Uptime: Wiki | v0.1.27 (Early Access)

Deliveries

New: a delivery robot at the Data Center. Load it up in the shipping zone, press E on the control panel on its pole, and pick a rack row. It drives itself down the aisle and waits there while you unload, then you send it back.
Its deck takes two server boxes side by side, and only stacks upward once the deck is full, so a mixed load of servers and switches makes the trip in one go.
Boxes come off in the order they went on, last on first off.
The panel lists what is on the deck, and has a Stop button to halt the robot where it stands if you want it somewhere else. It closes on its own once you walk away from the robot.
The robot takes the place of the hand trolley at the Data Center. The garage and basement are unchanged.

The processor breakdown no longer reports functions that are not there. It was working the figure out by subtraction, so a load balancer's work landed under "Functions" on a server with no serverless workloads on it.
Load balancers, Kubernetes and content delivery have their own line in that breakdown. They never had one, which is why their work had nowhere to go.
Live load can no longer exceed what the machine physically has. A load balancer's full traffic demand went into the total uncapped, so an eight-processor machine read 264%.
That figure was marking machines contended and cutting every other tenant on them to about a third. Servers with a busy load balancer should feel much less crowded.
The platform line no longer counts other workloads. It was total allocation minus virtual machines, so a load balancer got counted twice.

A rack with several problems shows one status strip, not one per problem. They were stacking, so the light got brighter and wider with every extra incident.
The strip sits on the rack's top edge again instead of floating above it. It was being sized from everything mounted in the rack, cables included, rather than the rack itself.

A database that came up while your servers were full now gets its replica once you make room. It kept the primary and owed the replica you were being billed for, and because something was there it was never looked at again, so buying another server changed nothing.
The retry places only the copies that are missing. It will not move a live database off the server it is already running on.
Start repair on a shortfall now retries the placement there and then, rather than waiting on the next sweep. When it still cannot place, it names the thing blocking it instead of saying to fix it directly.

You can designate machines for object storage again. Pick them in Segments, leave the customer as Any, and set the job to Object storage. It was being refused by a rule that stopped applying when reservations were split into who they are for and what they run, and the rule was never taken back out.
Once any machine is designated for it, the ones you have not designated stop taking shards. Object storage no longer lands on whatever happened to have free disk.
The designation survives loading. It was being wiped every time a save opened, so even where it did take, it did not last.
Reserving object storage for a particular customer or tier is still refused, and now says why: a shard has no network identity to match against, so leave the customer as Any and it will home.
As with any reservation, if the machines you are designating are running other work, that work needs somewhere to go and the change is refused if there is nowhere. The message names what is in the way.

Service control planes spread across your servers instead of all landing on the same one. Each class was ranking a fleet that looked untouched, so they all picked the same winner, and cordoning that server moved the whole pile to the next one. They now see what the others have taken, and count what is already running on a server when they rank it.

Breakout legs now run at what the trunk's lanes actually carry. A 40G QSFP+ port is four 10G lanes, so a 25G server port on one of them connects at 10G and all four lanes fill. Each leg used to claim the far port's full rate, so two 25G legs wanted 50G out of a 40G port and the second was turned away.
The fan-out limit counts lanes instead of legs, so a leg wide enough to need more than one lane costs what it actually uses.
Breakouts saved before this are re-rated when you load, so a trunk that was recorded as carrying more than it physically can stops refusing legs that fit.
Bonding ports now tells you it worked and names the bundle. It was sending the command and saying nothing, on the switch, gateway and appliance consoles alike.
Ports fed by a breakout leg can be bonded. They were offered as bondable and then dropped while the bond was assembled, so the key did nothing at all.
The switch console's unbond, mode and minimum-links controls are back. They looked for the bundle on a field that ports do not carry, so they never appeared.

Large sites no longer stutter. A dozen routines were walking every workload on the site once per customer, once per server, once per rack, so the cost grew with the square of how much you had built. Object storage moving into the same store in v0.1.26 is what tipped it over. A save with a thousand tenants spent 110ms preparing each update and now spends under 20ms.
Loading a save no longer freezes for minutes afterwards. The workload rebalance that runs on load capped how many machines it would move but not how many it would consider, so on a save with nothing left to improve it checked every virtual machine against every server, every tick, forever. It now works to a budget and stops after a set number of ticks, leaving anything it could not place better where it was.
Racks open on a large site again instead of reporting no data.
Servers in maintenance no longer drain continuously and log the same thing repeatedly. The drain re-ran every tick whether or not there was anything left to move.
Walking in and out of racks, and moving around a full site, holds its frame rate.

New: invert mouse look. It sits next to the sensitivity slider, in the Camera section of the Controls tab in Settings, and takes effect the moment you flip it. It covers the photo mode camera too. The right stick keeps its own orientation.

Quitting to the title from a loaded save works again.
Long headings wrap onto a second line instead of running off the edge. An incident's title was the visible case; the setting is shared by every heading on the laptop, so the rest follow.
Article names in the knowledgebase list wrap too. A long one used to shove its tag and category off the row.