This document describes some of the recommended practices that will improve the reliability of API clients.
The Modern Treasury API supports Idempotent Requests to prevent accidental duplication of API calls. This feature is particularly useful when initiating actions such as money transfers, entity creation, or resource modifications. For instance, if a network error outside of our control occurs while creating a Payment Order, you can safely retry the request using the same idempotency key to ensure that only one payment order is created.
We recommend building custom policies to retry requests that fail due to transient issues either due to rate limits (
429s) or non-terminal server-side errors (
504). We've outlined additional guidance in our Rate Limit section.
Our API servers use a timeout of 60s and will fail if our application takes longer than 60s to return a response. It is safe for API clients to use the same timeout.
You may require a shorter client-side timeout due to your application's requirements. If an API request incurs your client-side timeout, you should retry the request with the same idempotency key.
HTTP/2 and TLS 1.3 promise several performance and reliability benefits. If available, choose an API client that supports both HTTP/2 and TLS 1.3.
Unfortunately, several popular HTTP clients have yet to introduce native support for HTTP/2. We expect this to become more widely available over the next few years.
Here is an example for the
HttpClient in JDK 20 :
HttpClient client = HttpClient.newBuilder() .version(Version.HTTP_2) .build();
Each new encrypted TCP connection incurs some overhead. Creating many of these in a tight loop can consume valuable resources, lock threads, and cause timeouts. If you are using HTTP/1.1, we recommend adopting a mature connection pool and using HTTP Keep-Alive to reuse persistent connections. This improves connection reliability and decreases latency. Some clients e.g. Java's
httpx have built-in connection pools and already use them by default.
We recommend that API clients configure connection pools to use a keep-alive timeout (a.k.a. idle timeout) of between 2s and 30s. This is because AWS Global Accelerator uses an idle timeout of 30s. Applicable to both HTTP/1.1 and HTTP/2.
The relevant configuration parameters for some common languages and frameworks are listed below:
- This defaults to
infinite, which will cause connection errors due to a race condition that is described in Akka's documentation.
- This defaults to
- This defaults to 2 seconds, which works well.
- Older JDKs use a default of 20 minutes, which will cause connection errors.
- JDK 20 and above use a default of 30 seconds, which works well.
- This defaults to 5 seconds in httpx, which works well.
Updated about 2 months ago