Pinecone 429: rate limited, back off with jitter, batch your writes

TL;DR: 429 Too Many Requests means slow down. Retry with exponential backoff and jitter, batch upserts to cut request count, and watch metrics before you hit the wall. The error-handling guide includes a full implementation; the key principles are jitter, max retries, capped delay, and never retrying other 4xx errors.

The fix

1. Back off with jitter. Exponential backoff (base delay doubling, capped, plus random jitter) is the documented pattern. The error-handling guide includes a full implementation; the key principles are jitter, max retries, capped delay, and never retrying other 4xx errors.
2. Batch. Fewer, larger requests beat many small ones. Upsert in batches up to the 2 MB request limit instead of one vector per call.
3. Monitor. Watch request metrics and reduce rate proactively as you approach limits. Most limits can be raised on request; contact support with the use case rather than engineering around a cap in code.
4. 429 is the one 4xx you retry. Every other 4xx is a bug in your request; retrying it is a loop.
5. Put the backoff in the client layer once, not around every call site. Scattered sleep calls drift and compound.
6. For sustained high throughput, size it properly: dedicated read nodes remove read-unit limits for query, fetch, list, and full-text search. See the throughput guides before building a client-side throttle maze.

When to use this

• This covers exactly what the title says: Pinecone 429.
• You are setting this up for the first time, or auditing an existing setup.
• You want the key gotchas in one place before you start.

When not to use this

• You are doing a different workflow with Pinecone; these steps are specific to the title above.
• You need the full reference docs; this is the short path, not the manual.

Compatibility

• Not pinned to a specific version; follows current Pinecone behavior.

Pinecone's error-handling guide defines 429, prescribes exponential backoff with jitter, and notes most limits are adjustable on request and that dedicated read nodes lift read-unit limits. Verified against live docs 2026-09-26.