v0.1.15 (Early Access)

This patch adds customer network isolation, fixes how contention and placement read, makes saves rebalance when they load, and includes a cable-rendering fix.

New: network isolation (VXLANs)

You can now separate customers onto their own network segments. A segment is a VXLAN: a private network you place customers on, either by picking them one at a time or by tier. Everyone sits on the shared default network until you move them.

It lives in Operations, under Isolation. Your segments are a list on the left, and the selected segment's detail is on the right. Create a segment, add customers to it, and decide how far the isolation goes.

There are two levels, and the difference matters.

Logical (a segment on its own network). The customer's traffic is separated and a breach is contained. If another customer is compromised, the attacker cannot cross into this segment. The catch is that the customer still shares machines with others, so it can still compete for CPU and memory with whoever it sits next to. Use this to wall tenants off from each other on the network without buying them their own hardware.

Dedicated (reserve hardware for the segment). Reserving machines holds whole hosts for the segment. No other customer's workloads land there, so there is no resource contention from neighbours and you control exactly where those customers run. You reserve from the segment's detail pane: tick the machines you want, grouped by rack, and confirm. Every customer in the segment then moves onto that hardware.

The quick way in. Open a customer and choose Isolate to drop them onto their own segment. If you then want them on their own hardware, reserve machines for that segment in the Isolation view.

Why it matters. A breach only spreads between customers that share a segment. Give a customer its own segment, or its own hardware, and a compromise somewhere else stops at the boundary. The network map has a new Isolation view that colours the fleet by segment, so you can see who sits where and which hosts are reserved.

All of the isolation copy follows the plain and engineering language toggle.

Fixes

• Contention reads honestly now. After the last patch, some hosts still showed as contended when they were fine. Two things were being conflated. An allocation figure was drawn as if it were live load, so an overcommitted but idle host read as full. And hosts running background service coordination were flagged as busy when they had no real work on them. Utilisation now shows live load separately from committed capacity, a host that is only holding a reservation no longer reads as contended, and a contended host shows what is actually using its CPU, split across VMs, functions, and service overhead.
• Loading a save rebalances placement. If an older save had VMs stacked onto one host by a previous version's placement, that stack used to stay put on load. Now the fleet re-spreads those VMs across your idle hosts on the first tick after loading, so a save comes up balanced rather than frozen in its old shape.
• Segment membership repairs itself. A customer that leaves is now cleaned out of any segment it was pinned to, both when it happens and again on load, so an old save cannot carry stale references.
• QSFP cables no longer cross into an "X" through the cable managers. When two thick QSFP runs shared the same manager rail on a rack, they could twist past each other into a bowtie instead of sitting as clean, parallel runs. Cables sharing a manager column are now ordered by their vertical position, so they nest neatly (the "U" you would expect) rather than crossing. Most visible on QSFP because it is the fattest cable, but the fix applies to every cable in a manager column.

Notes

• The old host pools placement tool has been retired. Network segments cover placement now.
• No save or network protocol version change. Existing saves load normally.