Moving past TCP in the data center, part 2

lwn.net

19 points·chmaynard··2 comments

Osterhout tends to be prescient in his diagnoses, though not his solutions. Eg. His take on scripting languages filling a unique role that wouldn’t be replaced by system languages was pretty accurate; https://en.wikipedia.org/wiki/Ousterhout%27s_dichotomy , especially in the AI/ML community these days. His solution, Tcl/Tk, was niche though.

On the supposed death of TCP, Bruce Davie posted a more nuanced take: https://systemsapproach.substack.com/p/tcp-the-p-is-for-plat... And https://systemsapproach.substack.com/p/quic-is-not-a-tcp-rep...

In short, a lot of what we do on the web and data centre these days feels more like RPC, and HTTP/3 (QUIC) seems more fit for purpose. This doesn’t mean it replaces TCP for high throughput byte steams.

Another interesting development in the data centre is the rise of WireGuard for mesh secure connectivity, with Tailscale as its most recent herald. Hop-by-hop mTLS is notoriously hard to manage but Tailscale is a breeze. And identity based ACLs replacing CIDR-based firewalls? Yes please!

Homa's message-based API is different and he has been unable to find a way to "slide it in under the existing TCP sockets interface"

I suppose the fastest path to adoption would be something that worked like NFS, if your goal is to hit the data center first. Or maybe an ssh like client/server pair and other things could tunnel through it.

Reply on news.ycombinator.com