MCP is Deleting Its Own Handshake: Inside the Biggest Spec Rewrite Since Launch
I want to start with a bug, because this whole rewrite is really an answer to it.
Imagine I'm running an MCP server in production, three instances of it behind a load balancer that spreads incoming requests across them. A client connects and the first instance answers. Under the old design my server could issue a session ID, and if it kept that session's state in local memory it would expect every later request to come back to the same instance. With no shared store or affinity to fall back on, the client's next request lands on the second instance, which has no record of that session and returns a 404. The client fails.
One fix is sticky sessions: configure the load balancer to pin each client to whichever instance answered first. Another is to push the session's state into shared storage that every instance can read. Either works, at...
Copyright of this story solely belongs to hackernoon.com. To see the full text click HERE