Docker
How it works
1) An image is built from a Dockerfile — a layered, read-only template (each instruction is a separate, cached layer). 2) The image is pushed to a registry (docker push) and pulled on the host (docker pull). 3) docker run creates a container from the image — a host process isolated by namespaces (its own PID, network, and mount view) and constrained by cgroups (CPU, memory). 4) A union filesystem (OverlayFS) stacks the image layers with the container's writable layer. 5) The dockerd daemon manages the lifecycle, delegating execution to containerd and runc. Containers share the host kernel, so they start in seconds and are lighter than virtual machines.
Problem solved
Eliminates the "works on my machine" problem: discrepancies between development, test, and production environments. It provides portability and reproducibility without the overhead of full virtualization, plus process isolation on a single host.
Components
The daemon managing images, containers, networks, and volumes; serves the REST API.
A layered, read-only template with the application and dependencies, built from a Dockerfile.
A running instance of an image isolated by namespaces and cgroups.
A declarative file defining image build steps; each instruction creates a layer.
Image store; the public Docker Hub by default, with private registries possible.
Low-level container runtime; runc is the reference implementation of the OCI standard.
Implementation
Changes not written to a volume/bind mount are lost when the container is removed.
Containers share the kernel, so isolation is weaker than in VMs — relevant when running untrusted code.
Suboptimal layers and lack of multi-stage builds inflate image size and pull time.
Evolution
Solomon Hykes presents Docker at PyCon; the dotCloud project is open-sourced (March 2013) and dotCloud is renamed Docker, Inc.
Version 0.9 replaces the default LXC backend with its own Go component, libcontainer.
Docker and others announce the Open Container Initiative; Docker donates runc as the reference runtime implementation.
From version 1.11, Docker Engine is built on containerd and runc.
Execution paradigm
Not applicable to ML models — Docker is a containerization runtime, not a neural-network architecture.
Hardware requirements
Containerization is independent of the accelerator; isolation relies on the Linux kernel, not specific hardware.