[Mageia-discuss] Easyurpmi forMageia ?

[Michael Scherer]

Le dimanche 22 mai 2011 à 10:21 +0200, Wolfgang Bornath a écrit :
> 2011/5/22 Michael Scherer <misc at zarb.org>:
> > Le dimanche 22 mai 2011 à 05:51 +0200, Wolfgang Bornath a écrit :
> >> I haven't used EasyUrpmi, but I used the other (similar) web tool
> >> SmartUrpmi regularly.
> >>
> >> My reason was not to add PLF or any other repo (MUD or other). My main
> >> reason was my dislike of the %MIRRORLIST system of Mandriva, you could
> >> get connected to a very slow mirror or even to one which was not
> >> uptodate without urpmi being able to switch to another mirror during
> >> one session (a known bug). So I used SmartUrpmi to select a mirror
> >> which I knew to be fast, reliable and uptodate instead of playing
> >> lottery with %MIRRORLIST.
> >
> > There is likely stuff to improve with mirrorlist selection indeed ( I
> > would add AS number to it, as well as bandwidth, and push a smarter
> > system in urpmi one day to get nearer mirros, and add some weight based
> > on network related stuff ), but since everybody work around it, no one
> > is motivated to improve it or provides information for that. So because
> > people able to fix focus on not facing the problem and because there is
> > a know workaround, newer users face the problem.
> 
> As soon as the %MIRRORLIST system
>  - checks the fastest mirror before using it

So far, no one gave a working system to find which one is the fastest
without a exhaustive test of all mirrors ( such a test would be quite
wasteful from my point of view, and quite irresponsible for people with
metered download ). And in fact, that's not what you likely want. You
want a mirror that is fast enough to be limited by your current network
connexion. If the ADSL connexion is full, that the mirror is giving 10
or 100 mo directly to you doesn't matter much.

Now, as I said, better system for selecting mirrors should be used. But
for that, we need to know :
1) what is the current system ( I think it is based on the geographical
position, which is a very very broad assumption, it help mostly to avoid
brasil people fetching stuff from europa and thing like that )

2) when did it gone wrong ( ie, the mirror that was selected and that
shouldn't, and why, and how could we discriminate for a user )

In fact, from a network point of view and given current peering
business, the fastest may not be the cheapest one for your isp, so
that's even harder ( cause they could do some throttling, etc ).

Hence the suggestion of using AS numbers. 
First check if a mirror is on the same network, then on the same AS,
then in the same country, then in the same timezone, then in the same
continent. It will likely make people in Antarctic unhappy.
 

>  - checks the update status of the mirror before using it
usually, for non cooker/cauldron, this is likely not a issue. Release is
frozen, at worst, you have a security update being late. 

And the current system should check more aggressively if the mirror is
up to date ( but it seems to not be the case at the moment since ibiblio
was late due to a change on their side ). 

On the other hand, we cannot force people to sync every 2h, and if a
tier2 mirror sync every day, it can quickly become out of the list
depending on the limit. If someone sync on a tier2 mirror that is
already late ( for some network related reason, like in australia where
the country is not that well connected to the rest of the world ), it
will also be more late, and that would be a bad side effect.

Again, we need to have specific case of breakage to see how we could fix
anything, and more data on the mirrors to decide how and what to do,
what would be the impact, etc.

>  - switches to another mirror if the selected mirror is not reachable
> or not uptodate
I tried to fix that part, but unfortunately, the code is "suboptimal".
OTOH, using a manual mirror do not switch to a better mirror in case of
problem too ( IIRC ).

> I will gladly forget all webtools and dedicated mirror selection.
> (I remember writing almost the same at least 2 years ago in a same
> discussion at Mandriva)
> 
> > Worst, since people focus on specific mirrors ( d-c being the prime
> > example ), the mirror slowly become more and more overloaded with
> > requests, forcing admin of using complex scheme to divide the
> > ressources.
> 
> You can't avoid this. Users want to be connected to teh mirror which
> provides the best service, you can't force them to use a slow mirror
> just to protect better mirrors.


We do not force them, they do by themself. If everybody use the same
mirror, sooner or later, the mirror become the slowest. Maybe sooner if
the admin is a trafic shaping master like Olivier.

And after, we will see another mirror admin telling us "we removed the
tree because no one was using it" ( like switch.ch did for Mandriva ).

So if this is what we want and repeat the same problems over and over
because we think we cannot fix users behavior, yes, let's do nothing.

I think that the mirrors ecosystem also depend on the behavior of people
using it, and it is everybody responsibility to make sure it work fine.

-- 
Michael Scherer