[KristiforMilchev]

Second part of milestone M1: workers stay connected to the controller.

Controller

• Connect requires the worker's client certificate, and it must be the worker's current one. The worker is identified by its certificate, never by what it sends in Hello.
• Each connection is a session with heartbeats and a silence timeout. A registry lets the scheduler send messages to a worker. If the same worker connects again, its old session is replaced.
• controller worker revoke <id> disconnects a worker immediately, and it can't reconnect. A worker revoked through another controller is dropped at the next periodic check, which runs every 30s.
• controller worker list shows workers, pools and when each was last seen.

Worker kit (pkg/worker)

• Join generates the private key locally, joins, and saves the identity to disk.
• Client connects, sends heartbeats, delivers controller messages in order and reconnects with jittered backoff. It stops for good if the controller rejects it as revoked or unknown.

Tests: integration tests run a real controller with PostgreSQL and the real worker client over TLS. They cover:

• connect and status reporting;
• a controller → worker message;
• reconnecting after a drop;
• a silent worker being dropped;
• revocation, both direct and through the database;
• a hello for the wrong worker;
• a superseded certificate;
• a certificate from a different CA;
• 10 workers at once.

They passed 10 times in a row under the race detector.

Next: runs, jobs and the scheduler.

🤖 Generated with Claude Code