Re: [Duplicity-talk] (Feature Request) inotify integration?

[Aaron Whitehouse]

Hello all,

Jenn Hulley-Miller wrote:
>> Is it possible to provide duplicity with a list of changed files
>> [...]built via linux's inotify subsystem.

> I have looked at this before and it would be a good solution for
> detecting changed files.  [...]
> Please enter a request on the Launchpad site and we'll look at
> integrating this and retiring the old scanning method.
> ...Ken

I have been giving this a lot of thought over the last few days and 
cross-post it to the Deja Dup list as it is relevant to them as well.  I 
apologise that it has become longer than intended!

Conceptually, inotify (http://en.wikipedia.org/wiki/Inotify) is a dream 
addition to a backup program.  As Jenn says, scanning for changes is 
often the vast majority of the work for an incremental backup.

TimeVault talks about the move to event-driven backup here:
https://wiki.ubuntu.com/TimeVault/Restructure
"A snapshot delay is set for each directory to be watched (say, 1 min.). 
When a file changes in that directory, a snapshot is taken after the 
configured delay time. If the file changes before the snapshot is taken, 
then the timer is reset repeatedly until you're done fiddling with the 
file, or some other specified time runs out (say, 60 min.) and a 
snapshot is forced. This mini-snapshot is called a delta. It only 
involves the file in question and so should be relatively fast (my 
benchmarks put it at around 50-150ms for a 1MB file, less for smaller 
files), and can be done in the background."

This really could change the way that backups are done, as the notion of 
"scheduling" backups would become redundant if we could rely on events 
showing that files had changed.

However, if duplicity moved to inotify and it failed to notify the 
backup daemon that a file had changed, that file would never get backed 
up (unless it was later changed again and noticed).  The key ways in 
which I can see it failing are:
(a) if files to be backed up are on removable drives and changed on a 
different computer;
(b) if files to be backed up are on a partition modified by another OS 
install on the same computer (for example, one of our computers 
dual-boots Windows and deja dup monitors one folder on the Windows 
partition); or
(c) if we hit some bug in inotify (I have no idea how reliable inotify 
is in cases, say, of power cuts or files changed over SAMBA etc.)

The best generic option that I have been able to come up with would be 
to use inotify for changes to partitions used by the system (/, /home 
etc.) and to still do a system scan for removable drives and other 
partitions.  This still would not solve the use-case of someone who 
mounts and edits their /home partition files from another (say, 
Gnu/Linux) install on the same machine.

I would guess that inotify would be a better way to determine changes in 
99% of cases for 99% of users, but a backup program really needs to 
cater for those 1% cases.  If there are times that we cannot trust 
inotify, then duplicity really needs to be doing a system scan -- at 
least occasionally to pick up anything missed.  Potentially duplicity 
could run scans (as it does now) and compare the results to inotify 
results for the first x backups to try and profile the system and user. 
 If inotify is often wrong then full scans could be done each time and 
if inotify is always the same, then it could rely on this with a full 
scan every y backups/months.

A different approach, to obtain some of the benefit of inotify with no 
risk, would be to follow TimeVault's lead and instantly (after some 
delay to prevent 100 versions of a file while you are editing it) backup 
files noticed by inotify.  This would be better than the current 
approach as files would always be backed up, even if they changed 
between scheduled backups.  However, the current approach to scheduled 
backups (full scans) could be maintained and this would mop up anything 
missed by inotify -- though for 99% of users this scan would likely 
yield no changed files.

I think it would be great if we could come up with an option that would 
be best in nearly all cases, rather than adding additional configuration 
options.  This is something that a lot of users would not be able to 
make an informed decision about.

I thought it was worth discussing these issues on the list, so that the 
best approach can be embodied in a useful feature request.

Thanks again to all who have made such an excellent program.

Regards,

Aaron