Permalink
Switch branches/tags
Nothing to show
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
608 lines (483 sloc) 28.8 KB
title layout tags permalink
DevOps101 with SaltStack: Because salt goes EVERYWHERE
post
devops
saltstack
aws
2014/08/devops-101-on-saltstack-example/

I by all means am not a system administrator or SCM guy or DBA guy or whatever. I call myself Developer vulgaris. Still, I'm used to find myself in position I have locally working solution and little or no clue how the heck to get it out in window. Should I just throw it to support guys and let them handle that? Throwing code at admins Image source @mptron.com

It works perfectly well. Until... until you become small consultant venture with 3-5 team members and no additional separate support team to poke issues to.

Contents

Update The post topic and layout grew out of my control and present more or less chaotic nature of my thought flow. I've tried to concentrate the post around the following topics: the problem in hand, my take on it, strategy, goals and how to address them efficiently, evaluation and findings down the road... Hopefully you can make sense out of it.

Given the length, here's a helpful table of contents.

The Problem

I don't fanatically believe that system administrators are extinct creatures from Age of Reptiles. I also don't think that in-team operations experience is a MUST. In my humble opinion in some cases it makes more sense to outsource application maintenance or use PaaS. It depends on scope, budget, difficulty, level of expertise required, average Joe on the team, ... To sum up, in some cases it's up to external stuff you can't control (budget, client wishes..) and in some cases it's team internal decision.

Are operations part of your Core Competences? You don't want to outsource that.

In my latest project our team have found ourselves in need of system administration and maintenance. Project deliveries (amongst others) are legacy system migration to Amazon Web Services, support and operations takeover. As usual things become more and more difficult when you dig in deep enough. Even simple cases (and I consider that particular project small enough to give it M size complexity T-shirt) become nontrivial:

  • Mapping all different kinds of topology to constraints of new platform
  • Ensuring non-functional requirements still stand true
  • We want to take small confident steps and polish later. Thus we are unwilling to throw away all existing infrastructure, security configuration, instance intercommunication schemes, et cetera. We do it as we need to (or are forced to by two previous bullet points).

Recently I came across [@pukhalski's] 9 [tweet] 10

Architecture is constraint-based design. Art is design without constraints.

And I could not agree more. Sometimes we have to limit our creative side and do what we need to do in most safe/efficient way plausible given the constraints.

The Golden Hammer

The struggle for resolution in my practice quite often begins with search for mystic Golden Hammer - some particular practice, tool or set of tools which would do all the hard work instead of us and would make all hard problems to vanish away.

[DevOps] define:devops is such new sexy buzzword methodology glossing all around on top of i-cloudy i-thingy and claims to address exactly that.

tl; dr; In business there is no place for separate development and operations teams working independently in isolated environments. There is ONLY ONE PRODUCT TEAM. Period. Product team is solely and completely responsible for Product, there is no other group to blame for misdeployment or misconfiguration. Everybody needs to communicate, cooperate and work together.

I don't completely buy into DevOps universal swiss knife idea. It's clear that sane amounts of cooperation and communication are only welcome addition and natural fit into bigger product evolution picture, nothing new here. Also, one of DevOps interpretation is what development team empowered by set of tools and practices would prepare OR/AND perform most of operations tasks in automated, repeatable and testable manner. It does not necessary mean developers perform the operations themselves. There is still a place for operation guys if it needs to be there: e.g. monitoring, topology setup, hardware router configuration, Linux kernel parameter tuning, et cetera, et cetera. The DevOps scope is really about defining the framework for automation, testing and team internal communication. The goal is to judiciously automate business enablers, reduce change delivery turnout time and minimize one-sysadmin-know-how and human error factors.

The Goals

In our case our goals were more migration process related:

  • Minimize try-fail-repeat cycle turnout time
    Our initial assumption was we won't get into PROD-ready state right away. In the process we definitely would like to experiment with certain components AND/OR approaches more than few times. Think about multi node environment [provisioning] define:provisioning from scratch, DB backup restores, file system backup restores, etc... It's likely to take a while. EVERY. SINGLE. TIME. I wanted to automate as much as possible to reduce human error risks and speedup the whole procedure.

  • Enable local development
    Our goal was to enable every team member to contribute in right away from migration project start. The plan was to setup local [Vagrant] 11 virtual environment similar to current stage. That would allow us to remove dependency on AWS in local development environment, reduce initial ramp up to new team members, would make it easier to reproduce problems and provide fixes. And once again, experiments become more easily to setup and manage.

  • Document things
    One of the problems in early estimation was getting grasp of existing environment, topology and all inter-connections and inter-dependencies. Having our infrastructure in even semi automated way would be a big step forward in terms of platform internals documentation. Every developer would be able to sift through [Code Versioning System] define:vcs history and be able to make informed guesses about who, when, how and why introduced certain change, adjusted particular parameter or opened some port.

  • Single point of truth
    With infrastructure as data/code we would be able to answer that the state should be in every moment of time. Where should not be questions of whether John followed the Wiki instructions OR some ticket comment OR whatever else. There might be issues in implementation OR our assumptions about the platform universe might be wrong (e.g. what state revision was used to provision service A). However now there is only one place the state should come from - from Code Versioning System.

The Evaluation

I did my best in research what the tools available on the market are, that the trends are and what potential benefits I might get. In the matter of literally 30 minutes I have come up with the following list:

  1. [PuppetLabs Puppet] 1, [puppetlabs/puppet] 2
  2. [OpsCode Chef] 3, [opscode/chef] 4
  3. [Ansible] 7, [ansible/ansible] 8
  4. [SaltStack Salt] 5, [saltstack/salt] 6

I won't try to beat Google Search and repeat multiple various posts trying to do apples-to-oranges OR apples-to-apples comparisons. However, I'm wiling to share my subjective evaluation results and explain why I've chosen tool A over B in my particular use case.

The whole process was highly subjective - no measurable criteria, no Cartesian square comparisons etc... However, those are some of criteria I was paying attention to:

  • Access to and readability of source code and documentation.
  • Size of community and level of support it provides.
  • How quickly and easy it is to start doing something useful.
  • Ready to use templates, solutions OR solution repositories.
  • Best Practices OR similar tutorial documents availability.
  • Infrastructure around the community.

Please note I was limited in time to assess tools. Therefore the results most likely are misdirected or single sided in one way or another. If you, dear reader, find my notes misleading, wrong or incomplete - please comment here. I would like indeed have another look on this.

Chef

+ I liked

I think most programmers find imperative states are easy to reason about since it's something that they are used to do on daily basis - reading imperative code. Chef has one of the richest ecosystem, lots of available modules and templates, AWS OpsWorks. Chef also enables both master - slave and knife bootstrap modes.

- I didn't like

I know I've said earlier I find imperative code more natural choice for programmers, still I prefer declarative [idempotent] define:idempotence states over [imperative code] define:imperative any day. In my opinion idempotence hardly could be overrated. Also, in my experience declarative states are usually more concise and easier to communicate to non programmers (e.g. to consult with SysOps).

Another obstacle is Ruby DSL - no one in our team is neither familiar enough with the language, nor likes it. The language itself is the least of concerns. Our team has small experience maintaining RoR applications (Redmine mostly) and I have to say the way Ruby ecosystem behaves is out of my comfort zone and expertise.

Chef also has one of thickest indirection layers expressing target provisioning states, e.g. up to five folders with multiple files in each describing some particular piece of state. I'm not saying it's definitely the bad thing, it most likely is used to enforce certain structure, enables convention over configuration and provides ultimate modularity and extensibility ;) However it's something I would like to avoid. My preference is to have as flat structure as possible, with least files possible while maintaining certain style and conceptual integrity.

Having said that I find Chef to be excellent project, just maybe not the best fit for our team.

Puppet

+ I liked

In general I find [DSL] 12 more readable and concise than code. I also consider state idempotence nice feature. The project seems to have one of the biggest communities, especially in non-developers camps. By all means it is mature solution.

- I didn't like

It seems the project aggressively tries to monetize on "Enterprise" additions and services.

Also, in the process I had a look in what I consider to be a [simple case] 13 and I couldn't say I liked that. I like DSL and I liked hello world level examples. That being said I don't want to reinvent the wheel and learn another complex still useless XSLT/XML programming language.

Ansible

+ I liked

Lightweight. Almost no extra dependencies on target nodes. Remote provisioning over SSH. Ansible claims to be able to provision Cloud Computing resources (e.g. AWS, Digital Ocean etc..).

- I didn't like

Didn't find any kind of public playbook repository. Also I found documentation too brief. I lacked details, complex examples and best practices documentation.

Discovered later [Ansible Galaxy] 14 is publicly available playbook repository. I must say I most likely misjudged Ansible and this tool is definitely in will try on next small project category.

Salt

+ I liked

It's execution model is imperative but it's states are declarative and idempotent. Basically they are just [YAML] define:yaml files/data describing the end state of things. Some complex dynamic things could be expressed in state templates via [Jinja] 15 markup OR implemented as [Salt modules] 16 in Python programming language. To sum up, Salt provides some multi paradigm mix of things and it's your choice how to glue pieces together.

I particularly liked that Salt offered best practices and recommendations document.

Salt also has supports master - slaves, standalone masterless, salt over ssh, master of master of slaves modes. Again, you have a choice.

Salt also has some sort Cloud Computing resource provisioning support.

I was intrigued by fact Salt was one the most active OSS projects on Github last year. I've registered several documentation clarification and improvement issues to test it, occasionally visited IIRC and a little chat there. I have to admit I was astonished by level of support I received. If anyone from Salt community is reading this, kudos to you guys!

- I didn't like

Salt introduces a lot of it's own terminology. Not all of it is newcomer friendly.

The Final Choice

In the end I choose SaltStack Salt over other solutions. The list of reasons in order of importance:

  • I've found it easier to begin with. Most likely it has something with Getting Started tutorials and quality of documentation in general.
  • The astonishing community.
  • Idempotent declarative states.
  • Multi paradigm approach gave a feeling I could make it work one way or another.
  • Cloud Computing provisioning support

Things I wish I'd known before I start

In the process I have re-evaluated some of the evaluation results and findings.

Things I have discovered during implementation phase (mostly the hard way).

I. Community and enormous velocity has it's quirks

Salt community is huge and vibrant. At the moment of writing [saltstack] 6 has 320 watchers and 3,787 stars. And 1,462 open issues. No doubt, being an open souce gardener [is no an easy walk] 18. The community contribution are easy and kindly accepted, no strict procedures, no CLA agreements and long formal reviews. As a result some of functionality or documentation is incomplete or inconsistent.

I wouldn't say it's all bad OR black and white, I believe Salt has managed to be in compromise sweet spot - getting maximum from community, free of charge, while still be useful and mostly reliable.

II. Screw you GitFS, I'm going home

My initial plan was to leverage [GitFS] 19 to deliver environment configuration to Salt master and slave nodes (minions). Salt recommends to split states and sensitive data in separate things, e.g. [Salt states] 20 and [Salt pillar] 21.

Former one could be split in environments (each Git branch is a separate environment name) and multiple Git repositories with different branch names could be collected under one single environment (e.g. you own states under PROD branch, some Github repo with Apache provisioning formula under MASTER branch).

Pillar data however has drawback [#11575] 22 of not being able to redefine base environment data.

That, lack of transparency, no means to ensure pillar GitFS data and multiple state GitFS repositories are correctly inter-synchronized (each of GitFS repositories is potential point of failure, while most of data might be pulled correctly, however some other git repo might fail due to network connectivity and still refer to previous outdated version), etc. - effectively kill the idea, at least for use in production. A local copy of data on production instances in AWS seems much more better idea. Continuous Integration server with rsync to AWS master seems to provide more much more transparent and controlled setup. Those are two qualities I seek for in production setups.

III. Salt multi environment setup has no answer

[Environments] 23 seems to work best in master of masters of minions (aka [syndic] 24 mode. In such case it's convenient to control multiple environments from one super master.

However, in our case I would like to treat each particular environment as sandbox, each of being base or baseline in itself. E.g. I view STAGE environment as a separate sandbox version of PROD environment. They might not share the same topology, however all the roles/components are present in one way or another (e.g. we use RDS in AWS and PostgreSQL in local development environment). As it turns out, Salt does not have an answer on how to layout things in multi environment configuration. I have tried asking on user mailing lists and IIRC. People do things the way it works best for them, however there is not such thing as Recommended layouts for different types of environments document out there. There are certain means how to cover your needs one way or another (it's the place where Salt's enormous flexibility kicks in), still it feels as reinventing square wheel over and over, again and again.

IV. Salt orchestration is not worth it

Salt has certain means to define order and interdependencies between nodes. In Salt terminology it's called orchestration. The recommended way nowadays seems to be Salt's [orchestrate runner] 25.

I've spent fair amount of time trying to automate our environment initial roll out using that tool. However I've found it inadequately lacking for our needs. The nature of our project enforced multi-node deployment, multi-step restore from backups et cetera. That kind of logic, especially different kind of retry-wait-retry procedures, is difficult if even possible to express using that tool. I'm not saying that could not be done in safe reliable way. But, my outtake was it's quite time consuming to implement it and I would rather spend that time moving project further rather than staying still and doing time sinking exercises. I've decided to leave it as manual step-by-step procedure put in Wiki. It serves our needs, is much more reliable and doesn't bother us much.

V. Salt Cloud provisioning is insufficient

[Salt Cloud] 26 is Salt's way of how to describe and launch nodes (minions) on different target Cloud Computing providers. The list of supported providers covers most of mainstream cargo providers out there.

Initially I've tried to leverage node configuration to Salt Cloud, to keep things in one uniformed way. However I soon had a lot of questions with no answer: how to describe [VPC] 27, how to define security groups et cetera, et cetera.

I have found Salt Cloud inadequately lacking for my needs and not documented well enough (e.g. I've discovered some pieces of functionality from issue tracker, not from official documentation). Salt Cloud might be a handy tool for different use cases, e.g. to launch several nodes outside of VPC, no Elastic Load Balancer and no AutoScale groups, however we decided we will use different solutions more suited for our needs.

What worked well

I. AWS CloudFormation

AWS CloudFormation gives developers and systems administrators an easy way to create and manage a collection of related AWS resources, provisioning and updating them in an orderly and predictable fashion.

Technically speaking it's [JSON] define:json formatted AWS Cloud Computing resources template. See AWS CloudFormation [Template Reference] 31 for more information of what's possible to define using this solution.

What in particular I like about AWS CloudFormation:

  • [Version Control System] define:vcs friendly
    JSON files are perfect fit to be stored and versioned in VCS system. We still are able to dig into history, read comments and reason about changes, are able to rollback to previous versions or even bisect problems if we need to.

  • The [range] 32 of supported AWS resources
    AWS CloudFormation allowed us to define every bit of AWS infrastructure we were up to: Elastic Load Balancer, Instances, RDS, Security Groups, EBS volumes, ... To many to list. We didn't feel we are somehow limited in any sense by this solution.

  • The support of stack [updates] 33
    You modify AWS CloudFormation JSON template file and in most cases AWS is able to use it as new state to stack into (you still need to read each particular resource documentation to see whether the resource supports the update operation and what does it mean in each particular case).

  • Last, but not least - Atomicity of operations and treating stack as whole
    It either updates fully, or not. It isn't stuck in the middle (at least it shouldn't). And if something went wrong, the whole update could be cancelled. I also found wholism quite useful during early prototype phase. I would create POC environment, play with it for a while and then drop it completely with all associated resources. Very handy.

At the moment of writing we use hand-crafted AWS CloudFormation templates, mostly from the control point of view. We wanted to keep it as simple and plain as possible, with as little intermediate transformation, magic or whatever else involved in the middle.

Having said that, we are keeping eyes on alternative approaches, such as:

  • Use [Troposphere] 30 OR similar solutions to generate CloudFormation templates from meta-model. This hopefully will allow us to reduce boilerplate code and make our stack meta-model more concise, readable and documented (CloudFormation templates don't allow to have comments).
  • Use of [boto] 34 library to automate certain things, e.g. certain amount of orchestration AND production environment freeze (e.g. create AMI images from some instances and use them later on as Amazon Auto Scaling launch configuration image, ...).

II. Roles based configuration

We were highly influenced by [Roles based configuration] 29, the concept we overtook in our SaltStack based configuration as well.

In our case we have different STAGE and PROD topology configurations (different count of instances, no RDS in STAGE, etc..), mostly due to practical expense reasons.

In our case we use roles as means of specify resource location and as service discovery starting point.

We assign roles to instances via EC2 [UserData] 35 (in CloudFormation templates). We don't auto generate top files, instead we define that each particular role means in THIS particular environment (single node LDAP in STAGE might become Active-Active LDAP SyncReplica in PROD), e.g.

{% highlight yaml %} ... 'roles:www-frontend': - match: grain - fileserver.client-gfs - www.site1 - www.site2 ... {% endhighlight %}

This approach allowed us to concentrate on functionality and common denominators, with differences kept in role definitions in top files.

III. Service discovery

See previous point. We use (role, AWS region/Zone) tuple as service discovery ID.

It enables us to decouple components from physical topology removes need to hardcode things like DSN names, IP addresses etc..

At the moment we leverage self-made solution on top of [Salt Mine] 36, which seemed the quickest to setup at the time. However it's our feeling it adds quite a lot of [accidental complexity] 37 to the cake, which we don't like. We are keeping our eyes on [Consul] 38 or some clever [dnsmasq] 39 schemas.

IV. Plain layout

In general we like to keep things as simple as possible. We understand it as following:

  • Keep levels of indirection to meaningful minimum
  • Keep things plain
  • Prefer composition OR copy of things to any clever smart-ass inheritance / override schemas.

Plus, we had experienced certain annoyances, bugs and glitches with SalStack GitFS implementation. Having said that we have come to the following general approach how to layout things in our SaltStack based configuration.

{% highlight text %} /srv/salt ├── base │   ├── sanity.sls # shared │   ├── security-updates.sls # shared │   ├── ... ├── ldap │   ├── server-ha.sls # PROD uses it to define it's ldap-server role │   └── server.sls # DEV uses it to define it's ldap-server role ├── reverse-proxy │   ├── httpd │   │   ├── config │   │   │   ├── www-zyx │   │   │   │   ├── default │   │   │   │   ├── httpd.args.jinja # same role definition │   │   │   │   ├── httpd.conf.jinja # uses file.managed with multiple sources e.g. │   │   │   │   ├── prod-httpd.args.jinja # xyz: │   │   │   │   ├── prod-httpd.conf.jinja # file.managed: │   │   │   │   ├── prod-httpd.env.jinja # source: │   │   │   └── init.d.jinja # - salt://{ { grains['environment'] } }-xyz │   │   └── init.sls # - salt://xyz │   └── init.sls ├── stage-top.sls # STAGE top file └── prod-top.sls # PROD top file {% endhighlight %}

I'm not sure it's the best layout out there, but it works for us. Bear in mind that our salt universe might be considered small and simple enough. We definitely don't have hundreds of minions and thousands of services to provision, just a handful ten or so. So, I suspect we would come to different conclusion with different scale level.

Personal takeout

Update I've made a decision to time limit myself on this post and it's scope. So, instead of expanding on this I will limit myself to simple plain bullet points with no explanations whatsoever.

  • Automation saves huge amount of time

  • Don't buy into feature lists

  • Decouple first

  • Time limit proof of concept attempts

  • Make first, polish later