Uptime: Wiki | Uptime v0.1.10

Oversubscription

The capacity fit indicator now responds to your overcommit dial. Comfortable and Tight used to be measured against a host's raw physical CPU, so the dial never moved the reading. It now measures against the sellable, oversubscribed ceiling: turn the dial up and hosts read more comfortable, turn it down and they read tighter.
Turning the fleet overcommit dial down now rebalances your fleet instead of just changing a number. If lowering the density leaves a host carrying more than its new ceiling allows, the game sheds the excess virtual machines (newest first) onto hosts that have room. If the whole fleet is full there is nowhere to move them, so the over-packed host keeps serving under contention until you free up space.
The overcommit help now teaches the right thing. Reading up on CPU overselling used to open the unrelated network fabric oversubscription article. There is now a dedicated CPU overcommit explainer, and the "How far can I oversell?" and noisy-neighbour cross-links point to it.

Taking on a customer or an upgrade that would run your fleet oversubscribed now reads clearly. The request shows an "oversubscribed" verdict with a plain recommendation, and overselling is gated behind a soft "Approve anyway" confirmation rather than blocked outright. Overselling is a choice, not a wall.
A fleet with no free headroom now reads as such in the growth flow, so you can see at a glance when the answer is "buy more hardware" versus "dial up density."

The "Hide down" filter and the per-port "Select for bonding" control on the switch, gateway, and appliance consoles are now real sliding toggles instead of chip buttons.
The gateway and appliance consoles now have the per-port Select toggle and a visible "Bond ports" button, matching the switch. Marking ports for a bond no longer requires the keyboard.
The gateway's first tab is now called "Ports" instead of "Uplinks."
The Threats tab mitigation bar now reads correctly. It is coloured by status (green when protected, amber when absorbing, red when leaking) instead of always showing orange, and an empty bar means nothing is being blocked instead of showing a permanent sliver. The columns were relabelled so they mean what they say: "Residual" is now "Getting through," and "Capacity" is now "Flood budget," the budget for active flood/attack traffic, not a cap on the benign background traffic the box filters for free.

CPU contention alerts no longer pile up or cry wolf. A "CPU contended" incident now opens only when a host is genuinely over capacity and a customer on it is actually breaching its service level. A host that is busy but keeping up is shown as a state on the host console, not raised as an incident.
Contention incidents now clear on their own once the host recovers, load drops, or you migrate work off it. Previously they were opened but never resolved, so they accumulated forever.
Contention severity now reflects real impact (money at risk and service level), not how long the alert has been open. A minor, long-standing contention no longer escalates itself to the top "existential" severity.
Incident headlines now use the name you gave a server. A renamed host reads as "Server01 CPU contended" instead of "host 27," so you are sent to the box you actually named.

New workloads now balance across your fleet. New virtual machines are placed on the least-loaded eligible host instead of piling onto the first one that fits, so adding a server actually spreads load onto it rather than new customers landing on an already-busy host.
You can now migrate a virtual machine onto a host that still has room under its overcommit ceiling. The migration check used to ignore the overcommit dial and measure against raw physical CPU, so it refused hosts that clearly had space to sell.
A migration that cannot be placed is no longer silent. If there is no host with room (or the target is offline or incompatible), you get a clear message and the machine stays exactly where it was, instead of appearing to move and then snapping back.

The CPU resource bar now shows live load (actual demand against physical cores) rather than how much is reserved, so a genuinely busy host reads high and a packed-but-idle host reads its real, lower load.
The workloads list now includes a "Control plane + host services" row. The platform's own CPU use (the control plane and per-host agents) used to be invisible, which made the CPU figure look like it disagreed with the listed workloads.