Notifications only reach subscribers on the hostname used for the write #647
Description
Summary
A WebSocket subscriber is only notified if it subscribed through the same hostname that the write came through. When one pod is reachable under two names, such as localhost and a LAN address, a write through one name doesn't notify tabs on the other, even though they're watching the same file.
This looks deliberate rather than a bug, because the hostname is part of a resource's identity and subdomain pods rely on it. This issue records the behaviour and asks whether JSS should change it.
What happens
getRequestPaths in src/handlers/resource.js builds the change URL from the request hostname:
const resourceUrl = `${request.protocol}://${request.hostname}${urlPath}`;notifySubscribers in src/notifications/websocket.js then looks up that exact string in subscribers. checkSubscriptionAccess also requires the subscribed URL's origin to match the socket's origin, so a client can't subscribe across names.
Test
Version 0.0.225, started with --host 0.0.0.0. One subscriber on each address, then the same file was uploaded once through each with PUT:
| Upload through | localhost subscriber | 192.168.0.167 subscriber |
|---|---|---|
| localhost | Notified | Silent |
| 192.168.0.167 | Silent | Notified |
We found this when a script uploaded through localhost while browser tabs, including one on another LAN machine, had the map open through the LAN address. The tabs only updated after a refresh.
Behind a proxy this normally doesn't show, because the writer and readers all use the public name. A local writer behind a proxy would hit the same thing.
Other servers
From reading the code, not running it:
| Server | Matches on | Write on one name, subscriber on another | Trade-off |
|---|---|---|---|
| JSS | Full URL | Not notified | Writer and readers must share one name |
| NSS (solid-ws) | Path only (defaultToChannel) | Notified | With subdomain pods, the same path on two pods looks like it would share a channel |
| CSS | Full URL, one configured base URL | Other hosts get a 500 (E0001), unless fixedBaseUrl is set | One canonical name by configuration |
Options
-
Keep it and document it. Writer and readers use the same address. This needs no code, and it's what we're doing for now.
-
Match on the storage path. Key subscribers on what getEffectiveUrlPath returns, and send each socket the URL it subscribed with:
Request Storage path localhost/public/x /public/x 192.168.0.167/public/x /public/x (same file, both notified) alice.example.com/foo /alice/foo bob.example.com/foo /bob/foo (different file, kept apart) This hasn't been built or tested. The ancestor-container walk in broadcast and the ACL check on subscribe would need checking against it.
Either way, it would help to mention this in the notifications section of docs/configuration.md, since going from localhost to --host 0.0.0.0 is an easy way to run into it.
Activity