Once the cache has been filled all subsequent application deployments fail until the cache self-cleans. In CAS.log you see that the client is refusing to download the content because the cache, not the actual disk, is full.

This strikes me as a non-optimal design choice if the goal is to successfully install applications. If I want to install an application I do not want it to fail because of an artificial limit that until very recently was set at the time of client install. Most of the time the cache clears in 24 hours and there's no problem but there are two specific and prevalent cases where it is.

Task Sequences: During OSD we want every app and every update applied before the user gets the machine and we want it done quickly. It's very easy to go over a limit that day-to-day is ideal. This has forced SCCM admins to either abandon the application model during TS (a common MVP suggestion) or write custom scripts to temporarily increase the cache. Neither of these are ideal.

Large Applications: There are some applications (ex. CAD) that are just incredibly large (20+ GB) and all their own will exceed what an organization desires to reserve for cache size. Again, admins are either abandoning the app model entirely for these or running temporary cache modification scripts. This simply shouldn't be necessary for a system designed to deploy applicatons.

So, what to do? One option would to implement a rolling (first in, first out) cache where if cache space is needed the oldest non-persistent content for a compliant application is removed to make space. If the new app's content is larger than the cache itself then simply don't cache it. The only hard limit should be on the amount of free disk space to reserve before the SCCM client refuses to download content of any type.

TL;DR: Stop breaking application deployment just because the cache is full. No one wants to troubleshoot app failures.