IT Notes - restore

How we are migrating (many of) our servers from Linux to FreeBSD - Part 3 - Proxmox to FreeBSD

Stefano Marinelli — Tue, 14 Mar 2023 13:00:00 +0000

In recent years, we've been migrating many of our servers from Linux to FreeBSD as part of our consolidation and optimization efforts. Specifically, we've been moving services that were previously deployed using Docker onto FreeBSD, and it has proven to be a great choice for handling workloads efficiently.

To this end, we've also been migrating many of our virtual machines (VMs) to FreeBSD, deploying services within FreeBSD jails. In some cases, these jails have even replaced entire VMs and run bare metal. Although we prefer to move to native FreeBSD whenever possible, sometimes it's not the best option for all the services we offer. As a result, one of our most critical physical servers has been left behind for years.

This server was a Proxmox server that we installed many years ago and updated to version 6.4. It hosted some critical services, but upgrading to Proxmox 7.x posed some challenges. In particular, some of the LXC containers required tweaks.

Unfortunately, this server was quite old, with only four physical disks and 64 GB of RAM. It was located in an OVH data center and had been running well until one of the disks started to malfunction once a week, on Sundays. This would trigger a RAID reconstruction that kept the system busy for about two days.

Despite my preference for simple setups, this server had been deployed gradually over many years, and everything was tied together. As a result, unraveling the system to resolve the issues was not a simple task. Sometimes the combination of simple things can make everything complex.

The Proxmox Server

The Proxmox server was configured as the central hub for various services, including primary DNS, web hosting, VOIP, and more. It featured several bridges, each with its own specific purpose, and was connected to a virtual machine running MikroTik CHR. This machine was responsible for consolidating all incoming VPNs from the MikroTik devices we managed, both ours and those belonging to our clients. Additionally, it provided a series of bridges to manage these devices and all server management VPNs and other services. The Proxmox server also housed several virtual machines running Linux, FreeBSD, OpenBSD, and NetBSD, as well as LXC containers.

Over the last two years, we've been migrating most of these virtual machines and containers to FreeBSD-based VMs, which feature their own specific jails. Consequently, most of the VMs we've had to move were BSD-based, while only five Linux VMs remained. The LXC containers hosted a range of services, including servers managed by Virtualmin, a large installation of Zimbra (which was hosted within an LXC container running CentOS 7), as well as some minor Alpine Linux-based machines. We located all these virtual machines and containers in a LAN created and managed by CHR. All public IPs were managed by CHR, which relied on NAT mappings to establish communication between them. CHR had thus become the heart of our system, and if it experienced any issues, it could potentially take down the entire system. Fortunately, it remained stable for years.

Migration - first steps

The first step I took was to install FreeBSD on the new server. Easy peasy. The next step was to find a way for the CHR to migrate to the new server (under bhyve) and continue to manage all the public IPs of the original server. The problem is that OVH, with its failover IPs, ties a specific MAC address to each individual IP address. Therefore, the only way was to create a bridge on the FreeBSD server (on the Proxmox server, I already had the bridge on the physical network card) and create an L2 tunnel between the two servers - I used OpenVPN with tap interfaces, specifically inserted into the bridges. I could have used other methods and techniques, but I wanted to experiment with a setup that could allow, if necessary, to bridge a larger number of physical and virtual servers even if the IPs are all mapped to a single server. OVH does not allow, in fact, the splitting of classes, so a move must be made for the entire class, not for a single IP address.

Initially, MikroTik CHR 7 did not boot on bhyve. In the end, I managed to make it work, but I had other problems, probably related to the MTU of the interfaces. So I thought about taking the opportunity to unbind the LXC containers and VMs from CHR and remove MikroTik from the setup. With RouterOS version 7, in fact, Wireguard-based VPNs are also supported, so within a few days, it was possible to update the few routers still on 6.x and recreate some VPNs using Wireguard. I mapped both the VMs and LXC containers directly to their respective public IPs, greatly simplifying the steps. Everything worked perfectly.

The next step was to test the first migrations, starting from the VMs already on FreeBSD. For simplicity, I created a new FreeBSD VM in bhyve and copied (via zfs-send and zfs-receive) the datasets related to BastilleBSD. All services are installed in jails managed by Bastille, so this was enough to have, in a short time, a new operating server equivalent to the previous one. At that point, I shut down the original server, connected the VM to the bridge linked to the tunnel (after modifying its MAC address), turned on the new FreeBSD VM (on bhyve), and everything started to work correctly - but from the new physical server.

One by one, I moved all the FreeBSD VMs. For Linux, NetBSD, and OpenBSD, I simply copied the images and pointed bhyve to them. Some small specific configuration on vm-bhyve and everything started to work correctly. Where possibile, I replaced the “virtio” with “nvme” as it performs much better on bhyve.

Migration - LXC containers to Virtual Machines

For LXC containers, I initially thought of creating an Alpine Linux virtual machine, installing LXD, and copying each individual container. It worked for some of them, but for others, I started to encounter strange issues, similar to those that would have required manual intervention to upgrade from Proxmox 6.x to 7.x. As is often the case with Linux-based solutions, compatibility is not always preserved between updates, so I would have had to fine-tune all the containers, which I didn't feel like doing. The containers had been created (at the time) to optimize RAM usage on the Proxmox machine, but to date, they have caused more problems than benefits. In some cases, certain processes got "stuck," making it impossible to "reboot" the LXC container, requiring the entire physical node to be rebooted. If they had been virtual machines, I could have given a "kill" command from the virtualizer (to the respective KVM process, in that case) and restarted it.

For greater compatibility and ease of future management, I decided to convert the LXC containers into actual VMs on bhyve. The process was simple:

Creating an empty VM with vm-bhyve and booting the VM with SystemRescueCD.
Creating destination partitions and file systems in the VM, then doing a complete rsync of the original LXC container.
Adjusting the fstab file, installing the kernel on the destination VM, and creating the initrd (some containers were already copies of VMs, so the kernel remained installed and updated, even though it wasn't being used. The initrd, on the other hand, did not include the nvme or virtio drivers, so I had to regenerate it anyway.)
Adjusting the bhyve vm configuration file, doing one last rsync after shutting down the services, shutting down the original LXC container, and starting the bhyve VM.

Everything worked correctly, so one by one, I moved all the containers. The largest one ended up on another physical node (also FreeBSD with bhyve) temporarily because the space on the new server was not sufficient to contain it. It didn't need to be on this server, so no problem.

One by one, the LXC containers started on the new server. Apart from some minor adjustments to the destination VMs (different network interface names, etc.), I didn't encounter any particular problems even after several days. Everything works perfectly.

At the very end, I re-created the MikroTik CHR VM. I’ll keep this setup separate for now, as strictly tied to eoip interfaces. This was the main reason why I haven’t performed the migration before. Things were too tied together and I had to untie everything, step by step.

…and then one of the Linux VMs started to freeze

Several Linux VMs are just the basis on which Docker runs. One of them (not even among the busiest) started, every 12/15 hours, to completely freeze. It stopped responding to ping, and it was impossible to give any type of command from the console. In a word: stuck.

Searching the web, I found some references to this problem and, observing the errors of an ssh session that was left connected (stuck, but still showing the last error), I found it to be a problem similar to the one described in this post, namely:

"watchdog: BUG: soft lockup - CPU#0 stuck for 22s! [khugepaged:67]"

I tried various solutions such as changing the storage driver, the number of cores, the distribution (from Alpine to Debian), etc., but none of these operations solved the issue. I also noticed that the problem occurs with all Linux VMs, but only those with a recent kernel (> 5.10.x) freeze, while the others continue to work. The problem does not occur, however, with the *BSDs.

In the end, I:

Reduced the number of cores to 1 for the VMs that did not have a high load (some remained with multiple cores), hypothesising a problem with allocating cores that were too busy
Gave the command: "/usr/bin/echo 60 > /proc/sys/kernel/watchdog_thresh" to the VM.

The VM became stable, and I have not seen that error/warning on any other machine since. I will investigate further, but I believe it is a problem related to the Linux kernel, which, for some reason, generates a kernel panic if particular situations of CPU concurrency are generated.

The End…and a nice OOM!

After moving everything, I was finally able to migrate the entire class of OVH IPs from one physical server to another. The operation was quite quick, but in order to avoid problems, I notified all users and performed the operation on a Sunday and during off-peak hours. The whole process took about 10 minutes and there were no hitches of any kind.

For safety reasons, I kept the Proxmox machine active for a few more days, but there was no need to use it. However, after a couple of days, I encountered a problem: the largest VM, in some cases, was being "killed" because FreeBSD generated an OOM. I had never seen, from FreeBSD 13.0 onwards, any OOM related to "abuse" of RAM usage by ZFS, but in this case, it actually happened.

In the end, I understood that ZFS, on FreeBSD, is able to release memory, but not quickly enough to manage any "spikes" in individual VMs. In fact, the VMs do not know the situation of the physical host's RAM, so they will tend to occupy all the space allotted to them (even if only for caching). A sudden spike (i.e. if you create and launch a new VM) could cause a sudden increase in RAM usage by the bhyve process, and FreeBSD could be forced to kill it, even if part of the RAM is only ARC cache. While Proxmox supports HA (i.e., control over whether the VM is running), vm-bhyve only launches the VM (bhyve process). I should manage it with tools like monit, but for now, I preferred to simply set limits on ZFS RAM usage using "vfs.zfs.arc_max", and there have been no more problems.

Final considerations

The operation was long but linear. The most complex part was unraveling all the configurations related to MikroTik CHR and the VPNs linked to each individual LXC machine/container. Once everything was implemented on a dedicated VM, the operation was fairly straightforward.

The hardware specifications of the destination physical server are slightly better than the starting one, but the final performance of the setup has greatly improved. The VMs are very responsive (even those that were previously LXC containers running directly on bare metal) and, thanks to ZFS, I can make local snapshots every 5 minutes. In addition, every 10 minutes, I can copy (using the excellent zfs-autobackup) all the VMs and jails to other nodes both as a backup and as an immediate restart in case of disaster. I just need to map the IPs, and everything will start working very quickly. Proxmox also allows you to perform this type of operation with ZFS, but you still need to have Proxmox (in a compatible version) on the target machine. With the current setup, I only need any FreeBSD node that supports bhyve.

Proxmox is an excellent tool, well-developed, open-source, efficient, and stable. We manage many installations, including complex ones (ceph clusters, etc.), and it has never let us down. However, not all tools are ideal for all situations, and for setups like the one described, the new configuration based on FreeBSD has shown significantly interesting performance and greater management and maintenance granularity.

Virtualizing on vm-bhyve is not complex, but it is certainly not comparable, at the current state, to the simplicity of using a clean and complete interface like Proxmox's. A complete HA system is still missing (sure, it's achievable manually, but...), as well as complete management web interface. However, for knowledgeable users, it is undoubtedly a powerful tool that allows you to have excellent FreeBSD as a base. I'm totally satisfied with my migration and the result is far better than I expected.

How we are migrating (many of) our servers from Linux to FreeBSD - Part 2 - Backups and Disaster Recovery

Stefano Marinelli — Mon, 30 May 2022 03:03:52 +0000

After my post on why we’re migrating (most of) our servers from Linux to FreeBSD, I’ve started to write about how we’re doing it. After covering a basic installation (we’re doing a massive use of jails), I’m going now to describe how we’re performing backups.

Backup is not a tool. Backup is not a software you can buy. Backup is a strategy you need to study and implement to be able to solve your specific problems. You need to understand what you’re doing, otherwise you’ll always have a Schrödinger’s Backup - it may work or not and if you don’t test it well enough (i.e. restore) you’ll find out when it’s too late.

We’re performing backups in many different ways but, for our physical and virtual FreeBSD servers, we have a dual approach. We need both a “ready to use” backup (that will be described here, useful for a fast disaster recovery or prompt restore of specific jails) and a “colder”, more space efficient backup that can be kept for months (or years), more similar to the borg approach on previous posts. Generally speaking, we store our OS (and jails) on ZFS, so I’ll describe this kind of approach here.

Disaster recovery backup - ZFS send/receive

BastilleBSD creates its datasets and mounts them on /usr/local/bastille . There are no databases, all the jails’ configurations are inside that mountpoint so it’s quite easy to backup and restore all the jails or any single jail in one go. More, the “everything-in-a-jail” approach simplifies the restore process as you don’t need to restore the entire host OS, just install an empty FreeBSD server, install BastilleBSD and restore the jails. Or add the jails to already existing FreeBSD systems.

We normally use FreeBSD (or Linux with ZFS) backup servers, well protected and encrypted at rest. For the ZFS send/receive approach, our servers are NOT reachable from the outside. We can ssh into them only using a VPN - they’re too precious to be exposed on the World Wild Web - or, if strictly needed, we expose ssh only using keys, no passwords. We perform the backups using a pull strategy: the backup server connects to the production servers, gets the data, disconnects. The production servers have NO ACCESS to the main backup server. Should they ever be seriously compromised, the backup is safe.

There are many tools that can help to set up this kind of configuration. I’ve tried many of them and found that they all have some good and bad points. The one I decided to use for our servers is zfs-autobackup. It’s easy to use, everything can be set via command line and has a good cron (or Jenkins) output, useful to understand if everything is right.

Let’s consider two servers, one is called “ProdA” and the other is called “Bck” - we obviously want to backup the ProdA into Bck.

Installing ProdA has been covered on a previous post, Bck is quite simple and outside the scope of this post. We just need a protected zfs FreeBSD (or Linux) server. That’s all. Let’s assume that ProdA has a BastilleBSD zfs dataset (and children datasets), with jails and everything needed, as configured in the last post. We now need to install the needed software. On Bck:

pkg install py311-zfs-autobackup mbuffer

On ProdA:

pkg install mbuffer

mbuffer will be used as a ram buffer to avoid read/write spikes (or slowdowns) while sending/receiving the snapshots.

It’s time to prepare the destination dataset. Assuming that Bck has a zroot base dataset, we’ll be creating (as root) a zroot/backups/ProdA

zfs create -p zroot/backups/ProdA

Ok, let’s now go to ProdA

We want to create an unprivileged user that will send the data. We don’t want to allow Bck to connect as root, even if it’s trusted and secure. Let’s create a user called “backupper”. Then, we need to give backupper the right permissions:

zfs allow -u backupper send,snapshot,hold,mount,destroy zroot

Note: if you want Bck to be able to delete the snapshots on ProdA, backupper needs the destroy permission. That means this user can destroy the whole system as can ALSO destroy zroot (or any source dataset you decide). If you’re afraid of this, different approaches must be used (i.e.: local root performing snapshot/cleanups and Bck only transferring them, not hard to achieve with zfs-autobackup). Considering that the Bck is safe, secure and protected, we can tolerate this weakness. Just be sure nobody can break the “backupper” user. Do not use password, use ssh keys and treat this user with the same care you'd use with root.

Now, as root on ProdA:

zfs set autobackup:bck_server=true zroot

We’re setting a custom property, called “autobackup:bck_server”, allowing the zroot (and children) dataset to be backed up by zfs-autobackup. zfs-autobackup will search for all datasets with that property set to “true” (also on different pools) and will backup them. If there’s a specific dataset you don’t want to backup, just set it to “false”. Or if you don’t want to backup the entire zroot but, for example, only “zroot/bastille” (and children), just set autobackup:bck_server=true for that dataset.

ssh config

zfs-autobackup will connect via ssh and zfs-autobackup will try to connect as root. Moreover, even after exchanging the ssh key, Bck will connect many times to ProdA to send its zfs commands (one connection per command). Ssh session initiation is quite long, so there will be some latency. In order to (greatly) speed up this time,

“You can make your ssh connections persistent and greatly speed up zfs-autobackup:

On the server that initiates the backup add this to your ~/.ssh/config:

Host ProdA
User backupper
ControlPath ~/.ssh/control-master-%r@%h:%p
ControlMaster auto
ControlPersist 3600

(Taken from https://github.com/psy0rz/zfs_autobackup/wiki/Performance)

It's now time to go back to Bck and issue a command like this (one line):

/usr/local/bin/zfs-autobackup --ssh-source ProdA bck_server zroot/backups/ProdA --zfs-compressed --no-progress --verbose --buffer 32M --keep-source 0 --no-holds  --set-properties readonly=on --clear-refreservation --keep-target 1d1w,1w1m,1m6m  --destroy-missing 30d --clear-mountpoint

Bck will connect to ProdA, perform the snapshots and start transferring. The most interesting options I used here are:

 --keep-source 0 (only the last snapshot will be kept on ProdA)
 --set-properties readonly=on (be sure the Bck clone is read only, so we will be able to perform an incremental/differential backup next time)
 --keep-target 1d1w,1w1m,1m6m (keep one backup per day for one week, one per week for one month, one per month for six months)
 --destroy-missing 30d (if we've deleted a dataset, keep it for 30 days before removing it from Bck)
 --clear-mountpoint (do not mount the dataset in Bck, as it will cause problems sooner or later)

The first copy will be slow as it'll need to send all the data. The second will be quite fast as only the differences will be transferred.

How to perform a disaster recovery

Ok, your dataset (or datasets) has gone. You need to replace it with the last external backup. You have to retransfer the copy into ProdA (or another FreeBSD host, no difference). Connect to Bck and search for the snapshot you want to restore (zfs list -t snapshot will help). Once identified (one line):

zfs send -R zroot/backups/ProdA/zroot/bastille/bastille/jails/t1@bck_server@20220528005830 | mbuffer -4 -s 128k -m 32M | ssh root@ProdA "zfs receive -F -x canmount -x readonly zroot/bastille/bastille/jails/t1"

Note the -x canmount -x readonly flags. Remember that we altered the canmount and readonly properties of the transferred datasets during the backup, so we must restore them into a normal state.

Once finished, ProdA (or the other, restored host) will show t1 as an available jail and you'll be able to start it.

Efficient backup of lxc containers in Proxmox

Stefano Marinelli — Tue, 06 Oct 2020 08:22:44 +0000

Sometimes an lxc container can be a better alternative than a KVM virtual machine. It has a smaller overhead, easier and more efficient resource management and lower impact on the physical machine.

You get (potentially) less isolation from the physical machine and you are forced to use the host kernel. Moreover, you cannot run different a different OS.

Proxmox has a great support of lxc containers and they're often a good choice. I've recently converted some VMs to lxc containers and they're faster, lighter and the Ram occupation is quite lower.

There are two major drawbacks: you cannot migrate a running lxc container (it will be shut down, moved and restarted - if using a shared storage, it is often a matter of seconds) and backups to Proxmox Backup Server are slower than a VM's backup.

This is because a VM can monitor, thanks to KVM dirty bitmaps, its disk operations to know in advance which blocks should be checked and copied. When dealing with lxc containers it can't be done, so every backup will need to check every single file - this proved to be efficient in terms of space and deduplication, but quite slow in terms of time. One of my VMs (750 GB stored but with rare disk access) took just a few minutes to be backed up with dirty bitmaps, but more or less 2 hours to backup since it's been transformed to a lxc container.

borg backup proved to be quite efficient, so I tried to find a way to backup those containers performing a snapshot, backing up the snapshot and, then, releasing it.

We cannot use dattobd on the container, so we must snapshot and backup from the host. This has one big advantage: we don't need the container to even know its backup location as every operation will be done by the Proxmox host.

My containers generally are stored in Ceph rbd volumes or LVM thin pools.

I've created a script to deal with both those situations - and it doesn't depend on Proxmox, so it also works if you don't use it.

First of all, create a user and a borg repository on a remote machine and allow the Proxmox host to login via ssh using a key. Then, create a directory (/tobackup in the following scripts, on the host) where the snapshots will be mounted.

In the following example/scripts, no encryption will be used for the backup storage. If you want to encrypt your backup, you will have to perform small modifications.

Here's my rbd_borg_bck.sh :

#!/bin/bash
set -e

export SOURCE_DISK=$1
export USERNAME=$2
export HOST=$3

PATH="/usr/local/jdk/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/bin/X11:/usr/pkg/bin:/usr/pkg/sbin"
export PATH

export BORG_UNKNOWN_UNENCRYPTED_REPO_ACCESS_IS_OK=yes;

echo "Creating and mapping snapshot..."
rbd snap create $SOURCE_DISK@tobackup && DDISK=`rbd map $SOURCE_DISK@tobackup`

echo "Creating backup directory..."
mkdir -p /tobackup/$DDISK

echo "Mounting snapshot..."
mount -o noload $DDISK /tobackup/$DDISK

echo "DOING BACKUP..."

REPOSITORY=$USERNAME@$HOST:$USERNAME/
TAG=daily

borg create -v --stats --progress --compression zlib,9                           \
    $REPOSITORY::$TAG'-{now:%Y-%m-%dT%H:%M:%S}'          \
    /tobackup/$DDISK/                                       \
    --exclude '*/home/*/.cache*'                  \
    --exclude '*/home/*/.local*'                  \
    --exclude '*/home/*/.pki*'                    \
    --exclude '*/home/*/Virtualbox VMs*'          \
    --exclude '*/home/*/.vagrant.d*'              \
    --exclude '*/root/.cache*'                    \
    --exclude '*/var/swap*'

sleep 3s;

echo "Unmounting snapshot..."
umount /tobackup/$DDISK

sleep 3s;

echo "Unmapping snapshot..."
rbd unmap $DDISK

echo "Remove snapshot..."
rbd snap remove $SOURCE_DISK@tobackup

echo "Pruning..."
borg prune -v $REPOSITORY --stats --prefix $TAG'-' \
    --keep-daily=14 --keep-weekly=8 --keep-monthly=12

echo "Done!"

For example, to backup a disk stored in pool "vmrbdpool" named 'vm-101-disk-0' via ssh to server backupserver.mydomain.org, username on the server (and repository name, inside that user's home directory, called server01), you can launch the script this way:

/usr/local/bin/rbd_borg_bck.sh vmrbdpool/vm-101-disk-0 server01 backupserver.mydomain.org

This can also be scheduled and launched via cron (on the Proxmox host) or any other way.

For the LVM thin pools, the script is a bit different. Here's my lvm_thin_borg_bck.sh

#!/bin/bash
set -e

export SOURCE_DISK=$1
export USERNAME=$2
export HOST=$3
export LVM=$4
export DDISK=$4/$2-borgsnap

PATH="/usr/local/jdk/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:/usr/bin/X11:/usr/pkg/bin:/usr/pkg/sbin"
export PATH

export BORG_UNKNOWN_UNENCRYPTED_REPO_ACCESS_IS_OK=yes;

echo "Creating and mapping snapshot..."
lvcreate -n $2-borgsnap -s $1
lvchange -ay -Ky $DDISK


echo "Creating backup directory..."
mkdir -p /tobackup/$DDISK

echo "Mounting snapshot..."
mount -o ro /dev/$DDISK /tobackup/$DDISK

echo "DOING BACKUP..."

REPOSITORY=$USERNAME@$HOST:$USERNAME/
TAG=daily

borg create -v --stats --progress --compression zlib,9                           \
    $REPOSITORY::$TAG'-{now:%Y-%m-%dT%H:%M:%S}'          \
    /tobackup/$DDISK/                                       \
    --exclude '*/home/*/.cache*'                  \
    --exclude '*/home/*/.local*'                  \
    --exclude '*/home/*/.pki*'                    \
    --exclude '*/home/*/Virtualbox VMs*'          \
    --exclude '*/home/*/.vagrant.d*'              \
    --exclude '*/root/.cache*'                    

sleep 3s;

echo "Unmounting snapshot..."
umount /tobackup/$DDISK

sleep 3s;

echo "Remove snapshot..."
yes | lvremove $DDISK

echo "Pruning..."
borg prune -v $REPOSITORY --stats --prefix $TAG'-' \
    --keep-daily=14 --keep-weekly=8 --keep-monthly=12

echo "Done!"

In this example, we want to backup a disk "/dev/prox1/vm-120-disk-0", username on the server is server01 (same repository name), server is backupserver.mydomain.org and, as last argument, "prox1" which is the lvm pool.

/usr/local/bin/lvm_thin_borg_bck.sh /dev/prox1/vm-120-disk-0 server01 backupserver.mydomain.org prox1

Borg is efficient and fast and, thanks to this kind of setup, I can create a backup of the 750 GB container in 2 minutes.

Thanks to this kind of setup, backups can be done in a consistent and efficient way.

Proxmox Backup Server - hints for a perfect deployment

Stefano Marinelli — Sun, 23 Aug 2020 18:16:57 +0000

Proxmox Backup Server (PBS) has been released. It's still in beta but is already perfectly usable. After many years, it's now possible to perform incremental backups of the VMs and thanks to the qemu dirty bitmaps, backups are also fast.

The documentation is well-done and clear, so I suggest to have a look at it for all its options and features.

In short, it allows to backup any Proxmox VM or Container to a specific PBS server (which can also run Proxmox or not - it's not wise to backup on the same server where the VMs are running). It doesn't require any special file system nor setting as it splits files in chunks (so deduplication is possible and efficient) and can be stored in any supported FS.

I've been using it since the first day and it's quite efficient and reliable, I've backed up hundreds of servers and restored many of them without any problem.

Here's some hints (this post will be updated as needed):

The first backup will be slow - at least as slow as a traditional Proxmox backup. Don't worry, it's perfectly normal.
DO NOT TRIM THE VMs, but use the "discard" option at mount time (if Linux): trim will pass all the dirty blocks and zero them, filling the dirty bitmaps database and slowing down the next backup (even if it will not be larger).
For the same reason, if you want to keep multiple backup servers, remember to backup to ONLY ONE PBS server and then sync it to another one. Backing up the same VMs to multiple servers will confuse the dirty bitmaps and every backup will be slower (not larger), even if it will work.
Sometimes syncing to another server can fail and get stuck. At the moment (0.8.11-1), I found that stopping won't solve the issue and you have to restart the PBS, otherwise the task would seem to be running.

Proxmox Backup System has solved a long time problem: performing efficient, deduplicated, incremental Proxmox VM backups.