Curious: why choose something like Watson over using haproxy's pretty solid built in health checking mechanisms?
Watson checks the health of your local application, and propagates it to the (global) service discovery framework, so when other microservices want to connect to the service that Watson is monitoring, they know whether or not that service is available.
1. In the central LB setup, if the LB server dies, you're service dies. Not in this setup (HAProxy is deployed side-by-side with each instance). 2. Elasticity. Imagine you have a shopping cart microservice, a search microservice, and a users microservice. Each of these requires its own HAProxy instance. Every time you spin up or spin down a new instance of these microservices, you need to configure the central HAProxy to pay attention to these new instances. 3. Health checks don't work well in DNS. In the centralized load balancer infrastructure, you end up relying on DNS so that your users microservice can talk to the shopping cart microservice (for example). DNS requires client polling and has propagation delays, so if one of your shopping cart microservice load balancer dies, it takes time for all the other microservices to figure out where to connect to.
A central LB works well if you have just a single microservice. But when you have dozens of them, you're suddenly managing dozens of load balancers dynamically, and it gets to be pretty chunky to manage at that point.
Lots of ways to solve this problem; our big focus here is on simplicity. If you have the Chef-fu and time to do all the above, you could definitely make it work.