Systems-design questions at the level where the answer is an argument rather than a data structure — each worked end to end, with running code that measures whether the argument actually holds.
The format is one question answered independently in more than one language. Not a port: each track re-derives the design from its own language's cost model, and the tracks are allowed to reach different conclusions. Where they do, the disagreement is the finding.
| the question | tracks | |
|---|---|---|
| mpsc-event-pipeline-kata | Many producers push small events to a single consumer that batches them into a socket. Millions of events/sec, bounded memory, producers must never block indefinitely, and order is preserved per producer, not globally — the clause that permits an architecture with no shared coordination point at all. Follow-up: 64 producers but only 4–8 active, so stop polling 64 empty queues without reintroducing a contended global. | C++ · Go · Python |
| plugin-dispatch-kata (C++ answered) | 50–100 protocol handlers, some built in and some loaded from shared libraries, selected by linear scan at millions of packets/sec. Profiling blames dispatch; the team proposes four polymorphism mechanisms. The profile never said the virtual call was the problem, and all four options change how process is called rather than how many times matches is. Then the half it leads to: the right mechanism at the ABI boundary and the right mechanism on the hot path are not the same one, and nothing requires them to be. |
C++ · Go · Python |
- Your profiler said
dispatch. It did not sayvirtual. — the plugin-dispatch C++ answer walked through end to end, with generated figures and animations, including the three claims that did not survive being measured.
question.md— the question staged as a kata: numbered steps that build the argument in the order it has to be built, each with probes, a hint, and a self-assessable gate. The answer sketch is folded at the bottom, so the question is usable before you read the answer.<lang>/question.md— what the shared question cannot ask, because it only exists in that language.<lang>/solution.md— the worked answer and the measurements it rests on.<lang>/— the implementation, plus a benchmark binary that asserts the invariants under shared labels and exits non-zero if one breaks. Every number quoted in a solution comes from running it.
Each kata's README carries its own build, run, and race/thread-checking instructions.