Hello, had a curious question. We have some flows that are not doing a huge amount of computation in each step, but the setup time to initialize a container via batch takes up much more overhead than the time for the step to run. Is there anything that we can do to improve the instantiation time of the container for each step as we’re moving from step to step?

Hey Kevin 👋 can you speak a bit about your compute setup?

so we’re currently running a metaflow deployment on ecs+fargate, with all our workloads spun up on aws batch. with ec2 instances auto scaling as needed and based on resource requests.

one scenario for added latency would be the time incurred by AWS to bring in an additional instance to support additional workloads. They can be a source of massive latencies

we’re definitely incurring that today. We will be looking to reducing the # of steps in order to reduce the number of setups needed for each new bootstrapped container. The EC2 instance setup is additional overhead on top of what seems like the bootstrapping of the conda environment in each step