oneacrefund/superset-websocket repository overview

⁠Superset WebSocket Server

A Node.js WebSocket server that pushes realtime events from the Superset backend to the web frontend. It is a shared transport for authenticated realtime sockets: it broadcasts opaque entity-change nudges and fans targeted task-status events out to JWT-bound socket routing keys.

⁠Requirements

• Node.js 20.11+ to run the built bundle (the build targets node20, and import.meta.dirname requires 20.11). Building and developing this package follow the stricter engines range in package.json.
• Redis or Valkey, reachable for Pub/Sub

To use realtime push, enable it in the Superset backend (WEBSOCKET_ENABLE, WEBSOCKET_URL, WEBSOCKET_JWT_SECRET; see below) and run this server on the same browser-visible host.

⁠Architecture

⁠The realtime Pub/Sub channel and the message envelope

Realtime events are published to a single Redis Pub/Sub channel, <REALTIME_CHANNEL_PREFIX>realtime (just realtime with the default empty prefix), by the Superset Flask app (superset/tasks/manager.py). Pub/Sub is intentionally best-effort (fire-and-forget, at-most-once — no replay): it accelerates each feature's authoritative REST/poll path rather than guaranteeing delivery, so a message missed during a disconnect is reconciled by a frontend catch-up / REST refetch. See Superset Configuration⁠ for the prefix and its Redis ACL implications.

Each Redis message is a self-describing envelope that separates what a message is from who receives it:

{ topic, scope, routes?, payload }

• topic — the semantic stream the browser dispatches on: task.status, entity.changed, and future topics (notification.*, report.progress, …). A new surface adds a topic without inventing channel names or overloading payload shapes.
• scope — the delivery breadth the server routes by:
  • authenticated_global — broadcast to every authenticated realtime socket. This is authenticated-global, not public: anonymous users get no realtime principal, no JWT cookie, and therefore no socket, so they never receive these messages. (True anonymous/Public-role realtime would need a separate, restricted model and is out of scope.)
  • principal / tab — targeted to the routing keys in routes. The server treats the two the same way (deliver to routes); the distinction is only descriptive.
• routes — server-computed routing keys for a targeted scope: principal-grain user:<id> / guest:<hmac> (all of a principal's tabs), or per-tab user:<id>:<tabId> (one tab). Omitted for a broadcast. Never forwarded to the browser.
• payload — the feature-defined body forwarded verbatim to the browser.

The two topics in use:

1. entity.changed (scope: authenticated_global). An "an entity changed" broadcast whose payload carries only opaque ids ({entity_type, id}) — no status or sensitive data — so a list view can learn that an entity of a type it renders changed and re-fetch just the affected rows through the authorized API. Each client filters by entity_type and id.
2. task.status (scope: principal or tab). A targeted {task_id, status} message. It is published on a task's terminal completion (SUCCESS, FAILURE, ABORTED, TIMED_OUT), not on every intermediate transition — the chart-data client only needs to learn a task finished; intermediate progress for list views rides the entity.changed broadcast instead. Keys are principal-grain by default (reaching all of a principal's tabs) but a task type may narrow them to a per-tab channel (async chart-data does this), so only the tab watching a task is notified. A socket is bound to its per-tab channel only when the browser advertises a tab_id on the connect URL (see Connection); it is always also reachable on its principal channel.

The server forwards each browser message as {topic, payload} — the browser dispatches on topic and never sees the route it arrived by. The server treats each routing key as opaque; the producer is responsible for their correctness and validates every policy-supplied key against the task's own subscriber principals before publishing, so a key can never target another principal.

⁠Connection

When a user's browser connects, it does so over HTTP, including the JWT authentication cookie set by the Flask app (WEBSOCKET_JWT_COOKIE_NAME, default superset-ws-token). Because authentication is cookie-based, the WebSocket server must be served from the same browser-visible host as the web application. In Kubernetes this is expected to run cleanly as a separate Deployment/Service behind an ingress that supports WebSocket upgrades. The server verifies the JWT with the shared secret (jwtSecret / JWT_SECRET). Superset mints the token only after the request principal has can_read on the Realtime resource. The token carries aud, iss, sub, principal_type, channel, and exp; the server rejects tokens whose channel does not match the principal identity. The permission itself is not serialized into the token: a valid token signed with the websocket secret is the proof that Superset already authorized the transport. The socket is bound to the channel claim (its principal channel), which is how principal-grain task-status fanout selects recipient sockets. A principal may have multiple sockets (e.g. several browser tabs); a message to the principal channel is sent to all of them. When a browser also advertises a tab_id on the connect URL (wss://…/?tab_id=<id>), its socket is additionally bound to a per-tab channel (<channel>:<tab_id>, derived from the authorized principal channel so it can never cross principals), letting the producer target one tab. Because permission is checked when Superset mints the JWT and when the websocket server accepts the upgrade, revocation after minting is bounded by the token lifetime; the Superset default is 15 minutes. To keep a long-lived surface connected without waiting for a hard disconnect at expiry, Superset re-mints the cookie inside a sliding window (any request in the second half of the token's life gets a fresh cookie), and the browser client proactively refreshes the cookie and reconnects before expiry — so an active or idle-but-open realtime surface stays connected while revocation remains bounded by the same lifetime.

During websocket JWT secret rotation, the websocket service can accept both the current key (jwtSecret / JWT_SECRET) and one previous verify-only key (previousJwtSecret / PREVIOUS_JWT_SECRET). The Flask app does not need the previous key: it keeps minting new cookies with WEBSOCKET_JWT_SECRET, and any cookie signed by the old key is replaced on the next HTTP response. Keep the previous key configured on the websocket service until old cookies and open sockets have aged out, then remove it.

The service is stateless apart from process-local live socket handles. Each replica subscribes to the same Redis Pub/Sub channel, receives the same events, and forwards only to matching sockets connected to that replica. Sticky sessions are not required; reconnecting to another replica reuses the JWT cookie and binds the socket to the same principal channel. Short websocket JWT lifetimes keep connection lifetime bounded for pod recycling; if a pod is terminated before expiry, connected browsers reconnect and a one-shot catch-up fetch on reconnect reconciles any events missed while disconnected (chart-data waiters re-check /task/status_changes from their retained cursor; list views refetch their displayed rows). There is no recurring poll while the socket is connected.

⁠Connection Management

The server uses standard WebSocket ping/pong⁠ to detect dead connections. Active sockets are pinged regularly (config: pingSocketsIntervalMs) and the internal registry records the last pong timestamp; a socket that has not responded within socketResponseTimeoutMs is terminated. Sockets are also terminated after the JWT exp time passes. The channel registry is periodically cleaned (gcChannelsIntervalMs) to release stale references.

⁠Install

Install dependencies:

npm ci

⁠WebSocket Server Configuration

Copy config.example.json to config.json and adjust the values for your environment.

Configuration via environment variables is also supported which can be helpful in certain contexts, e.g., deployment. src/config.ts can be consulted to see the full list of supported values.

⁠Restricting WebSocket origins

To mitigate Cross-Site WebSocket Hijacking, set allowedOrigins (or the ALLOWED_ORIGINS environment variable, comma-separated) to the list of origins permitted to open WebSocket connections, e.g. the origin Superset is served from:

{
  "allowedOrigins": ["https://superset.example.com"]
}

The Origin header of each upgrade request must exactly match one of the configured values. When allowedOrigins is empty (the default) the check is skipped and any origin is accepted; a single "*" entry explicitly allows any origin. Setting this is recommended for production deployments, especially when the JWT cookie uses SameSite=None.

⁠Production hardening checklist

Because the websocket is now a shared realtime transport and ships in the main image, review these before exposing it publicly:

• ALLOWED_ORIGINS — set it to your Superset origin(s) (see above). The default (empty) accepts any origin.
• Secure, scoped JWT cookie — on the Flask side set WEBSOCKET_JWT_COOKIE_SECURE = True (HTTPS only) and choose an explicit WEBSOCKET_JWT_COOKIE_SAMESITE ("Lax" when the app and websocket share a site; "None" only for cross-site, which then requires Secure and an ALLOWED_ORIGINS allowlist). Scope with WEBSOCKET_JWT_COOKIE_DOMAIN if needed. Cookies are short-lived and JWT-bound, but these flags close the cross-site surface.
• Connection caps — the server accepts unlimited connections by default (0). Set maxTotalConnections and maxConnectionsPerChannel (and optionally maxSocketBufferBytes) to bound resource use and blunt abuse.
• Keepalive/timeouts — pingSocketsIntervalMs / socketResponseTimeoutMs govern dead-connection reaping; the defaults are sane, tune per environment.

⁠Superset Configuration

Enable realtime push in the Superset Flask app (in superset_config.py):

WEBSOCKET_ENABLE = True
WEBSOCKET_URL = "ws://<host>:<port>/"
WEBSOCKET_JWT_SECRET = "<a strong random secret, >= 32 bytes>"

The built-in Gamma role receives can_read on Realtime; grant that permission to any additional roles that should receive websocket notifications. Without that permission, Superset masks WEBSOCKET_ENABLE to False for the request and does not mint the websocket JWT cookie.

Note that the WebSocket server must be run on the same hostname (different port) for the JWT cookie to be shared between the Flask app and the WebSocket server.

Note also that localhost and 127.0.0.1 are not considered the same host. For example, if you're pointing your browser to localhost:<port> for Superset, then the WebSocket url will need to be configured as localhost:<port>.

The following values must be coordinated between the Flask app config and this server's config.json (or its environment-variable overrides):

The Redis connection (redis / REDIS_*) must point at the same Redis instance Superset publishes to (DISTRIBUTED_COORDINATION_CONFIG). The channel the server subscribes to is <REALTIME_CHANNEL_PREFIX>realtime — the prefix is empty by default (channel realtime), and the resulting name is a wire-protocol contract with the backend producer, so the prefix must be set identically on both sides (REALTIME_CHANNEL_PREFIX in Superset and REALTIME_CHANNEL_PREFIX / realtimeChannelPrefix here). The server forwards each envelope to browsers as {topic, payload}. A Redis ACL for this server must therefore allow subscribing to the resulting channel — realtime by default, or e.g. tenant-a:realtime when REALTIME_CHANNEL_PREFIX=tenant-a:. Set a per-deployment prefix to keep deployments that share one Redis/Valkey from cross-delivering realtime nudges (Redis pub/sub is not scoped by DB number).

⁠StatsD monitoring

The application is tracking a couple of metrics with statsd using the hot-shots⁠ library, such as the number of connected clients and the number of failed attempts to send a message to a client.

statsd can be configured with the statsd object in the configuration file. See the hot-shots⁠ readme for more info. The default configuration is:

{
  "statsd": {
    "host": "127.0.0.1",
    "port": 8125,
    "globalTags": []
  }
}

⁠Running

Running locally via dev server:

npm run dev-server

Running in production:

npm run build && npm start

⁠From the official Superset image (recommended)

The WebSocket server is bundled in the official Superset image and launched via an alternate entrypoint, so no separate image is required:

docker run --rm -p 8080:8080 \
  -e JWT_SECRET="<same value as the app's WEBSOCKET_JWT_SECRET, >= 32 bytes>" \
  -e PREVIOUS_JWT_SECRET="<old websocket JWT secret during rotation, optional>" \
  -e REDIS_HOST=<redis-host> \
  apache/superset:<tag> /app/docker/entrypoints/run-websocket.sh

Configure it with the same environment variables as the standalone server (see src/config.ts); JWT_SECRET / JWT_COOKIE_NAME must match the Flask app's WEBSOCKET_JWT_SECRET / WEBSOCKET_JWT_COOKIE_NAME, PREVIOUS_JWT_SECRET can hold the prior websocket JWT secret during rotation, and the Redis connection must point at the same instance as DISTRIBUTED_COORDINATION_CONFIG.

With docker compose, start it via the opt-in websocket profile:

docker compose --profile websocket up superset-websocket

⁠Health check

The WebSocket server supports liveness checks via one of:

GET /health

OR

HEAD /health

/health is a pure liveness probe: it returns 200 whenever the process is up, so a transient Redis blip does not churn pods.

For load-balancer draining, use the readiness endpoint instead:

GET /ready

/ready returns 200 only while the server's Redis Pub/Sub subscriber is connected and subscribed, and 503 when that subscriber has dropped (the server can't deliver messages until it reconnects). Wire connection draining to /ready, not /health, so a degraded pod stops receiving new connections while it recovers.

⁠Containerization

The server ships in the official Superset image (launched via /app/docker/entrypoints/run-websocket.sh, see "Running" above). A standalone Dockerfile is also provided in this directory for building the server on its own during development.