Should a Mock inside a synctest test follow synctest's clock? #35

Go 1.25 made testing/synctest stable. In a synctest test, the time package runs on a fake clock that jumps forward whenever every goroutine in the test is blocked. That overlaps with what quartz does, so the README needs to say when to use which (#34). Before writing that, we want to agree on one design question:

Should a *Mock created inside a synctest test follow synctest's fake clock, or keep its own time as it does today?

Today the two clocks are independent. Calling time.Sleep(time.Hour) inside the test does not move the mock, and mClock.Advance(time.Hour) does not move time.Now(). quartz.NewReal() already follows synctest's clock, because it calls the time package.

We prototyped a mock that follows synctest's clock: timers are real time timers, and Advance(d) is time.Sleep(d) followed by synctest.Wait(). Compared with today's *Mock:

• Better: ctx.Deadline() agrees with time.Now(), and Advance can move past several timers at once, firing them in order. Today Advance stops at the next timer.
• Worse, time moves without Advance: a test that forgot to call Advance and then waited for the code to start its next timer still passed, because synctest moved time forward by 10s on its own. Today's mock fails that test.
• Worse, real I/O: a 5s timeout never fired during a real TCP read. Advance(5s) blocked until the read finished, because synctest does not move time while a goroutine waits on the network. Today's mock canceled the read after 67µs.

We also tried keeping the mock's own timers but sleeping the same amount on every Advance, so both clocks stay equal. They drift apart as soon as code in the test sleeps on its own. After a 1-minute time.Sleep in library code, the mock was 1 minute behind.

The one thing a synctest-following mock adds over NewReal() is traps.

Our recommendation: keep the *Mock independent. The README would say: use synctest with NewReal() when the whole test can run inside synctest, and use a *Mock when it does real I/O or needs to control exactly when time moves. If traps turn out to be wanted with synctest, a separate option is to add traps to the real clock rather than changing the mock.

@spikecurtis, does that match your view?

🤖 This issue was created with the help of Coder Agents.