Important notes

  1. The Oxide CLI, Go SDK, and Terraform Provider remain unchanged from v20. If you are updating the system from v19 or an earlier release, please be sure to upgrade your API clients.

Installation

Oxide Computer Model 0 must be installed and configured under the guidance of Oxide technicians. The requirement may change in future releases.

Upgrade compatibility

Upgrade from version 20, 20.1, and 20.2 is supported. We recommend shutting down all running instances on the rack before software update commences. Any instances that aren’t stopped for software update are transitioned to the failed state when the control plane comes up. They can be configured to start automatically with auto-restart policy or they can be started manually by the user.

All existing setup and data (e.g., projects, users, instances) remain intact after the software update.

New features

Dense Turin support

The service processor and host operating systems have been enhanced to support a new rack specification that uses AMD Turin Dense (Zen 5c, 192-core). Note that the maximum vcpus per VM remains at 254 and the limit will be increased in the future.

Sled provisionable memory increase

The memory reserved for system use has been reduced to allow the provisionable memory per sled to be up to 91% of the amount of available DRAM on a sled. (omicron#10791)

Web console

We added more details to resource tables: instances now show external IP addresses, disks show their source image or snapshot, images show OS and version, and the VPC detail page shows the VPC’s ID and IPv6 prefix. Operators can now set an IP pool as the silo default in the same step as linking it. The image upload form now validates that the image size is a multiple of the block size before starting the upload. The Cmd/Ctrl+K command palette is more comprehensive.

Full console changelog

Other bug fixes and enhancements

Known behavior and limitations

End-user features

Feature AreaKnown Issue/LimitationIssue Number

Disk/image management

Disks in importing_from_bulk_writes state cannot be deleted directly. The procedure for unsticking a canceled disk import can be used as a workaround.

Disk/image management

Disk rejected by guest OS due to duplicate nvme device names. The issue is caused by a 20-character limit in applying the disk name to the device serial number. See the Troubleshooting guide for more information.

-

Disk/image management

The ability to modify image metadata is not available at this time.

Instance orchestration

Unable to start an instance that has a disk replica on a sled being updated.

Instance orchestration

Instance start API frequently times out when attached to local disks.

Instance orchestration

New instances cannot be created when the total number of NAT entries (private-to-external IP mappings) in the system exceeds 1024.

Instance performance

The tsc clocksource is treated as unreliable by guest, resulting in its fallback to use substantially slower timestamp syscalls. A workaround for this issue can be found in the Troubleshooting guide.

Instance performance

Linux guests unable to capture hardware events using perf record. A workaround for this issue can be found in the Troubleshooting guide.

VPC internet gateway

Changing a silo’s default IP pool causes some instances to lose their outbound internet access. This is due to a mismatch between the pool containing the instances' external IP (which are allocated from the new default pool) and the pool attached to the system-created internet gateways (which are linked to the old pool during creation time). Please see the Troubleshooting guide for some possible options for restoring instance outbound connectivity.

VPC routing

Subnet update clears custom router ID when the field is left out of the request body.

VPC routing

Network interface update clears transit ips when the field is left out of the request body.

-

Telemetry

VM instance memory utilization and VPC network/firewall metrics are unavailable at this time.

-

Operator features

Feature AreaKnown Issue/LimitationIssue Number

Silo management

The ability to modify silo and IDP metadata is not available at this time.

omicron#3400, omicron#3125

System management

Real-time availability status for sleds and physical storage is not yet shown in the inventory API or UI.

omicron#2035

System management

Operator-driven instance migration across sleds is currently unavailable.

-

System management

Some running instances transitioned to the "stopped" state after update.

omicron#9177