disco

Dead Simple COnfiguration management and continuous integration for linux like systems

DISCO is just now beginning development. Expect what you find here to do irreparable damage to any system you run it on. I ONLY run disco on a throwaway VM at current.

Why disco?

Because puppet, chef, cfengine, etc, are all great tools, but they all fall short of the mark, in terms of simplicity, ease of use, and reliability. None of them really follow the UNIX philosophy of "do one thing, do it well, and don't reinvent the wheel".

No, really, why did you name it "disco"?

I wanted an acronym based off of "Dead Simple Continuous Integration", and this was the closest I found.

Requirements

DISCO assumes that you:

- have at least one server capable of running rsyncd, sshd and bash 4+
- have one or more clients capable of running bash 4+, ssh, rsync, and fuse-unionfs

While that's a very simple requirements list, it currently restricts it to recent Linux systems. You may or may not be able to use this tool on FreeBSD or Mac OS X, I haven't tried. Due to the way it executes, this tool will probably never, ever execute properly on Windows.

Why focus so much on linux?

Because if we try to do everything and the kitchen sink, for every OS out there, we run the risk of falling short in the same ways the other CI tools have. By limiting our scope and problem space to recent GNU/Linux systems, we can write a much simpler tool in a much shorter amount of time that is much simpler to understand.

Is DISCO noop friendly (report all incoming changes)?

Yes, DISCO is noop friendly, with a caveat: The way we implement noop is through restricted bash shells. This is generally sufficient, and already proven and simple.

There are some questions around "is the NOOP really secure then?" Well, yes and no. The disco NOOP, like any (bash --restricted) shell, can be broken out of without a whole lot of work. Especially considering that the restricted shell runs as root. In this case, the disco NOOP, much like a barbed wire fence at a buffalo farm, is not expected to ACTUALLY keep anyone inside - it's just a gentle suggestion that you please not run all over the farm breaking things.

Unlike puppet's noop, which is implemented via a guaranteed safe DSL, DISCO assumes an existing trust network between your disco server and disco client; the goal of DISCO noop is to prevent well-meaning trusted sysadmins from accidentally doing stupid things. It does not try to secure your systems from malicious code. That security layer is moved up, onto the maintainer, who must verify the sanity of all code they are sending to client machines.

Specifically: Since DISCO use a unionfs and chroot for the restricted bash shell, I am confiden that your physical disk is safe from accidental modifications (everything is captured on the scratchpad, with the exception of writes to /proc and /sys, which are read-only and simply discarded). The bit that I can't promise is that one of your module maintainers won't find a way to kill a running process, or signal/restart a service in a way that DISCO can't trap/log/noop it. This may or may not be a big deal to you, but should be a consideration before you start migrating to DISCO - how much do you trust your module maintainers?

How do you establish the trust relationship?

DISCO uses rsync(+ssh) with SSH keys, so the answer is, "we don't establish it" - SSH handles that for us by the server allowing or denying the key.

How do you handle parameters (like puppet ENC, etc)?

DISCO uses a section of the filesystem to layout a tree of pathable, walkable parameters. This part of the filesystem is available to the client at execution time, so these variables can be used in scripts, templates, and definition files, to further customize execution based off of parameters. This lives on the SERVER, not the client.

From the server perspective, the parameters tree looks like:

/var/disco/parameters
___ ___ disco
___ ___ ___ client
___ ___ ___ ___ cmds
___ ___ ___ ___ ___ rsync
___ ___ server
___ ___ ___ uri
___ ___ NODE_NAME
___ ___ ___ modules
___ ___ ___ ___ ...
___ ___ ___ parameters
___ ___ ___ ___ ...

Think of it like a large JSON document expressed as a filesystem, with the document keys the filenames, and the values being their contents. This format was chosen because it can be easily created from any number of other existing datasources, and doesn't tie DISCO to any one particular tool (cobbler, etc). The admin is free to create this structure on the server however they please.

Given this, disco does not use a config file, all configuration parameters are present in this tree.

There are only two possible toplevel paths, /disco and /NODE_NAME. NODE_NAME is equal to the FQDN of the client making a request, and /disco is the internal client/server configuration. The parameter tree is transmitted from the server to the client via (yet another) rsync operation, and is accessible as a filesystem tree (or the disco-param command which is just a bash wrapper). These parameters appear in /var/disco/parameters on the client and server, and default values can be found there in the client/server install before the first run of the client.

/disco/client/cmds/rsync : The rsync command to use when synching
    files.
/disco/server/uri : The rsync URI from which to fetch module definitions.
/disco/NODE_NAME/modules : This list defines the modules to install 
    on a given node. 
/disco/NODE_NAME/parameters : This tree defines all configuration 
    parameters for the node not related to any module in particular.

Some special parameters are provided to the client, that do not exist on the paramters tree until runtime:

/disco/NODE_NAME/current_module : This parameter defines the full 
    name of the current module, such that a module definition file 
    can access its personal parameters via without knowing its name, e.g.:
    $(disco-param get $(hostname)/classes/$(disco-param get $(hostname)/current_module)/some/module/specific/path)

How to deploy stuff

DISCO uses rsync(+ssh) to distribute files, and bash to execute supporting scripts. It has a rudimentary dependency mechanism implemented via a topological sort.

Essentially, to deploy something, you need 3 things:

- some files and templates on an rsync server
- some scripts that may or may not do something with those files and templates
- a definition file saying where to get those files, templates and scripts, and which
  order to apply them in, as well as what other things you need deployed before this thing

Scripts

DISCO uses bash for a scripting and templating engine. Instead of writing a custom DSL that lets you specify operations (like Puppet did) or utilize a higher level language (like Chef did with ruby), DISCO just uses the proven bash shell.

Every time your module is executed (e.g. every time the disco client executes), all of the scripts are executed. The order of execution is determined by alphabetically sorting the filenames, rc.d style; so naming your scripts as 00-fix_perms.sh, 10-correct_nodes.sh, etc, will cause them to be executed in the proper order. This prevents you from having to create a separate file that describes the execution order.

Files vs Templates

Files and Templates are delivered exactly the same way - via rsync.

Files are static files who are delivered on to the disk, and no more operations are done to them. They are delivered with the same permissions that they are given by the rsync repository.

Templates are bash scripts who are delivered on to the disk, and then they are executed, with their file contents replaced by their output. Templates are subject to all the same restrictions as scripts (be mindful of the constraints of $NOOP), and in addition, they are ALWAYS interpolated in the safe NOOP execution environment (file modifications will be discarded, and only rudimentary bash builtins are enabled). Templates have access to all client parameters via the disco-param command. Templates will end up with the same permissions that rsync gives them.

Definition Files

Definition files are just a series of files that say what files on the disk should be templated, or executed as scripts, for this module; as well as defining module-level parameters, and dependency requirements, for this module.

Definition files can use node parameters via the $(disco-param /path/to/node/parameter) syntax. This interpolation is done on the client side, so the server does not execute any code for this. This is useful for when a module needs to pull different files or whatever depending on its branch, release name, whatever.

Module Layout

A disco module (also called a "disco ball" for fun) looks like this:

MODULE
___ requires
___ parameters
___ steps
├── files/
├── scripts/
└── templates/

Your module can theoretically pull files, scripts, and templates from any location that can be reached via rsync; however, it is generally considerd good form to include all things relevant to your module, inside its disco ball. The disco ball is then placed in an accessible location on the rsync server, and the disco client will pull all modules, files, scripts, and templates relevant to its execution, and run them.

ALL MODULE FILES, SCRIPTS, AND TEMPLATES ARE DELIVERED RELATIVE TO / ON THE CLIENT.

MODULE/requires

This file lists, one name per line, the names of other modules that must be installed on this node in order for this module to install correctly. This is used to create a dependency graph, and thereby determine execution order.

This file is optional.

MODULE/parameters

Each module can define default parameters which will be made available to all clients using the module. These parameters will be merged together on the client at module fetch time, and any node-specific parameters will override any default parameters specified here (they are rsync'ed over the top of each other). These parameters will be rooted at /MODULE_NAME/... .

This tree is optional.

Server Side Setup

The only server side setup required for DISCO is to setup an rsyncd and sshd server. This is outside the purview of this README.

We would recommend setting up the rsync server to allow your DISCO clients (which MUST run as root), to come in on a non-priveleged, non-root account. You can still use rsync's module definitions with non-root users by setting up ~/.rsyncd in that user's home directory, and adding "--rsh 'ssh -l USER_NAME'" to your /disco/client/cmds/rsync parameter on the clients. This will allow you to specify your rsync locations in your module definitions as USER@HOST::MODULE_NAME instead of having to specify a filesystem path, will give you all the benefits of an SSH key trust relationship, and no concern of incoming root access to the server. (Note that this also prevents the often mysterious and troublesome SSL certificate issues associated with other CI systems.)

The Gory Details ("how does it work?")

DISCO is a work in progress so not all of it is complete, but the general idea is this:

- DISCO client rsyncs its node configuration parameters from the server
- DISCO client performs topological sort of required modules, and for each one:
    - fetch all files, templates and scripts
    - resolve all templates
        - resolve all templates
    - execute all scripts
    - report all differences
- report overall success or failure (any piece of any module failing indicates failure)

DISCO is able to easily report all differences by executing all scripts and templates inside a restricted bash chroot, and on top of a read-only unionfs with a scratchpad on the top, some custom twiddly bits in the middle, and the existing running filesystem at the bottom (read-only). The scratchpad is not merged if there is a failure during live (non-NOOP) execution, to prevent from locking the system in a non-functioning state.

If the NOOP flag is set, then all the same operations are performed, but the restricted environment stops all potentially dangerous commands at the reporting level (presumably), and the fetched files are not merged out of the scratchpad onto the live filesystem.

See the client disco-fs-* and disco-exec-* scripts for more information on how this is done.