Request throttling
Cognite Data Fusion (CDF) returns an HTTP 429 Too many requests response status code in the following cases:
- A user sends too many requests in a given amount of time (rate limiting).
- A user sends too many concurrent requests (concurrency limiting).
- A user issues conflicting requests concurrently (conflicting requests). Each of these cases requires different solving measures. The response body explains the cause of a particular 429 response.
All requests contribute to throttling equally regardless of their source (extractors, functions, transformation, UI, SDK, etc).
When you register applications, do NOT share client IDs and secrets across multiple applications, even if the applications have common authentication requirements in CDF. It may cause the applications to be subject to rate-limiting and may also cause issues with audit logs, with events from multiple entities being identified under a common client ID.
The best practice is to avoid throttling, because throttled requests still consume some resources and therefore may be counted. For example, a request that was throttled due to concurrency may still consume rate capacity. However, if throttling happens, Cognite recommends using a retry strategy based on the truncated exponential backoff to handle sessions with the HTTP response codes 429 (or 503). The strategy lets clients slow down the request frequency to maximize productivity without having to re-submit failing requests.
Learn more about exponential backoff:
Rate limiting
Rate limiting happens when a user sends requests too fast. Refer to a particular resource's documentation for more details as different endpoints can have different rate limits. The limit scopes are the following:
- By user (identity)
- By project
Some endpoints consider request "weight" (for example, how many objects/items a particular request modifies or retrieves.
To avoid throttling due to rate limiting:
- Reduce the number of parallel workers sending these requests.
- Reduce "weight" of individual requests (for endpoints that have a "weighted" rate limit).
- Introduce delays between requests.
Concurrency limiting
Number of concurrent requests that a particular endpoint can process is limited. Refer to a particular resource's documentation for more details as different endpoints can have different concurrency limits. The limit scopes are the following:
- By user (identity)
- By project
To avoid throttling due to concurrency limiting:
- Reduce the number of parallel workers sending these requests.
Conflicting requests
Requests are considered conflicting when they're concurrently trying to modify the same resource. For example, if 2 concurrent requests update the same asset, they're considered conflicting, and one of them may get throttled.
To avoid conflicting requests:
- Instead of "change log"-style updates, where each modification made in the source system is replicated to CDF, replicate only the latest state.
- Reduce the number of parallel workers sending these requests.