Networking for distributed systems

Anycast and content delivery networks change where a user lands. They are tools for putting a copy close to demand. They are also another place a configuration can be wrong. Know which name is served from the edge and which must reach a specific region because of the data.

Private paths are still untrusted

A VPN, a dedicated interconnect, or a peered network gets you a path that is not the public internet. It does not get you permission. NIST’s zero-trust work exists because location, including “on our interconnect,” became a weak signal. Authenticate the workload. Encrypt anyway. Log the administrative changes to the path, because an interconnect is a high-value outage.

Failure is a routing event

Distributed systems fail when a path disappears or, worse, when it becomes slow and nobody’s health check notices. Timeouts are product decisions. A client that retries without a budget becomes the outage. Write the timeout next to the dependency. Draw the path for the one request you cannot afford to get wrong, including the name lookup.

Analysis

Networking is finished, as a discipline, only in slideware. In a system that spans accounts and regions it is the map of how a request is allowed to move. Keep the map smaller than the vendor’s feature list.

What this overview skips

It skips a tutorial on routing protocols. Those are real skills. The architectural point is upstream of them: names, intended paths, and a refusal to treat the path as proof of identity.