pnclENGINEGet API access

CAPACITY NOTES

Why you hit 429 with daily budget to spare

The plan ceiling is per second, not per day. A loop that fires every board at once bursts over it, and pacing is the free fix.

The arithmetic said you were fine: boards over interval, summed over the day, well inside the plan. Yet the logs show 429s. The confusion comes from mixing the two meters. Your daily volume is accounting; the ceiling that rejects you is per second. You can be rich in the first and broke in the second, and the fix is pacing, not a bigger plan.

The ceiling watches the second, not the day

Every plan carries a per-second allowance, and crossing it returns 429 with a Retry-After header, as the rate limits guide documents, checked 2026-10-25. Nothing in that response asks how your month looks. A burst that crosses the per-second line is rejected on the spot, even if the burst is the only thousand calls you make all day.

The typical shape of the problem is a loop that wakes on a timer and fires every board at once. Ten sports, live and prematch, all requested inside the same hundred milliseconds: that is 20 requests in a fraction of a second, against a ceiling that may be 10 per second. The loop then sleeps 29 seconds, well behaved on average, and still eats rejections every cycle.

Pacing is free headroom

Spread the same work across the cycle and the burst disappears. Twenty boards over a 30-second interval, started one every 1.5 seconds, never comes near the ceiling, and the day's request count is identical. The polling guide names even pacing as one of the habits that keep request volume flat, and this is where the habit pays: the plan's per-second line behaves like a much larger allowance when the load is smooth.

Two details make pacing stick. Add a small random offset to each board's schedule so cycles do not align after a restart. And keep the pacing in the scheduler, not in the retry handler, because a retry storm is just a burst with extra steps.

When it really is the plan

Pacing does not fix a total that is over the line. Sum your boards divided by their intervals, every worker included, and compare against the allowance. If the honest number exceeds the ceiling at your busiest minute, no smoothing saves you; that is a plan conversation, not a scheduling one.

The diagnostic is the 429 distribution. Rejections clustered at cycle starts, with quiet seconds between them, point at bursts. Rejections spread evenly through busy minutes, on a smooth schedule, mean the total itself is too high. The monitoring guide's request and error rates give you both views, checked 2026-10-25.

A ten-minute repair

  1. Log each request's timestamp for one busy hour.
  2. Bucket them by second and find the peak.
  3. If the peak crosses the ceiling while the average sits far below it, stagger the starts.
  4. Re-measure for a day. The 429 count should fall to noise without a single dropped board.

The daily number tells you what you bought. The per-second number tells you how to spend it. Keep both honest and the 429s take care of themselves.