Changelog

v2.0.0 (2026-11-19)

Cairn 2.0 is the first release under the Cairn name and the Cairn Source-Available License. It brings high availability across servers, full observability, project deletion, and the public site at cairnpaas.dev.

Breaking changes

  • License. Cairn is now distributed under the Cairn Source-Available License (see LICENSE). Running it needs a license key; the free Personal key covers personal use (up to 3 servers per install, 2 installs), and companies may evaluate it for 60 days. Earlier releases had no license file. See cairnpaas.dev/pricing and the Terms of Service.
  • Install diagnostics. Release builds of platformctl send anonymous install diagnostics after provision, upgrade and join; set PLATFORM_DIAGNOSTICS=off to stop them. What is sent is listed in the Privacy Policy.
  • Project delete really deletes. DELETE /api/projects/:id now backs up every database, removes the project's apps, namespaces, monitoring, tokens and data, and frees the name. It needs the devops role, project owner access, a recent MFA step-up and { "confirmName": "<project name>" }, and returns 202 with an operation id; databases are dropped only with deleteDatabases: true. See api-reference/platform-api/projects.md.
  • Auth errors. When credentials cannot be checked (for example while the database fails over), the API answers 503 with Retry-After instead of 401. Clients should retry on 503 rather than sign the user out.
  • nginx images move to the nginx/nginx-unprivileged repository (1.30.7); apps built from the Cairn templates pick it up on their next build.

Also in this release

  • Pre-delete database backups go to the off-site backup target when one is configured; an aborted delete returns the project within seconds.
  • Failover hardening found by forced server stops on a 3-server cluster: no sign-outs during database failover, k3s no longer re-imports unchanged image archives on restart (recovery in seconds instead of minutes), stuck Helm locks are cleared before every release, HA MinIO drives are sized from the disk, joining servers wait for the API before hardening.
  • Node SDK demo pins OpenTelemetry 1.x and CI checks that traces reach a collector.
  • Public site: feature clips, pricing, About, Terms of Service, Privacy Policy, structured data and llms.txt facts for search and AI assistants.
  • SECURITY.md: report vulnerabilities privately through GitHub advisories or [email protected].

The sections below list everything since v1.0.0 in detail.

Phase 4: observability, session replay, heatmaps, status pages

New

  • High availability (ha profile): multi-node k3s with embedded etcd, platformctl join --role server|agent, join token create/rotate/install, cluster firewall limited to PLATFORM_CLUSTER_NODES, etcd snapshots to S3 (#53); Longhorn replicated storage as the default StorageClass with S3 backups and volume alerts (#50); ingress HA with a kube-vip ARP VIP (INGRESS_HA_MODE=vip) or external LB/DNS failover (external), ingress-nginx DaemonSet with client-IP preservation, optional API_VIP for joins (#54). join prepares Longhorn prerequisites on every node and the storage skip file and ingress HA k3s fragment on servers; MongoDB replica set, Redis Sentinel (API and SDKs discover the master through Sentinel) and distributed MinIO on a longhorn-minio StorageClass whose replica count follows the node count (#52); PostgreSQL HA with CloudNativePG, continuous WAL archiving, daily base backups and point-in-time recovery (#51). See operations/ha.md.
  • Session replay (opt-in, consent-based) for the React and Angular SDKs: rrweb recording with input masking, unload delivery via sendBeacon / keepalive, portal player with error/navigation/network markers, audited access (replay.view, replay.chunk_view, replay.snapshot_view); agent tokens need the replays:read scope. See guides/frontend-rum.md.
  • Heatmaps: per-page click and scroll-depth aggregates with the newest replay snapshot of each page as the background; buffered, exact aggregation. See guides/heatmaps.md.
  • Status pages: public /status/<slug> pages fed by synthetic checks, incidents and scheduled maintenance, JSON/Atom/RSS feeds, light/dark themes. See guides/status-pages.md.
  • Observability views: fleet and per-project views (RED, services, logs, traces, RUM, sessions, synthetics, infrastructure, profiles), every query forced to the project scope. See guides/observability.md.
  • Database insights: PostgreSQL/MongoDB/Redis exporters on the full profile and a per-project Database tab with normalised top queries and slow operations. See operations/database-monitoring.md.
  • Dashboard shows live data instead of placeholders.
  • GDPR delete also removes error-tracking events, replay objects/sessions and submits Loki delete requests. See operations/telemetry-privacy.md.

Fixed (found by full-profile VM testing)

  • ha PostgreSQL failover after a server loss: on the 3-server power-off test, writes resumed 5m16s after the power-off with the bundled MinIO as the WAL archive, and 107 s with an external one. Two causes, both fixed: (1) the CloudNativePG operator does not reconcile, or fail over, a Cluster whose Barman Cloud plugin it cannot reach, and the plugin ran one replica, so failover waited until Kubernetes replaced a plugin pod on the failed server; the plugin now runs two spread replicas (a leader and a hot standby), and install longhorn and install postgresql spread both the operator and the plugin when all replicas share one node. (2) PostgreSQL acts on a promote request only between two WAL fetches from the archive, and a fetch from an object store that hangs (the bundled MinIO after losing quorum) waited 60 s, five times; WAL fetches now give up after 5 s and use standard retries (about 19 s at most). Measured after the fix: a new primary after 57.5 s and writes again after 61 s, with the primary, the active plugin and half of MinIO on the powered-off server. The client side was not involved: kube-proxy on every server watches its local API server, and the client node's rules followed the new endpoint within a second. Upgrading restarts each PostgreSQL instance once (the plugin sidecar's environment changes), with a switchover for the primary.
  • Offline Portainer: the Portainer chart pulls its image on every start (pullPolicy: Always), so on a server without internet Portainer stayed in ImagePullBackOff although the bundle had imported the pinned image, and provision --bundle failed at the portainer step. It now uses IfNotPresent.
  • Offline bundle image names: k3s marks the node Ready before it finishes importing the bundle's large image archive, so provision --bundle / join --bundle named the digest-pinned images too early and reported the ingress-nginx controller and admission images as "not in the bundle" (ImagePullBackOff offline). They now wait until k3s reports every staged archive imported (up to 20 minutes).
  • ha first server: provision on the first server of an ha cluster, before the others join, no longer waits 10 minutes per datastore for members that cannot start yet (their Longhorn volumes do not fit on one node, PostgreSQL standbys need a node each). With fewer than 3 Ready nodes it waits for the first MongoDB, Redis and PostgreSQL member and says that the rest start when the remaining servers join.
  • MinIO sign-in with the internal CA: when the domain resolves to a private address or Let's Encrypt cannot issue, MinIO could not fetch the platform's OpenID signing keys (x509: unknown authority), so its IAM never started and the storage-scoped and telemetry steps timed out. platformctl now publishes the internal CA certificate (never the key) as the ConfigMap storage/platform-ca, which the MinIO server and console mount through their chart values, before MinIO is installed and whenever the TLS step falls back to the internal CA; platformctl repair brings older installs up to date.
  • ha upgrades and kube-vip: platformctl upgrade and platformctl update refresh the local server's kube-vip manifest to the installed release; run sudo platformctl update on every other server after an upgrade (on joined servers it only refreshes local files). See operations/ha.md.
  • Offline bundle (#12, #179): provision --bundle no longer downloads Helm or apt packages: the release bundle carries the Helm binary and provision installs its prerequisites from the bundle's packages, as join --bundle does. The bundle also carries the release's Cairn images (API, portal, profile ingest, synthetics runner), which provision uses when PLATFORM_IMAGE_TAG is not set. Digest-pinned images (ingress-nginx controller and admission webhook) resolve on offline nodes, which failed with ImagePullBackOff. See getting-started/platformctl.md.
  • Browser RUM was blocked by CORS on every install: the API now publishes the Faro allow-list (monitoring/telemetry-origins).
  • Collector ingress returns 429/503 with Retry-After for throttling (was 500); ingress-nginx preserves real client IPs (externalTrafficPolicy: Local), so per-IP limits are per client again.
  • Upgrades from v1.2.0: loki-stack is migrated instead of failing; schema drift after a rollback is repaired; PLATFORM_IMAGE_TAG is persisted; new components and changed database values are reconciled by platformctl upgrade; project Redis users are re-created after a Redis restart. See operations/upgrades.md.
  • Alerting: bounded SMTP sends and concurrent delivery (one unreachable mail server no longer stalls every channel); SMTP secrets encrypted at rest; recovery titles on resolved notifications.
  • Retention that actually applies: error events, per-project logs (Loki runtime overrides), replays, heatmaps, alert deliveries and resolved incidents.
  • Browser synthetic checks: every navigation hop re-validated, dedicated runner API listener, missing runner reports become failed results, project env hosts reachable with the platform CA.
  • MongoDB standalone uses the Recreate strategy (rolling updates deadlocked on mongod.lock).
  • Portal: responsive layout (mobile drawer, tablet), readable Observability, checkbox/disabled-button styling, Errors page restyle, all docs pages available in the in-portal docs viewer.
  • Go toolchain 1.26.11 (GO-2026-6859).
  • Alerting per project (#151): project owners and devops configure their own project's channels, escalation policies, on-call schedules, silences, availability/latency SLOs and trace/resource monitors; raw PromQL/LogQL monitors and ratio SLOs stay with platform admin/devops. See operations/alerting.md.
  • Synthetics against the platform's own domain (#152): HTTP and browser checks of DOMAIN / *.DOMAIN no longer fail as private addresses; they are reached only through the ingress controller, trusting the platform CA.
  • NAT-friendly limits (#153): per-IP ingress limits are configurable (INGRESS_IP_RATE_LIMIT, INGRESS_IP_BURST, INGRESS_IP_CONNECTIONS; defaults 1000 r/s, burst 2000, 2000 connections) and the heatmap ingest limits are raised, so many visitors behind one office or carrier address are not throttled.
  • Replay on Firefox (#154): the page-hide upload is sent as a CORS simple request (sendBeacon, keepalive fallback), so the last chunk of short sessions is no longer lost.
  • Git branches (#155): the project page looks up branches through the API (authenticated, cached) and skips the lookup when the project has no repository.
  • Error grouping (#156): the same browser error is grouped into one issue across engines and builds (normalised v2 fingerprint; existing issues keep receiving events).
  • Firewall status (#157): platformctl marks harden-firewall done after hardening instead of reporting it as in progress.

End-to-end VM suite and release gate

  • Release gate: every tag is now published as a GitHub pre-release. A stable vX.Y.Z becomes the Latest release (what install.sh and releases/latest install, and the GHCR platformctl:latest copy) only after the e2e VM suite passes on the minimal and full profiles. Until then stable installs keep the previous release. See operations/e2e-ci.md.
  • e2e CI: nightly and on every tag, a fresh Ubuntu 24.04 EC2 VM is provisioned and checked (TLS, API, SDK, per-engine DB credentials, SSO to Grafana/Argo CD/Portainer, Argo CD sync, backup + restore test, tenant isolation, security probes), then destroyed. An hourly janitor removes leftovers.
  • platformctl provision also writes the step timings as JSON to /var/log/platformctl/timings.json.

Faster platformctl provision

  • Parallel install: cluster components install as a dependency graph, up to 4 at a time (--parallel N, PLATFORMCTL_PARALLEL; --parallel 1 keeps the old serial, live-output behaviour). Each step logs to /var/log/platformctl/<step>.log.
  • Step timings: provision prints a per-step timing table and saves it to /var/log/platformctl/timings.txt.
  • Image pre-pull: every image the install needs is pulled in the background right after k3s starts.
  • Pinned versions: k3s, Helm and all charts are pinned in platformctl/internal/versions; chart tarballs are prefetched while k3s installs. No more helm repo update; Bitnami charts come from OCI.
  • Docker CE is no longer installed on the server — k3s uses its own containerd. Private GHCR pulls now use GITHUB_TOKEN via /etc/rancher/k3s/registries.yaml.
  • Removed duplicate Helm upgrades (ingress-nginx, kube-prometheus-stack) and the ArgoCD reset on fresh installs.
  • Fix: Infisical was never actually installed (dead chart repo); it is now applied from its manifest.
  • New --profile minimal|standard|full and --bundle <path|auto> (offline bundle release asset).

v1.0.0 (2026-08-18)

First git-tagged 1.0.0 Platform release (prior production tags were v0.3.x).

  • MCP server (platform/mcp-server) with agent tokens (plat_agent_*), command guard, and human approval for destructive commands
  • User + agent MCP documentation (overview, setup, tools/policy, for-agents) and homepage curl | bash install
  • platformctl commands: provision, install, status, seed, update, recover, backup, version
  • Safer secret rotate / recover: passwords fsynced to /etc/platform before ALTER USER; JSON Patch rollout restart for platform-api
  • GitHub / GitLab login, profile, notifications, one-click rotate UI

The Caps-era notes below were documentation versions, not these git tags.

Caps docs v2.0.0 (2026-07-01)

Breaking Changes

  • Rebranded all packages from @caps/* to @mpratyush54/*
  • Renamed CapsClient → PlatformClient across all SDKs
  • Renamed CAPS_* env vars to PLATFORM_*
  • Renamed caps_sdk/ → platform_sdk/ in Python SDK

Features

  • OAuth2 / OpenID Connect support with RS256-signed tokens
  • Custom RBAC roles with permission presets (admin, devops, tech_lead, developer, viewer)
  • Preview environments with 72h auto-cleanup
  • Infisical-powered secrets management with AES-256-GCM encryption
  • Multi-SDK support (Node.js, Python, React, Angular)
  • Grafana dashboards + Loki log aggregation
  • Automated database backups to MinIO
  • GitLab/GitHub webhook integration
  • ClickUp task linking

Fixed

  • Angular @ template escaping (use &#64; instead of @)
  • ArgoCD subpath routing (--rootpath=/argocd)
  • Grafana subpath redirect (GF_SERVER_ROOT_URL fix)
  • MinIO PVC initialization ordering
  • DNS IPv6 timeouts (systemd-resolved config)
  • cert-manager staging/production issuer workflow

Caps docs v1.0.0 (2026-06-01)

  • Initial Caps-branded documentation snapshot (legacy; not git tag v1.0.0)