Cold Start Performance

A cold start includes finding capacity, loading the image, and initializing your app. Here are a few ways to reduce it.

Cache Dependencies and Model Files

Install dependencies in your image and cache model weights in a Volume. Point your model library’s cache path at the mounted directory.

Initialize Once Per Worker

Use on_start for initialization shared by requests handled by that worker. Access its return value through context.on_start_value.

from beam import endpoint


def initialize():
    # Replace with your model loading or expensive setup.
    return {"labels": ["positive", "negative"]}


@endpoint(on_start=initialize, keep_warm_seconds=300)
def predict(context):
    return context.on_start_value

Multiple workers initialize separately, so account for their combined memory use. See startup loaders.

Keep Capacity Warm

Increase keep_warm_seconds or set min_containers to keep containers running between requests. Warm containers are billed; traffic beyond their capacity can still cause cold starts.

Checkpoint Restore

Supported workloads can capture initialized state with checkpoint_enabled=True:

from beam import endpoint


def initialize():
    return {"ready": True}


@endpoint(on_start=initialize, checkpoint_enabled=True)
def predict(context):
    return context.on_start_value

Check lifecycle logs for capture and restore results. Support depends on the runtime and workload; reconnect external services after restoring. For containers, use beam deploy --help-all to find checkpoint and readiness options. For sandboxes, see snapshots.

Measure the Result

Compare cold and warm requests in the dashboard’s task timings. Use logs and metrics to find which startup step takes the longest.

Start Here

Sandboxes

Inference

Building Blocks

Container Images

Data and Storage

Scaling

Operations

Run in Your Cloud

Advanced Topics

Resources

Security

Cold Start Performance

Cache Dependencies and Model Files

Initialize Once Per Worker

Keep Capacity Warm

Checkpoint Restore

Measure the Result

​Cache Dependencies and Model Files

​Initialize Once Per Worker

​Keep Capacity Warm

​Checkpoint Restore

​Measure the Result

Cache Dependencies and Model Files

Initialize Once Per Worker

Keep Capacity Warm

Checkpoint Restore

Measure the Result