spatial-data-studio/docker-compose.yml at main · CirroBio/spatial-data-studio · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
services:
  spatial-data-studio:
    build:
      context: .
      dockerfile: docker/Dockerfile
    image: spatial-data-studio:latest

    ports:
      - "8080:8888"

    # Single read-write data mount (DESIGN §19.9): inputs + saved checkpoints +
    # snapshots share it. Defaults to ./test-data (ships with the repo) so demo
    # saves land alongside the sample datasets; override with SDS_DATA_HOST_DIR to
    # point at a dedicated folder in a real deployment.
    volumes:
      - type: bind
        source: ${SDS_DATA_HOST_DIR:-./test-data}
        target: /data

    # RAM-backed working set: the unpacked .zarr.zip extract dir and the per-session
    # normalized raster caches live on this tmpfs instead of disk, so tile/chunk reads
    # are served from memory. Its usage counts against the memory budget (see
    # SDS_WORK_DIR_IN_RAM), so admission trips a soft 503 before the tmpfs can grow past
    # the container limit and trigger the OOM killer. Durable checkpoints/snapshots stay
    # on the /data bind mount, never here.
    # The size=10g here is only a pre-resize fallback: on startup the work-tmpfs.sh
    # entrypoint remounts /work to SDS_WORK_TMPFS_PCT (85%) of the DETECTED memory limit,
    # so the tmpfs autoscales with whatever memory the container is given. That remount
    # needs CAP_SYS_ADMIN (cap_add below); without it the fallback size stands.
    tmpfs:
      - /work:size=10g,mode=1777

    # Required so the entrypoint can remount /work at the detected size. Scoped to this
    # single capability (not --privileged); it enables the mount syscall and nothing else.
    cap_add:
      - SYS_ADMIN

    environment:
      SDS_DATA_DIR: /data
      SDS_WORK_DIR: /work
      SDS_WORK_DIR_IN_RAM: "1"
      # Memory: container gets 12 GB (mem_limit below). SDS_CONTAINER_MEM_MB is left unset
      # so admission auto-detects the limit from the cgroup — the same value the entrypoint
      # sizes the tmpfs from — so raising mem_limit scales admission and the tmpfs together.
      # The /work tmpfs is folded into that admission accounting (SDS_WORK_DIR_IN_RAM).
      SDS_MAX_SESSIONS: "4"
      SDS_ADMISSION_PCT: "0.80"
      # Where a person watches the app (the 8080 mapping above) — quoted by the MCP
      # assistant (POST /api/mcp) when it directs the user to a session. Override
      # when the container is published at a different host/port.
      SDS_APP_URL: "${SDS_APP_URL:-http://localhost:8080}"
      MPLBACKEND: "Agg"
      PYTHONUNBUFFERED: "1"
      # Cirro upload. Users sign in from the browser with their own Cirro account
      # (device code), so no server-side credential — this only prefills the domain.
      CIRRO_BASE_URL: "${CIRRO_BASE_URL:-}"

    # Hard OS memory limit: the container is OOM-killed if it exceeds this. Keep
    # in sync with SDS_CONTAINER_MEM_MB above (12288 MiB == 12g) so the app's soft
    # admission control (refuses new work at SDS_ADMISSION_PCT of it) trips first.
    mem_limit: 12g              # docker compose / non-swarm
    deploy:
      resources:
        limits:
          memory: 12g          # docker stack deploy / swarm

    # Generous stop timeout so in-flight saves can flush (DESIGN §19.7)
    stop_grace_period: 120s

    restart: unless-stopped

    healthcheck:
      test: ["CMD", "curl", "-sf", "http://127.0.0.1:8888/api/healthz"]
      interval: 30s
      timeout: 10s
      start_period: 60s
      retries: 3