The problem, as I understand it (and Brion may come by to correct me) is essentially that the current dump process is designed in a way that can't be sustained given the size of enwiki. It really needs to be re-engineered, which means that developer time is needed to create a new approach to dumping.
The main target for improvement is almost certainly parallelizing the process so that wouldn't be a single monolithic dump process, but rather a lot of little processes working in parallel. That would also ensure that if a single process gets stuck and dies, the entire dump doesn't need to start over.
By way of observation, the dewiki's full history dumps in 26 hours with 96% prefetched (i.e. loaded from previous dumps). That suggests that even starting from scratch (prefetch = 0%) it should dump in ~25 days under the current process. enwiki is perhaps 3-6 times larger than dewiki depending on how you do the accounting, which implies dumping the whole thing from scratch would take ~5 months if the process scaled linearly. Of course it doesn't scale linearly, and we end up with a prediction for completion that is currently 10 months away (which amounts to a 13 month total execution). And of course, if there is any serious error in the next ten months the entire process could die with no result.
Whether we want to let the current process continue to try and finish or not, I would seriously suggest someone look into redumping the rest of the enwiki files (i.e. logs, current pages, etc.). I am also among the people that care about having reasonably fresh dumps and it really is a problem that the other dumps (e.g. stubs-meta-history) are frozen while we wait to see if the full history dump can run to completion.
-Robert Rohde
On Tue, Jan 27, 2009 at 11:24 AM, Christian Storm storm@iparadigms.com wrote:
On 1/4/09 6:20 AM, yegg at alum.mit.edu wrote: The current enwiki database dump (http://download.wikimedia.org/enwiki/20081008/ ) has been crawling along since 10/15/2008.
The current dump system is not sustainable on very large wikis and is being replaced. You'll hear about it when we have the new one in place. :) -- brion
Following up on this thread: http://lists.wikimedia.org/pipermail/wikitech-l/2009-January/040841.html
Brion,
Can you offer any general timeline estimates (weeks, months, 1/2 year)? Are there any alternatives to retrieving the article data beyond directly crawling the site? I know this is verboten but we are in dire need of retrieving this data and don't know of any alternatives. The current estimate of end of year is too long for us to wait. Unfortunately, wikipedia is a favored source for students to plagiarize from which makes out of date content a real issue.
Is there any way to help this process along? We can donate disk drives, developer time, ...? There is another possibility that we could offer but I would need to talk with someone at the wikimedia foundation offline. Is there anyone I could contact?
Thanks for any information and/or direction you can give.
Christian
Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l