Tor performance: what actually improved and what is still pending
For years, the running joke about the tor network was that it was slow by design. That was never entirely true - and it is getting less true every release. A quiet wave of performance engineering has been reshaping how the network handles load.
Congestion control finally shipped
The biggest change arrived with Tor 0.4.7: the network switched from a decades-old fixed-window algorithm to ART, the new congestion control proposal (prop-324). Instead of guessing how much data a circuit can carry, relays now measure feedback and adapt - much closer to how TCP itself behaves on the modern internet. The practical effect, confirmed by network-wide measurements, was a meaningful drop in queueing delays under load. Downloads that used to stall mid-transfer now throttle smoothly instead. Nobody writes press releases about queueing delays, but this is the kind of change users feel without being able to name it. Onion services got their own fixes
Hidden services historically suffered from a specific pathology: the rendezvous handshake is expensive, and busy services would choke on introduction-point overload. The 0.4.7 series also introduced rate limiting on introduction requests, which blunted the worst denial-of-service waves without operators resorting to captchas. Descriptor caching improvements and streamlined circuit building trimmed hundreds of milliseconds from first connections. On a network where every page load pays a handshake tax, that matters. What arti promises
The Rust rewrite, arti, is the long game. Memory safety aside, its architecture is designed for embeddability - tor as a library inside other applications - and for faster iteration on exactly the kind of protocol work described above. Arti is not yet a drop-in replacement for the C daemon, especially for relays. But client-side performance work lands there first now, and the feedback loop between bug report and fix has visibly shortened. What is still pending
Relay-side arti support remains the big unfinished item, and the C daemon will be maintained for years because of it. Bandwidth scanner modernization, better padding defaults, and onion service load-balancing beyond OnionBalance are all open work items on the Tor Project's roadmap. The honest summary: the network is faster than its reputation, improving release by release, and the remaining bottlenecks are known, named and actively being worked. Slow is a habit the network is finally breaking.
