This is the problem. The ways in which it can fail can vary (sometimes you can't even pull logs from the web interface), and if it fails during a production deploy you may be left at half or no capacity.

When something inextricably effs up, rebuild the environment and more often than not, the problem dissapears.