On 03/08/21 at 02:12pm, Allan McRae wrote:
This is a very incomplete and hacky patch... But it does do this:
allan@mando ~/arch/code/pacman (patchqueue)
pacman -Syy :: Synchronizing package databases... core 131.1 KiB 47.0 KiB/s 00:03 [######################] 100% extra 1645.7 KiB 490 KiB/s 00:03 [######################] 100% community 5.4 MiB 161 KiB/s 00:34 [######################] 100% error: failed retrieving file 'extra.db' from allanmcrae.com : The requested URL returned error: 404 error: failed retrieving file 'community.db' from allanmcrae.com : The requested URL returned error: 404 error: failed retrieving file 'core.db' from allanmcrae.com : The requested URL returned error: 404
allan@mando ~/arch/code/pacman (patchqueue)
sudo ./build/pacman -Syy [sudo] password for allan: :: Synchronizing package databases... core 131.1 KiB 51.2 KiB/s 00:03 [######################] 100% extra 1645.7 KiB 494 KiB/s 00:03 [######################] 100% community 5.4 MiB 172 KiB/s 00:32 [######################] 100% error: failed retrieving file(s) from allanmcrae.com : The requested URL returned error: 404
There are a few things to note. 1) we lose the file names for the failed downloads. And we lose information of how many downloads failed. 2) I only changed the messages for download failures so far. So (e.g.) having a file exceed its expected download size will still report the file name. 3) debug output still gives file names 4) note the "error" in the messages above, yet pacman was still successful as I have an actual mirror configured. I guess this (and others?) should be changed to warning. We get an overall error message if there is complete failure anyway. 5) I have been super lazy for this PoC, switching alpm_list_remove_dupes() to use string comparisons, rather than pointer comparisons. This will need fixed.
Things to discuss: - do we need to report the number of failures from a server. Currently 1 and 200 failures are treated the same. This involves doing more than a dedup of the error list. - converting errors to warnings. They are clearly not errors until they are... Is the final line enough: error: failed to synchronize all databases (failed to retrieve some files)
My main problem with this approach is that now alpm is spewing a bunch of identical errors to front-ends, so now unless a front-end dedups them the same way, we're actually making things worse. We could dedup in alpm, but then I don't think there's really an obvious place to actually send the message. Sending it at the end makes it impossible for the user to infer that they might have a bad mirror and should cancel the download and switch. Sending it on the first occurrence does the same because they don't know if it's just one file or a recurring problem. What about actually removing a bad mirror instead of deduplicating the resulting messages? Maybe start using the next mirror for subsequent payloads after n failures or immediately for things like an unresolvable host? If we do make the errors more generic, I think turning them into warnings and relying on the final "failed to retrieve some files" to indicate actual failure is fine.