Distributed is not necessarily more scalable than centralized

Centralized is not necessarily unscalable! 

Many people automatically associate centralized with unscalable, and distributed with scalable. And, this is getting ridiculous.

In the Spring semester, in my seminar class, a PhD student was pitching me a project for distributed storage: syncing from phone to work/home computers and other phones. The pitch started with the sentence "Dropbox is unscalable, because it is centralized". I was flabbergasted, and I asked a couple of times "Really? Do you actually claim that Dropbox is unscalable?". The student persisted and kept repeating that "Dropbox has a bottleneck because it is a centralized storage solution, and the distributed solution doesn't have that bottleneck". I couldn't believe my ears.

Dropbox already proved it is scalable: It serves files for more than 200 million users, who store 1 billion files every 24 hours. That it has a centralized architecture hosted in the cloud doesn't make it unscalable. As far as I can see there is no bottleneck caused by Dropbox having a more centralized architecture.

(For those who want to nitpick, I know Dropbox is not fully centralized; it uses AWS S3 for storage and Dropbox-company servers for metadata management. Also, it employs data parallelism in the backend for scalability, but, on the spectrum, it is closer to a centralized architecture than a fully decentralized one.)

Distributed is not necessarily scalable!

Some people when faced with a problem think, I know, I'll use distributed computing. Now they have N^2 problems. -- @jamesiry
Here is the second part. Distributing a system does not necessarily make it scalable. In fact, a fully decentralized architecture can sometimes be a disadvantage for scaling.

Consider Lamport's mutual exclusion (ME) algorithm presented in his seminal "Time, Clocks, and the Ordering of Events in a Distributed System". This ME algorithm is fully decentralized, and requires O(N) messages to be exchanged in response to one ME request. The Lamport ME algorithm employs broadcasts to keep all the nodes informed of all updates and get them on the same (more or less) state.

Now consider a centralized algorithm for ME: there is a centralized coordinator; the nodes send their request to the coordinator, and the coordinator assigns ME accordingly. (For the literalist: You can still have causal ordering in the centralized algorithm. Just use VC when nodes communicate and include VC in the request messages.) The centralized ME algorithm is more scalable: only 1 message is exchanged in response to one ME request. It has less drama and it is easier to maintain and build over.

Single point of failure?

A distributed system is one in which the failure of a computer you didn't even know existed can render your own computer unusable. -- Leslie Lamport
A common reflex argument about centralized solutions is that it constitutes a single point of failure (SPOF). But if a distributed solution is not designed carefully, it will have multiple points of failures (MPOF). Which one would you rather have?

Let's reconsider the Lamport ME and the centralized ME algorithms. The distributed algorithm does not offer any fault-tolerance advantages. Both algorithms are prone to getting stuck with one crash failure.

In fact, we can argue that it is easier to design fault-tolerance to a centralized solution: You can employ Paxos to replicate the centralized server. In contrast, it is often much harder to design and add fault-tolerance to a distributed system. Since a distributed system is complex, it is more prone to introduce corner cases that jeopardize fault-tolerance.


Distributed is not necessarily more scalable than centralized;
And centralized is not necessarily a scalability bottleneck.

As a distributed systems professor, I wouldn't imagine myself defending centralized solutions. But there it is.

To avoid potential misunderstandings, I am not saying fully distributed/decentralized solutions are bad and to be avoided. There are advantages to decentralization, like latency reduction. And some conditions necessitate decentralization, like geographic/political/corporate isolation. We know in the real world it is a mix of centralized, up to where that is manageable and has reasonable cost, and some distributed architecture. This also depends very much on the application/task.

PS: Maybe we should do an XtraNormal animation movie about this "centralized unscalable and distributed scalable" mania. Any takers?

PS2: I thank @tedherman for improvements to the 1st draft.

PS3: Optimistic replication is a great survey of more decentralized replication protocols, their advantages, and challenges.

Bonus Section: Paxos is a relatively centralized approach to distributed consensus

Consensus is usually not an all-hands-on process. That can be hard to scale. Consider our democratic system: It is pretty centralized; we only elect leaders to rule for us.

In the same sense, you can think of Paxos as the more centralized approach to distributed consensus and state machine replication. In Paxos, the participants do not interact with all other participants to decide the order of requests to be accepted, instead the leader dictates the order of requests and the participants just accept them. (A fully decentralized consensus algorithm would be like the synchronous rounds consensus algorithm where in every round each participant communicates with all other participants so that they can converge on the same state.)


Anonymous said…
What part of dropbox is centralized? It is massively distributed.
Anonymous said…
I think all the "magic" is done in AWS using S3. Dropbox is just some kind of wrapper of S3.
Bill Broadley said…
Shouldn't scalable include some idea of resources used, at a gross approximation you could use cost. So sure dropbox can scale, by using a very expensive global network of data centers run by amazon.

I agree that with enough resources anything can be done centrally, but not as efficiently.

Wouldn't you say that a distributed system scaled better than dropbox if it managed similar or better performance without using a very expensive global network of data centers?

So drop box's bottleneck isn't that they can't buy enough resources from amazon, but that they miss out on users who don't want to pay dropbox's rates.

Like say bittorrent's sync product, a relatively simple p2p application that provides similar functionality to dropbox, without Amazon. Yet it seems to outperform dropbox handily.

Details at this link:

I'd say that dropbox is a good example how centralized is not necessarily more scalable than distirbuted.

Popular posts from this blog

The end of a myth: Distributed transactions can scale

Hints for Distributed Systems Design

Foundational distributed systems papers

Learning about distributed systems: where to start?

Metastable failures in the wild

Scalable OLTP in the Cloud: What’s the BIG DEAL?

SIGMOD panel: Future of Database System Architectures

Dude, where's my Emacs?

There is plenty of room at the bottom

Distributed Transactions at Scale in Amazon DynamoDB