whisper-backup

I needed a handy way to backup my Graphite cluster to reliable storage such as Amazon S3 or OpenStack Swift. Also, the ability to restore that data in a sane away.

Example

$ whisper-backup --logfile /opt/graphite/storage/log/whisper-backup/whisper-backup.log \
        --bucket fitbit.graphite.$(hostname -s) \
        --retention 5 \
        --quiet \
        backup swift

Goals

• Compress WSP data. Space is cash and they compress well.
• Support storing multiple backups of the same WSP DB over a retention period.
• Be able to restore and backup all or part of an existing tree of metrics.
• Don't waste space on duplicate WSP files.
• Verify restored data.
• Allow for manual restores if needed, we simply store Gzip versions of the WSP files in our storage backend.
• Support multiple storage backends.
• Use flock() (the same locking method that Whisper uses) to lock each DB file before uploading it. This ensures we have a copy that wasn't in the middle of a file update procedure. You have your carbon-cache daemons set to use locking, right?
• On restore, if the WSP file already exists just backfill in the data rather than overwrite it.
• File space for temp file copies is limited and is definitely not large enough to fit an entire backup set into.
• Use multiprocessing to handle large backup sets faster.

Usage

I decided not to design this to store multiple servers worth of WSP files in a single bucket/container of the storage service. For large clusters this could be millions of files which may cause slowness with the API and other issues. So if you have multiple servers you should set each machine to backup to its own bucket/container.

$ python whisperbackup.py --help
Usage: whisperbackup.py [options] backup|restore|purge|list swift|s3 [storage args]

Options:
  -p PREFIX, --prefix=PREFIX
                        Root of where the whisper files live or will be
                        restored to, default /opt/graphite/storage/whisper
  -f PROCESSES, --processes=PROCESSES
                        Number of worker processes to spawn, default 4
  -r RETENTION, --retention=RETENTION
                        Number of unique backups to retain for each whisper
                        file, default 5
  -x PURGE, --purge=PURGE
                        Days to keep unknown Whisper file backups, -1 disables,
                        default 45
  -n, --noop            Do not modify the object store, default False
  -b BUCKET, --bucket=BUCKET
                        The AWS S3 bucket name or Swift container to use,
                        default graphite-backups
  -m METRICS, --metrics=METRICS
                        Glob pattern of metric names to backup or restore,
                        default *
  -c DATE, --date=DATE  String in ISO-8601 date format. The last backup before
                        this date will be used during the restore.  Default is
                        now.
  -d, --debug           Minimum log level of DEBUG
  -q, --quiet           Only WARN and above to stdout
  --nolog               Do not log to LOGFILE
  --logfile=LOGFILE     File to log to, default /var/log/whisperbackup.py
  --nolock              Do not use a lockfile
  --lockfile=LOCKFILE   Lock file, default /var/tmp/whisperbackup.py
  --locktimeout=LOCKTIMEOUT
                        Lock timeout in seconds, default 90
  --splay=SPLAY         Sleep a random time between 0 and N seconds before
                        starting, default 0
  -h, --help            show this help message and exit

Notes:

• Purge removes Whisper backups in the datastore for Whisper files not presently on the server. Such as deleted or moved Whisper files. A setting of 0 will immediately purge backups for metrics not on the local disk.

Requirements

Required python packages and the versions of which I've tested with.

• whisper >= 0.9.12
• carbon >= 0.9.12
• lockfile

For AWS S3 support:

• boto

For OpenStack Swift support:

• python-swiftclient >= 3.0.0 (timeout support)

Assumptions

• We assume that each WSP file will fit in memory.

To Do

• We use multiprocess.Pool for backups, but restores are still single process
• Purge bug: If a metric has been idle for 45 days then the backup date on that file in the object store hasn't changed. So once that metric is removed from local disk it will be immediately removed from the object store rather than 45 days after it was removed from local disk.