Week 8 Session: What does it talk to?

Inventory your program's network requirements, then find out which of them a browser can actually satisfy and what the rest would cost

19 April 2027

Inventory first, then sort by what can actually be tested.

The inventory. Every host, every port, every protocol. Does it listen? Does it need UDP? Does it hold a connection open, or is it request-response? Does it care about the exact bytes of its own request?

The HTTP endpoints you can test directly. Issue the request from a browser and see what comes back. The answer is immediate and binary: either the origin sends the header that permits you to read the response, or it does not.

Try a spread — an API that advertises support, a vendor API that does not, a package registry, and something of your own where you control the headers.

Everything else you cannot test that way, and pretending otherwise is the mistake this session exists to prevent. A raw TCP service, a UDP protocol, or anything that listens will not answer a browser request at all. For those the question is not “does it work” but “what shape of answer is available”:

• A relay you run. Tunnel the connection over a WebSocket to something holding the real socket. Not a hack of last resort — the x86 emulator you used in week 1 does exactly this for its networking.
• A relay somebody else runs. A VPN client compiled to WebAssembly puts the page on a private network. Note that it is still relaying over a WebSocket, because its own transport is UDP and the browser cannot send one; what you gain is an identity model, a routing table, and egress from machines you control rather than from the page.
• Change the protocol, if you control both ends.
• Move the requirement, if it turns out nobody needed UDP after all.

Notice that the first two are the same mechanism. Nothing below HTTP leaves a browser without being tunnelled by something, and the interesting differences are about ownership and abstraction rather than transport.

Each of those costs something — a server you now run, a network you now join, a protocol you now maintain. Price at least one of them.

Bring back: how many of your program’s endpoints are directly reachable, what you would do about the rest, and what that would cost somebody to operate.