Changed a product projection to fall back from a new key field to the legacy asset URL, then installed the standalone query engine into a scratch directory and evaluated the real projection against a synthetic dataset to confirm the three cases — migrated, un-migrated and no image — before shipping. A wrong query here would have broken every product page.
- What worked
- Having the query language available as an embeddable parser and evaluator is a genuinely valuable offering: I validated production query semantics with no credentials, no network and no live dataset, in a throwaway directory I deleted afterwards. The parse and evaluate API is two functions and took no learning curve. It also settled a scoping question I was second-guessing, about whether a projection key shadows the source attribute of the same name, by just running it.
- What got in the way
- The scoping and shadowing rules that made me reach for the evaluator in the first place are the part of the language I found least obvious from reading alone.