The Andrew Filesystem (or AFS) is a global distributed filesystem. With a single mountpoint, clients can access any site on the Internet which is running AFS as if it were a local filesystem.

OpenAFS is an open source implementation of the AFS services and utilities.

A collection of AFS servers and volumes that are collectively administered within a site is called a cell. The OpenStack project runs the AFS cell, accessible at /afs/

At a Glance

  • (a vldb and pts server in DFW)

  • (a vldb and pts server in ORD)

  • (a second vldb and pts server in DFW)

  • (a fileserver in DFW)

  • (a second fileserver in DFW)

  • (a fileserver in ORD)

  • (host running mirror update jobs)


OpenStack Cell

AFS may be one of the most thoroughly documented systems in the world. There is plenty of very good information about how AFS works and the commands to use it. This document will only cover the minimum needed to understand our deployment of it.

OpenStack runs an AFS cell called There are three important services provided by a cell: the volume location database (VLDB), the protection database (PTS), and the file server (FS). The volume location service answers queries from clients about which fileservers should be contacted to access particular volumes, while the protection service provides information about users and groups.

Our implementation follows the common recommendation to colocate the VLDB and PTS servers, and so they both run on our afsdb* servers. These servers all have the same information and communicate with each other to keep in sync and automatically provide high-availability service. As described in the Ubik protocol requires three servers to maintain availability; for that reason, two of our DB servers are in the DFW region, and the other in ORD.

Fileservers contain volumes, each of which is a portion of the file space provided by that cell. A volume appears as at least one directory, but may contain directories within the volume. Volumes are mounted within other volumes to construct the filesystem hierarchy of the cell.

OpenStack has two fileservers in DFW and one in ORD. They do not automatically contain copies of the same data. A read-write volume in AFS can only exist on exactly one fileserver, and if that fileserver is out of service, the volumes it serves are not available. However, volumes may have read-write copies which are stored on other fileservers. If a client requests a read-only volume, as long as one site with a read-only volume is online, it will be available.

Client Configuration

To use OpenAFS on a Debian or Ubuntu machine:

sudo apt-get install openafs-client openafs-krb5 krb5-user

Debconf will ask you for a default realm, cell and cache size. Answer:

Default Kerberos version 5 realm: OPENSTACK.ORG
AFS cell this workstation belongs to:
Size of AFS cache in kB: 500000

The default cache size in debconf is 50000 (50MB) which is not very large. We recommend setting it to 500000 (500MB – add a zero to the default debconf value), or whatever is appropriate for your system.

The OpenAFS client is not started by default, so you will need to run:

sudo service openafs-client start

When it’s done, you should be able to cd /afs/

Most of what is in our AFS cell does not require authentication. However, if you have a principal in kerberos, you can get an authentication token for use with AFS with:


If not running on Debian or Ubuntu you can install openafs client packages as well as Kerberos5 packages on your distro of choice.

Then to kinit, use your fully qualified user id:


Or for admin access:


Then aklog, specifying the cell:

aklog -cell


The following information is relevant to AFS administrators.

All of these commands have excellent manpages which can be accessed with commands like man vos or man vos create. They also provide short help messages when run like vos -help or vos create -help.

For all administrative commands, you may either run them from any AFS client machine while authenticated as an AFS admin, or locally without authentication on an AFS server machine by appending the -localauth flag to the end of the command.

Adding a User

First, add a kerberos principal as described in Adding A User Principal. Have the username and UID from puppet ready.

Then add the user to the protection database with:

pts createuser $USERNAME -id UID

Admin UIDs start at 1 and increment. If you are adding a new admin user, you must run pts listentries, find the highest UID for an admin user, increment it by one and use that as the UID. The username for an admin user should be in the form username.admin.


Any ‘/’ characters in a kerberos principal become ‘.’ characters in AFS.

Adding a Superuser

Run the following commands to add an existing principal to AFS as a superuser:

pts adduser -user $USERNAME.admin -group system:administrators

After this, you should update the system-config: playbooks/roles/openafs-server-config/files/UserList file to ensure the new username is authorized to issue privileged commands.

Deleting Files


This is a basic example of write operations for AFS-hosted content, so applies more generally to manually adding or changing files as well. As we semi-regularly get requests to delete subtrees of documentation, this serves as a good demonstration.

First, as a prerequisite, make sure you’ve followed the Client Configuration and Adding a Superuser steps for yourself and that you know the password for your $USERNAME/admin kerberos principal. Safely authenticate your superuser’s principal in a new PAG as follows:

pagsh -c /bin/bash
export KRB5CCNAME=FILE:`mktemp`
kinit $USERNAME/admin

If this is a potentially destructive change (perhaps you’re worried you might mistype a deletion and remove more content than you intended) you can first create a copy-on-write backup snapshot like so:

vos backup docs

When deleting files, note that you should use the read-write /afs/ path rather than the read-only /afs/ path, but normal Unix file manipulation commands work as expected (do _not_ use sudo for this):

rm -rf /afs/

If you don’t want to have to wait for a volume release to happen (so that your changes to the read-write filesystem are reflected immediately in the read-only filesystem), you can release it now too:

vos release docs -verbose

Now you can clean up your session, destroy your ticket and exit the temporary PAG thusly:


Creating a Volume


You should not run these operations on the fileservers where the volumes or replicas are to be created (afs01 or afs02). openafs `vos will resolve the ipv4 address of the fileserver host from the command-line. If you are using the tool on the fileserver, Debuntu’s use of for localhost and having the hostname in /etc/hosts can thus result in the vos tool not correctly filtering the loopback address and setting the server address for the volume as – making it effectively inaccessible. A similar problem can occur for NAT servers, if we were to use them. Running on an external host means the lookups shouldn’t return local addresses and avoids this issue. The other option is to specify the fileservers as the IP address, rather than the hostname, to avoid any lookup issues.

Select a fileserver for the read-write copy of the volume according to which region you wish to locate it after ensuring it has sufficient free space. Then run:


The a in the preceding command tells it to place the volume on partition vicepa. Our fileservers only have one partition and therefore this is a constant.

Be sure to mount the read-write volume in AFS with:

fs mkmount /afs/ $VOLUMENAME

You may want to create read-only sites for the volume with vos addsite and then vos release.

If the volume’s mountpoint lies within another volume, you may also need to vos release that parent volume before it will show up in the read-only path.

You should set the volume quota with fs setquota.

Deleting a Volume

Remove the mountpoint(s) of the volume:

fs rmmount /afs/

Be sure to release the parent volume (with vos release) if necessary after removing the mountpoint.

Run vos examine to see a list of volume sites. Identify the read-write and read-only sites.

Remove the read-only sites first; repeat this command for each one:

vos remove -server $FILESERVER -partition $PARTITION -id $VOLUME.readonly

Remove the read-write volume:

vos remove -id $VOLUME

Adding a Fileserver

Put the machine’s public IP on a single line in /var/lib/openafs/local/NetInfo (TODO: puppet this).

Copy /etc/openafs/server/* from an existing fileserver.

Create an LVM volume named vicepa from cinder volumes. See Cinder Volume Management for details on volume management. Then run:

mkdir /vicepa
echo "/dev/main/vicepa  /vicepa ext4  errors=remount-ro,barrier=0  0  2" >>/etc/fstab
mount -a

Finally, create the fileserver with:

bos create -server NEWSERVER -instance dafs -type dafs \
  -cmd "/usr/lib/openafs/dafileserver -L -p 242 -busyat 600 -rxpck 700 \
    -s 1200 -l 1200 -cb 2000000 -b 240 -vc 1200 \
    -udpsize 131071 -sendsize 131071" \
  -cmd /usr/lib/openafs/davolserver \
  -cmd /usr/lib/openafs/salvageserver \
  -cmd /usr/lib/openafs/dasalvager

It is worth evaluating these settings periodically

  • -L selects the large size, which ups a number of defaults

  • -p defines the worker threads for processing incoming calls. Since they block until there is work to do, we should leave this at around the maximum (which may increase across versions; see documentation)

  • -udpsize and -sendsize should be increased above their default

  • -cb defines the callbacks. For our use case, with a single mirror writer, this should be around the number of files the client is configured to cache (-dcache) multiplied by the number of clients.

Updating Settings

The helper script system-config: tools/ is a helper script to restart AFS servers, and optionally enable audit logging on the servers which is sometimes useful for debugging afs clients. You can edit settings in the script and run restart (or restart-auditing).

If you wish to update the settings for an existing server manually, you can stop and remove the existing bnode (the collection of processes the overseer is monitoring, created via bos create above) and recreate it.

For example

bos stop -server \
         -instance dafs \

Then remove the server with

bos delete -server \
           -instance dafs

Finally run the bos create command above with any modified parameters to restart the server.

Recovering a Failed Fileserver

If a fileserver crashes, take the following steps to ensure it’s usable after recovery:

  • Pause mirror updates and volume release cron jobs

  • Reboot the server; fix any filesystem errors and check the salvager logs

  • Check for any stuck volume transactions; remedy as appropriate

  • Perform a manual release of every volume from a terminal on a server using “-localauth” in case OpenAFS decides it can’t do an incremental update.

  • Re-enable cron jobs


We host mirrors in AFS so that we store only one copy of the data, but mirror servers local to each cloud region in which we operate serve that data to nearby hosts from their local cache.

All of our mirrors are housed under /afs/ Each mirror is on its own volume, and each with a read-only replica. This allows mirrors to be updated and then the read-only replicas atomically updated. Because mirrors are typically very large and replication across regions is slow, we place both copies of mirror data on two fileservers in the same region. This allows us to perform maintenance on fileservers hosting mirror data as well deal with outages related to a single server, but does not protect the mirror system from a region-wide outage.

In order to establish a new mirror, do the following:

  • The following commands need to be run authenticated on a host with kerberos and AFS setup (see afs_client; admins can run the commands on See the note above about not doing this on the actual fileservers. Firstly kinit and aklog to get tokens.

  • Create the mirror volume. See Creating a Volume for details. The volume should be named, where foo is descriptive of the contents of the mirror. Example:

    vos create a
  • Create read-only replicas of the volume. One replica should be located on the same fileserver (it will take little to no additional space), and at least one other replica on a different fileserver. Example:

    vos addsite a
    vos addsite a
  • Release the read-only replicas:

    vos release

    See the status of all volumes with:

    vos listvldb

When traversing from a read-only volume to another volume across a mountpoint, AFS will first attempt to use a read-only replica of the destination volume if one exists. In order to naturally cause clients to prefer our read-only paths for mirrors, the entire path up to that point is composed of read-only volumes:

/afs             [root.afs]
  / [root.cell]
    /mirror      [mirror]
      /bar       []

In order to mount the volume under mirror we need to modify the read-write version of the mirror volume. To make this easy, the read-write version of the cell root is mounted at /afs/ Following the same logic from earlier, traversing to paths below that mount point will generally prefer read-write volumes.

  • Mount the volume into afs using the read-write path:

    fs mkmount /afs/
  • Release the mirror volume so that the (currently empty) foo mirror itself appears in directory listings under /afs/

    vos release mirror
  • Create a principal for the mirror update process. See Adding A User Principal for details. The principal should be called service/foo-mirror. Example:

    kadmin: addprinc -randkey service/foo-mirror@OPENSTACK.ORG
    kadmin: ktadd -k /path/to/foo.keytab service/foo-mirror@OPENSTACK.ORG


    Each time ktadd is run, the key is rotated and previous keytabs are invalidated.

  • Add the service principal’s keytab to Ansible secrets. Copy the binary key to and then use hieraedit to update the files

    root@bridge:~# /home/zuul/src/ \
      --yaml /etc/ansible/hosts/host_vars/ \
      -f /path/to/foo.keytab KEYNAME

    (don’t forget to git commit and save the change; you can remove the copies of the binary key too). The key will be base64 encoded in the heira database. If you need to examine it for some reason you can use base64:

    cat /path/to/foo.keytab | base64
  • Ensure the values in this new variable are written to disk as the keytab on by adding it to the mirror-update role for the mirror scripts to use during update. You should check this with testinfra in testinfra/ (note this involves defining a “dummy” keytab for testing; see the other examples).

  • Create an AFS user for the service principal:

    pts createuser

Because mirrors usually have a large number of directories, it is best to avoid frequent ACL changes. To this end, we grant access to the mirror directories to a group where we can easily modify group membership if our needs change.

  • Create a group to contain the service principal, and add the principal:

    pts creategroup foo-mirror
    pts adduser foo-mirror

    View users, groups, and their membership with:

    pts listentries
    pts listentries -group
    pts membership foo-mirror
  • Grant the group access to the mirror volume:

    fs setacl /afs/ foo-mirror write
  • Grant anonymous users read access:

    fs setacl /afs/ system:anyuser read
  • Set the quota on the volume (e.g., 100GB):

    fs setquota /afs/ 100000000

Because the initial replication may take more time than we allocate in our mirror update cron jobs, manually perform the first mirror update:

  • In screen, obtain the lock on

    flock -n /var/run/foo-mirror/mirror.lock bash

    Leave that running while you perform the rest of the steps.

  • Also in screen on mirror-update, run the initial mirror sync. If using one of the mirror update scripts (from /usr/local/bin) be aware that they generally run the update process under timeout with shorter periods than may be required for the initial full sync. e.g. for reprepro mirrors

    /usr/local/bin/reprepro-mirror-update /etc/reprepro/ubuntu mirror.ubuntu

  • Log into and run screen. Within that session, periodically during the sync, and once again after it is complete, run:

    vos release -localauth

    It is important to do this from an AFS server using -localauth rather than your own credentials and inside of screen because if vos release is interrupted, it will require some manual cleanup (data will not be corrupted, but clients will not see the new volume until it is successfully released). Additionally, vos release has a bug where it will not use renewed tokens and so token expiration during a vos release may cause a similar problem.

  • Once the initial sync and vos release are complete, release the lock file on mirror-update.

Removing a mirror

If you need to remove a mirror, you can do the following:

  • Unmount the volume from the R/W location:

    fs rmmount /afs/
  • Release the R/O mirror volume to reflect the changes:

    vos release mirror
  • Check what servers the volumes are on with vos listvldb:

    VLDB entries for all servers
        RWrite: 536870934     ROnly: 536870935
        number of sites -> 3
           server partition /vicepa RW Site
           server partition /vicepa RO Site
           server partition /vicepa RO Site
  • Remove the R/O replicas (you can also see these with vos listvol -server afs0[1|2]

    vos remove -server -partition a -id
    vos remove -server -partition a -id
  • Remove the R/W volume:

    vos remove -server -partition a -id
Reverse Proxy Cache

Each of the region-local mirror hosts exposes a limited reverse HTTP proxy on port 8080. These proxies run within the same Apache setup as used to expose AFS mirror contents. mod_cache is used to expose a white-listed set of resources (currently just RDO).

Currently they will cache data for up to 24 hours (Apache default) with pruning performed by htcacheclean once an hour to keep the cache size at or under 2GB of disk space.

The reverse proxy is provided because there are some hosted resources that are not currently able to be practically mirrored. Examples of this include RDO (rsync from RDO is slow and they update frequently) and docker images (which require specialized software to run a docker registry and then sorting out how to run that on a shared filesystem).

Apache was chosen because we already had configuration management in place for Apache on these hosts. This avoids management overheads of a completely new service deployment such as Squid or a caching docker registry daemon.

No Outage Server Maintenance

We have redundant AFS DB servers. You can take one down without causing a service outage as long as the others remain up. To do this safely:

root@afsdb01:~# bos shutdown -wait -localauth
root@afsdb01:~# bos status -localauth
Instance ptserver, temporarily disabled, currently shutdown.
Instance vlserver, temporarily disabled, currently shutdown.

Then perform your maintenance on afsdb01. When done a reboot will automatically restart the bos service or you can manually restart the openafs-fileserver service:

root@afsdb01:~# service openafs-fileserver start

Finally check that the service is back up and running:

root@afsdb01:~# bos status -localauth
Instance ptserver, currently running normally.
Instance vlserver, currently running normally.

Now you can repeat the process against afsdb02 or afsdb03.

Taking down the actual fileservers is slightly more complicated but works similarly. Basically what we need to do is make sure that either no one needs the RW volumes hosted by a fileserver before taking it down or move the RW volume to another fileserver. When taking down we must also ensure that the vos releases that are performed on it by mirror-update are stopped.

To ensure nothing needs the RW volumes you can hold the various file locks on hosts that publish to AFS and/or remove cron entries that perform vos releases or volume writes.

If instead you need to move the RW volume first step is checking where the volumes live:

root@afsdb01:~# vos listvldb -localauth
VLDB entries for all servers

    RWrite: 536870934     ROnly: 536870935
    number of sites -> 3
       server partition /vicepa RW Site
       server partition /vicepa RO Site
       server partition /vicepa RO Site

We see that if we want to allow write to the mirror volume and take down we will have to move the volume to one of the other servers:

root@afsdb01:~# screen # use screen as this may take quite some time.
root@afsdb01:~# vos move -id mirror -toserver -topartition vicepa -fromserver -frompartition vicepa -localauth

When that is done (use listvldb command above to check) it is now safe to take down while having writers to the mirror volume. If operating on you should also hold all mirror update locks and the release-volumes lock. This ensures we do not interrupt any vos releases on that are run by mirror-update remotely. We use the same process as for the db server:

root@afsdb01:~# bos shutdown -localauth
root@afsdb01:~# bos status -localauth
Auxiliary status is: file server shut down.

Perform maintenance, then restart as above and check the status again:

root@afsdb01:~# bos status -localauth
Auxiliary status is: file server running.

DNS Entries

AFS uses the following DNS entries which indicate an even balance: 300 IN SRV 10 10 7002 300 IN SRV 10 10 7002 300 IN SRV 10 10 7002 300 IN SRV 10 10 7003 300 IN SRV 10 10 7003 300 IN SRV 10 10 7003

Be sure to update them if volume location and PTS servers change. Also note that only A (IPv4 address) records are used in the SRV data. Since OpenAFS lacks support for IPv6, avoid entering corresponding AAAA (IPv6 address) records for these so that it won’t cause fallback delays for other v6-supporting AFS client implementations.