Random Musings

O for a muse of fire, that would ascend the brightest heaven of invention!


Restic for fast backups & restores

Tuesday, 16 Jun 2020 Tags: backupsfreebsdrestic

restic is a high speed backup and recovery tool written in golang. It supports multiple remote storage options, and handles dedupe and encryption by default. It has a matching restserver remote backup daemon that enables you to run your own remote storage option, and can conveniently be used directly, for local backups as well.

It is purely command line driven, which you may consider as an asset.

architecture and configuration

The architecture is boring and standard - multiple backup clients, writing over network, to a single backup server with a large partition. Clients can work in parallel, and use a per-client encryption key, and are unable to remove backed up data themselves

  • an append-only configuration.

We’ll nominate a server, with a hostname of backup.example.net, as our backup server. It uses FreeBSD 12.1R and zfs by default, but this setup should work on any supported FreeBSD version, and you can just skip the zfs functions I use below.

Paranoid people can use zfs replication to a further offsite location, or to a tape drive, as required, for further safekeeping.

backup server

The backup server is shared both locally (direct file system), and over the network. This allows an admin to prune older backups, and prevents backup clients from doing so.

The restic(1) tool operates directly on the local respository, and restserver makes it available over the network to clients with appropriate credentials.

packages

Install packages sysutils/rest-server sysutils/restic, and note that it creates a user, group, and a directory /var/db/restserver where all restic backup blobs will be stored:

restserver:*:239:239::0:0:restic rest-server Daemon:/var/db/restserver:/usr/sbin/nolrgin

the dataset

The restserver daemon will store its tiny config files and secret key in its homedir, which by default is the same /var/db/restserver dir.

Let’s create a new zfs dataset, with a reasonably large record size, and clean up permissions:

# zfs create \
    -o recordsize=1m \
    -o exec=off \
    -o setuid=off \
    zroot/var/db/restserver
# chown root:restserver /var/db/restserver
# chmod 0750 /var/db/restserver

rc.conf settings

By default, we only allow remote backups to append data, it’s mildly hacky but gives us a little safety in case of a moderately malicious attacker, or incompetent sysadmin, and we use private repos which effectively puts each backup client in its own tidy world.

# /etc/rc.conf.d/restserver 
restserver_enable=YES
restserver_options="--no-auth --listen 127.0.0.1:8000 --private-repos --append-only"

Amend the IP address and port as you like, in my case I have spiped in front of restserver, to allow backup from my laptop via internet, and haproxy for more discerning backups via our zerotier mesh vpn. You can read up on these options in the restserver docs.

initialising the backup repo

The following steps should be done as the root user, to ensure secured permissions. We’ll generate a key from random data, use this for the repository, and ensure that the restic user can’t manipulate the key or secret.

# cd /var/db/restserver
# umask 077
# head -c 1000000000 /dev/random \
    | sha512 \
    | tee .restic
# restic init --repo /var/db/restserver --password-file .restic  
# chown root:restserver .restic config /var/db/restserver
# chmod 0750 /var/db/restserver
# chmod 0440 .restic config
# chown -R restserver:restserver /var/db/restserver

As noted, store this passphrase carefully, possibly duplicated – you cannot restore without it!

start the daemon

# service restserver start

You should confirm that the port (default 8000) is accessible over the network, amending firewall rules and docs as required.

From here on in, almost everything can be done as the restserver user or whatever backup account is appropriate for a given user.

adding a backup client

Similar to the backup server, we’ll make a random key, and create a user account specific to this backup client. During the server setup, we used a flag --private-repos which ensures that every client has its own namespace, and this is the other side of that configuration.

To ensure that the repository can de-duplicate across multiple clients, the master encryption key is used (from our backup server), to create a derived one specific to this client:

# cd /root
# umask 077
# head -c 1000000000 /dev/random \
    | sha512 \
    | tee .restic
# restic --verbose --no-cache \
    --repo rest:http://backup.example.net:8000/ \
    key add

enter password for repository: .....
repository e8831311 opened successfully, password is correct
enter password for new key: 
enter password again: 
saved new key as <Key of root@z01.example.net,
created on 2020-06-16 12:11:38.04829251 +0000 UTC m=+22.329739230>

On our server, we can see there’s a new user/host created:

$ sudo -u restserver \
    restic --repo /var/db/restserver \
    --password-file=/var/db/restserver/.restic \
    key list
repository e8831311 opened successfully, password is correct
 ID        User  Host                      Created
--------------------------------------------------------------
 4a06463e  root  z01.example.net           2020-06-16 12:11:38
 906d85ec  root  f01.example.net           2019-07-20 16:21:37
 ace94bba  root  f02.example.net           2019-07-20 17:49:31
 e2774077  root  f03.example.net           2019-07-15 17:40:12
 ff4224db  root  i09.example.net           2019-07-15 17:36:40
...

And now we can run a simple test backup:

# restic --repo rest:http://backup.example.net:8000/ \
  --password-file=/root/.restic \
  --no-cache \
  --verbose \
  --tag $(hostname -s) \
  --tag $(date -u +%Y%m%d-%H%M) \
  --tag config \
  backup /efi \
         /etc \
         /root \
         /usr/local/etc \
         /var/db/zerotier-one \
         /boot/loader.conf

open repository
repository e8831311 opened successfully, password is correct
lock repository
load index files
start scan on [/efi /etc /root /usr/local/etc /boot/loader.conf /var/db/zerotier-one]
start backup on [/efi /etc /root /usr/local/etc /boot/loader.conf /var/db/zerotier-one]
scan finished in 46.197s: 2166 files, 5.217 MiB

Files:        2166 new,     0 changed,     0 unmodified
Dirs:            5 new,     0 changed,     0 unmodified
Data Blobs:   1016 new
Tree Blobs:      6 new
Added to the repo: 2.871 MiB

processed 2166 files, 5.217 MiB in 0:49
snapshot 0c50950d saved

doing restores

There are numerous ways to restore, but I think this is the easiest, given in most of my cases I have a tarball that is smaller than the available RAM on the system - I can just mount the backup server via fuse integration, and then pipe the backup of choice into tar or zfs as required.

For the smaller /config tagged backups for each server, these can simply be copied out as usual, to get up and running quickly, and then a full recovery using whatever approach is best suited to the disaster at hand.

# cd /
# REPO=http://backup:8000/
# TOKEN=/root/.restic
# restic --repo rest:${REPO} \
>         --password-file=${TOKEN} \
>         --no-cache \
>         --verbose \
> mount /mnt
repository e8831311 opened successfully, password is correct
Now serving the repository at /mnt

When finished, quit with Ctrl-c or umount the mountpoint.
^C
signal interrupt received, cleaning up

scripts and automation

In all cases, there is a simple script /root/bin/restic.sh run via cron & emailing output daily.

The OpenBSD boxes back up using tar(1) and on FreeBSD I use zfs(4).

All systems grab the /efi partition and some useful configurations before proceeding, as a separate config backup, so that I can quickly recover the bare bones of a system including appropriate network and storage tunables, before recovering the whole data back.

I prefer to backup using OS native tools, and stream that to restic, as this allows anybody to handle the local recovery, even if they’re not familiar with restic itself.

All the backup operator needs to do, is to make available the tarball or zfs stream, and the other operator can do the rest.

Here are a couple of examples of more complicated backups. Note that the --stdin-filename is used to give both a PATH-like namespace for each streamed backup type, and to give the “file” a friendly name for recovery, that is not unique for each backup.

This allows later on to remove older snapshots, and to prune them, finally freeing up space in our disk repository. Use tags to make it easier to identify various backups, and try to keep the name of each backup the same, to simplify pruning and maintaining the repo.

OpenBSD using tar(1)

#!/bin/sh -eu
PATH=$PATH:/usr/local/bin:/usr/local/sbin
cd /
NOW=$(date -u +%Y%m%d-%H%M)
HOSTNAME=$(hostname -s)
REPO=http://backup.example.net:8000/
TOKEN=/root/.restic
# do the bare minimum to recover the system first
restic --repo rest:${REPO} \
        --password-file=${TOKEN} \
        --no-cache \
        --verbose \
        --tag ${HOSTNAME} \
        --tag ${NOW} \
        --tag config \
        backup /efi /etc /root /usr/local/etc /var/db/zerotier-one
tar cpf - / \
        | restic --repo rest:${REPO} \
        --password-file=/root/.restic \
        --no-cache \
        --verbose \
        --tag ${HOSTNAME} \
        --tag ${NOW} \
        --tag tarball \
        --stdin \
        --stdin-filename=/tarballs/${HOSTNAME}.tar \
        backup

FreeBSD using zfs(4)

#!/bin/sh -eu
PATH=$PATH:/usr/local/bin:/usr/local/sbin
cd /
NOW=$(date -u +%Y%m%d-%H%M)
HOSTNAME=$(hostname -s)
ZPOOL=zroot
REPO=http://backup.example.net:8000/
TOKEN=/root/.restic
# do the bare minimum to recover the system first
restic --repo rest:${REPO} \
        --password-file=${TOKEN} \
        --no-cache \
        --verbose \
        --tag ${HOSTNAME} \
        --tag ${NOW} \
        --tag config \
        backup /efi /etc /root /usr/local/etc /boot/loader.conf /var/db/zerotier-one
zfs snapshot -r ${ZPOOL}@${NOW}
zfs send -LvcpR ${ZPOOL}@${NOW} \
        | restic --repo rest:${REPO} \
        --password-file=/root/.restic \
        --no-cache \
        --verbose \
        --tag ${HOSTNAME} \
        --tag ${NOW} \
        --tag zfs \
        --stdin \
        --stdin-filename=/tarballs/${HOSTNAME}.zfs \
        backup

The output will be similar to this - restic startup and initialisation at the beginning, and the piped output from zfs following:

# /root/bin/restic.sh
open repository
repository e8831311 opened successfully, password is correct
lock repository
load index files
using parent snapshot 028cc821
read data from stdin
start scan on [/tarballs/z01.zfs]
start backup on [/tarballs/z01.zfs]
scan finished in 46.298s: 1 files, 0 B
full send of zroot@migration estimated size is 12.1K
send from @migration to zroot@20200129-1237:12.1-RELEASE-p1 estimated size is 624
send from @20200129-1237:12.1-RELEASE-p1 to zroot@20200320-1544:12.1-RELEASE-p3 estimated size is 624
send from @20200320-1544:12.1-RELEASE-p3 to zroot@20200421-1735:12.1-RELEASE-p3 estimated size is 624
...
send from @20200616-1318 to zroot/ROOT/12.1-RELEASE-p2@20200616-1607 estimated size is 624
TIME        SENT   SNAPSHOT

Files:           1 new,     0 changed,     0 unmodified
Dirs:            1 new,     0 changed,     0 unmodified
Data Blobs:  11896 new
Tree Blobs:      2 new
Added to the repo: 5.999 GiB

processed 1 files, 8.147 GiB in 37:11
snapshot 454b1981 saved
469.977u 112.802s 38:13.53 25.4%	6397+710k 53047+0io 0pf+0w

Pruning Snapshots

Over time, snapshots accumulate. Like most systems, there are some additional tasks to perform – forgetting, to mark a given snapshot as superfluous, which is a quick task, and pruning, which is, on my ~ 6TiB repo, a rather long one – over 80 minutes now. I run this weekly.

Forgetting Snapshots

We are tagging snapshots by short hostname, which makes cleaning up very easy. The restic docs cover this in great detail, so I’ll skip most of the explanation, but this pretty much does what you want – keep a handful of older backups around for ages, and a few more fresh ones just in case. Remember that we have ZFS for those little daily rm -rf transgressions.

$ time restic --repo /var/db/restserver \
    --password-file=/var/db/restserver/.restic \
    forget -v \
    --keep-monthly 3 \
    --keep-weekly 5 \
    --keep-daily 10 \
    --keep-last 5 \
    --tag flatline

I also have a config backup for each server - there’s really no need to keep these longer than a week or so as they rarely change, and we’re only concerned about an occasional rollback:

$ time restic --repo /var/db/restserver \
    --password-file=/var/db/restserver/.restic \
    forget -v \
    --keep-last 5 \
    --tag config

This does the right thing, and prunes all but the last 5 snapshots tagged as config for each server we backed up.

As a small annoyance, there’s no easy way to say “forget all the snapshots from this host/tag” – restic insists that you delete the final snapshot manually. This is really the only annoyance I’ve had so far with restic, and for a small number of slowly changing hosts, I can live with it.

Pruning the Forgotten Snapshots

As you’ll notice, forgetting is quick and easy, but the space doesn’t come back:

$ zpool list
NAME       SIZE  ALLOC   FREE  CKPOINT  EXPANDSZ   FRAG    CAP  DEDUP    HEALTH  ALTROOT
embiggen     7T  6.71T   300G        -         -    43%    95%  1.00x    ONLINE  -

So I run the very straightforward prune sub-command, and .. wait. Over an hour to complete, with a lot of disk activity.

$ time restic --repo /var/db/restserver \
    --password-file=/var/db/restserver/.restic \
    prune -v
repository e8831311 opened successfully, password is correct
counting files in repo
building new index for repo
...
removed <data/5209d03c2d>
removed <data/7797e1c567>
removed <data/fbfe0d9663>
removed <data/1faa3db324>
removed <data/47b49f75e6>
removed <data/b3199bdfc7>
removed <data/b13f44fe84>
removed <data/26a5648ffc>
removed <data/554ec26b70>
removed <data/038d2052af>
[7:28] 100.00%  1061893 / 1061893 files deleted
done
________________________________________________________
Executed in   86.14 mins   fish           external 
   usr time  352.00 secs    1.61 millis  352.00 secs 
   sys time  459.16 secs    0.56 millis  459.16 secs 

Once the pruning is done, the restic docs recommend running a check, which is, after the pruning, quite a bit faster:

$ time restic --repo /var/db/restserver \
    --password-file=/var/db/restserver/.restic \
    check -v

using temporary cache in /tmp/restic-check-cache-022959336
repository e8831311 opened successfully, password is correct
created new cache in /tmp/restic-check-cache-022959336
create exclusive lock for repository
load indexes
check all packs
check snapshots, trees and blobs
no errors were found

________________________________________________________
Executed in    4.33 secs   fish           external 
   usr time    3.96 secs  1263.00 micros    3.96 secs 
   sys time    0.19 secs    0.00 micros    0.19 secs 

However a quick check of my zpool shows that the space hasn’t yet been freed - which is obviously because we snapshotted our restic backup store as well.

For bonus points, here’s how I clean up that zfs space:

$ zfs list -t snapshot embiggen/var/db/restserver
NAME                                       USED  AVAIL     REFER  MOUNTPOINT
embiggen/var/db/restserver@20200615-0526   516K      -      177G  -
embiggen/var/db/restserver@20200615-1106     0B      -      177G  -
embiggen/var/db/restserver@20200615-1916     0B      -      177G  -
embiggen/var/db/restserver@20200616-1115   572K      -      177G  -
embiggen/var/db/restserver@20200622-1502     0B      -      367G  -
embiggen/var/db/restserver@20200622-1508     0B      -      367G  -
embiggen/var/db/restserver@20200715-0915   218M      -      805G  -
embiggen/var/db/restserver@20200805-0732   252M      -     1.12T  -
embiggen/var/db/restserver@20200827-0846   912K      -     1.63T  -
embiggen/var/db/restserver@20200828-1359     0B      -     1.63T  -
embiggen/var/db/restserver@20200828-2117     0B      -     1.63T  -
embiggen/var/db/restserver@20201001-1433     0B      -     2.80T  -
embiggen/var/db/restserver@20201001-1542     0B      -     2.80T  -
embiggen/var/db/restserver@20201022-1144   252M      -     3.46T  -
embiggen/var/db/restserver@20201022-2049   260M      -     3.64T  -
embiggen/var/db/restserver@20201103-1124     0B      -     4.01T  -
embiggen/var/db/restserver@20201103-1136     0B      -     4.01T  -
embiggen/var/db/restserver@20201105-0020  19.9M      -     4.01T  -
embiggen/var/db/restserver@20201109-0628     0B      -      163M  -
$ zfs destroy -vrn embiggen/var/db/restserver@%20201105-0020 
$ zpool list
NAME       SIZE  ALLOC   FREE  CKPOINT  EXPANDSZ   FRAG    CAP  DEDUP    HEALTH  ALTROOT
embiggen     7T  1.52T  5.48T        -         -     7%    36%  1.00x    ONLINE  -

Unanswered Questions

  • do the presumed bandwidth economies of local caching outweigh another local stash, and more wasted space?
  • what happens if a per-host key is compromised? Does this expose the whole main repo key, assuming this is somehow derived from the main key and the local host key?
  • does sending the entire tarball (or zfs or postgres etc) impact deduplication opportunities in any way?
  • can we do anything with zfs datasets to get better deduplication?

On Caching

You’ll notice above that I don’t use the local cache, and this may mean I transfer more data over the network than necessary. I may tweak this policy over time, but so far the duration of backups is not so high that I wish the additional penalty of managing a zfs cache dataset on every client. In addition, the convenience of backing up the whole of the zpool in one fell swoop (which would necessarily pick up a cache directory), in my experience, is far better than a highly optimised local backup strategy that uses an opt-in strategy for each dataset. The latter has a high risk of finding out, at the point of system recovery, that critical files and folders were never backed up.

A local cache can be added to any client thus:

# zfs create zroot/var/cache/restic
# chown -R /var/cache/restic
# sed -i '' -E -e \
    's/--no-cache /--cache-dir=/var/cache/restic /' \
    /root/bin/restic.sh

On Performance

My initial experimentation suggests that even a small 4 core backup client can saturate a midrange ADSL connection, but other than that I have not found any useful bottlenecks in network and storage yet.

It may be that large servers with 10Gib+ NICs can push enough data to force restserver to use multiple CPU cores when writing, but I have not found any significant limitations yet other than the generally shabby local internet.

On backing up ZFS streams

FreeBSD, and zfs, is my bread and butter. What we want is to rely on zfs snapshots for general “whoops lost some files” recovery, but have a full recovery option via a complete zroot/ROOT stream if needed.

To achieve this, we need to use only zfs send options that create a reproducible stream, so that restic can de-duplicate as much data as possible.

In old-school backups, this was called “synthetic full backups” or an “eternal incremental strategy”.

For example, here are 3 consecutive runs of backing up the same zfs datasets repeatedly, on a very inactive test server. We’re using exactly the same snapshot references, but there have been additional snapshots created on those datasets between and perhaps during the backups.

This backup command was executed 3 times in a row:

time zfs send -LvcpR zpool@20200616-1607 ''
    | restic --repo rest:http://backup.example.net:8000/ \
    --password-file=/root/.restic \
    --verbose \
    --no-cache \
    --tag z01 \
    --tag 20200616-1607 \
    --tag zfs \
    --stdin \
    --stdin-filename=/tarballs/z01.zfs \
    backup
restic --repo /var/db/restserver --password-file=/var/db/restserver/.restic snapshots --tag zfs,z01

repository e8831311 opened successfully, password is correct
snapshots for (host [z01.example.net]):
ID        Time                 Host             Tags                   Paths
----------------------------------------------------------------------------------------
82db07d8  2020-06-16 13:19:11  z01.example.net  z01,20200616-1318,zfs  /tarballs/z01.zfs
454b1981  2020-06-16 16:08:30  z01.example.net  z01,20200616-1607,zfs  /tarballs/z01.zfs
d7d6fb8e  2020-06-16 19:13:00  z01.example.net  z01,20200616-1607,zfs  /tarballs/z01.zfs
9b157847  2020-06-16 19:55:07  z01.example.net  z01,20200616-1607,zfs  /tarballs/z01.zfs
----------------------------------------------------------------------------------------
4 snapshots

This backup is ~ 9GiB in total, and as it’s purely zfs snapshots, is not changing. The streamed data should be an ideal option for restic’s sliding window block approach to shine, despite using zfs stream’s internal dedupe, and leaving blocks that are compressed inside zfs, still compressed in the stream. I would have expected better results than ~ only 1/3 of the data shared between 3 backups, but I cannot be sure how much impact the various zfs send parameters used -LDevcpR and many snapshots, have on the zfs stream itself. Maybe that is the greater issue.

I used restic’s diff function to show me what it sees as common between the snapshots, and it’s really not a lot. This explains why I have so much space on my storage array!

$ restic --repo /var/db/restserver \
    --password-file=/var/db/restserver/.restic \
    diff 454b1981 d7d6fb8e 
repository e8831311 opened successfully, password is correct
comparing snapshot 454b1981 to d7d6fb8e:

M    /tarballs/z01.zfs

Files:           0 new,     0 removed,     1 changed
Dirs:            0 new,     0 removed
Others:          0 new,     0 removed
Data Blobs:  12794 new, 12794 removed
Tree Blobs:      1 new,     1 removed
  Added:   6.455 GiB
  Removed: 6.455 GiB

$ restic --repo /var/db/restserver \
    --password-file=/var/db/restserver/.restic \
    diff d7d6fb8e 9b157847
repository e8831311 opened successfully, password is correct
comparing snapshot d7d6fb8e to 9b157847:

M    /tarballs/z01.zfs

Files:           0 new,     0 removed,     1 changed
Dirs:            0 new,     0 removed
Others:          0 new,     0 removed
Data Blobs:  12852 new, 12852 removed
Tree Blobs:      1 new,     1 removed
  Added:   6.483 GiB
  Removed: 6.483 GiB

$ restic --repo /var/db/restserver \
    --password-file=/var/db/restserver/.restic \
    diff 454b1981 9b157847
repository e8831311 opened successfully, password is correct
comparing snapshot 454b1981 to 9b157847:

M    /tarballs/z01.zfs

Files:           0 new,     0 removed,     1 changed
Dirs:            0 new,     0 removed
Others:          0 new,     0 removed
Data Blobs:  12853 new, 12853 removed
Tree Blobs:      1 new,     1 removed
  Added:   6.484 GiB
  Removed: 6.484 GiB

For a ~ 8GiB backup of the same data, zfs and restic are not behaving well together at all!

I expect with some alterations to how zfs generates the streams, particularly around command-line arguments, this may be a lot more efficient in future.

Closing Thoughts

I like it. The community forums and IRC channel are courteous & helpful, and I hope this post has given a little back for their time in writing and supporting this very neat piece of software.

restic is a pleasure to use, easy to set up, and screamingly fast. It fits my #1 use case which is having off-site backups of medium-size zfs streams (under 100GiB) in the very unlikely event of losing all servers and their replicated databases, and fast recovery of those streams.

I still use tarsnap for the business-critical data (invoices and accounting records, any legal stuff), which manages the compression and delta-blocks better than restic, from my somewhat cursory tests, and also has the advantage that it’s commercially supported by the very helpful Colin Percival, but tarsnap’s single large-file restore performance isn’t sufficient in this case.

I can see, in the unlikely event of a total multi-site remote server catastrophe, simply driving our offsite backup server to the nearest datacentre, plugging it into the internet, and recovering our stuff very very quickly, with minimal stress.

Additional Reading