Who am I? French software engineer living in California Joined Docker (dotCloud) more than 4 years ago (I was at Docker before it was cool!) I have built and scaled the dotCloud PaaS I learned a few things about running containers (in production)
Outline What are microservices? What are their challenges? How does Docker help? How can we do this today?
What are microservices?
Microservices: a style of software architecture Break big application down into many small services Example: e-commerce - web front-end - catalog of products - inventory/stock management - shipping calculator - payment processor - billing/invoicing - recommendation engine - user profiles
Why are we doing this? (1/2) Use different stacks for different services (use the right language/tool for the job) Replace or refactor individual services easily Less coordination required when deploying - deploy more often - less risk - more agility and speed
Why are we doing this? (2/2) Promotes many small teams instead of one big army - smaller teams = less communication overhead - see Jeff Bezos' "two-pizza rule" Enables effective ownership of services - a service always has an "owner" (=team) - the owner is responsible (=on call) - the owner is empowered (=can fix things)
What are the challenges associated with microservices?
Fast, efficient RPC calls Docker does not help with that. See instead: - ZeroRPC - Cap'n Proto - XMLRPC - SOAP - Dnode - REST - Queues (like AMQP), for long-running/async operations
Architecturing the application in small blocks Docker does not help with that. Some tips: - Pretend that you're outsourcing parts of your stack - Wrap database access; each object (or table) = one service
Efficient deployment system Instead of deploying one monolith once in a while, we deploy many services very frequently The manual, tedious, uniquely tailored deployment system has to be replaced with something simple and reliable It must be easy to add new components and deploy them Docker does help with that.
Network plumbing Instead of one monolith talking to a database, we now have 100s of services talking to each other Those services will often be scaled independently Instead of direct, deterministic library calls, we now have network calls across load balancers Scaling, load balancing, monitoring must be easy to do Docker does help with that.
How does Docker help us to implement microservices?
Building code, without Docker Deployment scripts are fragile Configuration management is hard (and not a silver bullet) Environments differ: - from dev to dev - from dev to prod
Building code, with Docker Write a Dockerfile for each component Builds are fast - each build step is saved - future builds will reuse saved steps Builds are reproducible - build environment can be well-defined - outside context is limited
Shipping code, without Docker Push code or build artifacts to servers Distro packages (deb, rpm...) are great, but hard to build (too generic) Artifact libraries are great, but tailored to specific languages (too specific) When deployment fails, rollback isn't guaranteed to work
Shipping code, with Docker Container images can be pushed to registries Container engines can pull from those registries Docker Inc. provides a free registry (the Docker Hub) Works out of the box in 2 minutes (includes account creation and engine setup) Can also be deployed on prem Rollbacks are easy (see: immutable infrastructure with containers)
Network plumbing, without Docker Application code must be modified to handle: - service discovery (e.g. lookup endpoint addresses into Consul, Etcd, ZK) - failover (same, but also needs to watch for changes) Development stack becomes either: - very complex (because of those extra components) - different from production (because the code path differs)
Network plumbing, with Docker Application code doesn't deal with plumbing Application code connects to services using DNS aliases DNS aliases are injected by Docker (as entries in /etc/hosts) In dev, DNS aliases map directly to service containers In prod, DNS aliases map to special-purpose containers, implementing service discovery, load-balancing, failover Those containers are called ambassadors
How can we do this today?
Developing on a single node Docker Compose Describe a stack of containers in a simple YAML file Start the stack with a single command Compose connects containers together with links Also provides simple scaling and log aggregation
Scaling with static resource scheduling Docker Compose + Docker remote API Deploy stateful services separately Replace them with ambassadors in the stack Instantiate the stack multiple times Fits well with existing IAAS auto-scaling models
Scaling with dynamic resource scheduling Docker Compose + Docker Swarm Swarm consolidates multiple Docker hosts into a single one Swarm "looks like" a Docker daemon, but it dispatches (schedules) your containers on multiple daemons Swarm speaks the Docker API front and back Swarm is open source and written in Go (like Docker) Swarm was started by two of the original Docker authors (@aluzzardi and @vieux) Swarm is not stable yet (version 0.3 right now)