Impact when changing instance count for asp.net web role in Azure

I'm having no luck trying to find out how channging the instance count for an ASP.Net web role affects requests currently being processed.

Heres the scenario:

1. An ASP.Net site is deployed with 6 instances
2. Via the console I reduce the instancecount to 4

Is azure smart enough to not remove instances from the pool if it is currently progressing requests or does it just kill them mid request?

I've been through the azure doco, goolge and a number of emails to MS tech support none of which were able to answer this seemingly simple question. I know about the events that get triggered by a sh开发者_C百科utdown etc but that doesnt really help in web site scenario with a live person waiting for a request to their response.

You cannot choose which instances to kill off. Primarily this is due to Windows Azure's instance allocation scheme, where your instances are split into different fault domains (meaning different areas of the data center - different rack, etc.). If you were to choose the instances to kill, this could leave you in a state where your remaining instances are in the same fault domain, which would void the SLA.

Having said that: You get an event when your role instance is shutting down (the OnStop() event). If you capture this event, you can do instance cleanup in preparation for VM shutdown. I can't recall if you're taken out of the load balancer at this point, but you could always force yourself out with a simple PowerShell command (Set-RoleInstanceStatus -Busy). This way your asp.net instance stops taking requests, and you can more easily shut down in a graceful manner.

EDIT: Sorry - didn't quite address all of your question. Since you get to capture OnStop(), you might have to implement a mechanism to make sure nothing's being processed in that instance. Since you're out of the load balancer, and assuming your requests are processed fairly quickly (2-5 seconds), you shouldn't have to wait long to clear out remaining requests. There's probably a performance counter to check, to see how many active requests are being handled.

Just to add to David's answer: the OnStop event happens when you are off the load balancer. For web apps, it is usually sufficient time to bleed out all requests after you are disconnected from the LB until the instance is shutdown. However, for long running or stateful connections (perhaps to a worker role), there would be an abrupt disconnect in some cases. While the OnStop method removes you from the LB, it does not terminate open connections. It simply prevents you from getting new connections. For web apps, this is usually enough time to complete the request (and you can delay the shutdown if necessary in the OnStop as well if you really want to).