Rework the application build/deploy pipeline for scalability, reproducibility,
and security:
- Build queue: deploys run through a bounded-concurrency Bull queue
(BUILD_CONCURRENCY, default 3) so concurrent user deploys can't flood the
cluster with Kaniko jobs. Build state (progress / cancel / session) moves from
in-memory Maps to Redis, so cancel + live logs work across backend replicas.
- Nixpacks + BYO Dockerfile: code runtimes build via Nixpacks (or the user's own
Dockerfile when present); the hand-written per-runtime Dockerfile generators
and runtime auto-detection are removed. WordPress keeps its templated path.
Build-time mirror env (NIXPACKS_BUILD_ENV) supports the Iran network.
- Source upload to MinIO: archives stream to in-cluster MinIO; build pods pull
via a presigned URL. Removes the PVC + helper pod + kubectl cp upload path.
- Report-only Trivy scan after build; per-severity summary stored on the
deployment and shown as a badge in the dashboard. Never gates a deploy.
- Registry GC: a Redis-locked daily job keeps the newest N image tags per app
(REGISTRY_KEEP_VERSIONS, default 3) and reclaims disk via garbage-collect.
- Hardening: git tokens are delivered via a per-build Secret + git credential
store instead of being embedded in the clone URL / Job manifest; build timeout
is configurable.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ingress / preview URLs:
- Default the app Ingress class and ACME HTTP-01 solver to Traefik
(k3s default) via a new INGRESS_CLASS env, instead of hardcoding nginx —
fixes 404s on clusters without ingress-nginx.
- Only put public, real-TLD hosts (custom domain + preview) in the TLS
block; the internal *.apps.cloudhost.local host no longer poisons the
Let's Encrypt order, so certs actually issue.
- Make the per-app preview number stable across redeploys so URLs stop
breaking, and let PREVIEW_BASE_DOMAIN configure the base domain.
Registry pulls:
- Point the k3s registries.yaml mirror endpoint at the registry NodePort on
loopback so node containerd never depends on cluster DNS (image pulls
survive node restarts).
Builds:
- Pin the Kaniko image, use IfNotPresent pull policy, drop the dead build
queue/processor, and retry transient Kubernetes API errors while polling
build jobs.
Logs & apps list:
- fluent-bit reads log files from head so startup output reaches
Elasticsearch.
- Order joined deployments newest-first so the apps list shows the latest
deployment status.
Allocation:
- Reserve in-flight (pending/building) capacity and stop globally degrading
the cluster on a single allocation failure, so concurrent deploys don't
starve or wrongly report "no healthy cluster".
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
keyhan