Application patterns

The most useful habits are to keep client ownership clear, preserve numeric types and distinguish a failed request from an unknown trading outcome.

Reuse clients and close resources

Create one client for related operations on an account. Reusing it allows HTTPX to reuse connections. Close clients with context managers; close a partially consumed stream with aclosing or the sync pricing iterator with closing.

Do not create an unbounded task for every price update. Keep a bounded queue or a latest-price cache, and decide whether dropping old prices is acceptable. Transaction processing needs reconciliation rather than silent loss.

Preserve financial values

Build decimal inputs from strings, for example Decimal("0.001"), instead of converting a binary float. Inspect instrument metadata before choosing quantities or prices: minimum_trade_size, trade_units_precision, display_precision and order-distance restrictions serve different purposes.

Rounding a value to display precision does not establish that an order is valid. Account rules, available margin, price bounds and current market state can still cause rejection or cancellation.

Read the response before changing local state

Most endpoint results are dictionaries containing models. Read response["account"].balance or response["orders"]; do not assume the enclosing response is itself the model. A write may return creation, fill, cancellation or reissue transactions. Inspect the actual keys and retain relevant transaction IDs.

Updating a Pydantic object changes local state only. To keep an account view current, use the API's account changes or transaction history instead of treating local objects as live references to server state.

Retry only when the operation permits it

The SDK already retries eligible reads. Adding another retry loop can multiply attempts and delay failure reporting. Set a total retry budget and account for the client's configured retries.

Do not put order creation, replacement or trade closure inside a generic retry decorator. A timeout can occur after a write has succeeded. Reconcile account and transaction state before deciding whether another write is necessary.

Token security

Use environment variables or a secret store and avoid printing extracted tokens. SecretStr protects routine display of configuration values; it does not secure application logs, exception attachments or stored files automatically.

Rotate credentials through the deployment's secret-management process and recreate clients when credentials change. Avoid logging full request headers or raw config.

Test the application boundary

Use HTTP mocks for error sequences and exact request assertions. Keep practice integration tests separate and opt-in. A successful practice run establishes only the scenarios actually exercised; it does not establish strategy performance or complete live-market equivalence.

See testing, connection failures and performance measurement for concrete patterns.