Enforce both limits with one account-level admission decision
Claid documents default image-processing ceilings of four requests per second and 120 per minute. Neither limit supersedes the other. A scheduler operating safely under one boundary can still exceed the other.
A conservative implementation maintains a shared ledger of request-start timestamps for all participating workers. Before admitting another request, count starts in the preceding one second and preceding 60 seconds. Admit only if the next start remains within both local thresholds.
The rolling-window ledger is an editorial scheduling policy, not a claim about Claid's exact server-side window algorithm. It deliberately avoids relying on a single worker's local count or assuming all other account traffic is visible.
Where unrelated clients use the same quota, reduce the scheduler's operating ceiling or coordinate their starts. A distributed fleet requires a genuinely shared admission mechanism; independent worker counters are not sufficient.
| Locally evaluated condition | Admission decision | Next action |
|---|---|---|
| Fewer than 4 starts in the local 1-second window and fewer than 120 in 60 seconds | Eligible under the local policy | Admit one start, record its timestamp, then process its response. |
| Short-window count at its local ceiling | Hold | Wait until enough recorded starts leave that local window. |
| Minute-window count at its local ceiling | Hold | Wait until the relevant earlier starts leave the 60-second window. |
| HTTP 429 received | Rate-limited | Record the reported reset interval, respect it and re-evaluate both local windows. |
| Response headers conflict with local capacity expectations | Reduce admission pressure | Reconcile observed provider limits and account traffic; do not infer unused capacity. |
Interpret provider headers as current-window feedback
Claid documents RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset. Its normal response example lists both configured limits in RateLimit-Limit, while the remaining and reset values explain the quota window currently being reported.
For example, the documented normal response reports a 120-request binding limit, 118 remaining and a reset in 32 seconds. That does not independently supply the remaining allowance in the one-second window. The documented 429 example instead reports a binding limit of four, remaining minus one and a one-second reset.
After a 429, retain the reset observation and hold the affected queue. Once the indicated interval has elapsed, consult the existing local timestamp ledger and any additional account-level throttling state. If the local gates permit it, cautiously admit the next request. Its subsequent response headers provide new feedback.
Do not require a fresh header before the first post-wait request: without a separate documented status mechanism, that requirement would be circular. Nor does a passed reset interval establish capacity in every other applicable window.
Separate scheduler policy from provider capacity
Keep a record of each admitted start, worker, response status and applicable rate-limit headers. Distinguish planned starts from actual attempts. If 429 responses recur despite conservative local counting, reduce the operating rate and investigate untracked account traffic or other restrictions.
Claid directs users to contact it regarding expanded limits or batch processing. The source does not specify a universal retry algorithm or guarantee that additional capacity will be granted.
Request pacing only decides whether a processing attempt may start. It does not establish processing duration, successful output delivery or product-image acceptance.
Illustrative scheduling decision: 180 images and three workers
A seller plans 180 independent image-processing requests across three workers. If each worker independently starts one request per second, the combined rate is three per second and potentially 180 starts in 60 seconds. That is below four per second but above the documented 120-per-minute ceiling.
The seller introduces one shared admission ledger and chooses an intentionally conservative combined pace of one start per second. Under that policy, the three workers collectively begin 60 requests in each 60-second planning interval, subject to any stricter observed provider restriction. The 180 requests require three such scheduling intervals; no completion-time claim follows.
If an actual response returns HTTP 429 and RateLimit-Reset: 1, the queue holds through that reported interval and then checks its local one-second and 60-second ledgers before cautiously releasing another request. A subsequent response can update the provider observations. The example does not invent separate returned capacity values for both windows.
Source and evidence scope
Prepared by the CartVisual editorial team with AI assistance. Competitor behavior is documented rather than independently executed for this guide.
Official source snapshot checked .
Claid's Rate limits documentation establishes default request ceilings, header meanings and illustrative normal and 429 responses. It does not specify the server's complete window algorithm, independent remaining counters for every window, automatic retries or guaranteed capacity expansion. The shared rolling ledger is editorial scheduler design.