Divy takes our sandbox networking post apart and asks which change actually caused the speedup. The staged numbers answer it: moving the dataplane hop from an L7 proxy to a plain L4 forwarder went from 1.12 to 2.07 GB/s and cut CPU from 0.90 to 0.50 CPU-seconds per GB, then kTLS and splice(2) added 0.43 GB/s on top and almost no CPU. We expected kTLS to be the win, so this is the more useful read: removing the parser was. The rest of the piece is the part a reader can reuse, on what an L7 hop buys you, what has to be rebuilt when it goes away (routing moves into a preamble, liveness becomes metered bytes), and why a workload of short lifecycle calls should expect nothing from any of it.
We didn’t write this one — it’s Divy Yadav’s piece, published on Towards AI. The note above is ours; the full article is theirs.