[Bug]: boto3 client recreated on every invocation

[leandrodamascena]

Expected Behavior

The Lambda SDK client created by Durable SDK should be reused across warm starts - replay behavior. It's a performance issue - every invocation is slower than it needs to be. And over time, that adds up to real cost.

src/aws_durable_execution_sdk_python/lambda_service.py - LambdaClient.initialize_client()

Actual Behavior

I'm building an app with durable functions and noticed that every time a Lambda invocation happens, the boto3 client gets recreated. This is wasteful because the client does a lot of setup work (credential resolution, endpoint config, connection pooling, unzip botocore models and stuff) that could be reused.

I think we could cache the client at module level using a global variable. Lambda might reuses the same sandbox for warm starts, so the cached client would stick around. Credentials shouldn't be an issue either - boto3 handles refresh automatically, which is how all AWS SDKs work.

I tested this code and got these results for a 128MB Lambda:

from aws_durable_execution_sdk_python import DurableContext
from aws_durable_execution_sdk_python.config import Duration
from aws_durable_execution_sdk_python.execution import durable_execution


@durable_execution
def lambda_handler(event: dict, context: DurableContext) -> dict:
    context.logger.info("Starting durable execution")

    # Wait for 5 seconds
    context.wait(Duration.from_seconds(5), name="wait_5_seconds")

    print("Wait completed, execution finished")
    return {
        "status": "success",
        "message": "Durable execution with wait completed",
    }

Creating new client

Duration: 183.866ms
Billed Duration: 184ms

Reusing the client

Duration: 102.735ms
Billed Duration: 103ms

Steps to Reproduce

Deploy the code above and test.

SDK Version

latest

Python Version

3.14

Is this a regression?

No

Last Working Version

No response

Additional Context

No response