mirror of
https://github.com/penpot/penpot.git
synced 2026-10-29 14:00:37 +00:00
b132939b71
The instance that holds a frozen plugin's connection may not hold the browser extension that can wake its tab. The session waker now asks wake agents through a WakeAgents interface. With Redis, an agent of this instance is asked first; otherwise the request goes out on a wake channel per user, to which an instance subscribes while it holds an agent of that user. Each instance answers on a response channel per request: with its agent's outcome, or with unknown-session at once when none of its agents holds the session. The first activation wins. A request that reaches no instance fails at once, so a frozen session without an agent still fails as before. When a forwarded task finds its plugin frozen, the instance holding the plugin connection reports the wake on the task's response channel. A waking notice lets the requester extend its timeout by the wake budget, and a sent notice restarts the plain task timeout, so the wake does not count against the task. Tasks that need no wake send no notices. A woken page reconnects through the load balancer, perhaps to another instance. With Redis the waker does not wait for a local reconnect: when the woken page has closed its socket, the owner sends a resend notice and the requester publishes the task again until an instance holds the session, for up to the wake wait. A closing connection also ends a pending liveness wait, so a page that closes without a last heartbeat is not waited for in vain.