Slashdot Mirror
Power Outage Takes Wikimedia Down
Baricom writes "Just a few weeks after a major power outage took out well-known blogging service LiveJournal for several hours, almost all of Wikimedia Foundation's services are offline due to a tripped circuit breaker at a different colo. Among other services, Wikimedia runs the well-known Wikipedia open encyclopedia. Coincidentally, the foundation is in the middle of a fundraising drive to pay for new servers. They have established an off-site backup of the fundraising page here until power returns."
What happened?
At about 14:15 PST some circuit breakers were tripped in the colocation facility where our servers are housed. Although the facility has a well-stocked generator, this took out power to places inside the facility, including the switch that connects us to the network and all our servers.
What's wrong?
After some minutes, the switch and most of our machines had rebooted. Some of our servers required additional work to get up, and a few may still be sitting there dead but can be worked around.
The sticky point is the database servers, where all the important stuff is. Although we use MySQL's transactional InnoDB tables, they can still sometimes be left in an unrecoverable state. Attempting to bring up the master database and one of the slaves immediately after the downtime showed corruption in parts of the database. We're currently running full backups of the raw data on two other database slave servers prior to attempting recovery on them (recovery alters the data).
If these machines also can't be recovered, we may have to restore from backup and replay log files which could take a while.
I was just in freenode joking with Jimbo about this. He said he thought was wondering how long it would be before slashdot ran a story about it (2 hours) and asked people to please stop with the consideracy theories. Meanwhile, the devs are working fairly furiously to get it back up (Kate hasn't slept in 27 hours. Jimbo just declared Feb 22 to be Kate-day) (--A wikipedia admin.)
To make laws that man cannot, and will not obey, serves to bring all law into contempt.
--E.C. Stanton
So far one of our database servers has completed a successful recovery (we're working through them all). On a gigabit link it takes something between 90 minutes and 4 hours to rsync from one to another. As soon as we have two database servers working, we'll be restoring service in read only mode. Likely to be that 90 minutes to 4 hours from now as worst case.
I'll post followups to this post later, as we're closer to being fully recovered.
No database can guarantee data integrity in the case of a power failure.
Think again. Techniques to do this have been around for years -- it's called stable storage. You just keep redundant copies of data that's changing, and use a neat and simple procedure to ensure that either they both get updated by a transaction, or the original data can be recovered. Certainly the most recent data might be lost, but there's no reason for the database to be corrupted or even in an inconsistent state.
Karma: Segmentation fault (tried to dereference a null post)
I just love stupid trolls that can't even use Google.
Tsearch2 - full text extension for PostgreSQL
DevX: Implementing Full Text Indexing with PostgreSQL - about Tsearch2.
Tsearch2 is included in the postgresql-contrib package of at least Debian and Novell/SuSE. Is that "out of the box" enough for a clueless MySQL user?
Yes. I wrote that cached page and it's now a bit out of date. IF, and it's not certain, local fire regulations permit the use of UPS systems in the racks we're going to be installing them. Decided on that after LiveJournal's unfortunate experience. But don't yet have them.
I find it an interesting coincidence the power outage happened so soon after that the Xenu article was featured.
Gee, you just had to mention the X-word! Now this thread won't load for most Scientologists because the keyword filters they were forced to install by their Church will see "Xenu" and block the site. After all the mere sight of the word could cause "pneumonia and death" if you haven't paid the Church of Scientology for the proper preparation.
Wikipedia's Xenu article has an interesting history if you look, as I did the other night when it was featured. Scientologists vandalize it regularly. You're supposed to pay them a half million (or some absurd sum of money) to find out about Xenu. After you find out, you're too embarrassed to admit to anybody that you paid a half million to learn that your problems are caused by bad science fiction, when you could have bought a house in Silicon Valley instead. So they obviously don't want a Wikipedia article giving away their half-million-dollar "trade secret" for free.
One trick I saw was to use HTML entities to spell out insults at the top of the article- like "only an idiot would believe this" or something. In the editor window, the entities weren't rendered and each letter appeared as a hex code.
A more effective attack took a different approach. The vandal in this case changed "Scientologists" to "Muslims", "Scientology" to "Islam", and inserted a boring-sounding sentence at the end of the first paragraph claiming that "Xenu" is another name that Muslims use for "Allah". It completely discouraged you from reading further. If you didn't know better you wouldn't find out how "Allah" distributed the thetans around volcanoes on various planets and blew them up with hydrogen bombs, and how their blown-up spirits cause problems in your personal life today.
This is OT, but what the hell, why not whack a beehive? Additional information on Xenu:
Operation Clambake (Hubbard maintained that humans are descended from clams)
The Xenu leaflet (all about Xenu- this information can save you lots of $$$$$)
The road to Xenu (authored by a woman who got suckered)
The Google cache of Wikipedia's Xenu article is also a must read.
I'm wondering if I'll get a lot of freaks, downmoderations, and hostile AC replies after I post this. After all, that's the kind of thing that Hubbard called "fair game". If it sinks below default visibility I'll repost it again with my karma bonus, so you theta-clear-wannabes out there can save your points for someone else.
http://www.postgresql.org/docs/8.0/interactive/wal .html
Posted on the mailing list wikipedia-l 32 minutes ago:
From: Brion Vibber
Reply-To: wikipedia-l@wikimedia.org
To: Wikipedia-l, Wikimedia Foundation Mailing List, Wikimedia developers
Date: Tue, 22 Feb 2005 04:47:56 -0800
Subject: Re: [Wikipedia-l] Wiki Problems?
Brion Vibber wrote:
> There was some sort of power failure at the colocation facility. We're
> in the process of rebooting and recovering machines.
The power failure was due to circuit breakers being tripped within the colocation facility; some of our servers have redundant power supplies but *both* circuits failed, causing all our machines and the network switch to unceremoniously shut down.
Whether a problem in MySQL, with our server configurations, or with the hardware (or some combination thereof), most of our database servers managed to glitch the data on disk when they went down. (Yes, we use InnoDB tables. This ain't good enough, apparently.)
The good news: one server maintained a good copy, which we've been copying to the others to get things back on track. We're now serving all wikis read-only.
The bad news: that copy was a bit over a day behind synchronization (it was stopped to run maintenance jobs), so in addition to slogging around 170gb of data to each DB server we have to apply the last day's update logs before we can restore read/write service.
I don't know when exactly we'll have everything editable again, but it should be within 12 hours.
Easily. See what those saying that MySQL can't do what MySQL does are promoting.:)
LiveJournal found that it had some disk systems which lied about having committed writes. The have a preliminary tool which copies what it's writing to disk to a networked system and then compares the after power off and recovery state to what the disk system said it could do. Are going to make it available to the community as time allows.
I expect we're going to find the same at Wikipedia. Here's a pretty typical error log, this one from the server which was master database server:
050222 5:11:12 InnoDB: Database was not shut down normally.
InnoDB: Starting recovery from log files...
InnoDB: Starting log scan based on checkpoint at
InnoDB: log sequence number 303 1283776146
InnoDB: Doing recovery: scanned up to log sequence number 303 1289018880
InnoDB: Doing recovery: scanned up to log sequence number 303 1294261760
InnoDB: Doing recovery: scanned up to log sequence number 303 1299504640
InnoDB: Doing recovery: scanned up to log sequence number 303 1304747520
InnoDB: Doing recovery: scanned up to log sequence number 303 1309990400
InnoDB: Doing recovery: scanned up to log sequence number 303 1315233280
InnoDB: Doing recovery: scanned up to log sequence number 303 1320476160
InnoDB: Doing recovery: scanned up to log sequence number 303 1325719040
InnoDB: Doing recovery: scanned up to log sequence number 303 1330961920
InnoDB: Doing recovery: scanned up to log sequence number 303 1336204800
InnoDB: Doing recovery: scanned up to log sequence number 303 1341447680
InnoDB: Doing recovery: scanned up to log sequence number 303 1346690560
InnoDB: Doing recovery: scanned up to log sequence number 303 1347688389
InnoDB: 1 transaction(s) which must be rolled back or cleaned up
InnoDB: in total 14 row operations to undo
InnoDB: Trx id counter is 1 935480064
050222 5:11:13 InnoDB: Starting an apply batch of log records to the database...
InnoDB: Progress in percents: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 InnoDB: Database page corruption on disk or a failed
InnoDB: file read of page 8617985.
InnoDB: You may have to recover from a backup.
050222 5:12:20 InnoDB: Page dump in ascii and hex (16384 bytes):
Observe that the database engine went back to its last checkpoint, noticed the partial transaction and undid it and was rolling ahead in the write-ahead log when it encountered a database page which failed its checksum test. That failed checksum test is why I think it's a problem with the disk system lying about what was written. You can get that when a database page spans two drives in a stripe set and one has committed the update while the other hasn't.
In more typical situations MySQL simply applies the updates and all is well. I've had a server set up to exceed RAM with swap turned off and get killed every ten minutes for hours and recover every time.
Just to be complete: