December 23, 2013

ESB's don't scale, which shouldn't be a problem because that is by design

Other posts in this trilogie:
Let's address that scalability issue. First of all, an ESB doesn't scale. That's by design. It was never intended to scale, it was intended to be the pivot in enterprise communications.

This is a post in a series about the Enterprise Service Bus, ESB, an acronym so easily to love or to hate. Parts of this post are also found in another post on the dimise of the ESB in a world of webservices.

ESB's are not buses at all, we only draw them as a bus or rather as a tube to which everything is connected. We draw them the same way we draw TCP/IP networks. But they are actually the hub in a hub-and-spoke model. They have to be, because they are the middleman handling all the Babylonical speak impediments within an enterprise.

An important aspect of the problem of not being scalable is state. ESB's have the crucial design flaw of being able to keep state.

Remember what the ESB was good for?

  1. Data Exchange
  2. Data Transformation
  3. Data Routing

When you talk to an ESB vendor about what the ESB can do for you, the vendor will tell you the three features I've listed before and an important 4th one. Well, the vendor thinks it's important, I believe it's a design flaw. This feature is "Data Enrichment". What the vendor means is that you can send a half-baked message to an ESB and before the ESB is delivering the message to its final destination, the recipient, it will call all kinds of other 'services' to enrich the original message with more information which is needed to deliver the message and have it understood, processable, by the recipient. This means that the ESB needs to keep state, while enriching the message. It also means that the ESB is no longer a mere intelligent data transport that routes and transforms on the fly, but it has become an application, a business application.
Because the ESB is designed to be able to do this, the ESB is designed to be able to keep state. And thus it doesn't scale. It scales as far as the box it is running on can scale. Scalability is all vertically.
There's another problem with the scalability and that is the dependency on physical resources. The queues of an ESB are physical, they're filesystems, databases or something else that's physical and allows for some persistence, or even complete persistence. This means again, that it doesn't scale because it needs coordination of accessing these resources because they guarantee sequence.

When scaling an ESB, it needs to be setup across nodes and there needs to be a vivid dialogue between the nodes about what each node is doing and has been doing, this is pure overhead. The busier the ESB is, the more chatter, ie the more overhead. This will require more nodes which requires more chatter. The result is a very disappointing performance curve.

Don't worry, this is all by design. The design here is that the ESB should be doing more than just be an intelligent data transport, because were it just an intelligent data transport, it would not have any added value in a SOAP ruled WS* compliant world, which is today's world. There's a whole blog post on this.
The ESB is designed to have added functionality to warrant the purchase of its licenses. This added functionality is allowing the sender (or client, or consumer) to not care about the sequence of its messages, because the ESB handles this. But that's a moo point since the client will either do a fire and forget (a feed) and not worry about sequence or it will wait for the response to a request before continuing processing. No added value at all. But the ESB, by queueing also ensures the sequence of requests or messages from several clients. So the recipient (or producer, or server) gets the requests or feed in the order the ESB receives them. Which means nothing because this may not at all be the same sequence the clients did in fact send them. Think about latency of different clients causing the ordering of messages going boinkers all over. Meanwhile this desparate need to sequence requires that there is a single point in the ESB doing all the sequencing. Which means dropping the desire to be scalable.

Ask the ESB vendor again why you should get yourself an ESB after reciting the previous, and your vendor will likely start talking about monitoring. How convenient it is to have the ESB doing all the monitoring for you, keeping track of an audit trail or at least a message trail because it is the central hub through which all the messages flow. Emphasizing the ESB's main flaw again, it being a central hub.

The ESB doesn't scale and it shouldn't be applied in an environment where it is meant to scale beyond just a few boxes. This is why the 'Enterprise' in ESB is misleading. It makes far more sense to have several ESB implementations in the enterprise, serving specific needs. There is a place for the ESB in the enterprise, trust me.

As always, I'm really interested in your views and ideas. More perspectives make for a richer understanding of the topic. So please share your thoughts, stay respectful and be argumentative.


No comments:

Post a Comment