Struct ShadowClassifier 

pub struct ShadowClassifier { /* private fields */ }

Serves the expensive adapter’s answers, unchanged, and asks the cheap one the same questions to measure agreement. The cheap adapter cannot change or fail a result: its errors are logged, recorded and counted as skipped.

Both adapters are asked concurrently, on the caller’s task, and the call returns when both have: its latency is the slower of the two, not the sum. A cheap call that stalls still holds the shadowed call, so the cheap adapter must enforce its own deadline. When the expensive adapter fails, its error is the call’s; the cheap call is still recorded.

profile is the expensive adapter’s: every answer returned is its.

Implementations

impl ShadowClassifier

pub fn new( cheap: (AdapterId, Arc<dyn Classifier>), expensive: (AdapterId, Arc<dyn Classifier>), accounting: Accounting, ) -> Self

A shadow measuring cheap against expensive.

pub fn report(&self) -> AgreementReport

The agreement measured so far — ungraded, since live traffic has no ground truth.

Trait Implementations

impl Classifier for ShadowClassifier

fn profile(&self) -> ClassifierProfile

What family of numbers this adapter’s confidence comes from, and how much state it will carry.

fn ask<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, state: &'life1 str, questions: &'life2 BTreeMap<String, Question>, ) -> Pin<Box<dyn Future<Output = Result<BTreeMap<String, Answer>, ClassifierError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,

Answers every question in questions about state, keyed the same way.