Reprepro

Debian package mirroring tool

At a Glance

Hosts:
Ansible:
Projects:
Documentation:
Bugs:

Overview

reprepro is the tool we use to mirror Debian repositories (including Ubuntu) to the AFS mirrors.

When updating package mirrors, it is undesirable for the index to be out of sync with the actual packages on disk, and vice-versa. This is generally achieved by syncing in two stages – firstly obtaining new files in the mirror and then secondly updating the indexes (and removing no-longer referenced files).

Problems occur if the upstream mirror updates itself during this process (which may happen up to 4 times a day). Debian, for example runs a “push” model where first-tier mirrors are notified of in-progress updates (see https://salsa.debian.org/mirror-team/archvsync/) and can restart any in-progress syncs to maintain consistency. OpenStack Infra is not suitable to apply for these notifications, as our mirror is not intended to be public and may be incomplete (we may not mirror all suites, or architectures, etc. as our needs dictate). This means if using other tools like ftpsync primarily intended for full replication we are very likely to have periods where our mirror gets out of sync (with subsequent job failures).

reprepro is more commonly used to build and manage private repositories, but has a number of features making it suitable for our use.

Rather than sync upstream indexes, it recreates them based upon files gathered from the upstream mirror. Since the upstream mirror remains consistent, reprepro will always download a consistent set of files. Then thanks to the release of the AFS mirror volume being atomic, we do not have any period where the repository package index doesn’t match the set of packages in the filesystem.

Since this does not require coordination with upstream, the same pattern is suitable across Ubuntu, Debian and other various apt repositories that may require integration (or perhaps do not provide facilities) for correct mirroring. Although reprepro can be more complicated to configure, it is consistent across these different distributions.

reprepro also makes it fairly easy to mirror only certain suites or architectures for a given repository, and to modify these configurations as required.

Repository signing

Note our repositories are not signed since reprepro recreates the indexes from scratch. This is actually somewhat helpful in avoiding the infra mirrors becoming de facto mirrors for a range of unrelated jobs (since we really do not guarantee contents for anything other than infra jobs).

apt will require --no-check-gpg or similar settings in configuration to use OpenStack mirrors.

Normal operation

Repository syncs are driven from cron on the mirror-update.opendev.org host using the /usr/local/bin/reprepro-mirror-update script. Repositories will update, remove old references and perform the vos release.

Removing components

As we drop support for various releases we may wish to remove them from the mirror. Firstly you should commit a change removing the distribution from the configuration files. It is then a one-time manual process to remove the actual files on disk, which should be done with the update lock held.

It is best to run this in a root screen session. This will look like:

# find the appropriate lock; config and volume;
# easy way is to look at the cron job arguments
$ flock -n /var/run/reprepro/<distro>.lock bash
$ K5START="k5start -t -f /etc/reprepro.keytab service/reprepro -- "
$ REPREPRO="${K5START} reprepro --confdir /etc/reprepro/<distro>"
# Clean up package indexes.  This takes a few minutes
$ ${REPREPRO} --delete clearvanished
# We have keepunreferenced, because we take a two-step approach
# of dumping the unreferenced files in one run, then deleting them in
# the next, to ensure any tests in progress can access the old files.
# In this removal situation, we should clear out the files now
# This can take a while depending on the size of what is removed
$ ${REPREPRO} --nokeepunreferencedfiles deleteunreferenced
# Perform an intermediate release as this may take some time.
# The ssh wrapper is the easiest way to release the volume
$ source /usr/share/mirror-update/functions.sh
$ vos_release <mirror.volume>
# Next we need to manually clear out some leftover lists/ and dists/
# content. Using your openafs account you want to rm the old distro
# release dist/ dir and lists/ files.
# After kinit and aklog (note no sudo):
$ rm -rf /afs/.openstack.org/mirror/<distro>/dists/<release>
$ rm /afs/.openstack.org/mirror/<distro>/lists/_<release>_lastseen \
     /afs/.openstack.org/mirror/<distro>/lists/*<release>_Release \
     /afs/.openstack.org/mirror/<distro>/lists/*<release>_Release.gpg \
     /afs/.openstack.org/mirror/<distro>/lists/*<release>_*_Packages
# The total number of files may vary from distro release to release
# depending on the number of architectures mirrored.
# Finally rerun the normal reprepro sync for the distro in order to
# ensure this still functions and we haven't deleted too much content.
# This will also perform a vos release.
$ reprepro-mirror-update /etc/reprepro/<distro> <mirror.volume> >> /var/log/reprepro/<logfile> 2>&1
# You can get the reprepro-mirror-update command from crontab -l.
# Just drop the flock prefix as we already hold the lock.
$ exit

Advanced Recovery Techniques

For a small repository, corruption is probably best handled by removing the entire repository and re-syncing. This is undesirable for larger repositories, however.

Note

Be careful with vos release which is done as part of /usr/local/bin/reprepo-mirror-update to avoid inadvertently releasing in progress work. Also be aware the commands in that script by default run under timeout which you may not want in recovery.

Corrupt reprepo databases will halt mirroring with often obscure symptoms. For example, this has been seen in production with reprepo ending up hung in an silent infinite loop. In this case, using strace revealed the last operation was on a file-descriptor related to a .db file, which gave a clue the databases were corrupt. Other failures may be possible, of course.

The following assumes you have a root shell with the correct AFS permissions for the mirror volumes, drop into something like:

k5start -t -f /etc/reprepro.keytab service/reprepro -- bash

We will use the Ubuntu repository as an example below.

In a crisis, you want to stop the cron job running to update the repo. You can either edit it out with crontab -e and put the host in the emergency file (so puppet doesn’t replace it) or, in a pinch, take the lock in a infinite loop like

flock -n /var/run/reprepro/ubuntu.lock bash -c while true; do sleep 1000; done

Firstly check in dmesg for AFS related errors. It is quite likely any corruption has happened due to issues at this layer, so ensure stability here before continuing to further recovery.

The databases are in the db directory in the mirror:

# ls /afs/.openstack.org/mirror/ubuntu/db
checksums.db  contents.cache.db  packages.db  references.db  release.caches.db  version

It is best to make backup copies before any recovery operations. Although AFS /should/ keep up, you should do any recovery of the db directory on a local copy to avoid any intermittent issues there further corrupting the database, then copy back the updated files when complete.:

# cp -r /afs/.openstack.org/mirror/ubuntu/db ~/db

For convenience, setup the common reprepo options for verbose logging, the configuration directory and to work on the local database:

# export REPREPRO="reprepro -VVV --confdir /etc/reprepro/ubuntu --dbdir ~/db"

From the upstream recovery document, the references.db can be removed and recreated quickly with:

$REPREPRO rereference

The checksums.db can also be recreated. You can rebuild with:

cd /afs/.openstack.org/mirror/ubuntu
find -type f -printf "pool/%P\n" > /tmp/file-list
$REPREPRO -b . _detect < /tmp/file-list

  • This will take several hours (~6 hours in 2017) as it touches all the repo files.

Note that if the .deb files on disk are corrupt, this may lead to errors on update about mismatching checksums which have been stored in the database. Likely you want to remove these files from disk and from the checksums database with a command similar to:

$REPREPRO _forget pool/main/p/package/the_package_1.2.3.deb
rm pool/main/p/package/the_package_1.2.3.deb

They should come back with the next update.

The contents.cache.db can be regenerated with the export command; this is run by the reprepro driver scripts automatically. It will take some time to regenerate, so this should be run manually.

In some situations where things are very out of sync, it may be easier to remove and replace an entire section of the repository. For example, if during updates files within xenial-security are seen to be corrupt, you can remove xenial-security from /etc/reprepro/ubuntu/distributions and run the following:

# remove old
$REPREPRO --delete clearvanished
# run an update
$REPREPRO update

You can then re-add the entries and run another update, which should resync everything from fresh.

You may also see errors relating to individual packages not being referenced correctly:

checking references to 'bionic|main|arm64' for 'texlive-latex-base': pool/main/t/texlive-base/texlive-latex-base_2017.20180305-1_all.deb
Missing reference to 'pool/main/t/texlive-base/texlive-latex-base_2017.20180305-1_all.deb' by 'bionic|main|arm64'
...
There have been errors!

In this case, the _addreference command can be useful. The parameters are the filekey, which is the path to the file, and the identifier, which is the tuple bionic|main|arm64 above. To restore the reference try:

# $REPREPRO _addreference pool/main/t/texlive-base/texlive-latex-base_2017.20180305-1_all.deb 'bionic|main|arm64'
Adding reference to 'pool/main/t/texlive-base/texlive-latex-base_2017.20180305-1_all.deb' by 'bionic|main|arm64'

Remember to put the databases back in place:

# mv /afs/.openstack.org/mirror/ubuntu/db /afs/.openstack.org/mirror/ubuntu/db.old
# cp -r ~/db  /afs/.openstack.org/mirror/ubuntu/

To stage a recovery prior to release, you can modify the mirror_root argument in openstack_project::mirror puppet to point Apache to the RW mirror /afs/.openstack.org/mirror where fixes are deployed, rather than the released RO /afs/openstack.org/mirror. This way you can avoid having to release the RO mirror and switch back quickly if things don’t work.

When fixed, you can either manually run vos release, or restore cron and let the next reprepro-mirror-update run do it.