Skip to content

Late May 2026: the network grows up

The busiest fortnight so far: a public looking glass and a public wiki, self-service DNS and BGP, nColo moving onto our own ASN, a fairer bandwidth model, and uptime history on the status page. It was also the fortnight we spent the most time finding our own mistakes, so the "what broke" list is long.

Written up afterwards

Written in September 2026 from our session notes and commit history, and dated to the end of the period it covers.

What shipped

16 May: back to one thing done properly. The homepage stopped advertising products that didn't exist yet and led with the one you could join: Torus. The dashboards got a calmer, high-contrast redesign, and the member wiki got real setup guides for nSolo and nColo in place of placeholders.

17 May: BGP route servers, and self-service BGP for nColo. Authenticated route servers now link the regions. nColo members can switch between static routing and a full BGP session themselves, and get a working sample router config.

22 May: DNS rebuilt around your tier. Self-service DNS is organised into zones that match what you actually have: a mesh address, a dedicated IP or a routed prefix. Downgrading no longer deletes records; they're archived for 90 days and can be restored in one click. From 23 May every connection also gets a hostname automatically, and nSolo and nColo addresses get matching forward and reverse DNS.

23 May: a web looking glass. Our live BGP state, routes, sessions and AS-path lookups for every region, in a browser.

23 May: the wiki goes public. Built as a separate site from the internal documentation, not just the same site with a login in front. The old wiki now redirects here.

23 May: nColo moves onto our own ASN. New routed prefixes are announced from AS61081 by default; existing ones followed.

24 May: burst speed plus a monthly allowance. Flat, always-on speed caps gave way to a fast peak speed with a monthly data allowance, and a soft slowdown rather than a cut-off if you go over.

24 May: uptime history. The status page tracks real per-service uptime and won't show an SLA figure until it has enough data behind it. No "100%" after an hour.

Also in late May:

  • Two ways of listing a hosted service merged into one: Neko Pages.
  • "Torus Shell" renamed Torus ZTNA, and "VXLAN" renamed nLAN.
  • Management traffic between the platform and the hubs now runs encrypted.
  • Signup checks for sanctioned jurisdictions, and a rewritten Terms of Service covering data handling, retention and acceptable use.
  • Members of irixnet, a long-standing retro computing community, can fast-track the signup review. Everyone else gets a predictable 72-hour window.
  • Account and password email now goes out through our own mail server.
  • Status changes post automatically to our Discord.
  • Your profile shows your declared country next to where we actually see you connect from, so regional service stays honest in both directions.

What broke, and what we fixed

  • Saves that quietly reverted. Some account changes appeared to save, then reverted after a routine restart. They now fail loudly instead.
  • Two security passes. An internal audit of one backend host found services reachable from outside that should only have been reachable internally. They were closed the same day, credentials rotated and the host patched; no member data was involved. A second pass on 23 May fixed a cross-site scripting bug in a listing editor, moved network hardware onto unique per-device credentials, and tightened the public front end.
  • A firewall rule in the wrong place, on every hub. A broad rule sat ahead of some per-account controls, so a few tier-based restrictions weren't taking effect. We found it in our own audit, not through an incident, and fixed and verified it on every hub on 28 May.
  • A gallery that deleted itself. One tier-downgrade path deleted a member's photo gallery instead of hiding it. Galleries now survive any tier change, and we restored the affected photos from backup.
  • Infrastructure on the member map. A couple of infrastructure connections were showing on the public status map as if they were members. Filtered out.
  • An outage nobody noticed. The backend was down for about 50 minutes and monitoring didn't see it, because it only checked the static homepage. Alerting now watches the real service.
  • Pro speeds. New Pro connections were being set up slower than advertised. Corrected.
  • Stale egress paths. Switching between shared regional egress and a dedicated nSolo address could leave traffic on the old path. The two are now strictly either/or.

By the numbers

234 changes to the platform repository between 16 and 31 May, the busiest fortnight to date.