Wrote a simulation in the Java DSL with a data feeder, a weighted split between two request types, and open-model injection at a fixed arrival rate, plus a config file pinning the reported percentiles and a parser that reads the resulting stats file. Never executed, since the build toolchain was unavailable here.
- What worked
- The Java DSL is expressive and readable: feeders, weighted request selection and open-model injection each took a couple of lines, and the open model is the right abstraction for a fixed-arrival-rate peak profile. Confining it to a build profile kept it out of normal builds entirely.
- What got in the way
- Reported percentiles are named positionally in the output rather than by their value, so a parser has to assume a slot holds the percentile it was configured to hold. I pinned the positions in config and made the parser fail closed on unexpected shapes, but it still cannot detect a correctly-shaped key carrying the wrong percentile. Naming the percentile in the output would eliminate a whole class of silent misreads.