prole/deploy/gcp/gke/knoe-db.yaml

apiVersion: postgresql.cnpg.io/v1
kind: Cluster
metadata:
  name: knoe-db
  namespace: knoe-db-0
spec:
  instances: 3
  enablePDB: false
  # Image pulled from GCP Artifact Registry — set ARTIFACT_REGISTRY in conf/prod/gcp.cfg
  # e.g. us-central1-docker.pkg.dev/<project>/knoe-system/knoe-db:<pg-release-tag>
  imageName: "${ARTIFACT_REGISTRY}/knoe-db:${KNOE_DB_IMAGE_TAG}"
  postgresUID: 100
  postgresGID: 101
  maxSyncReplicas: 1

  affinity:
    enablePodAntiAffinity: true
    # Keep spread as a preference for small dedicated Standard DB clusters so 3 pods can still
    # schedule while nodes reconcile; strict topology can be enforced in later rollout.
    podAntiAffinityType: preferred
    topologyKey: kubernetes.io/hostname        # physical node boundary (not zone)
    tolerations:
      # Allow scheduling on GKE Spot nodes when explicitly enabled for this DB cluster.
      # Without this toleration the cluster-autoscaler predicate simulation fails
      # for any MIG whose nodes carry the spot taint, blocking scale-up entirely.
      - key: "cloud.google.com/gke-spot"
        operator: "Equal"
        value: "true"
        effect: "NoSchedule"
    # nodeSelector removed: knoe-cnpg-0 is a dedicated DB cluster — all nodes are
    # available to CNPG. A workload label selector here causes scheduling failures
    # when CNPG v1.28 translates it into requiredDuringScheduling nodeAffinity.

  postgresql:
    parameters:
      shared_buffers: 64MB   # ~25% of 256Mi request; restore to 128MB when resources increase
      pg_stat_statements.max: '10000'
      pg_stat_statements.track: all
    shared_preload_libraries:
      - pg_stat_statements
      - pg_tde
    pg_hba:
      # Local Unix-socket connections (CNPG default + knoe role)
      - local all postgres trust
      - local all knoe scram-sha-256

      # postgres / knoe-db / knoe roles: cluster-internal (RFC1918) only.
      # Cluster pod CIDRs: db cluster 10.24.0.0/14, app cluster 10.84.0.0/14;
      # node subnet 10.180.0.0/16. 10.0.0.0/8 covers all of those.
      - host all postgres 10.0.0.0/8 scram-sha-256
      - host knoe knoe-db 10.0.0.0/8 scram-sha-256
      - hostssl knoe knoe-db 10.0.0.0/8 scram-sha-256

      # PHASE 1 EXTERNAL ACCESS — any member of `knoe_developer`, over TLS+SCRAM.
      # `+rolename` in pg_hba matches role membership (not just literal name),
      # so `etc/onboard_engineer.sh` adds new engineers via `GRANT knoe_developer
      # TO <user>` without ever editing pg_hba — that's the reusable property.
      # Phase 2 (queued for Junie) replaces this with libpq OAUTHBEARER:
      #   hostssl all all 0.0.0.0/0 oauth issuer=https://accounts.google.com validator=knoe_oauth scope="openid email"
      - hostssl all +knoe_developer all scram-sha-256

      # Internal cluster (RFC1918) — all roles, SCRAM (allows the supabase
      # services in app cluster knoe-dev-0 to reach the DB cluster).
      - host all all 10.0.0.0/8 scram-sha-256
      - hostssl all all 10.0.0.0/8 scram-sha-256

      # Block any plaintext from external (TLS required for the public LB)
      - hostnossl all all 0.0.0.0/0 reject

      # Catch-all reject for anything not matched above
      - host all all 0.0.0.0/0 reject
      - hostssl all all 0.0.0.0/0 reject

  bootstrap:
    initdb:
      database: knoe-db
      owner: knoe
      localeCollate: 'en_US.utf8'
      localeCType: 'en_US.utf8'
      secret:
        name: knoe-db-user
      postInitTemplateSQL:
        # Supabase convention: relocatable extensions live in `extensions`, not
        # `public`. Studio's Database Advisor flags `public.pg_stat_statements`
        # as a Security warning the moment a user opens the dashboard. Without
        # an explicit SCHEMA clause `CREATE EXTENSION` lands the relocatable
        # extension in the first writable schema in the connecting role's
        # search_path, which for `postgres` is `public`.
        - CREATE SCHEMA IF NOT EXISTS extensions;
        - CREATE EXTENSION IF NOT EXISTS pg_stat_statements SCHEMA extensions;
      postInitSQL:
        - DO $do$ BEGIN IF NOT EXISTS (SELECT FROM pg_roles WHERE rolname = 'knoe') THEN CREATE ROLE knoe LOGIN NOSUPERUSER NOCREATEDB NOCREATEROLE INHERIT; END IF; END $do$;
        - DO $do$ DECLARE owner_password text; BEGIN SELECT rolpassword INTO owner_password FROM pg_authid WHERE rolname = 'knoe'; IF owner_password IS NOT NULL THEN EXECUTE format('ALTER ROLE knoe PASSWORD %L', owner_password); END IF; END $do$;
        - DO $do$ BEGIN IF NOT EXISTS (SELECT FROM pg_roles WHERE rolname = 'knoe_catalog_executor') THEN CREATE ROLE knoe_catalog_executor NOLOGIN; END IF; END $do$;
        - CREATE SCHEMA IF NOT EXISTS knoe AUTHORIZATION knoe;
        - ALTER SCHEMA knoe OWNER TO knoe;
        - REVOKE ALL ON SCHEMA knoe FROM PUBLIC;
        - ALTER ROLE knoe SET search_path TO knoe, public;
        - CREATE EXTENSION IF NOT EXISTS pg_tde SCHEMA knoe;
        - CREATE EXTENSION IF NOT EXISTS pgcrypto SCHEMA knoe;
        - CREATE EXTENSION IF NOT EXISTS postgis SCHEMA knoe;
        - CREATE EXTENSION IF NOT EXISTS postgis_topology;
        - ALTER SCHEMA topology OWNER TO knoe;
        - CREATE EXTENSION IF NOT EXISTS vector SCHEMA knoe;
        - CREATE EXTENSION IF NOT EXISTS tds_fdw SCHEMA knoe;
        - GRANT USAGE ON SCHEMA knoe TO knoe;
        - GRANT EXECUTE ON ALL FUNCTIONS IN SCHEMA knoe TO knoe;
        - GRANT USAGE ON SCHEMA knoe TO knoe_catalog_executor;
        - GRANT EXECUTE ON ALL FUNCTIONS IN SCHEMA knoe TO knoe_catalog_executor;
        - ALTER DEFAULT PRIVILEGES FOR ROLE knoe IN SCHEMA knoe GRANT EXECUTE ON FUNCTIONS TO knoe_catalog_executor;
        - GRANT EXECUTE ON ALL FUNCTIONS IN SCHEMA topology TO knoe;
        - CREATE SCHEMA IF NOT EXISTS storage;
        - CREATE SCHEMA IF NOT EXISTS graphql_public;
        - DO $do$ BEGIN IF NOT EXISTS (SELECT FROM pg_roles WHERE rolname = 'anon') THEN CREATE ROLE anon NOLOGIN; END IF; END $do$;
        - DO $do$ BEGIN IF NOT EXISTS (SELECT FROM pg_roles WHERE rolname = 'authenticator') THEN CREATE ROLE authenticator LOGIN; END IF; END $do$;
        - GRANT USAGE ON SCHEMA public TO anon;
        - GRANT USAGE ON SCHEMA storage TO anon;
        - GRANT USAGE ON SCHEMA graphql_public TO anon;
        - GRANT anon TO authenticator;
        # demo schema for guest read-only access (evolves over time)
        - CREATE SCHEMA IF NOT EXISTS demo;
        - DO $do$ BEGIN IF NOT EXISTS (SELECT FROM pg_roles WHERE rolname = 'guest') THEN CREATE ROLE guest NOLOGIN; END IF; END $do$;
        - GRANT USAGE ON SCHEMA demo TO guest;
        - ALTER DEFAULT PRIVILEGES IN SCHEMA demo GRANT SELECT ON TABLES TO guest;
        # knoe.user — identity registry (Knoey Users)
        - CREATE TABLE IF NOT EXISTS knoe.user (id SERIAL PRIMARY KEY, username TEXT NOT NULL UNIQUE, realm TEXT NOT NULL DEFAULT 'PROLE.LOCAL', email TEXT, display_name TEXT, tenant_realm TEXT, is_realm_admin BOOLEAN DEFAULT false, created_at TIMESTAMPTZ DEFAULT now(), updated_at TIMESTAMPTZ DEFAULT now());
        - CREATE TABLE IF NOT EXISTS knoe.user_role (user_id INT NOT NULL REFERENCES knoe.user(id) ON DELETE CASCADE, role TEXT NOT NULL, granted_at TIMESTAMPTZ DEFAULT now(), PRIMARY KEY (user_id, role));
        - GRANT SELECT, INSERT, UPDATE ON knoe.user TO knoe;
        - GRANT SELECT, INSERT, UPDATE ON knoe.user_role TO knoe;
        - GRANT USAGE, SELECT ON SEQUENCE knoe.user_id_seq TO knoe;

  managed:
    roles:
      - name: admin
        ensure: present
        login: true
        superuser: true
        comment: "Admin principal — full cluster database access"
      - name: guest
        ensure: present
        login: true
        superuser: false
        comment: "Guest principal — read-only access to demo schema"
      - name: developer
        ensure: present
        login: false
        superuser: false
        comment: "Developer group role — granted to knoe-system user accounts"

  resources:
    requests:
      cpu: "100m"
      # 512Mi (was 128Mi) — postgres baseline working set is ~290Mi on the
      # primary (shared_buffers + wal_buffers + per-backend memory + a small
      # OS page cache visible to cgroups), so 128Mi caused the cnpg-grafana
      # "Resource Pressure" tile to flag Memory at working_set / request ≈ 2x,
      # which the dashboard maps to a red "Data Loss" label (>0.98 ratio).
      # Right-sizing to 512Mi puts the steady-state ratio in the green
      # "Healthy" zone (<0.8) and gives the scheduler an accurate signal for
      # spreading replicas across nodes. Pods still have plenty of headroom:
      # 2Gi limit is unchanged.
      memory: "512Mi"
    limits:
      cpu: "500m"
      # 2Gi (was 512Mi) — barman-cloud-backup is single-threaded gzip + GCS
      # upload buffering and the throughput tops out at the memory ceiling.
      # 2Gi cuts a 9 GB DB backup from 30–90 min down to 5–10 min.
      memory: "2Gi"

  enableSuperuserAccess: true

  # CNPG-issued server cert is auto-rotated by the operator. Listing
  # pg.0.knoe.dev as an alt DNS name lets engineers connect with
  # `sslmode=verify-full` after fetching the CNPG-issued CA cert from the
  # `knoe-db-ca` Secret. Phase 1 of the per-engineer psql access plan; replaced
  # by libpq OAUTHBEARER + Let's Encrypt in Phase 2.
  certificates:
    serverAltDNSNames:
      - pg.0.knoe.dev

  storage:
    size: 50Gi
    pvcTemplate:
      accessModes:
        - ReadWriteOnce
      resources:
        requests:
          storage: 50Gi
      storageClassName: premium-rwo   # pd-ssd; 3×50Gi PGDATA + 3×50Gi WAL = 300Gi total (fits 300GB SSD quota)

  walStorage:
    size: 50Gi
    pvcTemplate:
      accessModes:
        - ReadWriteOnce
      resources:
        requests:
          storage: 50Gi
      storageClassName: premium-rwo   # pd-ssd; restore to premium-rwo after quota increase (matches PGDATA above)

  monitoring:
    # NOTE: enablePodMonitor + podMonitorRelabelings are CNPG-deprecated and
    # will be removed in a future release. Migration path: manually manage
    # the PodMonitor resource. Tracked in docs/TODO.md as a follow-up.
    enablePodMonitor: true   # kube-prometheus-stack is installed before CNPG → PodMonitor CRD present
    # Add a `cluster` label to every scraped sample, sourced from the
    # cnpg.io/cluster pod label. The cnpg-grafana dashboards filter every
    # panel by `cluster="$cluster"`, so without this relabel only 2 of 85
    # CNPG metrics (the operator-collector ones) render.
    podMonitorRelabelings:
      - sourceLabels: ["__meta_kubernetes_pod_label_cnpg_io_cluster"]
        targetLabel: cluster
        action: replace