# Internal package + image cache (`PKG_CACHE`)

Opt-in build support for the LAN package cache **CT130 `pkg-cache` = `192.168.1.40`**
(Proxmox node `pve`, storage `shared-storage-tn`). Without `PKG_CACHE`, builds use
the upstream repositories exactly as before.

Why it exists: during the 2026-10-21 PHP security release, builds failed because
Remi mirrors lagged hours behind origin, Remi's origin reset or timed out HTTPS
under load, and an EPEL mirror had a broken certificate.

## What the cache serves

| URL on the cache | Upstream tiers, tried in order (stale copy last) |
|---|---|
| `http://192.168.1.40/remi/` | `https://rpms.remirepo.net`, then `http://rpms.remirepo.net` (origin only, no mirrors) |
| `http://192.168.1.40/epel/` | `https://dl.fedoraproject.org/pub/epel`, `https://mirror.math.princeton.edu/pub/epel`, `https://mirrors.kernel.org/fedora-epel` |
| `http://192.168.1.40/almalinux/` | `https://repo.almalinux.org/almalinux`, `https://mirror.math.princeton.edu/pub/almalinux` |
| `http://192.168.1.40/ubuntu/` | `https://archive.ubuntu.com/ubuntu`, `https://us.archive.ubuntu.com/ubuntu` |
| `http://192.168.1.40/ubuntu-security/` | `https://security.ubuntu.com/ubuntu`, `https://archive.ubuntu.com/ubuntu` |
| `http://192.168.1.40/litespeed/` | `https://rpms.litespeedtech.com`, then `http://rpms.litespeedtech.com` |
| `192.168.1.40:5000` (plain http) | Docker Hub pull-through cache (`registry-1.docker.io`) |

Cache rules:
- **Metadata** has a 3-minute TTL. After that it is revalidated upstream (`If-Modified-Since`). Metadata here means `repomd.xml*`, repodata files whose names don't start with a hash, apt `dists/` except `by-hash/`, `*-latest*.rpm`, `remi-release-N.rpm`, `RPM-GPG-KEY*` and directory listings.
- **Packages** (`*.rpm`, `*.deb`, hash-named repodata, `by-hash/`) are kept for 30 days, and evicted after 60 days without use. The cache is capped at 110 GB (`min_free` 15 GB).
- On a connect error, timeout or 5xx, a request falls through to the next tier. Only the **last** tier answers from a stale copy, so metadata goes primary → fallback → stale. A fresher fallback always wins over stale.
- `proxy_cache_lock` is on: concurrent misses on one file make a single upstream fetch.
- Each response carries `X-Cache-Status` (`HIT`/`MISS`/`REVALIDATED`/`EXPIRED`/`STALE`/`BYPASS`) and `X-Pkgcache-Tier`. On a HIT, the tier header names the location, not the upstream that filled the entry.

## Building with the cache

```bash
docker buildx build --build-arg PKG_CACHE=http://192.168.1.40 --build-arg PHPVER=85 -f Dockerfile .
```

`scripts/pkg-cache-repos.sh` (copied to `/usr/local/sbin/`) is wired into
`Dockerfile`, `Dockerfile.fpm`, `Dockerfile.shared-httpd` and `Dockerfile.shared-ols`:

- `rpm`: rewrites the AlmaLinux, EPEL and Remi `baseurl`s to `<cache>/<prefix>/... <origin>/...`, and comments out that section's mirrorlist/metalink. dnf treats the second URL as a failover, so a dead cache falls back to the origin by itself.
- `apt`: rewrites the Ubuntu archive/security and LiteSpeed URIs to the cache. apt has no failover, so CI must probe the cache first (see below).
- `release-urls`: the EPEL and Remi release RPM URLs, from the cache when it is set.
- `restore`: puts the original repo files back. Every RUN that rewrites ends with `restore`, so **no shipped layer contains the cache address**. Images run on datacenter hosts that cannot reach it.

`Dockerfile.litespeed` and `Dockerfile.lsphp` are **not wired yet**. Their
multi-stage `lsphp-apt-upgrade.sh` flow needs two changes first:
- a `PKG_CACHE` ARG in each stage;
- the madison check accepting the cache URL. It currently requires `*rpms.litespeedtech.com*`, and through the cache `apt-cache madison` shows `http://192.168.1.40/litespeed/debian`.

The PHP gates are unchanged and still apply: `EXPECT_PHP_MIN`,
`EXPECT_LSPHP_PKG`/`EXPECT_PHP_VERSION`.

## Why plain HTTP on the LAN (trade-off)

- The cache fetches upstream over **HTTPS with certificate verification**. Remi and LiteSpeed also have a plain-http origin fallback, which mirrors what the Dockerfile already did.
- Client to cache is plain HTTP. dnf still verifies **package GPG signatures**, and apt verifies the signed `InRelease`/`Release.gpg` plus package hashes. A LAN attacker therefore cannot inject packages. On 2026-10-21 apt measurably refused the cached LiteSpeed index when given the wrong key (`NO_PUBKEY 011AA62DEDA1F085`).
- The residual risk is **metadata replay or freeze**: an attacker on the LAN path, or on the plain-http fallback leg, serves an *old* but validly signed `repomd.xml`. That would let a build install old PHP. Remi/EPEL/AlmaLinux do not enable `repo_gpgcheck`, so dnf does not check metadata signatures. The **php.net version gates** cover this for our images: a build fails if it gets PHP older than php.net's current release.
- WHP **hosts** have no such gate. They should keep the origin as the second `baseurl` and not rely on the cache for freshness on security-release day. Run `pkgcache-refresh` on the cache instead.
- Signing keys must **not** be fetched through the cache. Take them from the base image or directly over HTTPS.
- HTTPS to the cache (internal CA or a Let's Encrypt DNS-01 certificate for `pkg-cache.internal.cloud-hosting.io`) would close the LAN leg. It needs a CA in every build image or a public DNS name plus certificate. Not done; Josh to decide.

## CI integration plan (NOT applied — needs Josh)

1. **Network (blocker).** The ephemeral runners (`eph-pve-*` on `pve` vmbr2 10.99.98.0/24, `eph-pve3-*` on `pve3` vmbr1 10.99.99.0/24) are NATed, and `iptables FORWARD` **drops all RFC1918 destinations**, including 192.168.1.40. A narrow exception is needed on each node, placed before the DROP lines in `/etc/network/interfaces.d/*`:
   ```
   # pve  (/etc/network/interfaces.d/ vmbr2 stanza), before the 192.168.0.0/16 DROP:
   post-up   iptables -A FORWARD -i vmbr2 -o vmbr0 -d 192.168.1.40 -p tcp -m multiport --dports 80,5000 -j ACCEPT
   post-down iptables -D FORWARD -i vmbr2 -o vmbr0 -d 192.168.1.40 -p tcp -m multiport --dports 80,5000 -j ACCEPT
   # live, without ifreload:
   iptables -I FORWARD 1 -i vmbr2 -o vmbr0 -d 192.168.1.40 -p tcp -m multiport --dports 80,5000 -j ACCEPT
   # pve3: same with vmbr1
   ```
   `gitea-runner.internal.cloud-hosting.io` resolves to 10.10.14.20, which is not on this LAN. Whether it can reach 192.168.1.40 is untested.
2. **Probe, then opt in** (no `${{ }}` inside `run:`, per repo rule):
   ```yaml
   - name: Probe package cache
     id: pkgcache
     run: |
       if curl -fsS --max-time 5 -o /dev/null http://192.168.1.40/ ; then
         echo "url=http://192.168.1.40" >> "$GITHUB_OUTPUT"
       else
         echo "url=" >> "$GITHUB_OUTPUT"; echo "::warning::pkg-cache unreachable - building from upstream"
       fi
   # in docker/build-push-action `build-args:` add
   #   PKG_CACHE=${{ steps.pkgcache.outputs.url }}
   ```
3. **Docker Hub mirror for BuildKit.** The workflows use `docker/setup-buildx-action` (docker-container driver). BuildKit **ignores dockerd's `registry-mirrors`**, so configure the builder instead:
   ```yaml
   - uses: docker/setup-buildx-action@v3
     with:
       buildkitd-config-inline: |
         [registry."docker.io"]
           mirrors = ["192.168.1.40:5000"]
         [registry."192.168.1.40:5000"]
           http = true
   ```
   This exact config was used for the test builds on 192.168.1.193. Only Docker Hub images go through it; `repo.anhonesthost.net/...` is unaffected.
4. **dockerd clients** (plain `docker pull`, e.g. the build server or test hosts). Untested, because applying it needs a daemon restart. Add to `/etc/docker/daemon.json` and restart docker:
   ```json
   { "registry-mirrors": ["http://192.168.1.40:5000"], "insecure-registries": ["192.168.1.40:5000"] }
   ```

## WHP hosts (plan only)

The production hosts (whp01/whp02/sdbees) are in the datacenter and **cannot reach 192.168.1.40**. Only LAN hosts such as 192.168.1.148 can. Options for Josh:
- (a) a second identical cache CT in the datacenter;
- (b) expose this one over NetBird;
- (c) LAN hosts only.

For a host that can reach a cache, use the same idea as `pkg-cache-repos.sh rpm`, keeping the origin second so the cache is never a single point of failure:

```ini
# /etc/yum.repos.d/remi-safe.repo [remi-safe]   (AL9: enterprise/9)
baseurl=http://<cache>/remi/enterprise/$releasever_major/safe/$basearch/ https://rpms.remirepo.net/enterprise/$releasever_major/safe/$basearch/
#mirrorlist=...
```

## Operating the cache (CT130)

- Config: `/etc/nginx/conf.d/pkgcache-http.conf` (cache zone, maps, log format) and `/etc/nginx/pkgcache/{server-common,meta,pkg}.conf`. `/etc/nginx/sites-available/pkgcache` is **generated** by `/usr/local/sbin/pkgcache-gen-nginx`: edit its `UPSTREAMS` table, then run `pkgcache-gen-nginx > /etc/nginx/sites-available/pkgcache && nginx -t && systemctl reload nginx`.
- Registry: `/etc/docker/registry/config.yml` (Debian `docker-registry` 2.8.3, the same code as `registry:2`), service `docker-registry`.
- Data: nginx in `/srv/cache/nginx`, registry in `/srv/cache/registry` (150 GB mount point `mp0`, `backup=0`).
- **Security-release day:** run `pkgcache-refresh` to re-fetch every cached metadata entry now (or `pkgcache-refresh remi` for one prefix; `--list` shows them; `--purge` deletes them instead). A refresh bypasses the cache from localhost only and replaces each entry in place. If upstream is down, the old copy stays. Without it, metadata is at most 3 minutes old anyway.
- Monitoring: `pkgcache-stats` shows disk, hit ratio, fallback-tier use, STALE count and upstream errors. Logs are in `/var/log/nginx/pkgcache.access.log` (with `cache=`, `tier=`, `up=` fields) and `journalctl -u docker-registry`. Suggested Uptime Kuma checks: `http://192.168.1.40/` and `http://192.168.1.40:5000/v2/`.
- Docker Hub: anonymous pulls are rate-limited per egress IP, and every client now shares this cache's IP. To raise the limit, add `username`/`password` (a Hub access token) under `proxy:`.