idea: build layers with docker instead of spinning up ec2 instances

[n1ru4l]

There are several docker images available that mimic ec2/lambda environments. You could use those instead.

https://hub.docker.com/_/amazonlinux/

[danielmorvay]

Well, a simple image recognition task on lambda (max ram/cpu) takes only 1.3-2.5 sec. Doing the same with docker containers would speed this (more CPU) up but anyway you might need to pay for 1 mins. I think, for me, the lambda is a better solution as I'm having big burst and long idle periods (EC2 not good at all for this pattern) and my model fits into memory. Generally, for bigger tasks I'd also use docker. Maybe I missed an argumentum you have, please feel free to share your thoughts and experience.

[n1ru4l]

This issue is not about runtime. It is about building the layers.

[antonpaquin]

I think this is a good idea -- an EC2 instance probably doesn't match the environment as well as one of those containers can. I'll start on the implementation

[danielmorvay]

      This issue is not about runtime. It is about building the layers.

Ok, I misunderstood you. Now I see your point.