On Tue, Dec 13, 2016 at 4:47 PM, Andrew Martin <amar...@xes-inc.com> wrote:
> > > ----- Original Message ----- > > From: "Mark Phippard" <markp...@gmail.com> > > To: "amartin" <amar...@xes-inc.com> > > Cc: "users" <users@subversion.apache.org> > > Sent: Tuesday, December 13, 2016 3:35:37 PM > > Subject: Re: Backup using ZFS Snapshots > > > On Tue, Dec 13, 2016 at 4:17 PM, Andrew Martin <amar...@xes-inc.com> > wrote: > > > >> Hello, > >> > >> I am running a Subversion 1.9.3 server on Ubuntu 16.04. I currently use > >> svnadmin hotcopy to safely backup the SVN repositories, but since the > >> repositories are now hosted on a ZFS dataset, I would like to utilize > ZFS's > >> snapshot capabilities to create atomic, point-in-time backups of the > >> repositories. My plan to do this is as follows: > >> > >> 1. create zfs snapshot > >> 2. clone zfs snapshot and mount at a temporary location > >> 3. run svnadmin hotcopy from the mounted clone to safely create a backup > >> 4. umount and destroy clone > >> > >> My only concern is if a commit is in-progress when the zfs snapshot > occurs, > >> would svnadmin hotcopy still be able to safely handle creating the > backup? > >> > > > > svnadmin hotcopy will not have a problem with an in-process commit. That > > is kind of one of the points of that command. > > > > > > > >> Is this a safe procedure for creating backups? > >> > > > > I know very little about ZFS but I do not understand why you would do > > this. If it can take snapshots, then why wouldn't you just take a > snapshot > > of the actual live repository, why would you want to copy it? > > > > Setting that aside, I think hotcopy is a very good way to do backups ... > > but it is not clear what value would come from taking a snapshot of the > > backup given that version control history is immutable etc. Also, once > you > > create the initial hotcopy, you can use the hotcopy --incremental option > to > > just copy the newer revisions. Do that from a post-commit hook and you > can > > always have a backup up to the latest commit. > > > > To me snapshots would seem like a way to do backups without using > hotcopy. > > I am not sure the value of combining the two ... but it should work > > technically if you have "reasons". > > > > Mark > > The reason I want to do it this way is that I want to "pull" the backups to > the backup server rather than "pushing" them from the Subversion server. In > order to do that, I plan on mounting the clone over NFS on the backup > server > and then using "svnadmin hotcopy" to pull any new data onto the backup > server. > Yes, I could just mount the live repository, but I have a number of > repositories > to backup and using the ZFS snapshot ensures that they are all from a > single > point-in-time. > > It sounds like "svnadmin hotcopy" will just ignore any in-progress commit, > so > even if the ZFS snapshot happened in the middle of a commit, it would still > result in a consistent backup. > > Yes, hotcopy will leave you with a consistent repository at the state it was in when the backup started. Any commits that are in process will not be copied/included in the backup. The problem with hotcopy, is that unless you craft a method where you can use --incremental, you are copying the entire repository every time you do a backup even though only a small number of new files may have been created. The SVN repository format is write-once. Only new files get added as new revisions get created. So the repository can be backed up very efficiently if you just keep capturing the new revisions being created. FWIW, I am not sure how concerning it really is if you took a snapshot in the middle of a commit. From what I recall, there is a very tiny race condition on the file that says what the HEAD revision is -- db/current. But in the worst case and there is a problem, you can just run svnadmin recover to reset it. It is not like you will have to deal with total file corruption or something. There is also a SQLite DB cache now as well which might add some other edge case problems on a live repository, but there would be solutions for any problems there as well. We have used NetApp snapshots to backup repositories for over a decade and I am not aware of any problems we have ever had when recovering a repository. I assume ZFS snapshots could work similarly where the entire FS could be snapshotted and then presumably it has features like NetApp to synch those snapshots to a DR location. -- Thanks Mark Phippard http://markphip.blogspot.com/
