Docker & CI/CD Patterns — DoclingOptimisation

Dockerfile Standards

• Multi-stage build: builder (compile deps, pre-fetch models) + runtime (slim, non-root)
• Base image: python:3.14-slim-bookworm (pinned, not :latest)
• Non-root user: docling:docling
• uv for package management — uv sync --frozen --no-dev
• Models pre-fetched at build time (no network at runtime)
• CPU-only PyTorch from https://download.pytorch.org/whl/cpu
• .dockerignore excludes .git, .claude, .history, .firecrawl, logs, research docs

Docker Compose (Local Dev)

• Compose-native limits (mem_limit, cpus) for Docker Desktop
• Volume mounts: ./input:/home/docling/input:ro, ./output:/home/docling/output
• Default threads: DOCLING_NUM_THREADS=10 for local benchmarking

Azure Container App Job

• This is a batch/one-shot workload — use Container App Jobs, NOT revision-based apps
• CPU profile: 4 vCPUs / 8 Gi, set DOCLING_NUM_THREADS=4
• GPU profile: NC8as-T4 / 8 vCPUs / 56 Gi, requires CUDA PyTorch build, set DOCLING_DEVICE=cuda
• Container registry: Docker Hub (migration to ACR planned)

GitHub Actions

• Separate CI (lint, format-check, type-check) from CD (build, push to Docker Hub)
• Use astral-sh/setup-uv@v4 for uv in CI
• Use --frozen for reproducible installs
• Pin action versions
• Use concurrency groups to cancel redundant runs
• For Docker Hub: use docker/login-action@v3 + docker/build-push-action@v6

Image Tagging

• latest for main branch
• sha-<short-sha> for every push
• Semantic version tags for releases