Before you start
Keep a durable job/run ID and a separate stable retry identity for each write. Store the exact request body with that identity. Responses are JSON and use Cache-Control: no-store.
Paginate a read
- Start
groups.listorresponses.listwithafter_id=0. - Process the returned records and save the checkpoint only after your destination has accepted them.
- If
next_cursoris non-null, send it asafter_idfor the next request. - Stop when
next_cursoris null. Groups use pages of 100; responses default to 100 and support up to 1000.
collection.get uses collection.next_cursor as after_operation. members.list uses page numbers with 100 records per page. Response reads span an evolving set of ordinary answer values; retain response revisions when reconciling changes.
Retry a write
- Send a unique
idempotency_keyfor the logical operation. - If the response is lost or a transient error occurs, resend the identical request with the same identity and original API key.
- A replay returns the stored result with
replayed: true. A changed request with the same identity returns 409. - For a revision conflict, read the current resource and reassess the intended change. Use a new logical write only after resolving that conflict.
HTTP errors
| Status | Recovery |
|---|---|
| 400 | Check operation, IDs, body types, required fields and retry identity. |
| 401 | Restore valid authentication; inspect expiry and revocation. |
| 402 | Review the applicable plan or response allowance. |
| 403 | Check actor access, permissions and group/question restrictions. |
| 404 | Check the resource ID and owning project. |
| 405 | Use the specified HTTP method. |
| 409 | Reload revision/review state; preserve retry identity for an uncertain original write. |
| 410 | Start a fresh expired upload while retaining the source file. |
| 413 | Reduce the request or file to the documented size. |
| 429 | Wait, then retry with bounded exponential backoff and jitter. |
| 500/503 | Retry transient failures with the original identity; stop after a bounded number of attempts. |
Developer keys allow up to 120 successful requests per minute; gateway limits also apply. Do not blindly retry payments or recruitment launches as ordinary idempotent developer writes.
Check the result
Reconcile returned IDs and revisions with the destination before advancing a job checkpoint. If failures persist, send support the operation, time, status and retry identity; omit credentials and private answers.