[Summary, full note in Shaarli]
The beginner's guide to Infrastructure Automation (Medium, @rdsubhas, via cron weekly). "Encore un bon article qui explique où se situe Docker vis à vis de la CM."
- Coordinated releases and snowflake servers: manual deployments over a weekend, every server set up by hand, easy mistakes (one forgotten server in the 2012 trading disruption cost over $400M), and nobody can rebuild a server that dies.
- Task runners (Bash, Perl, Python scripts): automate the steps but are neither idempotent ("user already exists" on a second run) nor portable (apt-get vs yum, coreutils differences).
- Configuration management (CFEngine, Puppet, Chef, Ansible, Salt): describe the end state (user www in group nobody, /var/www owned by www, OpenJDK 1.6) and the tool works out the diff. Idempotent and cross-platform. Pitfalls: admins writing execute "apt-get ..." everywhere, developers writing endless include_recipe; you need both ops and dev skills, sometimes the automation code is bigger than the app.
- Immutable infrastructure: always throw the old server away and start from a fresh, identical OS, so plain scripts are enough. Made practical by VMs, cloud and Vagrant. Pros: no patching live machines, no soft upgrades, zero-downtime switch. Cons: slow to provision, service discovery as IPs change, no persistent data on the server (hard for databases, NFS), co-hosting apps.
- Containers: kernel features (process groups with CPU/memory limits, chroot, virtual ethernet, mounts) give a sandbox close to a VM that starts in seconds; Unix zones, then LXC made it mainstream.
- Container images: Docker made repeatable images (Dockerfile = a "fat application") shipped from dev to prod; hosts only need Docker (CoreOS, Rancher).
- State isn't dead: it shrinks to environment variables per environment.
- DevOps: devs own their application's image, ops build the platform that runs containers (scalability, performance, resilience, uptime). Devs do ops, ops do dev.