Unlike CPU workloads, GPU workloads can currently only process one task at a time. If you’re running workloads on a GPU, your tasks are going to be processed one-by-one unless autoscaling is enabled.

Autoscaling can be passed as an argument to REST APIs or Task Queues.

Autoscaling is only available in apps that are deployed. You can’t use autoscaling in deployment previews or runs.

Scaling out horizontally (adding more containers)

When you deploy a Task Queue or REST API, Beam creates a queueing system that manages each task that’s created when your API is called.

You can configure how Beam will scale based on how many things are in the task queue (queue depth) or how quickly to process a task in the queue (request latency).

If you aren’t sure which strategy to use, we suggest starting with QueueDepthAutoscaler.

Scale by Queue Depth

Our simplest autoscaling strategy allows you to scale by the number of tasks in the queue.

This allows you to control how many tasks each replica can process before scaling up. For example, you could setup an autoscaler to run 30 tasks per replica. When you pass 30 tasks in your queue, we will add a replica. When you pass 60, we’ll add another replicas (up until the max_replicas setting).

from beam import QueueDepthAutoscaler

autoscaling_config = QueueDepthAutoscaler(

def your_function():

Scale by Request Latency

We also offer a more advanced autoscaling strategy called MaxRequestLatency. This allows you to scale based on the length of time you’d like to wait before a request starts processing.

For example, if you want your app to process each incoming request in under 30 seconds, you can set a desired_latency of 30. Beam will spin up additional resources when (average time per task) * (queue size) exceeds 30 seconds.

from beam import RequestLatencyAutoscaler 

autoscaler = RequestLatencyAutoscaler(desired_latency=30)

def your_function():

It’s important to note that max_replicas will always be prioritized over desired_latency. If the replica limit is reached, your API will continue running but some requests will take longer than the desired_latency you’ve set.

from beam import RequestLatencyAutoscaler 

autoscaler = RequestLatencyAutoscaler(

Real-World Examples

Max Request Latency

Let’s say you’ve deployed a Task Queue with a MaxRequestLatency autoscaler. You’ve configured it to scale up to 5 replicas if your request latency exceeds 30 seconds.

from beam import App, Runtime, Image, Volume, RequestLatencyAutoscaler

app = App(
        commands=['apt-get update', 'apt-get install -y libgomp1']

# Deploys function as async task queue
def train():

Here’s an example scenario:

  • You start with a single GPU replica.
  • You get 3 simultaneous requests to your API.
  • Each GPU can only process a single task at a time, and each take 1 minute to run.
  • Because you’ve added an Autoscaling rule with a desired_latency of 30 seconds, Beam spins up 3 GPU replicas to run these requests on parallel GPUs.

If you didn’t have autoscaling enabled, these tasks would run one-by-one, and take a total of (1+1+1) minutes to run.

With autoscaling, the tasks run on 3 parallel machines, so all 3 task are completed in 1 minute.

Queue Depth

Suppose you have a task queue where each task takes 10 seconds to process.

from beam import App, Runtime, QueueDepthAutoscaler

app = App(name="model-training", runtime=Runtime())

    autoscaler=QueueDepthAutoscaler(max_tasks_per_replica=3, max_replicas=3),
def my_task_queue():
    import time

    print("I'm done!")

Here’s how this would autoscale:

  • After we have a backlog of >3 tasks, we add a replica (up to 3 replicas).
  • If you have >= 3 tasks in the queue, you can expect it to take up to 30 seconds for that 3rd task to process.

Increasing throughput in a single container

You can increase throughput for your workloads by configuring the number of workers to launch per container. For example, if you have 4 workers on 1 container, you can run 4 tasks at once.

Workers are especially useful for CPU workloads, since you can increase throughput by adding workers and additional CPU cores, rather than using autoscaling to additional replicas.

from beam import App, Runtime, Image

app = App(
        cpu=5, # Each worker runs on its own CPU core 

# Launch on 5 workers per container to increase throughput
def predict():
    # These workloads will share compute resources with each other 

Workers are always orchestrated together. When the container launches, all the workers launch.

This can result in higher throughput than using multiple containers with horizontal autoscaling.

Worker Use-Cases

Workers allow you to increase your per container throughput, vertically. Autoscaling allows to scale the number of containers and increase throughput horizontally

Each worker will share the CPU, Memory, and GPU defined in your app. This means that you’ll usually need to increase these values in order to utilize more workers.

For example, let’s say your model use 3Gi of GPU memory, and your app is deployed on a T4 GPU with 16Gi of GPU memory. You can safely deploy with 4 workers, since those will fit within the 16Gi of GPU memory available.

It’s not always possible to fit multiple workers onto a single machine. In order to use workers effectively, you’ll need to know how much compute is consumed by your task.

When you’ve added multiple workers, you’ll be able to see each time a new worker is started in your logs: