Add app start differences to the upgrade guide.

[#172094705]

Co-authored-by: Weyman Fung <wfung@pivotal.io>
Co-authored-by: Teal Stannard <tstannard@pivotal.io>
Co-authored-by: Mona Mohebbi <mmohebbi@pivotal.io>

Author: 3 people

docs/v3/source/images/start_diagram.png

@@ -140,6 +140,28 @@ Here are some examples of implications for clients:
 2. An application can have multiple processes, each with their own start command, scale, and stats
 3. An application might not be running with its most recent package or droplet
 
+#### Starting Apps
+
+In the V2 API, the start endpoint (`PUT /v2/apps/:GUID` with `state`: `STARTED` in the request) was responsible for converting source code into a running executable in the cloud. In V3, [starting an app](#start-an-app) will only start the app processes with the current droplet.
+
+To reproduce the V2 start behavior in V3:
+
+![Start Diagram](start_diagram.png)
+
+1. [List packages](#list-packages) and filter on package state with value `READY` and order by recency.
+1. If a package has been turned into a droplet [(see this endpoint)](#list-droplets-for-a-package) this means it has been staged already. In V2 workflows, this would mean this package is what the current droplet is running.
+1. Stage the package by [creating a build](#create-a-build). This turns your package into a droplet.
+1. Update the app’s [current droplet](#set-current-droplet) to the selected droplet. This droplet will be run when the app starts.
+1. Change the app’s state to [started](#start-an-app)
+
+This gives V3 users more flexibility when managing applications. The following
+diagram shows many different flows for starting an app.
+
+![Start Flow](start_flows.png)
+
+ Apps can upload multiple packages, stage multiple droplets, roll back to older droplets, and other complicated workflows.
+
+
 ### Asynchronous Operations
 
 Unlike V2, clients cannot opt-in for asynchronous responses from endpoints. Instead, endpoints that require asynchronous processing will return `202 Accepted` with a Location header pointing to the job resource to poll. Endpoints that do not require asynchronous processing will respond synchronously.