Let me brief about my project, I'm building containerized security training environments (Future - Randomized Security Environments), aimed to help local students, organizations on their Information security training needs.
Current working - I have an instance group which auto scales according to load running a script to add and remove nodes from swarm, I use pub/sub topics to cater deployment needs which are deployed through (docker stack deploy command). It was tested by 4-5 people and was thought to be working perfectly until, we started trails on my own college students.
It got issues such as port numbers not being assigned to new deployments after 20-25 people deployed onto swarm, I am not understanding why, I mean resource usage is optimal, but swarm isn't assigning ports, after restarting the whole instance, swarm was updated with ports.
I knew there was a swarm option for task-history-limit which default set to 5, maybe that was the issue and it that's why it wasn't able to concurrently deploy. Later same thing happened (After 40 deployments) even after setting it to a higher number (Upgraded infra, low utilization in logs). Even now I'm getting nightmares on not knowing the correct reason of why this is happening.
Sample Deployment stack - https://gist.github.com/Mre11i0t/d16ed39e543094b50019d58d7e4bff99, Aim is to deploy this environment on-demand basis, each environment being isolated to respective user.
I'm jumping from a local docker-compose building, to a production environment, in which I have 4 vps. The first (the manager) is the one with the least resources. The other 3 have the same and are biggers (the workers). I decided to use docker swarm, to manage these infrastructure. My doubt is, Should I be concerned about which host x container is running on? Or this is a bad concept of mine? I mean, is the docker swarm meant for me to abstract from the underlying nodes, and create the services and containers trusting that docker will manage the resources successfully?
Answer is... both!
The goal is to let Docker Swarm manage things for you as much as possible, but also add constraints in order for your application to deploy on the hardware that matches best its requirements.
For example, if you have a reverse proxy and machine learning models, you might want to deploy your reverse proxy on a CPU optimized server, and your machine learning models on a memory optimized instance.
You need to label your nodes properly, and then add constraints so that services are only deployed to the nodes that match your labels. For example in the example above you could add 2 labels: reverse-proxy and ml.
I am explaining how to do this more precisely in this article in case you're interested: https://juliensalinas.com/en/container-orchestration-docker-swarm-nlpcloud/
We have a bare metal Docker Swarm cluster, with a lot of containers.
And recently we have a full stop on the physical server.
The main problem, happened on Docker startup where all container tried to start on the same time.
I would like to know if there is a way to limit the amount of starting container?
Or if there is another way to avoid overloading the physical server.
At present, I'm not aware of an ability to limit how fast swarm mode will start containers. There is a todo entry to add an exponential backoff in the code and various open issues in swarmkit, e.g. 1201 that may eventually help with this scenario. Ideally, you would have an HA cluster with nodes spread in different AZ's, and when one node fails, the workload would migrate to another node and you do not end up with one overloaded node.
What you can use are resource constraints. You can configure each service with a minimum CPU and memory reservation. This would prevent swarm mode from scheduling more containers on a node than it could handle during a significant outage. The downside is that some services may go unscheduled during an outage and you cannot prioritize which are more important to schedule.
I am trying to wrap my head around CoreOS and I perused their official docs, some random articles, and even watched this excellent presentation by their CTO.
My understanding of CoreOS is that its a stripped down, bare bones Linux distribution that requires anything running on it to be an OCF-compliant container, not just a Docker container.
My understanding of fleet is that its systemd at the cluster level
My understanding of flannel is that its a network layer that is used by both etcd and fleet to route network requests to containers living in the cluster
So first off, if my above assertions are incorrect or misled in any way, please begin by correcting me! Assuming that I'm more or less on track, I have a few concerns here:
What concrete benefit(s) does CoreOS offer Docker-contained apps that is not present with other Linux distros, such as Ubuntu or Debian? In other words, what objective benefits do I gain by going Docker/CoreOS vs. Docker/Ubuntu?
Fleet just seems like a scheduling engine, like Mesos or Kubernetes. Is it a direct competitor to these projects, or do they handle scheduling at different "layers" (different types of responsibilities)? If so, what are these distinctions?
There are a lot of moving parts here. The answer already posted is very good. I think there are going to be opinions in any answer you get. I thought I'd go through your punch list in my attempt at 100 bounty points :-)
I've been using CoreOS/Flannel/Kubernetes/Fleet everyday now for about 6 months. When you posted the url to the introduction I decided to watch it. Wow, great presentation. I think Brandon Philips is a very good teacher. I like the way he built upon each technology as he introduced it. I would recommend that tutorial to anyone.
CoreOS is a linux based operating system. It is very stripped down, nothing extra running. For me, it does these things:
Auto updates. Does this well. Dual partitions, updates non-active, swaps active, falls back (I think, I have never experienced a fallback). The have tackled the 'how to update your operating system after you deploy' issue and made it relatively painless.
systemd init system. This one took me a bit longer to like (being a /etc/init.d guy) but, after a while it grows on you. There is a pretty steep learning curve. Once you get what is going on you will like how systemd keeps the machine running specific things, dependencies, restarts (if you want), listening on sockets (like super initd) and spawning processes, d-bus (although I don't know much about this part yet). systemd lets you specify 'units' and units can have dependencies, pre and post processes, etc.
basic services. I've copied the brief description line from each of the services that are running on my CoreOS system.
systemd - It provides a system and service manager that runs as PID 1 and starts the rest of the system
docker - Docker is an open source project to pack, ship and run any application as a lightweight container
etcd - etcd is a distributed, consistent key-value store for shared configuration and service discovery
sshd - sshd (OpenSSH Daemon) is the daemon program for ssh(1). Together these programs replace rlogin and rsh, and provide secure encrypted communications between two untrusted hosts over an insecure network.
locksmithd - locksmith is a reboot manager for the CoreOS update engine which uses etcd to ensure that only a subset of a cluster of machines are rebooting at any given time. locksmithd runs as a daemon on CoreOS machines and is responsible for controlling the reboot behaviour after updates.
journald - systemd-journald is a system service that collects and stores logging data.
timesyncd - systemd-timesyncd is a system service that may be used to synchronize the local system clock with a remote Network Time Protocol server
update_engine
udevd - systemd-udevd listens to kernel uevents. For every event, systemd-udevd executes matching instructions specified in udev rules. See udev(7).
logind - systemd-logind is a system service that manages user logins.
resolved - systemd-resolved is a system service that manages network name resolution. It implements a caching DNS stub resolver and an LLMNR resolver and responder.
hostnamed - This is a tiny daemon that can be used to control the host name and related machine meta data from user programs.
networkd - systemd-networkd is a system service that manages networks. It detects and configures network devices as they appear, as well as creating virtual network devices.
CoreOS doesn't necessarily require that everything that you want to run must be a container. It will run anything that a unix box will run. yum and apt-get are conspicuously missing, but wget is included. So, you can 'install' programs, libraries, even apt-get via wget and be on your way to polluting the CoreOS base. That wouldn't be good, though. You really do want to keep it pristine. To that end, they include a 'toolbox' which lets you run a container like sandbox to do your work that goes away when you log out of it.
My favorite part of CoreOS is the cloud-config. On first boot you can provide user_data called a cloud-config. It is a yaml file which tells the base CoreOS what to do when it boots the first time. This is where you install things like fleet, flannel, kubernetes, etc. It is a real easy way to get a repeatable install of a combination of your choosing on a VM. In a typical cloud-config I will write configuration files, copy files from other machines to install on the new machine, and create unit files that control the other processes we want CoreOS' systemd to manage (like flannel, fleet, etc). And it is completely repeatable.
Here is another interesting thing about CoreOS. You can modify the dependency and configuration of existing units. For example, CoreOS starts docker. But, I want to modify the startup sequence of docker, so I can add a drop-in configuration that augments the existing system docker configuration. I use this to drop-in the dependency for flannel before docker starts, so I can configure docker to use a flannel provided network. This isn't necessarily CoreOS, but, it does make it all fit together.
I think you can use cloud-config with Ubuntu as well as CoreOS, and you can do the same things. So, I think the benefit you get from CoreOS over Ubuntu would be that you get a new release often, the operating system is auto-updated, and you don't have anything 'extra' running (it's lean, and a reduced attack vector is fallout). CoreOS is tuned for docker (it is already running) and ubuntu doesn't have it already running. Although, you can create a cloud-config file that will make ubuntu run docker... In summary, I think you have CoreOS understood.
Another thing that you can get with CoreOS is support, directly from the company, either paid or unpaid. I have had many questions answered by the people at CoreOS via this forum and CoreOS Dev/CoreOS User Google groups.
Your fleet description is also pretty good. Fleet manages a cluster. A cluster is one or more CoreOS machines. So, if you are going to use fleet you must use CoreOS, I guess this would be another of those benefits of CoreOS over Ubuntu.
Much like how a Unit File for systemd controls running a process on a host, a Unit File for fleetd controls running a process on a cluster. There is a bit of syntactic sugar, but a Unit file for fleet is about the same as a unit file for systemd. They fit very well together. Fleet's unit files are saved in etcd's database, so once ingested the unit is persistent, even if the machine(s) that host the unit service go down, the unit description exists in etcd.
Fleet also has commands for listing my machines in my cluster, listing a unit file, showing the units that are running, etc. Basically you can submit units to run on the cluster (or all machines, or on a specific kind of machine (like with ssd drives), or on the same machine as something else is running (affinity), etc, etc).
Fleet keeps it running. If the machine goes away its units are going to be run on some other machine in the cluster.
In the tutorial you reference Brandon uses Fleet to launch Kubernetes. It is very simple to do. By making the Fleet unit files place Kubernetes on all machines in the fleet cluster, as machines are added and subtracted from the fleet cluster Kubernetes automatically uses that machine and schedules the Kubernetes to run on them. I have run my Kubernetes cluster like this as well. However, I don't do that much anymore. I am sure there is a benefit that I don't see, but, I feel like it is not necessary in my environment. Since I already boot my machines with a cloud-config file, it is trivial to put the Kubernetes node services directly in there. In fact, with cloud-config, if I wanted to use Fleet to boot the Kubernetes stuff, I would have to write the Fleet unit files, start Fleet, submit the unit files I wrote to Fleet, when I could just write a unit file to start the Kubernetes node. But I digress...
Fleet is a scheduling mechanism, just like Kubernetes. However, Fleet can start any executable just like systemd via a unit file, where Kubernetes is geared towards containers. Kubernetes allows definition of:
replication controllers
services
pods
containers
(other stuff as well).
So, the assertion that Fleet is just a different 'layer' of scheduling is a good one. You might add that Fleet schedules different things. In my work I don't use the Fleet layer, I just jump directly to the Kubernetes because I am working only with containers.
Finally, the assertion about flannel is incorrect. Flannel uses etcd for its database. Flannel creates a private network for each host that it routes between them. The flannel network is handed to docker, and docker is told to use that network to assign ip addresses from. So, docker processes that use flannel can communicate with each other over ip. All of the port mapping stuff can be skipped since each container gets its own ip address. These docker processes can communicate infra and intra machine on the flannel network. I could be wrong, but I don't think there is any connection between Fleet and flannel. Also, I don't think etcd or Fleet use flannel to route their data. Etcd and Fleet route whether or not flannel is being used. Docker containers route their traffic over flannel.
-g
Yes, your understanding is pretty much correct.
Coreos is designed as a more secure operating system that autoupdates itself by default and runs the bare minimum in services to lessen any attack vector. http://www.activestate.com/blog/2013/08/alex-polvi-explains-coreos
Everything needs to either run in a container, be statically compiled (golang binaries as an example), or be a shell script. There is no python or ruby installed.
Containers/systemd units started by Fleet can be rescheduled on another node should its server fail (assuming you container is ephemeral) and should keep the requested number of instances running over the cluster, whilst obeying deployment constraints. https://coreos.com/using-coreos/clustering/
Mesos is more of a framework for a scheduler, you still need something else (chronos/marathon) to provide jobs to execute, but it is very flexible in that regard and handles utilising server resources better.
I don't have much experience with Flannel, but the new networking plugins coming in a future version of Docker may give you more options for container networking. http://blog.docker.com/2015/06/networking-receives-an-upgrade/
what objective benefits do I gain by going Docker/CoreOS vs.
Docker/Ubuntu?
Technical benefits
The features that attract me to CoreOS are:
It's a cluster, not a single-machine, OS
It's built with failure in mind
It's self-updating
CoreOS is a cluster, while Ubuntu is a single machine. With CoreOS when the machine a container is on disappears, the cluster starts the container somewhere else. When that Ubuntu server fails, its containers go down with it. CoreOS allows the machine to be disposable, which is a good thing.
With that said, keep in mind CoreOS does not handle data persistence; data stored in a container does not exist! ;) In my case I dynamically attach EBS volumes where needed.
Design benefits
To me more importantly, the technical benefits above bring along design benefits. Going into a system designing knowing, "this process will randomly disappear," is great for building resiliency. From the beginning, services are stateless and because you literally have no idea what system a dependent service is on, they must also be discoverable. CoreOS's etcd, a distributed configuration store, can be used to discover where a service is located. Finally, because processes may not be on the same machine, network-accessible services -- a must for horizontally-scalable systems -- are the only way to go.
All in all, I find CoreOS a great for building Twelve-Factor Apps and you get Chaos Monkey for free.
Fleet just seems like a scheduling engine, like Mesos or Kubernetes.
Is it a direct competitor to these projects, or do they handle
scheduling at different "layers" (different types of
responsibilities)? If so, what are these distinctions?
Yes, Fleet schedules a container and determines where in a cluster it runs. If that machine disappears, Fleet also takes responsibility for re-launching it on a working machine.
I haven't taken a deep dive into Kubernetes, but there does appear to be overlap. The way I understand it thus far is that Fleet handles running a single container (a "unit"), while Kubernetes is complementary and orchestrates multiple units comprising a system. For example, Fleet ensures Postgres stays running; Kubernetes ensures your application, e.g. comprised of Postgres, Redis, and Django, are all humming away.
I installed an 8-node kubernetes cluster (1 master + 7 minion) but I faced a networking problem among minions.
I installed my cluster according to this step-by-step Fedora manual, so I use Fedora 20 with its testing repository to get kubernetes binaries.
After installing, I wanted to try the guestbook example, but it seems to me there is a problem with the inter-container networking.
Although containers/PODs are in running state and I can reach my 3 frontend containers (via browser) and the redis containers as well (via natcat), but the frontend, which not on the same host with the redis, cannot reach redis master. The frontend's PHP give back network exception.
Can anybody help me why the containers cannot reach each other among the hosts?
I hope I could describe my setup enough accurately and thanks in advance.
The Fedora guide you followed will only get you running on a single machine. It avoids the issues around setting up networking across nodes.
For kubernetes to work, the following network set up must be satisfied:
Every container should be able to talk to every other container, even across nodes. This means also that the bridge IP range for those containers must not overlap.
Code running on any node that isn't in a container should be able to reach every container (and vise-versa), even across nodes.
It is not necessary (but useful) if computers on the network that aren't part of the cluster can reach the containers directly.
There are a lot of ways to achieve this -- for instance the set up for vagrant sets up GRE tunnels between each node. On GCE we use features of the platform to do the routing. If you are on physical machines on a switch you can probably just do a big layer 2 network w/ bridges. A bulletproof way to get started (but perhaps not the most performant, depending on your set up) is to use something like flannel.
We are working on making this stuff easier to start up (without using a mess of shell scripts) and are thinking of building something like flannel in so that there is a reasonable default.