Motivated by Florian Haas I’ve looked into Ansible and Salt Stack. After spending most of the day reading the docs, I’ve achieved a certain grasp of the functionality of both systems. Here’s my summary.
Ansible and Salt use some common concepts and share a surprising number of technical details that I’ve put them up here, to be able to concentrate on the unique features below.
- Implementation language: implemented in python.
- Syntax and templating: YAML files and the Jinja templating language are used as basic building blocks for their instructions. The underlying data model is totally different though.
- Abstraction layer: A library of modules provides target environment independent commands to manage resources like services and packages. This is equivalent to puppet’s resource abstraction layer.
- Execution: Both provide central and decentral execution modes similar to puppet’s agent and apply faces.
- Ad-hoc commands: Contrary to puppet, both allow arbitrary commands to be executed immediately on the managed hosts. Both can use their respective command language to build these commands.
- Facts Discovery/Inventory: Both provide ways to collect and use information from the managed systems.
- Homogeneous Systems: Both documents focus on the homogeneous system use-case. Surely both do have the possibility to react to differences of the underlying system via facts and conditionals, but are light on the topic of actually implementing this. Given this is my first reading, I might just have missed it.
Ansible implements a scripting based approach. Every task is a series of
steps that have target-specific implementations. For example, there is a
service task, which has similar functionality to the
service type of
Puppet. Contrary to Puppet, Ansible’s Playbooks are evaluated top-to-bottom
and executed in order, except for so-called handlers which can be triggered
by certain events in the execution. They have the same purpose as
subscribe in Puppet.
Ansible executions are always triggered from a central server and require only a ssh connection to, and a python installation on the target hosts. I guess local-only executions are possible too, but would not make use of Ansible’s power. Ansible can use arbitrary users to connect to hosts and use sudo as required to affect the managed host. Together with the possibility to input passphrases during the Ansible run, this could possibly used to manage certified, audited, or qualified systems where Ansible is not covered by the certification process, while still fulfilling the auditing requirements. That makes my head spin. Also I’d question the sanity of organizations/certifications who require such hoop-jumping, but I digress!
Playbooks can be modularized and parameterize easily. Puppet’s (artificial?) distinction between classes and defines is not required as the complete playbook is executed serially and can intentionally modify the same resource many times. Task include files in particular have the role of defines. They can have parameters and be included multiple times with different arguments on the same host. This leads nicely to something puppet modules (and, no, I do not think that’s a language issue, just something I noticed that the Ansible docs introduce as a “normal” use case) are notoriously bad at, namely multi-instance installations.
Finally, something that is really nifty, is Ansible’s delegations. Each task can individually be executed not on the current target host, but delegated to some other host. For example, a Wordpress task running on the php backend host could delegate the database setup to a mysql host and configure SSL offload on a http frontend server. The whole process would still be a unified top-to-bottom task description. Perhaps some includes, but nevertheless a single flow, executed and sequenced properly.
The documentation is well organized and has a good flow to it. Reading it cover-to-cover went smoothly from introductory material to advanced topics.
Salt has the severe disadvantage of having to contend against the very solid impression Ansible’s docs made earlier. Also, leading the reader astray from the documentation frontpage with a string of “Getting Started” and “Continue reading here” links that leave you stranded in the middle of the documentation while skipping some basic material, didn’t help my reception of the content.
Did I mention that there is a link to the Table of Contents in the navigation sidebar that is not always available? What?
Salt has an agent (minion is a great name for that) running on each managed host. Communication happens over a ZeroMQ messagebroker hosted in the central master process. Contrary to Ansible this causes a whole chapter about properly firewalling the various Salt parts. On the other hand, it presumably reduces communications overhead by orders of magnitude.
Salt lets you define states (SLS) which can then be applied as necessary. This feels much more puppet-like than Ansible’s sequential execution model. Since this is state based and not execution based, the same dependency management as puppet is required.
Configuring multiple instances of the same thing is explicitly implemented by using jinja templating and loops over arguments or other data.
Cross-node configurations are implemented by querying the various datastores implemented in salt, which feed from the inventorised data, external inputs and applied states.
Conclusions and personal opinion
This can only be a very shallow review, as I was only reading the official docs of both projects. Still, I think it already shows the fundamental differences between the systems.
Personally, I like Ansible’s documentation, syntax, and execution model better.