docker-compose.prod.yml

# AiSOC — production compose.
#
#     docker compose -f docker-compose.yml -f docker-compose.prod.yml up -d
#
# This is the file `apps/docs/docs/deployment/docker.md` has pointed at since
# the deployment page was written. It did not exist, which is what discussion
# #629 reported: the documented production command failed on a missing file,
# and the only stack an operator could actually start was the development one,
# where `ENVIRONMENT` defaults to `development` and
# `services/api/app/api/v1/dev_auth.py` resolves an unauthenticated request to
# a demo user whose role is `admin`.
#
# It is an overlay over `docker-compose.yml`, so there is exactly one
# definition of every service and this file holds only what production changes.
# A standalone copy would have drifted from the base within a release, which is
# the failure this repository keeps finding.
#
# It was written with the Compose Spec `include` directive and a single `-f`.
# That does not work: `include` imports a model, and overriding a service it
# imported is rejected with `services.<name> conflicts with imported resource`.
# `infra/compose/docker-compose.dev.yml` had the same shape and has since
# moved to a two-`-f` overlay for the same reason — it could never have
# carried an override, which is why it carried nothing.
# Measured rather than assumed — `-f docker-compose.prod.yml` alone fails on
# Compose v2.29.7 and v2.39.4 and resolves only on 5.1.3, so the documented
# production command worked on almost no installation. Two `-f` flags are the
# ordinary override form and resolve identically on all three.
#
# WHAT THIS FILE GUARANTEES
#
#   * The authentication bypass is unreachable. `ENVIRONMENT` and `AISOC_DEV_MODE`
#     are fixed values here, not interpolated, so no `.env` can re-enable them.
#   * Nothing starts on a default credential. Every secret uses `${VAR:?...}`,
#     which makes Compose refuse to start and name the variable rather than
#     starting on nothing — and, because `:?` rejects an *empty* variable and
#     cannot look at a value, `preflight-secrets` below runs ahead of every
#     other service and refuses the literals this repository publishes. Until
#     it existed this bullet was false: `${POSTGRES_PASSWORD:?… the
#     development default is a published literal}` accepted `aisoc_dev_secret`,
#     which is what `.env.example` shipped.
#   * No datastore is reachable from the host. Postgres, Redis, Kafka and
#     Qdrant publish no ports at all; the console and the ingest endpoint are
#     the only things bound, and they bind to loopback unless you set
#     AISOC_BIND_ADDR deliberately (put a TLS-terminating proxy in front —
#     see docs/runbooks/HARDENING.md section 2).
#
# WHAT IT DOES NOT DO
#
#   TLS, secret rotation, network policies, audit-log forwarding and backups
#   are not in a compose file. `docs/runbooks/HARDENING.md` is the checklist
#   and this file is the starting posture, not the finished one.


services:
  # ---------------------------------------------------------------------------
  # The value check `${VAR:?}` cannot do.
  #
  # Compose's `:?` operator rejects an unset or empty variable. It cannot look
  # at a value, so `${POSTGRES_PASSWORD:?… the development default is a
  # published literal}` accepted `aisoc_dev_secret` — the exact literal its own
  # message names, shipped in `.env.example` and defaulted to in
  # `docker-compose.yml`. Every guard below had the same hole, so the guarantee
  # in this file's header held for none of them.
  #
  # This runs first and every other service waits for it to exit 0.
  # `scripts/check_published_secrets.py` derives the refused values from the
  # tree — the `${VAR:-literal}` defaults in the compose files, the non-empty
  # values in `.env.example`, and INSECURE_SECRET_KEY_DEFAULTS — so a default
  # added later is refused without anyone remembering to list it here.
  #
  # Every `${VAR:?}` variable in this file is repeated below. That is what
  # gets it into the container's environment, and `--check` fails when a guard
  # exists that the preflight never sees.
  # ---------------------------------------------------------------------------
  preflight-secrets:
    image: ghcr.io/beenuar/aisoc-core-api:${AISOC_VERSION:-latest}
    pull_policy: missing
    container_name: aisoc-preflight-secrets
    restart: "no"
    entrypoint: ["python3", "/repo/scripts/check_published_secrets.py", "--check-env"]
    # Read-only, and the whole tree rather than the four files the harvest
    # happens to read today: a hand-listed mount is another list that stops
    # matching the code, and the failure mode is the check quietly reading
    # less than it thinks it does. `AISOC_REPO_ROOT` is how `gate_toolkit`
    # is pointed at a tree without shelling out to git, which this image
    # does not carry.
    volumes:
      - .:/repo:ro
    environment:
      AISOC_REPO_ROOT: /repo
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:?set POSTGRES_PASSWORD — the development default is a published literal}
      AISOC_APP_DB_PASSWORD: ${AISOC_APP_DB_PASSWORD:?set AISOC_APP_DB_PASSWORD — the development default is a published literal}
      REDIS_PASSWORD: ${REDIS_PASSWORD:?set REDIS_PASSWORD}
      SECRET_KEY: ${SECRET_KEY:?set SECRET_KEY — run `make env` to generate it}
      AISOC_CREDENTIAL_KEY: ${AISOC_CREDENTIAL_KEY:?set AISOC_CREDENTIAL_KEY — run `make env` to generate it}
      AISOC_SERVICE_TOKEN: ${AISOC_SERVICE_TOKEN:?set AISOC_SERVICE_TOKEN — run `make env` to generate it}
      AISOC_ACTIONS_SERVICE_TOKEN: ${AISOC_ACTIONS_SERVICE_TOKEN:?set AISOC_ACTIONS_SERVICE_TOKEN — run `make env` to generate it}
      AISOC_AGENTS_SERVICE_TOKEN: ${AISOC_AGENTS_SERVICE_TOKEN:?set AISOC_AGENTS_SERVICE_TOKEN — run `make env` to generate it}
      AISOC_REALTIME_JWT_SECRET: ${AISOC_REALTIME_JWT_SECRET:?set AISOC_REALTIME_JWT_SECRET — run `make env` to generate it}
      REALTIME_INTERNAL_TOKEN: ${REALTIME_INTERNAL_TOKEN:?set REALTIME_INTERNAL_TOKEN — run `make env` to generate it}
      METRICS_TOKEN: ${METRICS_TOKEN:?set METRICS_TOKEN — guards /metrics outside development; `openssl rand -hex 32`}
      JWT_SECRET: ${JWT_SECRET:?set JWT_SECRET — signs SAML/OIDC sessions}
      NEO4J_PASSWORD: ${NEO4J_PASSWORD:?set NEO4J_PASSWORD}
      CLICKHOUSE_PASSWORD: ${CLICKHOUSE_PASSWORD:?set CLICKHOUSE_PASSWORD}
      GRAFANA_ADMIN_PASSWORD: ${GRAFANA_ADMIN_PASSWORD:?set GRAFANA_ADMIN_PASSWORD — required for the monitoring profile}

  # ---------------------------------------------------------------------------
  # Data plane. No published ports: in the development stack these are bound to
  # 127.0.0.1 so a developer can attach a client, which is reasonable there and
  # is an attack surface here. `ports: []` overrides the base list rather than
  # appending to it.
  # ---------------------------------------------------------------------------
  postgres:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:?set POSTGRES_PASSWORD — the development default is a published literal}
      # The DML-only runtime role the services connect as, so the row-level
      # security policies actually apply to them. Separate from the owner
      # password above and rotated the same way.
      AISOC_APP_DB_PASSWORD: ${AISOC_APP_DB_PASSWORD:?set AISOC_APP_DB_PASSWORD — the development default is a published literal}

  redis:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  zookeeper:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  kafka:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  qdrant:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  # ---------------------------------------------------------------------------
  # Application services. `ENVIRONMENT` and `AISOC_DEV_MODE` are literals on
  # purpose: an interpolated value here would let a stray `.env` put a
  # production deployment back into the environment where the auth bypass
  # lives, which is the defect this file exists to make impossible.
  # ---------------------------------------------------------------------------
  api:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      ENVIRONMENT: production
      AISOC_DEV_MODE: 0
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:?set POSTGRES_PASSWORD}
      AISOC_APP_DB_PASSWORD: ${AISOC_APP_DB_PASSWORD:?set AISOC_APP_DB_PASSWORD}
      DATABASE_URL: postgresql+asyncpg://aisoc_app:${AISOC_APP_DB_PASSWORD:?set AISOC_APP_DB_PASSWORD}@postgres:5432/aisoc
      DATABASE_MIGRATION_URL: postgresql+asyncpg://aisoc:${POSTGRES_PASSWORD:?set POSTGRES_PASSWORD}@postgres:5432/aisoc
      REDIS_URL: redis://:${REDIS_PASSWORD:?set REDIS_PASSWORD}@redis:6379/0
      # `make env` generates all of these. They are required rather than
      # defaulted because a service that starts without one either serves
      # unauthenticated or signs with a key published in this repository —
      # both were reported vulnerabilities, fixed in v12.0.0.
      SECRET_KEY: ${SECRET_KEY:?set SECRET_KEY — run `make env` to generate it}
      AISOC_CREDENTIAL_KEY: ${AISOC_CREDENTIAL_KEY:?set AISOC_CREDENTIAL_KEY — run `make env` to generate it}
      AISOC_SERVICE_TOKEN: ${AISOC_SERVICE_TOKEN:?set AISOC_SERVICE_TOKEN — run `make env` to generate it}
      AISOC_ACTIONS_SERVICE_TOKEN: ${AISOC_ACTIONS_SERVICE_TOKEN:?set AISOC_ACTIONS_SERVICE_TOKEN — run `make env` to generate it}
      # Verified by `alert_writeback.service_token_valid`, which reads this
      # and only this. Unset, every internal context route the agents service
      # calls refuses it, and triage silently runs without organisation
      # memory, tenant skills, dispositions, identity or runbooks.
      AISOC_AGENTS_SERVICE_TOKEN: ${AISOC_AGENTS_SERVICE_TOKEN:?set AISOC_AGENTS_SERVICE_TOKEN — run `make env` to generate it}
      AISOC_REALTIME_JWT_SECRET: ${AISOC_REALTIME_JWT_SECRET:?set AISOC_REALTIME_JWT_SECRET — run `make env` to generate it}
      REALTIME_INTERNAL_TOKEN: ${REALTIME_INTERNAL_TOKEN:?set REALTIME_INTERNAL_TOKEN — run `make env` to generate it}
      # The remaining two secrets `enforce_secure_defaults` hard-fails on, and
      # the reason this stack could not boot at all: `ENVIRONMENT: production`
      # above turns every entry in `warn_if_insecure_defaults` into a refusal,
      # and these two were the only ones no service block passed through. The
      # API crash-looped on
      #
      #   InsecureProductionDefaultsError: Refusing to boot in production with
      #   insecure defaults:
      #     - METRICS_TOKEN is empty in a non-development environment
      #     - JWT_SECRET is empty or set to the well-known placeholder
      #
      # while compose reported the stack up. Declaring them here is what makes
      # the refusal arrive the way the rest of this file's do — before anything
      # starts, naming the variable — rather than as a restart loop behind a
      # console that is already listening.
      #
      # `make env` deliberately does not generate METRICS_TOKEN: a token set
      # anywhere overrides the development exemption on /metrics, so writing
      # one into every `.env` would close the scrape endpoint on development
      # stacks whose Prometheus sends no credential. It is a production secret,
      # set deliberately here.
      METRICS_TOKEN: ${METRICS_TOKEN:?set METRICS_TOKEN — guards /metrics outside development; `openssl rand -hex 32`}
      JWT_SECRET: ${JWT_SECRET:?set JWT_SECRET — signs SAML/OIDC sessions; the same value ingest-worker takes below}

  ingest-worker:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !override ["${AISOC_BIND_ADDR:-127.0.0.1}:8081:8080"]
    restart: unless-stopped
    environment:
      # Both keys, because `envmode.Current()` reads ENV first and the base
      # file's ENV is what shadowed an operator's ENVIRONMENT before this.
      ENV: production
      ENVIRONMENT: production
      # Required outside development by
      # services/ingest/internal/config/config.go. This is the service that
      # accepts events off the network, so it is the last one that should be
      # allowed to start without a signing secret.
      JWT_SECRET: ${JWT_SECRET:?set JWT_SECRET — ingest refuses to start without it outside development}
      AISOC_SERVICE_TOKEN: ${AISOC_SERVICE_TOKEN:?set AISOC_SERVICE_TOKEN}
      AISOC_NEO4J_PASSWORD: ${NEO4J_PASSWORD:?set NEO4J_PASSWORD}

  fusion:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      ENVIRONMENT: production
      AISOC_DEV_MODE: 0
      AISOC_SERVICE_TOKEN: ${AISOC_SERVICE_TOKEN:?set AISOC_SERVICE_TOKEN}
      CLICKHOUSE_PASSWORD: ${CLICKHOUSE_PASSWORD:?set CLICKHOUSE_PASSWORD}

  agents:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      ENVIRONMENT: production
      AISOC_DEV_MODE: 0
      AISOC_SERVICE_TOKEN: ${AISOC_SERVICE_TOKEN:?set AISOC_SERVICE_TOKEN}
      AISOC_ACTIONS_SERVICE_TOKEN: ${AISOC_ACTIONS_SERVICE_TOKEN:?set AISOC_ACTIONS_SERVICE_TOKEN}
      # The same value the api service declares above. Eight modules here
      # present it; the API verifies it without falling back.
      AISOC_AGENTS_SERVICE_TOKEN: ${AISOC_AGENTS_SERVICE_TOKEN:?set AISOC_AGENTS_SERVICE_TOKEN}
      NEO4J_PASSWORD: ${NEO4J_PASSWORD:?set NEO4J_PASSWORD}

  actions:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      ENVIRONMENT: production
      AISOC_DEV_MODE: 0
      # Without this the service answers 503 on every mutating route, which is
      # the intended failure introduced in v12.0.0 — it previously served
      # isolate_host and run_script to anything that could reach the port.
      AISOC_ACTIONS_SERVICE_TOKEN: ${AISOC_ACTIONS_SERVICE_TOKEN:?set AISOC_ACTIONS_SERVICE_TOKEN}

  connectors:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      ENVIRONMENT: production
      AISOC_DEV_MODE: 0
      AISOC_CREDENTIAL_KEY: ${AISOC_CREDENTIAL_KEY:?set AISOC_CREDENTIAL_KEY}
      AISOC_SERVICE_TOKEN: ${AISOC_SERVICE_TOKEN:?set AISOC_SERVICE_TOKEN}
      SECRET_KEY: ${SECRET_KEY:?set SECRET_KEY}

  threatintel:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      ENVIRONMENT: production
      AISOC_DEV_MODE: 0
      NEO4J_PASSWORD: ${NEO4J_PASSWORD:?set NEO4J_PASSWORD}

  realtime:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      # Unset, the edge rejects every connection and /internal/* answers 503.
      # Both used to fall back: the ticket verifier to a constant committed to
      # this repository, and the internal guard to "authorized".
      AISOC_REALTIME_JWT_SECRET: ${AISOC_REALTIME_JWT_SECRET:?set AISOC_REALTIME_JWT_SECRET — run `make env` to generate it}
      REALTIME_INTERNAL_TOKEN: ${REALTIME_INTERNAL_TOKEN:?set REALTIME_INTERNAL_TOKEN — run `make env` to generate it}

  litellm:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  ollama:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  # Declared here only to take the preflight dependency. It is the one service
  # the base file defines that this overlay otherwise has nothing to say about,
  # and a single service that starts ahead of the credential check is the hole
  # the check exists to close.
  ollama-pull:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully

  # The MCP server (depth 9.1), a `--profile mcp` opt-in that is off by
  # default. It publishes 8096 in the base file for local clients; in
  # production it is unpublished like every other upstream, because the
  # console does not proxy it and an operator who wants an external MCP
  # endpoint exposes it deliberately. It authorizes every call from the
  # credential in the request (no static key — `check_mcp_over_http.py`
  # enforces that), so the credential check still gates its start.
  mcp:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  # The console. Demo mode is inlined at build time by Next, so this asserts
  # the runtime side only; pull the `latest` tag rather than a `-demo` one.
  web:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    restart: unless-stopped
    environment:
      NODE_ENV: production
      AISOC_DEMO_MODE: "false"

  # ---------------------------------------------------------------------------
  # `full`-profile stores. They do not run in a CORE deployment, but the base
  # file gives every service a default that points at the published dev
  # credential, so `--profile full` would start Neo4j and ClickHouse on a
  # password anyone can read. Required here rather than defaulted, and required
  # on the services that dial them so the two cannot disagree.
  # ---------------------------------------------------------------------------
  neo4j:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      NEO4J_AUTH: neo4j/${NEO4J_PASSWORD:?set NEO4J_PASSWORD — required for the full profile}

  clickhouse:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      CLICKHOUSE_PASSWORD: ${CLICKHOUSE_PASSWORD:?set CLICKHOUSE_PASSWORD — required for the full profile}

  opensearch:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  # Kafka UI browses every topic with no authentication of its own. It is
  # useful while developing and is not something to run beside production data,
  # so it is moved onto a profile nothing here activates: `--profile full` no
  # longer starts it. An operator who wants it can still ask for it by name.
  kafka-ui:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    profiles: !override ["debug-ui"]
    ports: !reset []
    # `no`, not `unless-stopped`: this is a tool you bring up to look at
    # something and shut down again. Restarting it on boot would quietly leave
    # an unauthenticated topic browser running next to production data.
    restart: "no"

  # ---------------------------------------------------------------------------
  # HARDENING.md section 2 — transport and network surface.
  #
  # Profiled services that bind a host port in the development file. The
  # console and the ingest endpoint are the only things that need to be
  # reachable: apps/web/next.config.js proxies every upstream server-side, so
  # a browser only ever talks to `web`. Reach these with `docker compose exec`
  # for debugging, or publish one deliberately in a local override.
  # ---------------------------------------------------------------------------
  alertmanager:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  enrichment:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []

  honeytokens:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []

  osquery-tls:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []

  otel-collector:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  prometheus:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  purple-team:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []

  slack-bot:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []

  teams-bot:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    # Same treatment as its sibling: the base file publishes a port for
    # local work, and in production only the console and ingest are
    # reachable from the host. `!reset` rather than an empty list, because
    # a plain override appends.
    ports: !reset []

  tempo:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped

  ueba:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []

  # HARDENING.md section 5 — observability and audit.
  #
  # Grafana defaults to admin/admin in the development file. Required here, so
  # `--profile monitoring` cannot start a dashboard over production telemetry
  # on a password printed in its own documentation.
  grafana:
    depends_on:
      preflight-secrets:
        condition: service_completed_successfully
    ports: !reset []
    restart: unless-stopped
    environment:
      GF_SECURITY_ADMIN_PASSWORD: ${GRAFANA_ADMIN_PASSWORD:?set GRAFANA_ADMIN_PASSWORD — required for the monitoring profile}