On Sun, 2009-12-13 at 03:31 -0500, Qadri wrote: > On Sun, Dec 13, 2009 at 2:07 AM, Brendan Long <korin43@xxxxxxxxx> wrote: > > > On Sun, 2009-12-13 at 11:16 +0800, Ng Oon-Ee wrote: > > > On Sat, 2009-12-12 at 17:08 -0700, Brendan Long wrote: > > > > 2009/12/11 Ng Oon-Ee <ngoonee@xxxxxxxxx> > > > > > > > > > On Sat, 2009-12-12 at 02:13 +0100, Heiko Baums wrote: > > > > > > Am Sat, 12 Dec 2009 08:58:17 +0800 > > > > > > schrieb Ng Oon-Ee <ngoonee@xxxxxxxxx>: > > > > > > > > > > > > > Because sometimes all the mirrors listed in mirrorlist will not > > have > > > > > > > the file, if its just been uploaded. Also not everyone stays > > > > > > > up-to-the-minute with updates, judging by the "updated after a > > month" > > > > > > > posts we see once in a while. > > > > > > > > > > > > > > I'm concerned about the last bit, if a package was just uploaded > > and > > > > > > > only exists on one mirror, everyone who updates and has that > > package > > > > > > > in the period between its uploading and its appearance on their > > local > > > > > > > mirrors will 'fall-back' on varying mirrors (lengthening the > > update > > > > > > > process) and all end up on the poor main server (or Tier 1/2 > > mirrors). > > > > > > > Bad for both the mirror bandwidth as well as most probably much > > slower > > > > > > > for the user, who could probably just wait a day or so for the > > update > > > > > > > to come to his (faster, presumably) local mirror. > > > > > > > > > > > > > > > > > > Wouldn't it be possible to first upload the packages and update the > > db > > > > > > files when the packages on the mirrors (at least on several > > mirrors) > > > > > > are updated? > > > > > > > > > > > > If I have such a "problem" that a package is on no mirrors, which > > > > > > doesn't happen often, I usually abort the system update and wait > > one > > > > > > day. I think that's the normal and easiest way of solving this > > issue. > > > > > > > > > > > > Greetings, > > > > > > Heiko > > > > > > > > > > The few mirrors which sync first would have quite much higher > > bandwidth > > > > > usage =). > > > > > > > > > > The concern then is that in the period of time between uploading of > > > > > packages and updating of db, the db would point to a package > > (foo-1.3) > > > > > while the mirror would only have the new version (foo-1.4), since I > > > > > don't think many mirrors keep multiple copies of the same package > > > > > (schlunix I know off, any others?). So that would break updating as > > > > > well, just in a different direction, and this would not be > > recoverable > > > > > from. > > > > > > > > > > > > > > Maybe a better idea would be to make pacman keep track of the last time > > it > > > > got an updated package list, and if it's beyond a certain point, it > > starts > > > > checking other mirrors (maybe optional "No updates have been found in 5 > > > > days, would you like to scan other mirrors for updates?"). > > > > > > The difference I would see is that in the current system an out-of-date > > > mirror is still use-able (no mismatch between db and package as we're > > > discussing). Some would manually change mirrors, but others would not, > > > so the net effect of automating fallover would be to increase load above > > > what it currently is. > > > > > > Not everyone needs packages right on the day they're uploaded, or runs > > > pacman everyday (I do, though =p) > > > > > > > Not everyone runs pacman every day, but when you do, you expect to get > > up to date packages. > > > > > > So should it be a function of the program to make sure that happens? Or is a > responsibility of the user? Should the functionality be programmed into > pacman to make sure that happens, or should we be asking that users be aware > of what repos they're using? Well said, I agree. I believe that if separate db and package downloads are implemented it should not be so users can be 'up-to-the-minute' in packages, but for greater security. In fact, now that I think about it, having two dbs (one on the mirror with all packages as available on that mirror and one 'master' with a list of authoritative checksums) would make sense, as it fulfils the security aspect well while avoiding the problem of db/package mismatch. The 'master' db would have to have a history of previous checksums as well.
