docker-compose.dev.yml

# Endpoints on localhost (host ports are offset from the in-container ports
# to avoid clashing with a Prometheus/VM/Grafana stack already bound to the
# conventional ports on the dev host):
#   VictoriaMetrics  :8429  (-> container :8428)
#   Prometheus       :9080  (-> container :9090)
#   node-exporter    :9100
#   forecaster       :9081  (-> container :9091; /metrics, /healthz, /-/reload)
#   Redis            :6378  (-> container :6379; query-cache backend; HA isn't
#                            wired in compose since leader election needs a
#                            kube API server)
#   Grafana          :3001  (-> container :3000; admin / admin) — "load" profile only
#
# Two profiles:
#   default (`make dev-up`)      — VM, Prometheus, Redis, node-exporter,
#                                  forecaster. Minimum for forecaster
#                                  development.
#   load    (`make dev-up-load`) — adds http-demo, http-demo-loadgen,
#                                  Grafana. Full demo experience.

name: promforecast-dev

services:
  victoriametrics:
    image: victoriametrics/victoria-metrics:v1.142.0
    container_name: pf-victoriametrics
    command:
      - "--storageDataPath=/storage"
      - "--httpListenAddr=:8428"
      - "--retentionPeriod=30d"
      - "--futureRetention=2d"
      # Deduplicate samples sharing a (series, timestamp). The remote_write
      # sink rewrites the full forecast curve every refit, so each future
      # step is re-emitted by ~horizon/refresh successive runs (here 30m/5m
      # ≈ 6x) at the *same* grid timestamp with an updated value. The sink's
      # documented contract is last-write-wins per (labels, timestamp) (see
      # docs/operations/backfill.md); VictoriaMetrics only honours that when
      # dedup is enabled, otherwise the overwrites pile up as raw duplicate
      # samples and inflate count_over_time. 15s matches the Prometheus
      # scrape_interval in docker/prometheus.yml so raw 15s resolution on the
      # underlying signals is preserved while the curve overwrites collapse.
      - "--dedup.minScrapeInterval=15s"
    ports:
      - "8429:8428"
    volumes:
      - vm-data:/storage
    healthcheck:
      test: ["CMD", "wget", "-qO-", "http://127.0.0.1:8428/health"]
      interval: 5s
      timeout: 3s
      retries: 20
    restart: unless-stopped

  http-demo:
    # Tiny FastAPI app with a handful of endpoints + /metrics. Driven by
    # the loadgen below so the forecaster has http_* time series to fit on.
    # Behind the "load" profile so `make dev-up` (minimum dev stack)
    # doesn't start it; `make dev-up-load` does.
    profiles: ["load"]
    build:
      context: ./docker/http-demo
      dockerfile: Dockerfile.app
    image: promforecast-http-demo:dev
    container_name: pf-http-demo
    ports:
      - "8001:8000"
    restart: unless-stopped

  http-demo-loadgen:
    # ~20-30 req/min with a slow sinusoidal modulation so the rate has shape.
    # Same profile as http-demo since one without the other is useless.
    profiles: ["load"]
    build:
      context: ./docker/http-demo
      dockerfile: Dockerfile.loadgen
    image: promforecast-http-demo-loadgen:dev
    container_name: pf-http-demo-loadgen
    environment:
      TARGET: http://http-demo:8000
      BASE_RPM: "25"
      AMPLITUDE_RPM: "15"
      PERIOD_SECONDS: "600"
    depends_on:
      - http-demo
    restart: unless-stopped

  node-exporter:
    image: prom/node-exporter:v1.11.1
    container_name: pf-node-exporter
    command:
      - "--path.procfs=/host/proc"
      - "--path.sysfs=/host/sys"
      - "--path.rootfs=/rootfs"
      - "--collector.filesystem.mount-points-exclude=^/(sys|proc|dev|host|etc)($$|/)"
    pid: host
    volumes:
      - /proc:/host/proc:ro
      - /sys:/host/sys:ro
      - /:/rootfs:ro
    ports:
      - "9100:9100"
    restart: unless-stopped

  redis:
    # Lets the dev forecaster exercise the Redis-backed query cache path
    # against a real backend instead of in-memory only. HA mode (leader
    # election) needs a Kubernetes API server, which docker-compose does
    # not provide — so the snapshot cache stays unused here. The query
    # cache validates fine in single-replica mode.
    image: redis:7-alpine
    container_name: pf-redis
    command: ["redis-server", "--save", "", "--appendonly", "no"]
    ports:
      - "6378:6379"
    healthcheck:
      test: ["CMD", "redis-cli", "ping"]
      interval: 5s
      timeout: 3s
      retries: 20
    restart: unless-stopped

  prometheus:
    image: prom/prometheus:v3.11.3
    container_name: pf-prometheus
    command:
      - "--config.file=/etc/prometheus/prometheus.yml"
      - "--storage.tsdb.path=/prometheus"
      - "--storage.tsdb.retention.time=2h"
      - "--web.enable-lifecycle"
    ports:
      - "9080:9090"
    volumes:
      - ./docker/prometheus.yml:/etc/prometheus/prometheus.yml:ro
      - prom-data:/prometheus
    depends_on:
      victoriametrics:
        condition: service_healthy
    restart: unless-stopped

  forecaster:
    build:
      context: ./forecaster
    image: promforecast:dev
    container_name: pf-forecaster
    command: ["--config", "/etc/promforecast/config.yaml"]
    ports:
      - "9081:9091"
    volumes:
      # Dev-tuned config: short lookback / step / min_points so the first fit
      # happens minutes after `make dev-up` instead of needing days of data.
      # examples/configs/* are production-grade and require a long-running TSDB.
      - ./docker/dev-config.yaml:/etc/promforecast/config.yaml:ro
    depends_on:
      victoriametrics:
        condition: service_healthy
      redis:
        condition: service_healthy
    restart: unless-stopped

  grafana:
    # Grafana is in the "load" profile because the bundled dashboards are
    # tied to the http-demo data; a minimum `make dev-up` for forecaster
    # development doesn't need it. `make dev-up-load` brings it up.
    profiles: ["load"]
    image: grafana/grafana:13.0.1
    container_name: pf-grafana
    environment:
      GF_SECURITY_ADMIN_USER: admin
      GF_SECURITY_ADMIN_PASSWORD: admin
      GF_AUTH_ANONYMOUS_ENABLED: "true"
      GF_AUTH_ANONYMOUS_ORG_ROLE: Viewer
      GF_USERS_DEFAULT_THEME: dark
    ports:
      - "3001:3000"
    volumes:
      - ./docker/grafana/provisioning:/etc/grafana/provisioning:ro
      - ./dashboards/grafana:/etc/grafana/dashboards:ro
      - grafana-data:/var/lib/grafana
    depends_on:
      - victoriametrics
      - prometheus
    restart: unless-stopped

volumes:
  vm-data:
  prom-data:
  grafana-data: