Should I use AWS Elastic Beanstalk or the Amazon EC2 Container Service (ECS) to scale Docker containers?

I've developed a Docker based application comprised of multiple microservices. It has to consume Amazon SQS messages and processes them. At first I wanted to use AWS Elastic Beanstalk, but then I fell over the EC2 Container Service. Now I don't know which one to choose.

As of now, Elastic Beanstalk supports Multi-Container-Environments. That's great because every microservice has its own application server inside a docker container. The next problem is scaling:

I don't know how the scaling mechanism works. For example: I have 5 docker containers in my Elastic Beanstalk Environment. Now only the fifth docker container is under heavy load, because it has a huge amount of SQS messages to process, the other four are nearly idle, because they don't need much CPU or maybe don't have a lot of SQS messages. Let's assume the 5th container runs a JBoss application server. As far as i know, the server only can consume a limited amount of parallel requests even if there is enough CPU/memory available.

If the JBoss Docker container isn't able to handle the amount of requests, but there is enough CPU/memory available, of course I want to automatically start a second Docker/JBoss container on the same instance. But what happens, if I don't have enough CPU/memory? Of course I want to spin on a second instance, which is configurable through a auto-scaling group in EB. Now a second instance spins up, but every container except of the 5th is nearly idle, of course I don't want them to spawn 4 unnecessary at the second instance too, which would be a waste of resources. Only the 5th should spawn and the others should scale like the 5th scale based on configurable parameters like e.g.: CPU/memory/SQS.

I don't exactly know if Amazon ECS is doing that, or if it's possible at all, but I really can't find any source on the internet about this topic, which is in general said, scaling based on instances/containers.

EB vs ECS really comes down to control. Do you want to control your scaling and capacity or do you want to have that more abstracted and instead focus primarily on your app. ECS will give you control, as you have to specify the size and number of nodes in the cluster and whether or not auto-scaling should be used. With EB, you simply provide a Dockerfile and EB takes care of scaling your provisioning of number and size of nodes, you basically can forget about the infrastructure with the EB route.

Here's the EB documentation on Docker: http://docs.aws.amazon.com/elasticbeanstalk/latest/dg/create_deploy_docker.html

With ECS you'll have to build the infrastructure first before you can start deploying the the Dockerfile so it really comes down to 1) your familiarity with infrastructure and 2) level of effort that you want to spend on the infrastructure vs the app.

Not to resurrect a dead question, but hopefully this helps someone.

The accepted answer is not clear enough: based on what OP described, OP wants ECS, not Multi-Container Elastic Beanstalk (MCEB). As far as I can tell, MCEB never attempts to efficiently pack containers into instances. OP asks in a comment, "if only one is under load, it scales only this one, or does it always scale up the instances and start all containers, no matter under what load they are ?" And the answer is "the latter"; MCEB scales up the instances and starts all containers, no matter what load they are under.

Edit

Don't use the architecture you're imagining.

How micro are your microservices? Would it be ridiculous to give them each a t2.nano? Then make them each a single-container Docker EB app - EB worker applications can be driven by SQS messages. Or use apex.run.

Edit 1/31/18:

AWS Fargate seems pretty cool.

Edit 6/5/19:

Use EKS if you need to orchestrate containers, to satisfy an itch. But really, try to avoid this. Distributed systems are hard.