I pretty much abandoned the use of asserts in production code once I was introduced to test driven development. Amongst their failings, assert statements are only suited to testing very localised conditions, such as within a particular function - it's not clear to me how they can be used to simply verify the behaviour of a number of objects working in concert with one another.
Some recent papers have touched on this issue. For example 'Assessing the Relationship between Software Assertions and Code Quality' by Gunnar Kudrjavets, Nachiappan Nagappan and Thomas Ball of Microsoft: "We observe from our case study that with an increase in the assertion density in a file there is a statistically significant decrease in fault density. Further, the usage of software assertions in these components found a large percentage of the faults in the bug database."
However, both assertion density and low defect rates also correlate with higher levels of developer experience. So, nothing definitive is proven, but the study has got me thinking. One of the recurring themes in the Coders at Work book that I enjoyed recently, is that several of the expert practitioners interviewed in the book mentioned how useful they found thinking about code in terms of its invariants. The use of assertions does seem to lend itself to testing invariants like this - some quantity which can be calculated at the start of a function call, and then asserted to be unchanged by the end of it.
So can assertions be useful even in code that is well unit-tested? Are there any rules we can use to choose which things might be more productively tested using such asserts, rather than in unit tests?
I don't know. This is as far as I've got. Thoughts, speculations and anecdotes welcome.