This is why we need better application profiling yesterday. Not "I need write to x location" but "I am highly sensitive to read latency in x location, and highly sensitive to available space (append only) in y location." cgroups is the beginning. Eventually, a much-evolved container notion will solve this. I expect embedded and mobile distributions/OSs to follow suit.
The reason is simple: some applications write huge temporary files, and lots of people don't have much free memory to spare.
And you can solve the pileup by simply cleaning it up regularly using a scheduled task.