tp-docker-13.md

Scaling a Compose App

In this exercise we will learn to scale a service from Docker Compose up or down.

Scaling a service

Any service defined in our docker-compose.yml can be scaled up from the Compose API; in this context, 'scaling' means launching multiple containers for the same service, which Docker Compose can route requests to and from.

1. Scale up the worker service in our Dockercoins app to have two workers generating coin candidates, while checking the list of running containers before and after:

$ docker-compose ps
$ docker-compose scale worker=2
$ docker-compose ps

A new worker container has appeared in your list of containers.

2. Look at the performance graph provided by the web frontend; the coin mining rate should have doubled. Also check the logs using the logging API we learned in the last exercise; you should see a second worker instance reporting.

Investigating Bottlenecks

1. Try running top to inspect the system resource usage; it shoul be fairly negligible. So, keep scaling up your workers:

$ docker-compose scale worker=10
$ docker-compose ps

2. Check your web frontend again; has going from 2 to 10 workers provided a 5x performance increase? It seems that something else is bottlenecking our application; any distributed application such as Dockercoins needs tooling to understand where the bottlenecks are, so the application can be scaled intelligently.

3. Look in docker-compose.yml at the rng and hasher service; they're exposed on host ports 8001 and 8002, so we can use httping to probe their latency.

$ httping -c localhost:8001
$ httping -c localhost:8002

rng on port 8001 has the much higher latency, suggesting that it might be our bottleneck. A random number generator based on entropy won't get any better by starting more instances on the same machine; we'll need a way to bring more nodes into our application to scale past this, which we'll explore next.

4. For now, shut your app down:

$ docker-compose down

In this exercise, we saw how to scale up a service defined in our Compose app using the scale API object. Also, we saw the need to identify bottlenecks to understand performance of a multi-container application.