Testing against a hostile network¶
Level: 301 · deep dive
Stub — an outline, not a lesson. There is no runnable example behind this page yet, so nothing on it has been through the check that backs every other claim in this library. The bullets below are the questions the finished page has to answer.
One line: turmoil ↗ runs the server and its clients as simulated hosts on a simulated network, so latency, dropped connections and partitions become deterministic inputs to an ordinary unit test — possible here because the server already takes its IO and its durability through abstractions it can swap.
What it has to cover¶
- A simulation with one server host and several client hosts, and the simulated
TcpListenerandTcpStreamthat replace Tokio's - Swapping the network in: why the handler generic over
AsyncRead + AsyncWrite(chapters 8 and 10) and the durability trait (chapter 13) make this a change at the edges rather than a rewrite - Injecting failure: added latency, dropped messages, a partition between hosts and its repair, and a server host that crashes and restarts — replaying its log from chapter 9
- Determinism: a seed that makes a failing run repeat exactly, and how to turn that run into a regression test
- The properties worth asserting: no acknowledged write is lost across a crash; a client that retries after a timeout does not apply a write twice
- What a simulation cannot find: the real kernel, the real disk's sync behaviour, and timing on real hardware
The trap it exists for¶
A test suite that only ever runs on a network that never fails. Every chapter before this one handled failures; this is the first place the tests make any of them happen.
What minidb gains¶
A test that crashes the server mid-write under a partitioned network, and passes.
See also¶
- Testing async code — paused time and in-memory streams, which this extends to the whole network
- Mocking a server — the client-side version of a fake network, and what it cannot prove
- Deterministic testing ↗ and nondeterminism ↗ — in the Concurrency library
Po polsku¶
turmoil uruchamia serwer i klientów jako symulowane hosty w symulowanej sieci, więc opóźnienia, zerwane połączenia i podział sieci (partition) stają się powtarzalnymi danymi wejściowymi zwykłego testu jednostkowego. To możliwe tylko dlatego, że serwer korzysta z wejścia-wyjścia i z trwałości danych przez abstrakcje, które da się podmienić. Test ma sprawdzać własności — na przykład że żaden potwierdzony zapis nie ginie po awarii — i dawać się odtworzyć z tego samego ziarna losowości.
Szukaj po polsku: testowanie z symulacją sieci · testy deterministyczne · rust turmoil network simulation · deterministic simulation testing