Does tRPC automatically retry or chunk failed httpBatchLink payloads in production?

Answer

No, tRPC’s httpBatchLink does not automatically retry or split a failed batched request.

Confirmed facts

• The link sends a single POST containing an array of procedure calls and relies entirely on the underlying HTTP client (fetch or XMLHttpRequest) for transport.
• tRPC provides no built‑in retry mechanism, no automatic chunking, and no size‑based splitting.
• If the request fails (network error, timeout, HTTP error such as 413 Payload Too Large), the link treats the entire batch as a failure and returns an error for each item; it does not resend or resend smaller batches.

Likely explanation

When a batched payload exceeds the limits of a serverless or edge platform (e.g., AWS Lambda 6 MB, Vercel Edge 4 MB, Cloudflare Workers 1 MB), the request fails before tRPC can process any of its items. Because the link does not inspect or modify the payload after construction, there is no opportunity for it to retry or to break the batch into smaller pieces.

Steps to handle this case

1. Measure the typical size of a batched request in your dev environment (e.g., log JSON.stringify(batch).length).
2. Compare that size to the provider’s documented body‑size limit.
3. If the size approaches or exceeds the limit, choose one of the following:
• Disable batching for serverless deployments by using httpLink instead of httpBatchLink.
• Manually limit the batch size (e.g., split procedures into groups of N before calling the link).
• Wrap the fetch used by httpBatchLink in a retry middleware (e.g., retry-fetch) that respects idempotency of your procedures.
4. Deploy and verify that only a single POST is sent per link call (see network tab) and that no automatic retries occur unless you added them.

Missing diagnostic detail

To decide whether to disable batching or implement a custom size threshold, confirm the exact maximum body size enforced by your serverless/edge provider (e.g., via its documentation or a quick test that returns 413). Knowing this limit changes the recommendation from a generic "reduce payload size" to a specific batch‑size target.