Skip to main content
We limit how fast you can send requests, in two places. Our servers throttle each API key, and an edge filter in front of them limits each IP address.

​
Limits on our servers

Each throttle is a bucket of tokens. A request spends tokens from its bucket. The bucket holds up to its capacity, which is your burst, and refills at its refill rate, in tokens per second.
ThrottleBurstCapacityRefill
place Hz
cancel Hz
read Hz
account Hz
stream Hz
history Hz
The history bucket meters the record reads: fills, transactions, settled orders, and the history routes. Each such read costs a base plus one token per page of rows, so a larger page spends more. GET /v3/limits reports every bucket’s numbers.

​
Websocket subscriptions

The stream bucket meters how fast you subscribe. A second limit caps how much you hold: a connection may watch up to markets at once. An event counts as the markets it contains. A subscribe that would take you over that answers SUBSCRIPTION_LIMIT_EXCEEDED, and it subscribes to nothing. GET /v3/limits reports the cap. The key routes and the subaccount routes spend the account bucket, one token a request. A transfer spends from it too.

​
Limits at the edge

The edge is a filter in front of our servers. It checks every request on every route before the request reaches them.
The edge refuses a request with a 403 and an HTML body, with no code and no message. The request never reaches our servers.

​
Pace your requests

A throttle refills over time, up to its capacity: tokens(t+Δt)=min⁡(C,  tokens(t)+r Δt)\text{tokens}(t + \Delta t) = \min\bigl(C,\; \text{tokens}(t) + r\,\Delta t\bigr) Here CC is the capacity, rr is the refill rate, and Δt\Delta t is the time that has passed.
  1. Model each throttle in your client.
  2. When you send a request, subtract its cost from your model.
  3. While your model holds too few tokens, queue the request instead of sending it.

​
Error response

When you go over a limit, we answer 429 with a Retry-After header:
  • Wait Retry-After seconds before you retry. Retrying sooner hits the same empty throttle.
  • Split your work across keys. The throttle counts per key, not per connection.
  • Handle a 429 even if you model the throttles. Our servers share token counts with each other gradually, so your model is never exact.