couchcopy 0.1.1

Backup and restore CouchDB cluster

Create an archive of a running CouchDB node, saving CouchDB files data/.shards, data/_dbs.couch and data/shards in this order. To allow backup of a running CouchDB, files are copied before archive creation.

Restore an archive of a CouchDB node to a new CouchDB. The new CouchDB can be a cluster of multiple nodes. The new CouchDB configuration should already be done before using couchcopy, however, all existing data will be deleted. During restoration, CouchDB will be stopped and restarted on each cluster nodes.

Limitations

Tested with CouchDB 3.1.1 only.

To restore an archive, couchcopy needs to stop and start CouchDB. It assumes that CouchDB is controlled by systemd. If you don’t use systemd you can change parameters --couchdb-start="sudo systemctl start couchdb" and --couchdb-stop="sudo systemctl stop couchdb".

Your CouchDB n value should be higher or equal to the number of nodes in your CouchDB cluster. Otherwise saving shards from one node would not be enough to save and restore all databases. See CouchDB documentation for more details on replicas and nodes

The number of shards per database, i.e. the value of q, should be the same for the origin CouchDB and the destination CouchDB. Otherwise, tree /data/shards is not the same.

couchcopy assumes you have read access on CouchDB data folders, and read and write access on couchcopy create --rsync-reuse-folder FOLDER folder. It assumes you do not have write access on CouchDB data folders during restoration and thus will use sudo to write. You can use sudoers timestamp_timeout if you don’t want to re-type passwords too often.

Get started

pip install --user couchcopy
couchcopy backup old-server,/var/lib/couchdb backup.tar.gz
couchcopy restore backup.tar.gz admin:password@cluster_vm1,/var/lib/couchdb \
    admin:password@cluster_vm2,/var/lib/couchdb \
    admin:password@cluster_vm3,/var/lib/couchdb

To quickly access data from a backup, by spawning a CouchDB instance:

couchcopy load backup.tar.gz

For more options:

couchcopy -h
couchcopy backup -h
couchcopy load -h
couchcopy restore -h

On Fedora, CouchDB can be installed and configured with the following :

sudo dnf copr enable -y adrienverge/couchdb
sudo dnf install couchdb
sudo sh -c 'echo "admin = password" >> /etc/couchdb/local.ini'
sudo systemctl restart couchdb

If you work with remote machines, CouchDB needs to listen to remote IPs on each machine. You can enable it with the following (for security, revert it afterwards):

sudo sed -i 's/;bind_address = 127.0.0.1/bind_address = 0.0.0.0/g' /etc/couchdb/local.ini

Implementation details

During restoration, if the new CouchDB nodes names are not the same as the old CouchDB, nodes names are updated using CouchDB /_node/_local/_dbs endpoint. See CouchDB /_node/_local/_dbs endpoint documentation.

During restoration, couchcopy first updates one CouchDB node metadata (i.e. the list of nodes names) then it lets CouchDB itself synchronize metadata to the other nodes. couchcopy exits when the synchronization is finished for all nodes, using undocumented CouchDB /_dbs endpoint to monitor CouchDB nodes synchronization. You can skip that part if you want, i.e. you can exit couchcopy safely when the following log trace is displayed [Waiting for CouchDB cluster synchronization...]. For a CouchDB of 10^5 database, updating the first node metadata takes 35 minutes, then metadata synchronization to the other nodes takes 6 minutes. For a CouchDB of 100 databases only, both operations are nearly instantaneous.

Developer notes

To speed up CouchDB nodes synchronization it is possible to:

• Disable compaction daemon during synchronization (for 10^5 database, nodes synchronization goes from 6 minutes down to 4 minutes).

• Copy the saved _dbs.couch on every machine, but it sounds dangerous, it sounds better to let CouchDB rebuild these files itself (for 10^5 database, nodes synchronization goes from 6 minutes down to 0 seconds).

• Machines disk IOPS consumption is around 1200 IOPS during restoration.

• Sometime, the nodes synchronization, instead of taking 6 minutes for 100kDB, takes more than 3 hours. I wasn’t able to find the cause or eliminate this bad performance reliably. I advise using fast machines on the same local network, and disable compaction.

• Interesting discussions on CouchDB:

  • On nodes renaming after data/* copy for a backup restoration:

    https://github.com/apache/couchdb/discussions/3436#discussioncomment-494504

  • On CouchDB cluster internal backfill for a backup restoration:

    https://www.mail-archive.com/[e-mail]/msg30003.html

  • Un-answered questions about nodes renaming speed, and backup feasibility:

    https://www.mail-archive.com/[e-mail]/msg29982.html https://github.com/apache/couchdb/discussions/3383

Build and publish

python setup.py sdist
twine upload dist/*

License

This program is licensed under the GNU General Public License version 3.