Mailinglist Archive: opensuse-features (365 mails)

< Previous Next >
[openFATE 306379] Use rsync when refreshing repositories
  • From: fate_noreply@xxxxxxx
  • Date: Tue, 2 Mar 2010 15:46:04 +0100 (CET)
  • Message-id: <feature-306379-18@xxxxxxxxxxxxxx>
Feature changed by: Christoph Thiel (cthiel1)
Feature #306379, revision 18
Title: Use rsync when refreshing repositories

openSUSE-11.2: Rejected by Michael Löffler (michl19)
reject date: 2009-08-11 16:01:45
reject reason: too late for 11.2, moved to 11.3
Priority
Requester: Important

- openSUSE-11.3: Evaluation
+ openSUSE-11.3: Rejected by (cthiel1)
+ reject date: 2010-03-02 15:45:59
+ reject reason: This feature is out of scope for 11.3.
Priority
Requester: Important

Requested by: Piotrek Juzwiak (benderbendingrodriguez)

Description:
It would be a great idea to use rsync when refreshing repositories, one
of the bad things is the refresh speed. It gets worse when people have
many repositories. I'm not sure but it already compares if something
has changed in the repo but to speed things up it would be great to use
rsync. For example big repositories like Packman for example download
every time the default 10 minutes are over (in zypp settings) while
nothing great changes there.

Discussion:
#1: Roberto Mannai (robermann79) (2009-05-08 14:57:32)
The best way to download incrementally only the diff of a binary file,
for my best knowledge, is using the GDIFF protocol, who was submitted
ten years ago to the W3C consortium:
http://www.w3.org/TR/NOTE-gdiff-19970901
(http://www.w3.org/TR/NOTE-gdiff-19970901)
I know for sure that a commercial product of Configuration Management
(Marimba, now buyed by BMC - see http://www.marimba.com
(http://www.marimba.com/) ) use it, implemented in Java: it is very
useful in low bandwidth nets, when downloading a service pack, for
example. I don’t know if one person could use that Java algorithm
implementation, anyway, being a commercial application.
Other implementations are in PERL and RUBY:
http://search.cpan.org/~geoffr/Algorithm-GDiffDelta-0.01/GDiffDelta.pm
(http://search.cpan.org/%7Egeoffr/Algorithm-GDiffDelta-0.01/GDiffDelta.
pm)
http://webscripts.softpedia.com/script/Development-Scripts-js/gdiff-gpatch-18695.html

(http://webscripts.softpedia.com/script/Development-Scripts-js/gdiff-gpatch-18695.html)
An open source .NET (C#) implementation: http://gdiff.codeplex.com/
(http://gdiff.codeplex.com/) with MPL license
I cannot understand why that algorithm is not widely used, given its
quality; it shoud be useful if it was available when downloading large
files like ISOs or VM images, or repositories information

#2: Roberto Mannai (robermann79) (2009-05-08 15:08:01) (reply to #1)
In your usecase, the repository could provide a GDIFF file of content
metadata variation, the delta between two known "versions" of it in the
time.

#3: Piotrek Juzwiak (benderbendingrodriguez) (2009-06-18 23:16:55)
Hmm, i guess packman wouldn't implement that only for me ;) Though it
would speed things up as it is a widely known and spoken that
refreshing the repo in openSUSE is slow.

#4: Luc de Louw (delouw) (2009-07-04 18:21:31)
Why not rsync? Because it does not work with http(s). This is important
since in many companies the only way to get data from the internet is
via http proxy.
The GDIFF approach sounds promissing

#5: Roberto Mannai (robermann79) (2009-07-04 18:39:56) (reply to #4)
For a "GDIFF on HTTP" implementation, see
http://www.w3.org/TR/NOTE-drp-19970825

#6: Jan Engelhardt (jengelh) (2009-07-05 15:23:26)
Making use of rsync would bring zypper the checksumming, automatic
download resuming/repairing at no cost ;-)

#7: Roberto Mannai (robermann79) (2009-11-30 21:58:19) (reply to #6)
"delouw" says that rsync does not support HTTP. This is a real blocking
problem.

#9: Jan Engelhardt (jengelh) (2009-12-05 13:30:59) (reply to #7)
Even if the rsync *program* could do HTTP, it would not help you much,
because HTTP does not implement that rolling checksum and all the other
fluffy things of rsync.
Also, it seems obvious to me that use of rsync is an optional extra
feature that you can chose to ignore when refreshing your
repositories.

#10: Robert Davies (robopensuse) (2009-12-17 11:27:59) (reply to #6)
There would be a cost, but it would be born by Mirrors and their
Admins.  Currently Mirrors can offer HTTP and trad. ftp, Mirror Brain
distributes downloads transparently and re-uses general cacheing
infrastructure.  Gentoo have used a seperate infrastructure for rsync-
ed portage data, rather than the usual high bandwidth/storage
traditional mirrors because rsync support was niche.  It would be rare
to enable the rsync checksum for daemon access on a public server,
because of the high CPU load of that feature, and potential for DOS
attack.
Implementing optionally, risks new ways for refresh to be slow eg)
rsync protocol requests are tried and dropped silently by uncooperative
firewalls.
If checksums & delta's are desirable additions for repo format, then a
more general solution which worked with HTTP would be better and
benefit more users and automatically re-use local proxy caches.

#8: Robert Davies (robopensuse) (2009-11-30 22:23:41)
Why not make the transfered refresh file, by definition based on
deltas?  First time through, you have current v empty file, the repo
can save a delta against the empty file &  and monthly / weekly
changes, with delta's for changes made against those, then a refresh
can check for updates in last week if it has an current week file
(download only if exists), monthly if the weekly is out of date and
fall backto delta v empty ifle if local montly & weekly are both out of
date.  Some sanity check, based on server's idea of date can prevent
clients getting things too horribly wrong.
Then most of the time it's a small file, that's easily cached for short
time; the monthly can be cached for longer with predictable TTL, and
current v empty could be cached for a day say.
I'm not sure about the repo format, but if delta handling involved
compression then wouldn't simple text format for the repo transfer
contents be natural and efficient, any binary file should likely be a
cache generated locally.
Though this sounds much more complicated, presumably there's tools for
generating the repo contents file, and processing the downloaded repo
file, so it ought not be so difficult in principal.



--
openSUSE Feature:
https://features.opensuse.org/306379

< Previous Next >
This Thread
  • No further messages