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.'"

Comments

Sounds inefficient

[martin-boundary]

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.

Re:Sounds inefficient

[iONiUM]

I read TFA (I know, that's crazy). They don't come right out and say it, but I believe what they did it put a MapReduce type system (MapReduce splits the elements into subtasks for faster calculation) on database triggers. So basically this new system is spreading a database across their file system, across many computers, and allows incremental updates that, when occur, will trigger a MapReduce type algorithm to crunch the new update.

This way they get the best of both world. At least, I think that's what they're doing, otherwise their entire system would.. stop working.. since MapReduce is the whole reason they can parse such larger amounts of information.

Re:Sounds inefficient

[kurokame]

No, that's not it.

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 new framework is] completely incremental," he says. When a new page is crawled, Google can update its index with the necessarily changes rather than rebuilding the whole thing.

There are still cases where Caffeine uses batch processing, and MapReduce is still the basis for myriad other Google services. But prior the arrival of Caffeine, the indexing system was Google's largest MapReduce application, so use of the platform has been significantly, well, reduced.

They're not still using MapReduce for the index. It's still supported in the framework for secondary computations where appropriate, and it's still used in some other Google services, but it's been straight-up replaced for the index. Colossus is not a new improved version of MapReduce, it's a completely different approach to maintaining the index.

Re:Sounds inefficient

[kurokame]

Sorry, Colossus is the file system. Caffeine is the new computational framework.

I made the same error in several posts now...but Slashdot doesn't support editing. Oh well! Everyone reads the entire thread, right?

Re:Sounds inefficient

[onefriedrice]

Wait... you really have two friends named Irony and Sarcasm? That's incredible! What are the chances...

Re:Sounds inefficient

[maraist]

BigTable scales pretty well (go read it's white-papers) - though perhaps not as efficiently as map-reduce for something as simple as text to keyword statistics (otherwise why wouldn't they have used it all along).

I'll caveat this whole post with - this is all based on my reading of the BigTable white-paper a year ago, but having played with Cassandra, Hadoop, etc occasionally since then. Feel free to call me out on any obvious errors. I've also looked at a lot of DB internals (Sybase, Mysql MyISAM/INNODB and postgresql).

What I think you're thinking is that in a traditional RDBMS (which they hint at), you have a single logical machine that holds your data.. That's not entirely true, because even with mysql, you can shard the F*K out of it. Consider putting a mysql server on every possible combination of the first two letters of a google-search. Then take high density combinations (like those beginning with s) and split it out 3, 4 or 5 ways.

There are drastic differences to how data is stored, but that's not strictly important - because there are column-oriented table stores in mysql and other RDBMS systems. But the key problem of sharding is what's focused on Mysql-NDB-Cluster (which is a primitive key-value store) and other distributed-DB technologies that best traditional DBs at scalability.

BUT, the fundamental problem that page-searches are dealing with is that I want a keyword to map to a page-view-list (along with meta-data such as first-paragraph / icon / etc) that is POPULATED from statistical analysis of ALL page-centric data. Meaning you have two [shardable] primary keys. One is a keyword and One is a web-page-URL. But the web-page table has essentially foreign keys into potentially THOUSANDS of keyword records and visa-versa. Thus a single web-page update would require thousands of locks.

In map-reduce, we avoid the problem. We start off with page-text, mapped to keywords with some initial meta-data about the parent-page. In the reduce phase, we consolidate (via a merge-sort) into just the keywords, grouping the web pages into ever more complete lists of pages (ranked by their original meta-data - which includes co-keywords). In the end, you have a maximally compact index file, which you can replicate to the world using traditional BigTable (or even big-iron if you really wanted).

The problem of course, was that you can't complete the reduce phase until all web pages are fully downloaded and scanned.. ALL web pages. Of course, you do an hourly job which takes only high-valued web-pages and merges with the previous master list. So you have essentially static pre-processed data which is over-written by a subset of fresh data.. But you still have slowest-web-page syndrome. Ok, so solve this problem by ignoring web-load requests that don't complete in time - they'll be used in the next update round.. Well, you still have the issue of massive web-pages that take a long time to process. Ok, so we'll have a cut-off for them too.. Mapping nodes which take too long, don't get included this round (you're merging against you last valid value - so if there isn't a newer version, the old one will naturally keep). But the merge-sort itself is still MASSIVELY slow. You can't get 2-second turn-around on high-importance web-sites. You're still building a COMPLETE index every time.

So now, with a 'specialized' GFS2 and specialized BigTable, either or both with new fangled 'triggers', we have the tools (presumably) to do real-time updates. A Page load updates its DB table meta-data. It see's it went up in ranking, so it triggers a call to modify the associated keyword's table (a thousand of them). Those keywords have some sort of batch-delay (of say 2 seconds) so that it minimizes the number of pushes to production read-servers.. So now we have an event queue processor on the keyword table. This is a batch processor, BUT, we don't necessarily have to drain the queue before pushing to production. We only accept as many requests as we can fit into a 2 second time-slice. Presumably

There is another...

[bosef1]

So does that mean Microsoft is developing a competeing distributed computing system called "Guardian"? And how does that possibly seem like a good idea?

Awesome choice of name.

[Scytheford]

"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/

Re:Awesome choice of name.

[Anonymous Coward]

Colossus is also the name of the computers Bletchley Park used to crack the German Lorenz cipher.
http://en.wikipedia.org/wiki/Colossus_computer

I have to say...

[tpstigers]

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.

Re:I have to say...

[martin-boundary]

No. Eric's only a half-a-fruitbat.

Re:I have no idea

[icebike]

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.

Summarizing...summarizing...

[kurokame]

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.

Re:Well

[kurokame]

No, the old system was transactional as well. The problem was that it was transactional across a very large number of operations being run in parallel, and any failure could cause the entire transaction to fail. The new system is incremental rather than monolithic. While it may not be quite as fast across a large number of transactions, it doesn't risk major processing losses either. Such failures are very unlikely, but the Google index has grown large enough that it is probably running into unlikely problems all the time.

MapReduce is also staged, and the first stage must complete before the second can start. At Google's scales, this adds up to quite a lot of wasted power.

Processing a batch of data with Colossus is probably slower than using MapReduce under ideal circumstances. But failures don't incur a major penalty under Colossus, and MapReduce ties up CPU cycles with waits which aren't wasted under Colossus. Even if Colossus is slower under ideal circumstances, it's more reliable and more efficient in practice.

Re:Well

[kurokame]

Statistics: making the unlikely happen every day if you roll the dice enough times.

Re:I have no idea

[A Friendly Troll]

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.

Mod Offtopic, please

[Khyber]

This is going to give my Camfrog name a new meaning, as I *LOVE* screwing around with file systems. Colossus Hunter, indeed!

Re:Well

[TheRaven64]

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.

Re:I have no idea

[bitflip]

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=

Re:I have no idea

[Runaway1956]

Bing probably redirects the search to Google, then displays the results on their own page. Bleahhh.

Re:It is quick

[Surt]

I assume google polls sites, and polls faster every time it finds a change, slower every time it does not find a change. Eventually it gets to a wobbly around the probable update speed of the site. Otherwise they'd have to trust sites to call their API with updates, and that would let any search engine which DID employ a wobbly poll strategy to beat them in results.