name: nostrsearch

services:
  # ── Unified node ───────────────────────────────────────────────────────────
  # ONE process: search API + archive HTTP + nostr archive relay + live
  # firehose, all over a shared Tantivy index, one archive database, and one
  # stats/WoT engine. Writes are serialized through a single writer task, and
  # the search reader auto-reloads on commit, so relay/firehose events become
  # searchable within the commit interval.
  server:
    image: voidic/nostrsearch:latest
    build:
      context: .
      dockerfile: Dockerfile
    restart: unless-stopped
    ports:
      - "8080:8080"
    # Archive downloads and relay websockets are descriptors too, on top of
    # RocksDB SSTs and Tantivy writer locks. The process raises its own soft
    # limit at startup, but it cannot exceed the hard limit set here.
    ulimits:
      nofile:
        soft: 65536
        hard: 65536
    environment:
      INDEX_ROOT: /data/index
      BIND: 0.0.0.0:8080
      RUST_LOG: nostrsearch=info,tower_http=info
      # Publish the raw .jsonl.zst corpus at /archive and archive into it.
      ARCHIVE_DIR: /data/archive
      STATE_DIR: /data/stats
      WOT_OUT: /data/wot.bin
      # Accept inbound writes at / (nostr archive relay). Relay events are
      # archived AND indexed AND folded into stats by the same process.
      ENABLE_RELAY: "1"
      # Comma-separated upstream relays enable the live firehose in-process.
      RELAYS: "wss://relay.damus.io,wss://nos.lol,wss://relay.primal.net"
    volumes:
      - index-data:/data
    # Search is mmap-heavy; give the page cache room. Tune to the host.
    deploy:
      resources:
        limits:
          memory: 24G
        reservations:
          memory: 4G
    healthcheck:
      test: ["CMD", "wget", "-qO-", "http://127.0.0.1:8080/healthz"]
      interval: 30s
      timeout: 5s
      retries: 3
      start_period: 10s

  # ── One-shot archive ingest ────────────────────────────────────────────────
  # Unified pipeline in backfill mode: streams a directory of Nostr JSONL dumps
  # (.jsonl/.json/.zst/.gz/.bz2) into BOTH the Tantivy index and the stats/WoT
  # engine, then exits. Run on demand:
  #
  #   DUMPS_DIR=/path/to/dumps docker compose --profile ingest run --rm ingest
  #
  # Writes into the same index volume the server reads. Analysis state persists
  # under /data/stats so a later run (or the firehose) warm-starts with WoT.
  ingest:
    image: voidic/nostrsearch:latest
    build:
      context: .
      dockerfile: Dockerfile
    profiles: ["ingest"]
    restart: "no"
    entrypoint: ["./bin/ingest"]
    command:
      - "--index-root"
      - "/data/index"
      - "--input-dir"
      - "/dumps"
      - "--heap-mb"
      - "64"
      # Total writer heap is max-open-shards x heap-mb. Keep the cap at or
      # above the number of months the corpus spans: dumps are often not
      # date-ordered, and a smaller cap evicts writers that are immediately
      # needed again, paying a commit + fsync each time.
      - "--max-open-shards"
      - "64"
      - "--commit-docs"
      - "200000"
      # Stats/WoT are computed from the same stream and hot-swapped into scoring.
      - "--state-dir"
      - "/data/stats"
      - "--wot-out"
      - "/data/wot.bin"
    environment:
      RUST_LOG: nostrsearch=info,nostr_archive_cursor=info
    volumes:
      - index-data:/data
      - ${DUMPS_DIR:-./dumps}:/dumps:ro
    deploy:
      resources:
        limits:
          memory: 16G

volumes:
  index-data:
    driver: local