Groups | Search | Server Info | Login | Register

Groups > comp.sources.testers > #4

filesystem content tracking

Cross-posted to 3 groups.

Followups directed to: comp.unix.misc, comp.sources.testers

Show all headers | View raw

	[Cross-posting to news:comp.unix.shell, for the related
	questions are raised there from time to time.]

    More than occasionally, there's a need to create and maintain an
    "inventory" of the filesystem contents, to facilitate backup, or to
    maintain a history of file addition and removal, or for other
    purposes.  Below, I describe my current view on a possible design
    for a tool aiming to solve such a task (tentantively named FCCS, for
    Filesystem Content Checking (Control?) System), and relate my
    experiences with implementing and using of the versions of this tool
    relying on SQLite and a custom ASN.1-based file format for history
    storage.


    First try: SQLite

	Somewhat recently, I've given this task one more try.  My first
	approach was to use a SQLite database to record the current
	state and the history of filesystem changes.  The basic "units"
	of the filesystem state being tracked are:

	* the results of the stat(2) system call;

	* the message digests computed over (regular) file contents.

	Each time a "filesystem scan" is performed, the software records
	a "session", which binds filenames to stat(2) results and
	contents' digests.  Should there be a "previous" session, its
	data is "linked" to the newly created one, and is then used to
	avoid computing the digests for the unchanged files.

	Specifically, the digest is computed if the stat(2) results
	obtained for the file (sans the st_atime and st_dev fields) are
	/not/ bound to any (filename, digest) pair within the current or
	"previous" session.

	On the other hand, if there /is/ such a binding, the digest may
	be recomputed anyway if the binding is itself too old.  For this
	reason, the bindings are themselves timestamped.  (And so are
	the sessions.)

	(For anyone wishing to try this version, its sources could be
	found at [1, 2].)


    Trying a bunch of ASN.1-based files instead

	Unfortunately, despite my attempts to implement a
	space-efficient database schema, the volume of the resulting
	database is not only enormous by itself (about 360 bytes per
	file being tracked for the first session) but also grows rather
	rapidly (about 86 bytes per file being tracked per session.)

	Also, as the database grows, the addition of new records begins
	to take growing amounts of time.  For the third session of the
	database tracking about 300000 files, the rate could be as slow
	as a few regular files per second on a fairly decent hardware.
	Indeed, such a session may take a whole day to complete!

	This, combined with that there seem to be no free software
	extensions for transparent SQLite database file compression,
	prompted me to change my strategy.

	Namely, I've decided that instead of using a single database
	file for all the sessions, I'd use one binary "log" file per
	session, while allowing the tool to both read the files
	containing previous sessions (so to compute digests only for the
	new and changed files) and to /reference/ them from within the
	resulting file (so to conserve space.)

	As there's no need anymore to allow concurrent access to a
	single file (files containing previously recorded sessions can
	safely be examined while the new file is being written), the
	in-file database indices are becoming less relevant.

	Additionally, as the files in question are to be read mainly
	sequentially, it becomes feasible to compress them even when
	they're still used.  Also, as there're distinct files for
	distinct sessions, it's easy to compress those no longer needed,
	or move them to some cheaper storage (slow HDD's and DVD+R's.)

	Now, to the numbers.  For the test run, I've processed 47151
	files (41285 regular, 2.6 GiB in total), and it took only 6:21
	to produce the corresponding 251978 records (10 MiB in total.)
	The time is comparable to raw sha1sum(1) (about 4:20), and the
	space demands are roughly 222 bytes per file on average (compare
	to the SQLite-based version's 360 above.)

I: file_bind=251978, file_rec=251977, digest=251976
132.12user 13.32system 6:21.35elapsed 38%CPU (0avgtext+0avgdata 233008maxresident)k
4096210inputs+8outputs (4major+14613minor)pagefaults 0swaps

	Unfortunately, the use of the previously recorded sessions is
	not implemented at this moment, so I don't have the space
	requirements for the "incremental update" case at hand, but I
	expect them to be times lower than those for the SQLite-based
	version.  Combined with the ability to easily compress or move
	away the older sessions, I hope this may finally get the tool to
	a usable and useful state.

	(For anyone wishing to try this version, its sources could also
	be found at [1, 2], under the fccs-2012-03-asn.1 branch.)


    Miscellaneous facts

	The software in question is written in Perl and was primarily
	tested on Debian GNU/Linux 6.0 "Squeeze" systems.  It depends on
	a few Perl packages, such as Digest::SHA and UUID, and also on
	either DBD::SQLite and DBI (SQLite-based version), or
	Convert::ASN1 (ASN.1-based one.)

	It's my intention to make the software available under the
	GNU General Public License version 3 or later, and the related
	data schemata will probably be released into the public domain.


    References

[1] http://gray.am-1.org/~ivan/archives/git/gitweb.cgi?p=fc-2012.git
[2] http://gray.am-1.org/~ivan/archives/git/fc-2012.git/

-- 
FSF associate member #7257

Back to comp.sources.testers | Previous | Next — Next in thread | Find similar

Thread

filesystem content tracking Ivan Shmakov <oneingray@gmail.com> - 2012-03-14 21:53 +0700
  Re: filesystem content tracking Ivan Shmakov <oneingray@gmail.com> - 2012-03-19 18:03 +0700
    Re: filesystem content tracking Ivan Shmakov <oneingray@gmail.com> - 2012-03-20 23:37 +0700
    Re: filesystem content tracking Ivan Shmakov <oneingray@gmail.com> - 2012-03-22 11:41 +0700

csiph-web