Project development has moved to: https://github.com/mesos/docker-compose-executor.
compose-executor project aims to enable Mesos frameworks to launch a pod of docker containers. Kubernetes/K8 introduced the notion of a collection of docker containers that share namespaces and treat the collection as a single scaling unit. Brendan Burns talked about some design patterns/use cases for pods in DockerCon'15.
Docker Compose is a cherished tool used in docker community that helps us model a collection of docker containers. The specification is very flexible: containers can be launched both in multiple hosts or a single host. Furthermore, you can also model a pod collapsing namespaces (net, IPC supported in docker, PID support coming). The spec is also gaining popularity in 3rd party platforms/IDE like Azure (baked in Visual Studio support). AWS also added support behind it since DockerCon'15.
The project goal is to model a pod of containers with docker-compose and launch it with your favorite Mesos frameworks like Marathon, Apache Aurora, etc. One does not need to switch to Kubernetes on Mesos if all that they are looking to do is launch pods (model pod like workloads). With network and storage plugins supported directly in Docker, one can model advanced pods supported through compose. Furthermore, instead of using some different spec to define pods, we wanted to build around the compose spec that is well accepted in the docker community.
- Write a general purpose docker-compose executor so that it can be used by any Mesos frameworks (We are delivering a patch in Apache Aurora to have custom executor support other than thermos. Marathon can readily support it). The executor will launch a pod of containers per task and maintain the lifecycle.
- DockerContainerizer is not best suited to run it because it can only launch one container per task and has a few other shortcomings. Thus, we plan to use MesosContainerizer with cgroups isolator and then with the newly supported cgroups-parent flag, put the docker containers in the pod under the executor cgroup so that it’s a hierarchy. We are delivering a patch in docker-py and compose to support this docker flag.
- Using Mesos hooks, we plan to catch rare conditions where the executor crashes and containers can leak, giving us the ability to clean them up.
- The compose spec file can be downloaded in the mesos sandbox through the mesos task URL.
- We will make sure the container names haves proper prefix to reflect the mesos task.
- Multiple pods can be launched in the single host without conflicts.
- Container stdout/stderr logs should be redirected to sandbox stdout/stderr.
- We will be writing a standalone tool that can parse the compose file to add up the resources (cpu/mem) of the containers in the compose spec because for Mesos it’s a single task. This is needed before submitting the task to Marathon/Aurora. In the future, there can be frameworks which consume the compose file in the framework and handles it in the server side with local pod or distributed containers treated a unit (compose with docker network namespace say) and spread across multiple tasks. We are not modelling that for now.
- We are going to write a sample compose file trying to model like a pod collapsing net namespace. Once the pid namespace is collapsed, one can run tools like wetty in a separate container in the pod and give terminal emulator to the pod (avoid ssh/docker exec access). We will show a wetty integration merged into a container rather than a separate container for now. Other monitoring/debug tools can be launched in a separate container in the pod.
This is a sample compose file modelling a pod (collocated set of containers treated as a unit):
baseC:
container_name: baseC
image: centurylink/wetty-cli
net: "bridge"
ports:
- "5000:5000"
- "8000:3000"
web:
container_name: web
build: .
net: "container:baseC"
volumes:
- .:/code
links:
- redis
redis:
container_name: redis
image: redis
net: "container:baseC"
In real life use cases, as Brendan also explained in the talks, the pod will comprise of the main app container with side car containers such as amsaddor/proxy containers, log aggregators, monitoring/debug containers etc. For the sake of simplicity and common example, we are putting both web and redis in the same pod. Useful for development clusters and demo :)
In the compose file above, baseC is the primary container getting the IP assigned from docker bridge and is also responsible for advertising ports for other containers in the pod (docker does not allow for other containers joining the network namespace expose ports at that point. Hence, the primary container exposing all the ports and secondary containers in the pod will bind to those).
We are also running wetty (https://github.com/krishnasrinivas/wetty ) inside the primary baseC container, to have terminal emulator to access the pods. Port 5000 is app port, 8000 is wetty port, redis port not exposed outside pod.
The redis and web container share the network namespace of the baseC container.
Here is the output of running netstat –plant from wetty terminal:
This shows information about all the bind ports in different containers in pod.
We cannot share pid namespace yet and that is coming soon moby/moby#10163
With Wetty, one can avoid ssh/docker exec into the pod and use the browser terminal emulation.
Instead of static host port mapping, we can leverage dynamic ports as mesos resource and map it in compose file. However, if the network model assigns a routable IP (instead of private host ip’s, default bridge model), we don’t need to expose host ports as such.
1.Create a marathon app POST /v2/apps
Sample payload:
{
"id": "docker-compose-demo",
"cmd": "",
"cpus": 1.0,
"mem": 64.0,
"ports":[0,0,0],
"instances": 1,
"executor":"/code/docker-compose-executor.sh",
"labels": {
"fileName": "docker-compose-example/docker-compose.yml"
},
"uris":["https://dl.dropboxusercontent.com/u/26009359/docker-compose-example.zip"]
}
executor contains path to the executor. Below screenshot shows taskId appended to container names and
mapping host ports provided by mesos to container ports.
2. Scale up & Scale down PUT /v2/apps/docker-compose-demo
Sample payload:
{
"instances": 3
}
- Patch work for custom executors: https://reviews.apache.org/r/36289/
Steps to run:
$ git clone https://github.com/rdelval/aurora/tree/compose-demo
$ cd aurora
$ vagrant up
$ vagrant ssh
optional: $ aurorabuild all
$ aurora create dcomp docker-comp ""
Open web browser and go to 192.168.33.7:5050 to see task running in Mesos
Point browser to 192.168.33.7:8081/scheduler/www-data to see task running in Aurora
Point browser to 192.168.33.7:5000 to navigate to redis running inside of Docker
Point browser to 192.168.33.7:8000 to navigate to Wetty running inside of Docker
*Note: The first time this is run it could take some time to get the docker images depending on your internet connection, see sandbox stderr and stdout for progress
- docker-py project pull-request: docker/docker-py#716
- Work in progress repo of compose project: https://github.com/mohitsoni/compose
compose-executor is licensed under the Apache License, Version 2.0. See LICENSE for the full license text.