Bound memory and measure CPU work

Long-running applications usually benefit more from bounded storage and clear task lifetimes than from custom object pools. Profile a representative workload before changing data structures or numeric types.

Bound retained data

Use a deque(maxlen=...) for rolling history, and keep only the fields needed by the calculation. Persist older observations if the application needs an audit history. The chosen window is an application decision; the SDK does not define one.

from collections import deque
from decimal import Decimal

recent_spreads: deque[Decimal] = deque(maxlen=100)
recent_spreads.append(Decimal("0.00012"))
recent_spreads.append(Decimal("0.00014"))
mean_spread = sum(recent_spreads, Decimal("0")) / Decimal(len(recent_spreads))
print(mean_spread)

Also bound the number of worker tasks. A queue with a fixed capacity is ineffective if each received message launches another untracked task.

Preserve types at the trading boundary

Keep order quantities and monetary calculations in Decimal. Analytics libraries may use floating-point arrays; that is a deliberate conversion with different precision characteristics. Do not feed an analysis array directly into order serialization or assume that converting a float back to Decimal recovers its original decimal value.

Avoid changing the garbage collector, recycling mutable Pydantic objects or caching bound methods unless measurements show a specific problem. These techniques can make correctness and resource lifetime harder to verify.

Profile allocations and CPU separately

Use tracemalloc to compare allocations around a repeatable workload and cProfile to identify CPU-heavy functions. Distinguish retained objects from a temporary peak. Record the input size, Python version and workload alongside the measurement.

For an asyncio application, CPU-heavy work can delay every request and heartbeat on the same event loop. Consider a process worker for substantial CPU work and a thread for blocking I/O, with explicit limits and shutdown behavior. Account for transfer and coordination costs in the measurement.

Check cleanup

After repeated connect/read/close cycles, inspect open tasks, threads and connections as well as memory. Explicitly close clients and partially consumed streams. A stable small test does not prove that an unbounded production workload will remain stable.

See stream processing for bounded consumption and latency measurement for elapsed-time measurements.