testing-patterns（测试模式）

Testing Patterns & Effect Abstraction

Short version: model your “effects” as traits, inject them, keep core logic pure, and provide real + fake implementations. That’s the idiomatic Rust way; free monads aren’t a thing here.

Pattern

• Define algebras as traits (ports).
• Implement adapters for prod (HTTP, DB, clock, FS) and for tests (fakes/mocks).
• Inject via generics (zero-cost, monomorphized) or trait objects (dyn Trait) when you need late binding.
• Keep domain functions pure; pass in effect results or tiny capability traits.

Minimal sync example

use std::time::{SystemTime, UNIX_EPOCH};

pub trait Clock {
    fn now(&self) -> SystemTime;
}

pub trait Payments {
    type Err;
    fn charge(&self, cents: u32, card: &str) -> Result<String, Self::Err>; // returns ChargeId
}

pub struct Service<P, C> {
    pay: P,
    clock: C,
}

impl<P, C> Service<P, C>
where
    P: Payments,
    C: Clock,
{
    pub fn bill(&self, card: &str, cents: u32) -> Result<String, P::Err> {
        let _ts = self
            .clock
            .now()
            .duration_since(UNIX_EPOCH)
            .unwrap()
            .as_secs();
        // domain logic… (e.g., time-based rules)
        self.pay.charge(cents, card)
    }
}

// --- prod adapters ---
pub struct RealClock;
impl Clock for RealClock {
    fn now(&self) -> SystemTime {
        SystemTime::now()
    }
}

pub struct StripeClient;
impl Payments for StripeClient {
    type Err = String;
    fn charge(&self, cents: u32, _card: &str) -> Result<String, Self::Err> {
        // call real API
        Ok(format!("ch_{cents}"))
    }
}

// --- test fakes ---
#[cfg(test)]
mod tests {
    use super::*;
    use std::cell::RefCell;
    use std::time::{Duration, SystemTime};

    struct FixedClock(SystemTime);
    impl Clock for FixedClock {
        fn now(&self) -> SystemTime {
            self.0
        }
    }

    struct FakePayments {
        pub calls: RefCell<Vec<(u32, String)>>,
        pub next: RefCell<Result<String, String>>,
    }
    impl Payments for FakePayments {
        type Err = String;
        fn charge(&self, cents: u32, card: &str) -> Result<String, Self::Err> {
            self.calls.borrow_mut().push((cents, card.to_string()));
            self.next.borrow_mut().clone()
        }
    }

    #[test]
    fn happy_path() {
        let svc = Service {
            pay: FakePayments {
                calls: RefCell::new(vec![]),
                next: RefCell::new(Ok("ch_42".into())),
            },
            clock: FixedClock(SystemTime::UNIX_EPOCH + Duration::from_secs(123)),
        };

        let id = svc.bill("4111...", 4200).unwrap();
        assert_eq!(id, "ch_42");
    }
}

Prod wiring stays simple:

let svc = Service { pay: StripeClient, clock: RealClock };

Trait objects (dynamic dispatch when needed)

pub struct Svc<'a> {
    pay: &'a dyn Payments<Err = String>,
    clock: &'a dyn Clock,
}

Ensure traits are object-safe (no generic methods, no impl Trait returns).

Async Effects

1. async-trait macro – ergonomic, small overhead:

use async_trait::async_trait;

#[async_trait]
pub trait Http {
    async fn get(&self, url: &str) -> Result<String, anyhow::Error>;
}

1. RPITIT (return-position impl Trait in traits) for macro-free, low-overhead code:

use core::future::Future;

pub trait Http {
    fn get(&self, url: &str) -> impl Future<Output = Result<String, anyhow::Error>> + Send;
}

Pick #1 for simplicity, #2 if you want zero-macro builds and control over allocations.

Mocks vs. Fakes

• Prefer hand-rolled fakes/stubs or in-memory adapters.
• If you need expectation-based mocks:

- mockall for general traits. - wiremock / httpmock for HTTP.

• For FS/DB, lean on temp dirs (tempfile, assert_fs) or in-memory backends.

Tips

• Don’t over-abstract; put traits only at IO boundaries (time, network, FS, DB).
• In async code, wrap shared deps in Arc<dyn Trait + Send + Sync> when needed.
• Return owned data (Vec<T>) from trait methods to avoid lifetime tangles.
• Keep domain logic as pure functions over data; invoke effects at the edges.
• For CLI flows, lean on tests/support/mod.rs (CliFixture, RemoteRepo, and helpers that pre-wire SK_CACHE_DIR/SK_CONFIG_DIR) so every integration test spins up the same deterministic temp repos.

Testing Standards

• Coverage gate: 45% (cargo llvm-cov). CI currently enforces cargo llvm-cov --fail-under-lines 45. Treat that as the floor, not the ceiling—once main sits comfortably above a higher percentage, ratchet the workflow file and avoid ever lowering the bar without a written justification.
• Business logic ⇒ property tests. Use proptest for any non-trivial domain rule (scheduling, diffing, parsing, state machines, etc.). Unit tests that check a couple of examples aren’t enough; capture invariants as properties.
• Structure: keep property tests in tests modules alongside unit tests, e.g.:

#[cfg(test)]
mod prop_tests {
    use super::*;
    use proptest::prelude::*;

    proptest! {
        #[test]
        fn fee_is_never_negative(amount in 0u64..) {
            let fee = compute_fee(amount);
            prop_assert!(fee >= 0);
        }
    }
}

• Make generators realistic. Compose any::<T>(), prop::collection, or custom strategies so you’re exercising edge cases (empty, max values, random ordering).
• Integration tests still matter. Use harnesses under tests/ or crates/*/tests/ to cover end-to-end flows (e.g., env lifecycle, DB migrations) but keep them deterministic—no real network calls.

When in doubt, assume reviewers will ask “where’s the property test?” and “what’s the coverage delta?” Bake both answers into the PR.

Property-Based Testing Workflow (proptest)

When you add or refresh property tests, approach the work like a mini bd task—not a plan-tool exercise. Claim/track the effort via bd (ready → update ... --status in_progress → close), and keep the following loop tight:

1. Identify high-value properties. Start with the public API or core modules. Look for invariants (round trips, idempotence, ordering guarantees, etc.) that actually buy us something. Skip trivial wrappers.
2. Study how the code is used. Before writing a property, grep the repo to see how that function/struct is consumed so your strategy stays within real-world preconditions.
3. Write precise proptest cases. Small number of high-signal tests beats shotgun suites. Favor clear strategies (e.g., prop::collection, any::<T>(), from_regex) and only add bounds when the code truly requires them.
4. Lean on real generators. Model inputs with strategies (vecs, maps, enums) instead of manual loops like for skip in 0..5. Let the generator produce arbitrarily long lists/arrays (only constrain them when the production code has a hard limit) so burn-in runs and shrink output stay meaningful.
5. Run and reflect. cargo test (or the specific crate) with the new property tests. If a proptest failure exposes a gap, either fix the bug or constrain the strategy with a documented reason.

Keep the tests maintainable: name the property after the behavior it documents, describe why the invariant matters in a short comment when it isn’t obvious, and prefer deterministic shrink-friendly strategies. The expectation is that every non-trivial business rule eventually has a companion proptest! block living next to its unit tests. When you record notes or TODOs for these efforts, put them in the bd issue itself so the history stays alongside the task—no side trackers, no plan tool usage, no Hypothesis snippets.

Analogy: Traits + adapters ≈ Haskell typeclasses + interpreters. Stick to this pattern instead of free monads.