(The below issues are very much on my mind as I think about how to apply for another NIH grant to fund continued development on the khmer project.)
Imagine that we have a graph of novel functionality versus software engineering effort for a particular project, cast in the shape of a tower or pyramid, i.e. a support structure for cool science.
The more novel functionality implemented, the taller the building, and the broader the software engineering base needs to be to support the building. If you have too much novel functionality with too little software engineering base, the tower will have too little support and catastrophe can ensue - either no new functionality can be added past a certain point, or we discover that much of the implemented functionality is actually unstable and incorrect.
Since everybody likes novel functionality - for example, it's how we grade grants in science -- this is a very common failure mode. It is particularly problematic in situations where we have built a larger structure by placing many of these individual buildings on top of others; the entire structure is not much stronger than its weakest (least supported) component.
Another possible failure mode is if the base becomes too big too soon:
That is, if too much effort is spent on software engineering at the expense of building novel functionality on top of it, then the building remains the same height while the base broadens. This is a failure for an individual project, because no new functionality gets built, and the project falls out of funding.
In the worst case, the base can become over-wrought and be designed to support functionality that doesn't yet exist. In most situations, this work will be entirely wasted, either because the base was designed for the wrong functionality, or because the extra work put into the base will delay the work put into new features.
Where projects are designed to be building blocks from the start, as opposed to a leap into the unknown like most small-lab computational science projects, a different structure is worth investing in -- but I'm skeptical that this is ever the way to start a project.
Supporting this kind of project is something that Dan Katz has written and presented about; see (for example) A Method to Select e-Infrastructure Components to Sustain.
And, of course, the real danger is that we end up in a situation where a poorly engineered structure is used to support a much larger body of scientific work:
The question that I am trying to understand is this: what are the lifecycle stages for research software, and how should we design for them (as researchers), and how should we think about funding them (as reviewers and program officers)?
To bring things back to the title, how do we make sure we mix the right amount of software development (cold porridge) with novel functionality (hot porridge) to make something edible for little bears?