning your interval will ever give you.
Cycles that finish early are free. Cycles that finish late are a queue you chose not to build. Write the skip rule once and the interval becomes a promise again.