How to manage RAID 10 on copr-backend

There are currently six AWS EBS sc1 volumes used for hosting Copr Backend build results. Four disks are forming one 24T raid10, two more disk form 16T raid1. These two arrays are used as “physical volumes” for the copr-backend-data LVM volume group, and we have a single logical volume on it with the same name copr-backend-data (ext4 formatted, mounted as /var/lib/copr/public_html).

Everything is configured so the machine starts on its own and mounts everything correctly. We just need to take a look at /proc/mdstat from time to time.

Manually checking/stopping checks

Commands needed:

echo idle > /sys/block/md127/md/sync_action
echo check > /sys/block/md127/md/sync_action

Detaching volume

It’s not safe to just force detach the volume in AWS EC2, it could cause data corruption. Since there are several layers (volumes -> raid -> LVM -> ext4) we need to go the vice versa while detaching.

stop apache, copr-backend, cron jobs, etc.
unmount: umount /var/lib/copr/public_html
disable volume group: vgchange -a n copr-backend-data
stop raids: mdadm --stop /dev/md127
now you can detach the volumes from the instance in ec2

Attaching volume

attach the volumes in AWS EC2
start raid and volume group mdadm --assemble --scan. In case the --assemble --scan doesn’t reconstruct the array, it is OK to add the volumes manually mdadm /dev/md127 --add /dev/nvme2n1p1.
mount the /dev/disk/by-label/copr-repo volume

There’s a ansible configuration for this, and list of volumes.

Adding more space

Create two gp3 volumes in EC2 of the same size and type, tag them with FedoraGroup: copr, CoprInstance: production, CoprPurpose: infrastructure. Attach them to a freshly started temporary instance (we don’t want to overload I/O with the initial RAID sync on production backend). Make sure the instance type has enough EBS throughput to perform the initial sync quickly enough.
Always partition the disks with a single partition on them, otherwise kernel might have troubles to auto-assemble the disk arrays:
```
cfdisk /dev/nvmeXn1
cfdisk /dev/nvmeYn1
```
Create the raid1 array on both the new partitions:
```
$ mdadm --create --name=raid-be-03 --verbose /dev/mdXYZ --level=1 --raid-devices=2 /dev/nvmeXn1p1 /dev/nvmeYn1p1
```
Wait till the new empty array is synchronized (may take hours or days, note we sync 2x16T). Check the details with mdadm -Db /dev/md/raid-be-03. See the tips bellow how to make the sync speed unlimited with sysctl.

Note

In case the disk is marked “readonly”, you might need the mdadm --readwrite /dev/md/raid-be-03 command.
Place the new raid1 array into the volume group as a new physical volume (vgextend does pvcreate automatically):
```
$ vgextend copr-backend-data /dev/md/raid-be-03
```

Extend the logical volume to span all the free space:

$ lvextend -l +100%FREE /dev/copr-backend-data/copr-backend-data

Resize the underlying ext4 filesystem (takes 15 minutes and more!):
```
$ resize2fs /dev/copr-backend-data/copr-backend-data
```
Switch the volume types from gp3 to sc1, we don’t need the power of gp3 for backend purposes.
Modify the https://github.com/fedora-copr/ansible-fedora-copr group vars referencing the set(s) of volume IDs.

Other tips

Note the sysctl dev.raid.speed_limit_max (in KB/s), this might affect (limit) the initial sync speed, periodic raid checks and potentially the raid re-build.

While trying to do a fast rsync, we experimented with a very large instance type (c5d.18xlarge, 144GB RAM) and with vm.vfs_cache_pressure=2, to keep as many inodes and dentries in kernel caches (see slabtop, we eventually had 60M of inodes cached, 28M inodes and 15T synced in 6.5hours). We had also decreased the dirty_ratio and dirty_background_ratio to have more frequent syncs considering the large RAM.