Like everything, using code generation is a tradeoff. Basically, you want codeGen when the cost of writing the templates is less than the costs of writing what the template generates. For example, codeGen rocks at creating and maintaining tedious plumbing code - data access plumbing is the canonical example.
You should probably consider code generation when the target code:
- Has similar patterns and deterministic rules, and little deviation from those rules. If there's an MS word doc or wiki page providing detailed instructions on how to write certain code, you may be able to send those instructions to the code generator instead.
- Is large and brittle, and it can be described easily (i.e. 1 line in an xml config fiile to describe = 20 lines of C# and SQL coding).
- Requires syncing with external data sources (like ADO.Net plumbing based on the database schema, such as var type, size, parameter order, count).
- Requires multiple files to be kept in sync (like a SQL stored proc being in sync with C# ADO.Net wrapper).
- Requires in-depth knowledge of a problem domain (like ADO.Net for data-access plumbing).
- Is continually updated (like adding new classes to your DAL).
- May need to be expanded in the future (like adding a whole new layer of webservices, or Audit triggers, to your DAL. Or, perhaps even migrating to a new language).
Code Generation is not a "Golden Hammer" - while it's great, it's not the perfect solution for everything. codeGen may not be the best solution if the target code:
- Is very custom with no general pattern. If you can't abstract a pattern out of the target code, then you won't be able to write a generation template.
- Is too small and trivial - In general, codeGen should decrease your total lines of code. So if you're writing a 50 line template to produce a single 30 line C# file, it's probably a bad ROI.
- Can be refactored away, and doesn't need to exist in the first place.
Like everything else, in some contexts, it just isn't profitable - but in other contexts it's awesome. Some problems are cheaper to solve using other techniques, like unit testing, automation, open-source, DSLs, or other techniques; but every advanced developer should have code generation in their tool belt.