How is redundancy built when two direct-attached disks have no hardware array?

Confirm the disks are direct-attached

After sign-in, read the block devices before creating a filesystem. On Linux, lsblk -o NAME,SIZE,TYPE,MOUNTPOINT,MODEL lists two NVMe drives of about 1TB each. TYPE is disk. Partitions are absent, or only the ones written when the system was installed. No md device and no hardware-array console is the expected state.

If one disk is already the root mount and the other is empty, the installer used a single disk. The empty disk can be partitioned on its own, or joined to a software array after a copy exists off the machine. The disk that holds root is not stripped and rebuilt into an array before that copy exists.

Usable capacity is the two disks added

Direct-attached means the operating system sees both disks at once. One can hold the system and the other the data, or both can hold virtual-machine disk files. The sum is close to 2TB, after the partition table and the filesystem take a small share. Reading the pair as a RAID-1 mirror of 1TB does not match the two disk rows in lsblk.

With no hardware array, a disk that goes offline leaves the files on the other disk in place. A directory that spans both disks is not rebuilt by itself. The mount point decides which disk holds a file. No array card in the chassis makes that decision.

Three layouts trade different things

Separate mounts fit a machine that keeps the system disk and the data disk distinct. Database files, a backup directory, and virtual-machine disks can sit on the second disk, with root on the first. That layout has no mirror, so the backup is written somewhere other than this machine.

Software RAID 1 spends capacity on a mirror. Two 1TB disks yield about one disk of space. Writes land on both. Reads can be shared. The mirror covers a single disk failure. It does not cover a mistaken delete, and it does not cover both disks failing together. Creating it clears the partitions on the disks that join.

Work in this order

1. Record both disk names

   lsblk and nvme list produce names such as /dev/nvme0n1 and /dev/nvme1n1. Order is not a guess. A reboot can swap the numbers.

2. Copy data that must be kept

   Creating an array rewrites the disks that join. When the system is already on one disk, partition only the empty disk. The disk that holds root stays out of a new array.

3. Pick one layout, then partition

   For separate use, partition the empty disk, run mkfs, and write the UUID into /etc/fstab. For a mirror, create the array on two idle disks, then put a filesystem on the array device.

4. Rehearse a restore

   After the mirror exists, the array tool should show both members in sync. Restore one small file from the backup into a temporary directory and confirm that copy can be read.

The default disks are delivered direct-attached

On AMD EPYC 7543P, whether memory is 128GB or 256GB, the default disks are 2 × 1TB NVMe, with no hardware RAID, each disk direct-attached. Monthly rent is $455 or $612. Memory capacity does not change how the disks attach. More space is one or two extra 2TB NVMe drives at $84 each per month. Those disks also stay outside any hardware array.

The facility is in Tokyo, Japan. The preinstalled system can be Ubuntu, Debian, AlmaLinux, CentOS Stream, FreeBSD, or Proxmox VE. When Proxmox places virtual-machine disks on a directory or on a separate disk, the second NVMe is still either independent storage or a software array. The installer does not choose a hardware array, because that card is not present.

When to leave the disks unjoined

When root already fills one disk and no copy exists off the machine, leave the layout as it is and only check lsblk and the mount points. Disks of different model or size are not assumed to form a symmetric mirror. An added 2TB disk is handled in the other note. That disk is direct-attached as well and does not join these two 1TB drives on its own.

After redundancy is in place, access stays the same. Linux still uses SSH, and the address is in the console. A change to the disk layout does not change the host name or the port.

A mirror and a backup cover different failures

Software RAID 1 only keeps the same data on both members. A mistaken delete, a wrong format, or a file written empty appears on both disks at once. The mirror is not a backup. A database, a virtual-machine disk, and site files that must be kept also exist somewhere other than this machine. From time to time, one small file is pulled from that copy to confirm it still opens.

With separate mounts, the system disk and the data disk can be copied on different schedules. The system disk keeps the installed configuration. The data disk follows how much the database or the virtual-machine disks write. Neither copy is written back onto the NVMe that is about to be replaced or repartitioned.

Check capacity again before creating an array. Two disks near 1TB suit a symmetric mirror. If one is clearly smaller, usable mirror space follows the smaller disk and the extra sectors stay unused. That pair is better mounted apart: the larger disk for data, the smaller one for the system. An added 2TB disk is not grouped into a mirror with these 1TB drives. The added-disk note explains why.

After the array exists, write the array name or UUID into the boot configuration and into fstab. An assembly that lives only in the current session can leave the device present after reboot while the filesystem does not return to the previous directory. Reboot once. Root or the data directory should still be where it was expected, and lsblk should show the members in sync rather than stuck degraded. While the set is degraded, new workload data waits until both members are back.

If Proxmox stores virtual-machine disks in a directory, that filesystem is the layout chosen above. A mirror, a separate second disk, or a later 2TB drive is only a different storage directory from the virtual machine’s point of view. The installer does not create a hardware volume. Moving a disk after the wrong directory was chosen means stopping the virtual machine first, then moving the disk file, so the copy is not taken while writes continue.