Files
computer-playbook/docs/contributing/environment/cache.md
kevinveenbirkenbachandClaude Opus 5 8d8701d9b3 feat(apt): one Ubuntu mirror list for the package cache and image builds, with failover, a short timeout and a health test
The ubuntu CI image build (job 103277644636) hung on Canonical's
archive: apt fetched 26 MB at 417 kB/s, then 61.6 kB in 29 s, then sat
on the package download until the job was cancelled after 23 minutes,
while download.docker.com delivered 42.8 MB in under a second on the
same runner. Locally archive.ubuntu.com and security.ubuntu.com took
20-26 s to the first byte against 2 s for a university mirror.

INFINITO_APT_UBUNTU_MIRRORS in default.env is now the one list of
Ubuntu mirrors, in failover order: de.archive.ubuntu.com, ftp.fau.de,
ftp.halifax.rwth-aachen.de, archive.ubuntu.com.

- Image builds: the Dockerfile sources the list alongside the install
  scripts and runs scripts/install/apt-mirrors.sh first. In an image
  with ubuntu.sources it points both stanzas at apt's mirror+file
  method, priorities in list order, and sets Acquire::http::Timeout to
  10 s, so apt moves on from a mirror that fails or goes silent; other
  distros are untouched. In ubuntu:latest apt-get update took 8 s
  against 110 s on the stock sources, a dead first mirror was left
  behind within 25 s against 62 s on apt's defaults, and the swarm node
  packages (80 debs) came in 17 s. The Dockerfile's own RUN lines ran
  clean on ubuntu, debian, fedora and arch.
- Package cache: package.sh proxies the first entry as the primary
  Ubuntu repos and the second as their fallback. Nexus revalidates
  metadata on every request here (metadataMaxAge 0), which costs a
  full download from a mirror without an ETag; both de.archive and
  ftp.fau.de send one, and the 15 MB universe index now comes back
  from the cache in 0.02 s.
- tests/external/repository/test_apt_ubuntu_mirrors.py keeps the list
  honest: the first mirror must answer InRelease within 5 s (median of
  three) or the test names a faster later one, no mirror's -updates may
  lag the freshest by more than 24 h, and every entry must be the
  primary archive, a Canonical country alias or an enabled official
  mirror in Launchpad's registry. It caught the first draft of the
  list: mirrors.edge.kernel.org is not an official Ubuntu mirror.

After pulling, an existing checkout needs `make dotenv-force`: plain
`make dotenv` re-reads the old .env through BASH_ENV and keeps its
values.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-11 17:13:54 +02:00

11 KiB
Raw Permalink Blame History

Local Caches 📦

Three local cache services accelerate re-deploys, reduce upstream traffic, and harden the development workflow against transient upstream outages. They share one compose profile (cache) and one override file (compose/cache.override.yml) and are gated together by profile.py: active on developer machines, inactive on CI runners.

Services 🧩

Registry Cache 🐳

registry-cache runs rpardini/docker-registry-proxy and transparently MITMs every Docker image pull from the inner dockerd, regardless of upstream registry (docker.io, ghcr.io, mcr.microsoft.com, …). The runner trusts the proxy CA via registry-ca.sh.

Pulls only. Pushes are not intercepted.

Package Cache 📚

package-cache runs sonatype/nexus3 and exposes pull-through proxies for package-manager downloads:

Format Repo names Upstream
apt apt-debian, apt-debian-security, apt-ubuntu, apt-ubuntu-security deb.debian.org; both Ubuntu repos take the first entry of INFINITO_APT_UBUNTU_MIRRORS
apt (mirrors) apt-debian-mirror, apt-debian-security-mirror, apt-ubuntu-mirror, apt-ubuntu-security-mirror ftp.debian.org, security.debian.org; both Ubuntu repos take the second entry of INFINITO_APT_UBUNTU_MIRRORS
pypi pypi-proxy pypi.org (incl. files.pythonhosted.org)
npm npm-proxy registry.npmjs.org
rubygems gem-proxy rubygems.org
go go-proxy proxy.golang.org
helm helm-bitnami charts.bitnami.com/bitnami
yum yum-rocky, yum-fedora download.rockylinux.org, download.fedoraproject.org
raw raw-githubusercontent, raw-codeload-github, raw-packagist, raw-alpine raw.githubusercontent.com, codeload.github.com, repo.packagist.org, dl-cdn.alpinelinux.org

A proxy repo that holds a cached copy serves it when its remote is unreachable, regardless of metadataMaxAge (measured). autoBlock is therefore off: while a remote is auto-blocked Nexus answers 404 Remote Auto Blocked instead of falling back to that cached copy. The mirror repos below only come into play when nothing is cached either.

Every apt suite exists twice, once per upstream mirror. Nexus has no group repository type for apt (only maven, raw, docker, yum, npm, pypi, rubygems, go and friends), so the failover lives in the frontend: a request that the primary repo answers with 502/503/504 is re-run against the -mirror repo, which proxies a different host. 404 deliberately does not retry, because a missing file is missing on both mirrors and apt probes for missing files often. The frontend waits at most 8 s for a primary before it re-runs the request against the mirror; the -mirror locations keep the global 300 s.

Bootstrap is idempotent and runs from package.sh once the stack is healthy.

Package Cache Frontend 🔐

package-cache-frontend runs nginx:alpine and reverse-proxies upstream package-manager hostnames onto the matching Nexus repo path. Combined with extra_hosts DNS-hijack on consumers, package managers can hit their real upstream URL and still flow through the cache.

Two listener layers:

  • HTTPS (port 443): per-hostname server certs signed by a dedicated CA. Used by the infinito runner (Ansible-driven pip install, gem install, composer install, curl https://…). The runner trusts the CA via package-frontend-ca.sh.
  • HTTP (port 80): plain mirrors for deb.debian.org, archive.ubuntu.com, security.ubuntu.com, dl-cdn.alpinelinux.org. Used by inner-dockerd Dockerfile builds via build.extra_hosts DNS-hijack. No CA-trust required in the build container.

Cert generation runs in a throw-away alpine container driven by package-frontend-certs.sh before the frontend starts.

Activation 🎚️

The cache decision is exposed via Profile.cache_stack_enabled() in profile.py. When active:

  • compose/cache.override.yml is layered on top of the base compose.yml by compose.py and down.py via common.pycompose_file_args.
  • The cache services are added.
  • The runner's infinito service receives:
    • bind-mounts for the registry-cache CA, the package-cache client snippets (pip.conf, npmrc, apt/${INFINITO_DISTRO}.list), and the frontend CA file
    • extra_hosts entries DNS-hijacking the HTTPS upstream hostnames to the frontend's static IP
    • INFINITO_CACHE_PACKAGE_FRONTEND_IP env var for the inner compose wrapper
  • Cert generation, Nexus repo bootstrap, and runner trust-store install run from compose.py.

When inactive, the override is omitted: cache services do not exist, the runner has no cache mounts or DNS-hijack, package managers go direct to upstream.

CI signals (GITHUB_ACTIONS=true, INFINITO_RUNNING_ON_GITHUB=true, CI=true) deactivate. Fresh runner disks per CI job give no cross-run amortization.

Coverage Matrix 📋

Traffic Mechanism Cached
Image pulls (any registry, inner dockerd) registry-cache MITM via HTTP_PROXY env on dockerd ✓
apt-get install (Ansible task in runner) apt/${INFINITO_DISTRO}.list URL-rewrite to package-cache:8081 ✓
pip install (Ansible task in runner) pip.conf URL-rewrite ✓
npm install (Ansible task in runner) .npmrc URL-rewrite ✓
gem, composer, go, curl https://… (Ansible task in runner) DNS-hijack + frontend HTTPS + runner CA-trust ✓
RUN apt-get install (inner Dockerfile build, Debian/Ubuntu) build.extra_hosts DNS-hijack + frontend HTTP ✓
RUN apk add (inner Dockerfile build, Alpine ≤3.17 HTTP) build.extra_hosts DNS-hijack + frontend HTTP ✓
RUN pip/npm/gem/composer/go install (inner Dockerfile build) none today ✗
RUN apk add (Alpine ≥3.18 HTTPS) none today ✗
App container runtime traffic (HTTPS to external API) none today ✗

The HTTPS-only inner-build gap requires per-image CA-trust bootstrap, which is out of scope for the current minimal-invasive design.

Compose Wrapper Auto-Detection 🪄

The runner's compose wrapper at roles/sys-svc-compose/files/python/compose.py auto-detects compose files when invoked from per-app directories under /opt/compose/<app>/:

File When
compose.yml always
compose.override.yml when present (role provides it)
compose.ca.override.yml when present (TLS self-signed CA-inject runs)
compose.cache.override.yml generated on the fly when INFINITO_CACHE_PACKAGE_FRONTEND_IP is set; emits build.extra_hosts for every service that has a build: key

pull.py delegates to the same wrapper so pull and build --pull operations see the identical -f set.

Environment Variables 🌳

The host-side env-vars (INFINITO_CACHE_REGISTRY_*, INFINITO_CACHE_PACKAGE_*) are declared in default.env and the dynamic ones (cache sizes from df//proc/meminfo, sha256 admin password) are computed by utils/env/builder.py. make dotenv writes them into .env, which Compose auto-loads and which BASH_ENV makes available to every Makefile recipe through scripts/meta/env/load.sh. See variables.md.

Per-variable defaults and purposes are in compose.yml.md under the cache section.

INFINITO_CACHE_PACKAGE_MAX_AGE_MIN (default 8640 = 6 days) controls how long every Nexus proxy repo holds an upstream response before revalidating. The bootstrap helper applies it as contentMaxAge, metadataMaxAge, and negativeCache.timeToLive. The default is intentionally one day below the 7-day Valid-Until window Debian and Ubuntu generate in their apt Release files, so Nexus always refreshes the manifest before apt-get update would see it expired. Raise it for slow-moving upstreams when staleness is preferable to upstream load; lower it further for fast-moving feeds (apt-debian-security, branch-pinned tarballs) when staleness becomes visible.

Operations 🛠️

Action Command
Start the stack with caches make compose-up
Stop the stack make compose-down
Wipe local cache state make clean-cache
Manually re-bootstrap Nexus repos bash scripts/docker/cache/package.sh (after make dotenv or sourcing scripts/meta/env/load.sh)
Manually regenerate frontend certs bash scripts/docker/cache/package-frontend/certs.sh
Reload nginx in the frontend docker exec infinito-package-cache-frontend nginx -s reload
Inspect cache hits docker logs -f infinito-package-cache and docker logs -f infinito-package-cache-frontend

Cache state persists under /var/cache/infinito/core/cache/. Paths are configurable via the env scripts above.

Adding a New Upstream ➕

When a new package manager or upstream needs caching:

  1. Register a Nexus proxy repo in package.sh.
  2. If the upstream uses HTTPS, add it to HOSTNAMES in package-frontend-certs.sh so a leaf cert is issued.
  3. Add a server-block in upstreams.conf that reverse-proxies onto the new Nexus repo path (rewrite if upstream URL prefix differs from the Nexus repo path).
  4. Add an extra_hosts entry on the infinito service in compose/cache.override.yml for runner-side traffic.
  5. If the upstream is HTTP-only and inner-dockerd builds need it, also add it to _CACHE_HTTP_HOSTNAMES in compose.py so the per-app compose.cache.override.yml includes it in build.extra_hosts.

Background 📚