Backup Alternatives

rfx77

The problem we have with XOA Backup is that it does not support Deduplication like Kopia, Borg or Restic. So we have to do our VM Backups with a script.

The problem we faced with a simple solution like dump xva or ova files to kopia/restic/borg is that the tar file-format and the metadata (directory name Ref...) is different on every export (even when you export the same snapshot twice) and because of that dedup does not work (well). we had to do separate export for metadata and VDIs but because of this a restore is not unattended.

I looked a little bit into the implementation of the XOA OVA export code and in my opinion, it should be possible for Vates to alter the format so that is could be deduplicated well. you than can easily stream it via stdin to kopia. kopia can also dump to stdout so restore should also be possible.

olivierlambert

Is deduplication such a big deal? I mean, the cost of it is not negligible and (at least with SMAPIv1) the gain isn't great at all.

rfx77

@olivierlambert

With dedup i mean dedup of the backup repo. And its an extremely big deal for backup repos. Backup vendors put a lot of effort in this. CommVault and NetBackup have it natively and Veeam does recommend deduplicated storage (Data Domain,...).

In our case we use an S3 backend and if you pay for a month worth of full backups 2-3 times a full backup size or 30 times a compressed full backup size is really a big deal. And we keep out backups sometimes for month.

We are doing bi-weekly fulls with our homebuilt kopia strategy of a 1.3 TB SR (no empty disks) and are using apx. 1.37 TB S3.

its from 0.64 TB up to 1.37 TB. Green is size used and black are the blocks which could be deleted.

if someone is interested why we use kopia instead of restic. restic is super slow on dumps to stdout and on mount. we use mount in combination with libguestfs to do file-level-recovery.

this is the script. not super professional and not the best error-handling but it may be useful:

#!/bin/bash



VMNAME=$1
SERVER=$2
SERVER=${SERVER:=172.25.10.2}

BACKUPDESC="Backup $(date +'%Y-%m-%d %H:%M:%S') of VM $VMNAME"

echo "$(date +'%Y-%m-%d %H:%M:%S') Starting Backup of vm $VMNAME"


echo SERVER:$SERVER


PASS=$(cat /root/.pass)
export XE_EXTRA_ARGS="server=$SERVER,port=443,username=root,password=$PASS"



VMID=$(xe  vm-list name-label=$VMNAME | grep uuid | sed "s/.*: //g")

if [ -z "$VMID" ]
then
  echo VMID not found
  exit 1
fi

echo VMID: $VMID

SNAPID=$(xe vm-snapshot uuid=$VMID  new-name-label=backup_$VMNAME)

if [ -z "$SNAPID" ]
then
  echo SNAPID not found
  exit 1
fi


echo SNAPID: $SNAPID

declare -a DISKID
declare -a DISKNAME


TEMPD=$(mktemp -d)
mkdir $TEMPD/data

echo "$(date +'%Y-%m-%d %H:%M:%S') Temp directory: $TEMPD"
xe vm-export metadata=true uuid=$VMID filename=$TEMPD/data/metadata.ova.tar
xe snapshot-disk-list uuid=$SNAPID vdi-params=all vbd-params=all > $TEMPD/data/vdi.txt
xe vm-list name-label=$VMNAME params=all > $TEMPD/data/vm.txt

ls -l $TEMPD/data



DISKID=( $(xe snapshot-disk-list uuid=$SNAPID vdi-params=uuid,name-label vbd-params=false | grep uuid | awk '{print $5}') )
DISKNAME=( $(xe snapshot-disk-list uuid=$SNAPID vdi-params=uuid,name-label vbd-params=false | grep name-label |  awk '{print $4}') )
echo DISKID: ${DISKID[@]}
echo DISKNAME: ${DISKNAME[@]}

for i in ${!DISKID[@]}; do
  DN=${DISKNAME[$i]}
  DID=${DISKID[$i]}
  echo "$(date +'%Y-%m-%d %H:%M:%S') tagging disk $i is ${DID} ${DN}"
  xe vdi-param-set name-description="$BACKUPDESC" uuid=$DID
done



for i in ${!DISKID[@]}; do
  DN=${DISKNAME[$i]}
  DID=${DISKID[$i]}
  echo "$(date +'%Y-%m-%d %H:%M:%S') backup disk $i is ${DID} ${DN}"

  xe vdi-export uuid=$DID filename= format=raw | kopia snapshot create --tags=vm:$VMNAME,data=$DN --no-progress --log-level=warning --override-source=/vms/$VMNAME/$DN --stdin-file=${DID}.vhd - || echo "ERROR"
  sleep 5
  xe vdi-destroy uuid=$DID
done



tar cvpf $TEMPD/metadata.tar -C $TEMPD/data .
kopia snapshot create  --tags=vm:$VMNAME,data=meta  --no-progress --log-level=info --override-source=/vms/$VMNAME $TEMPD/metadata.tar || echo "ERROR"
rm -rf $TEMPD
echo "$(date +'%Y-%m-%d %H:%M:%S') waiting 30s"
sleep 30

echo "$(date +'%Y-%m-%d %H:%M:%S') deleting snapshot $SNAPID of VM $VMNAME"
xe snapshot-uninstall uuid=$SNAPID force=true || echo ERROR
sleep 20
xe snapshot-uninstall uuid=$SNAPID force=true > /dev/null 2>&1 || true



echo "$(date +'%Y-%m-%d %H:%M:%S') Done Backup of vm $VMNAME"

olivierlambert

We tried dedup for the backup repo with ZFS, to see the ratio kept. Due to VHD fragmentation, the gain was between 10 to 30% tops. Far far from being useful. That's why I think it would makes sense with SMAPIv3 and not v1 (because of the VHD format and its fragmentation)

edit: also a simple script like this won't be able to deal with chain protection and many other important features.

rfx77

@olivierlambert The problem is what you dedup.

We dedup raw VDI exports and they are nearly perfect dedup candidates.
I attached our script in the previous post.

With XVA i also only get 25% dedup ratio because of the tar metadata-problem. We decided that this is useless.

olivierlambert

We tested to dedup the VHD exported. Your dedup ratio will diminish inexorably with time, as your VHD will become more and more fragmented. In raw, it should be similar at some point, since blocks aren't removed after usage (no trim support in SMAPIv1)

rfx77

To clarify: In a XVA the folder names are of the kind Ref... and the number in the folder-name is random. This kills the dedup. If we could force the folder-name to be like Disk_[1,2,3,4] XVAs would be very good deduplicatable

rfx77

@olivierlambert
We are using dedup backup-stores for years now and yes they grow if you keep very long time-ranges. but they never grow as much as compressed fulls. and if you only keep shorter ranges (1-2 month) than the dedup ration will be nearly constant after some month.

I just checked the CommVault-Repos of our customers and they are all above 70% dedup saving and about 50% of the repos are above 90% dedup saving.

Depending on the size of their backup NAS they all have retention seetings between 60 days and one year.

Our own XenBackup S3 repo for Kopia will stop to grow when we delete older snapshots. for now we just kep backing up and this is the reason why it grows.

i removed the first column because it would show confidential customer information.

nikade

Dedup is a huge deal, in our SAN we're seeing a 6.7:1 ratio which means we're saving a ton of money. With that said, Veeam for example also has dedup but it totally sucks.
I think it might be hard to get a good ratio with backup-data.

olivierlambert

@rfx77 I don't know how you can keep shorter range if you do forever incremental The fragmentation will only grow

KPS

Is there any possibility to disable compression?
I think, it could be a good option to fully give ZFS control over the remote:

ZFS can do compression
Without "external compression", deduplication could be much more efficient

olivierlambert

We benched ZFS dedup without compression and the result was bad. I'm convinced everything will change with a proper data path that will support trim.

rfx77

@olivierlambert dedup has nothing to do with forever incremental.

if you are doing backups, only the changed blocks are stored and the blocks which are not used in any backup are deleted. all blocks are backuped but the software stores only the blocks which are new

with our kopia solution we do always full backups. kopia manages the dedup. with commvault we do fulls and incs

i think that most people here are thinking to much backup as the way veeam does it with reverse-inc and forever inc. veeam does it in a very specific way. with commvault, netbackup,... we do classic fulls and incs.

i dont know how to describe it better. if you do an incremental backup you only backup the changed blocks. when you do a full afterwards these blocks are already in the backup repo so they dont have to be written. you can do incs and fulls and you will only have the blocks in the repo which are necessary to restore these backups. so one year of daily fulls or incs may not use more then 3-4 times a full in the backup repo.

rfx77

To be clear. i never talked about deduplicated storage. i am always talking about deduplicated backup store. the edup is not done by the fs but by the backup-software

deduplicated storage is a completely different topic which can be addresse with sapiv3 but its not the think my posts are all about

rfx77

@olivierlambert we are talking about different things. i dont mean dedup storage. i am talking about deduplicated backup store. the dedup is done by the backup-software (kopia,restic,borg, commvault,...) they support various backends like s3, ftp, sftp, local disk,...

rfx77

@nikade You can see the ratios in the picture above. they are all around 90%

olivierlambert

@rfx77 That's exactly what I'm talking about. As I said, we tested dedup ratio with exported VHD files and it tends to diminish pretty quickly as you have more and more fragmentation, at a point it's worthless. But, yes it works pretty well if you have similar templates underneath and little fragmentation.

wtdrisco

@olivierlambert this backup looks really good. I currently in my VMware space use veeam. It's ok. I mainly just run backups of my virtuals and move them to an offline storage so if an encryption hack gets in, they can't encrypt the disks. So I run a script to move the file through a link to a Synology. That is just the VMs. For all my SQLs I run smss jobs to back up dbs and then transaction logs every 15 min. Again moving to offline storage. That gets me at least 15 min recovery to the crash. So I really do not focus too much on the backup software but to get me a VM back but my data i do separately so I have to ability to restore locally or move it to external resource. People here have some good discussions on the backup. Good thread

rfx77

@olivierlambert fragmentation does not matter with deduplication (at least how dedup is used in backup repos). i dont know why it does not work in your case. we do raw and not vhd. maybe thats a point. there can also be a problem when the guest-filesystem is zfs or btrfs because how those filystem alain their blocks.

in general (with non ROW/COW guest fs) fragmentation should not do any harm. at least thats what we see at our customers backup repos. and they have windows and different linux guests. we do not use base templates so there should not be any hidden benefit of something.

olivierlambert

We tested with VHD format, not with raw. We could do a test, but with raw you can't do deltas.