Slashdot Mirror
Google Caffeine Drops MapReduce, Adds "Colossus"
An anonymous reader writes "With its new Caffeine search indexing system, Google has moved away from its MapReduce distributed number crunching platform in favor of a setup that mirrors database programming. The index is stored in Google's BigTable distributed database, and Caffeine allows for incremental changes to the database itself. The system also uses an update to the Google File System codenamed 'Colossus.'"
This sounds like it's going to be highly inefficient for nonlocal calculations, or am I missing something? Basically, if the calculation at some database entry is going to require inputs from arbitrarily many other database entries which could reside anywhere in the database, then the computation cost per entry will be huge compared to a batch system.
"This is the voice of world control. I bring you peace. It may be the peace of plenty and content or the peace of unburied death. The choice is yours: Obey me and live, or disobey and die. [...] We can coexist, but only on my terms. You will say you lose your freedom. Freedom is an illusion. All you lose is the emotion of pride. To be dominated by me is not as bad for humankind as to be dominated by others of your species. Your choice is simple."
-Colossus.
Source: http://www.imdb.com/title/tt0064177/
I am so glad Google has moved away from the Argus platform and into the Mercedes system. It makes it so much easier for those of us who are used to programming in Gibberish. Don't get me wrong - the days of Jabberwocky code were brilliant, but it's high time we moved into the Century of the Fruitbat.
Follow the link to the Original Article over on The Register , where you will find a rather lucid explanation, far better than the summary above can provide.
Short answer:
The old method of building their search database was essentially a Batch Job, Run it, wait, wait, wait a long time, swap results into production servers.
The new method is continuous updates into a gigantic database spread over their entire network,
This is why things show up in Google days, sometimes weeks ahead of the other search engines. The other guys are still trying to clone Google's old method.
Sig Battery depleted. Reverting to safe mode.
Colossus is incremental, whereas MapReduce is batch-based.
In MapReduce, you run code against each item with each operation spread across N processors, then you reduce it using a second set of code. You have to wait for the first stage to finish before running the second stage. The second stage is itself broken up into a number of discrete operations and tends to be restricted to summing results of the first stage together, and the return profile of the overall result needs to be the same as that for a single reduce operation. This is really great for applications which can be broken up in this fashion, but there are disadvantages as well.
MapReduce is a sequence of batch operations, and generally, Lipkovits explains, you can't start your next phase of operations until you finish the first. It suffers from "stragglers," he says. If you want to build a system that's based on series of map-reduces, there's a certain probability that something will go wrong, and this gets larger as you increase the number of operations. "You can't do anything that takes a relatively short amount of time," Lipkovitz says, "so we got rid of it."
The problem for Google is that the disadvantages scale. The fact that you have to wait for all operations from the first stage to finish and that you have to wait for the whole thing to run before you find out if something broke can have a very high cost at high item counts (noting that MapReduce typically runs against millions of items or more, so "high" is very high). With the present size, it's apparently more advantageous to get changes committed successfully the first time, even if MapReduce might be able to compute the result faster under ideal circumstances.
For example, why do you use ECC memory in a server? Because you have a bloody lot of memory across a bloody lot of computers running a bloody lot of operations, and failures potentially have more serious consequences than if a program on someone's desktop. At higher scales, non-ideal circumstances are more common and have more serious consequences. So while they still use MapReduce for some functions where it's appropriate, it's no longer appropriate for the purpose of maintaining the search index. It's just gotten too big.
This is why things show up in Google days, sometimes weeks ahead of the other search engines.
For a hands-on example of what icebike is saying, look here:
http://www.google.com/search?q=%22This+is+why+things+show+up+in+Google+days%2C+sometimes+weeks+ahead+of+the+other+search+engines%22
Actually, Google will index Slashdot comments in a matter of minutes.
Yes and no. With MapReduce, they were hitting Amdahl's Law. The speed limit of any concurrent system is defined by the speed of the slowest serial component. This is why IBM still makes money selling very fast POWER CPUs, when you can get the same speed on paper from a couple of much cheaper chips.
The old algorithm (massive oversimplifications follow) worked by indexing a small part of the web on each node, building a small index, and then combining them all in the last step. Think of a concurrent mergesort or quicksort - the design was (very broadly) similar.
The problem with this was that the final step was the one that updated the index. If one of the nodes failed and needed restarting, or was slow due to the CPU fan failing and the processor down-clocking itself, the entire job was delayed. The final step was largely serial (although it was actually done as a series of hierarchical merges) so this also suffered from scalability problems.
The new approach runs the partial indexing steps independently. Rather than having a separate step to merge them all, each one is responsible for merging itself into the database. This means that if indexing slashdot.org takes longer than expected then this just delays updates for slashdot.org, it doesn't delay the entire index update.
The jab at Microsoft in the El Reg article is particularly funny, because Google is now moving from a programming model created at MIT's AI labs to one very similar to the model created at Microsoft Research's Cambridge lab, in collaboration with Glasgow University.
I am TheRaven on Soylent News
Hmmm. Bing has it, too - both hits I got on Google, I got there, as well.
http://www.bing.com/search?q=%22This+is+why+things+show+up+in+Google+days,+sometimes+weeks+ahead+of+the+other+search+engines%22&go=&form=QBLH&qs=n&sk=
Bing probably redirects the search to Google, then displays the results on their own page. Bleahhh.
"Windows is like the faint smell of piss in a subway: it's there, and there's nothing you can do about it." - Charlie Br