timvw_rust_runtime/
retry.rs

1//! Retries for HTTP APIs: [`RetryPolicy`], a tower retry policy, and [`RetryLayer`], which adds to it a look at the
2//! body of a `429` to tell a rate limit (retried) from an exhausted quota (not retried).
3//!
4//! `Client::default()` already retries with `RetryLayer::new(RetryPolicy::default())`. For another policy, start from
5//! the bare service:
6//!
7//! ```no_run
8//! # #[cfg(feature = "reqwest")]
9//! # fn demo() {
10//! use timvw_rust_runtime::retry::{self, RetryLayer, RetryPolicy};
11//! use timvw_rust_runtime::{Client, ReqwestService};
12//!
13//! let client = Client::from_service(ReqwestService::default())
14//!     .layer(RetryLayer::new(RetryPolicy::new(5)).permanent_429(retry::openai_insufficient_quota));
15//! # }
16//! ```
17
18use std::fmt;
19use std::sync::Arc;
20use std::task::{Context, Poll};
21use std::time::Duration;
22
23use bytes::{Bytes, BytesMut};
24use http::header::RETRY_AFTER;
25use http::{HeaderMap, StatusCode};
26use http_body_util::{BodyDataStream, BodyExt};
27use tower::retry::Policy;
28use tower::{Layer, Service};
29
30use crate::maybe::BoxFuture;
31use crate::timer::{since_epoch, Instant, Sleep};
32use crate::{Body, BoxError, MaybeSend, MaybeSync};
33
34/// The header OpenAI (and others) send next to `Retry-After`, in milliseconds.
35const RETRY_AFTER_MS: &str = "retry-after-ms";
36
37/// How much of a `429` body [`RetryLayer`] reads by default to classify it: 64 KiB.
38pub const DEFAULT_BODY_LIMIT: usize = 64 * 1024;
39
40/// Retries `429 Too Many Requests`, `408 Request Timeout` and `5xx` responses, and failures to connect.
41///
42/// - It waits as long as the server asks, with `retry-after-ms` (milliseconds) or `Retry-After` (seconds or an HTTP
43///   date), up to [`RetryPolicy::max_retry_after`] (60 s by default). The wait is read once into a deadline, counted
44///   from when the response arrived, and the policy sleeps until it: time spent in between (reading a `429` body for
45///   [`RetryLayer::permanent_429`]) is not added. A server that asks for longer than `max_retry_after` gets its
46///   response returned at once: waiting less than it asked would likely fail again, and waiting an hour is rarely
47///   what a caller wants. A date in the past means retry now.
48/// - With [`RetryPolicy::server_directed_429`], a `429` is retried only when the server says when.
49/// - Without those headers it backs off exponentially from [`RetryPolicy::base_delay`] (200 ms), doubling on each retry
50///   up to [`RetryPolicy::max_delay`] (8 s), with jitter: each delay is picked at random between half and all of it, so
51///   that clients that failed together do not retry together. [`RetryPolicy::jitter`] turns that off.
52/// - It never retries a response marked [`Permanent`] (see [`RetryLayer::permanent_429`]), nor a request whose body is
53///   a stream (a streamed file upload), which cannot be sent again.
54#[derive(Clone, Debug)]
55pub struct RetryPolicy {
56    max_retries: usize,
57    attempts: usize,
58    base_delay: Duration,
59    max_delay: Duration,
60    max_retry_after: Duration,
61    jitter: bool,
62    server_directed_429: bool,
63}
64
65impl RetryPolicy {
66    /// A policy that retries at most `max_retries` times (after the first attempt).
67    pub fn new(max_retries: usize) -> Self {
68        Self {
69            max_retries,
70            attempts: 0,
71            base_delay: Duration::from_millis(200),
72            max_delay: Duration::from_secs(8),
73            max_retry_after: Duration::from_secs(60),
74            jitter: true,
75            server_directed_429: false,
76        }
77    }
78
79    /// The first backoff delay, doubled on each retry (default 200 ms).
80    pub fn base_delay(mut self, delay: Duration) -> Self {
81        self.base_delay = delay;
82        self
83    }
84
85    /// The longest backoff delay (default 8 s). It does not limit how long a server can ask to wait with
86    /// `Retry-After`: that is [`RetryPolicy::max_retry_after`].
87    pub fn max_delay(mut self, delay: Duration) -> Self {
88        self.max_delay = delay;
89        self
90    }
91
92    /// The longest wait a server can ask for with `Retry-After` or `retry-after-ms` (default 60 s). A response that
93    /// asks for longer is not retried but returned.
94    pub fn max_retry_after(mut self, delay: Duration) -> Self {
95        self.max_retry_after = delay;
96        self
97    }
98
99    /// Whether each backoff delay is picked at random between half and all of it (default `true`). The waits a
100    /// server asks for are never jittered.
101    pub fn jitter(mut self, jitter: bool) -> Self {
102        self.jitter = jitter;
103        self
104    }
105
106    /// Whether a `429` is retried only when the server says when, with `Retry-After` or `retry-after-ms` (default
107    /// `false`: every `429` is retried, with backoff when the server does not say when).
108    ///
109    /// Some APIs answer a `429` without advice when waiting will not help (a quota, a request too large for a rate
110    /// limit), and with it when it will. `408`, `5xx` and connect failures are retried either way.
111    pub fn server_directed_429(mut self, server_directed: bool) -> Self {
112        self.server_directed_429 = server_directed;
113        self
114    }
115
116    /// The number of retries allowed.
117    pub fn max_retries(&self) -> usize {
118        self.max_retries
119    }
120
121    /// When to retry: the server's deadline, or `now` plus the backoff; `None` when the server asks to wait longer than
122    /// `max_retry_after`.
123    fn deadline(&self, retry_after: Option<Instant>, now: Instant) -> Option<Instant> {
124        if let Some(deadline) = retry_after {
125            return (deadline.saturating_duration_since(now) <= self.max_retry_after)
126                .then_some(deadline);
127        }
128        Some(now + self.backoff())
129    }
130
131    /// The backoff before the next retry, jittered unless turned off.
132    fn backoff(&self) -> Duration {
133        let backoff = self
134            .base_delay
135            .saturating_mul(1u32 << self.attempts.min(16))
136            .min(self.max_delay);
137        if self.jitter {
138            jittered(backoff, random_unit())
139        } else {
140            backoff
141        }
142    }
143}
144
145/// Three retries.
146impl Default for RetryPolicy {
147    fn default() -> Self {
148        Self::new(3)
149    }
150}
151
152/// `backoff` scaled to between half and all of it, by `unit` in `[0, 1)`.
153fn jittered(backoff: Duration, unit: f64) -> Duration {
154    backoff.mul_f64(0.5 + 0.5 * unit)
155}
156
157/// A random number in `[0, 1)`, from the standard library's randomly keyed hasher: good enough to spread retries.
158fn random_unit() -> f64 {
159    use std::hash::{BuildHasher, Hasher};
160    let hash = std::collections::hash_map::RandomState::new()
161        .build_hasher()
162        .finish();
163    (hash >> 11) as f64 / (1u64 << 53) as f64
164}
165
166/// When a response arrived, set by [`RetryLayer`] on a `429` before it reads the body, so that the time spent reading
167/// is not added to the server's `Retry-After`.
168#[derive(Clone, Copy, Debug)]
169struct ReceivedAt(Instant);
170
171/// The clocks a `Retry-After` is read against: `Retry-After` in seconds counts from when the response arrived, an HTTP
172/// date against the wall clock (the time since the Unix epoch), read together with the monotonic one.
173#[derive(Clone, Copy, Debug)]
174struct Clocks {
175    received_at: Instant,
176    now: Instant,
177    wall: Duration,
178}
179
180impl Clocks {
181    fn now<B>(response: &http::Response<B>) -> Self {
182        // The wall clock first, so that a pause between the two reads cannot shorten an HTTP date's deadline.
183        let wall = since_epoch();
184        let now = Instant::now();
185        let received_at = response
186            .extensions()
187            .get::<ReceivedAt>()
188            .map_or(now, |received| received.0);
189        Self {
190            received_at,
191            now,
192            wall,
193        }
194    }
195}
196
197/// Whether a response or error is worth retrying, and the deadline the server set, if any.
198fn retryable<B>(
199    result: &Result<http::Response<B>, BoxError>,
200    server_directed_429: bool,
201) -> Option<Option<Instant>> {
202    match result {
203        Ok(response) => {
204            let status = response.status();
205            let transient = status == StatusCode::TOO_MANY_REQUESTS
206                || status == StatusCode::REQUEST_TIMEOUT
207                || status.is_server_error();
208            let permanent = response.extensions().get::<Permanent>().is_some();
209            if !transient || permanent {
210                return None;
211            }
212            let deadline = retry_after(response.headers(), Clocks::now(response));
213            if server_directed_429 && status == StatusCode::TOO_MANY_REQUESTS && deadline.is_none()
214            {
215                #[cfg(feature = "tracing")]
216                tracing::debug!(
217                    target: "timvw_rust_runtime::trace",
218                    status = status.as_u16(),
219                    "not retrying: a 429 without Retry-After (server_directed_429)"
220                );
221                return None;
222            }
223            Some(deadline)
224        }
225        Err(error) => is_connect(error.as_ref()).then_some(None),
226    }
227}
228
229/// The deadline a server sets: `retry-after-ms` in milliseconds, else `Retry-After` in seconds (both may have a
230/// fraction) or as an HTTP date. A date in the past is now.
231fn retry_after(headers: &HeaderMap, clocks: Clocks) -> Option<Instant> {
232    let value = |name| Some(headers.get(name)?.to_str().ok()?.trim());
233    let delay = |value: &str, scale: f64| {
234        let delay = Duration::try_from_secs_f64(value.parse::<f64>().ok()? * scale).ok()?;
235        clocks.received_at.checked_add(delay)
236    };
237    if let Some(deadline) = value(RETRY_AFTER_MS).and_then(|value| delay(value, 0.001)) {
238        return Some(deadline);
239    }
240    let value = value(RETRY_AFTER.as_str())?;
241    delay(value, 1.0).or_else(|| {
242        let date = httpdate::parse_http_date(value).ok()?;
243        let date = date
244            .duration_since(std::time::UNIX_EPOCH)
245            .unwrap_or_default();
246        let delay = date.saturating_sub(clocks.wall);
247        clocks.now.checked_add(delay)
248    })
249}
250
251/// Whether the request failed to connect, so it was not sent and can be sent again. Never on WebAssembly: `fetch`
252/// fails the same way whether or not the request reached the server (and for CORS), so a failure there is not retried.
253#[allow(unused_variables)]
254fn is_connect(error: &(dyn std::error::Error + 'static)) -> bool {
255    #[cfg(all(feature = "reqwest", not(target_family = "wasm")))]
256    {
257        let mut source = Some(error);
258        while let Some(error) = source {
259            if error
260                .downcast_ref::<reqwest::Error>()
261                .is_some_and(reqwest::Error::is_connect)
262            {
263                return true;
264            }
265            source = error.source();
266        }
267    }
268    false
269}
270
271impl<B> Policy<http::Request<Body>, http::Response<B>, BoxError> for RetryPolicy {
272    type Future = BoxFuture<'static, ()>;
273
274    fn retry(
275        &mut self,
276        _request: &mut http::Request<Body>,
277        result: &mut Result<http::Response<B>, BoxError>,
278    ) -> Option<Self::Future> {
279        if self.attempts >= self.max_retries {
280            return None;
281        }
282        let retry_after = retryable(result, self.server_directed_429)?;
283        let now = Instant::now();
284        let Some(deadline) = self.deadline(retry_after, now) else {
285            #[cfg(feature = "tracing")]
286            tracing::debug!(
287                target: "timvw_rust_runtime::trace",
288                retry_after_ms = retry_after
289                    .map_or(0, |deadline| deadline.saturating_duration_since(now).as_millis() as u64),
290                max_retry_after_ms = self.max_retry_after.as_millis() as u64,
291                status = result.as_ref().ok().map(|response| response.status().as_u16()),
292                "not retrying: the server asks to wait longer than max_retry_after"
293            );
294            return None;
295        };
296        self.attempts += 1;
297        #[cfg(feature = "tracing")]
298        tracing::debug!(
299            target: "timvw_rust_runtime::trace",
300            attempt = self.attempts,
301            max_retries = self.max_retries,
302            delay_ms = deadline.saturating_duration_since(now).as_millis() as u64,
303            status = result.as_ref().ok().map(|response| response.status().as_u16()),
304            "retrying"
305        );
306        Some(Box::pin(Sleep::until(deadline)))
307    }
308
309    fn clone_request(&mut self, request: &http::Request<Body>) -> Option<http::Request<Body>> {
310        let body = request.body().try_clone()?;
311        let mut clone = http::Request::new(body);
312        *clone.method_mut() = request.method().clone();
313        *clone.uri_mut() = request.uri().clone();
314        *clone.version_mut() = request.version();
315        *clone.headers_mut() = request.headers().clone();
316        Some(clone)
317    }
318}
319
320/// A response extension that tells [`RetryPolicy`] not to retry a response it would otherwise retry.
321///
322/// [`RetryLayer::permanent_429`] sets it on a `429` whose body says the error is permanent. A layer of your own, inside
323/// the retries, can set it too: `response.extensions_mut().insert(Permanent)`.
324#[derive(Clone, Copy, Debug, Default, PartialEq, Eq, Hash)]
325pub struct Permanent;
326
327/// Whether the body of a `429` says the error is permanent, such as [`openai_insufficient_quota`].
328type Classifier = Arc<dyn Fn(&[u8]) -> bool + Send + Sync>;
329
330/// OpenAI's "quota exhausted": a JSON body whose `error.code` or `error.type` is `insufficient_quota`.
331///
332/// OpenAI answers `429` both when requests come too fast, which passes, and when the account's quota or credit is
333/// used up, which does not pass by waiting. Use it with [`RetryLayer::permanent_429`].
334pub fn openai_insufficient_quota(body: &[u8]) -> bool {
335    const INSUFFICIENT_QUOTA: &str = "insufficient_quota";
336    let Ok(value) = serde_json::from_slice::<serde_json::Value>(body) else {
337        return false;
338    };
339    ["code", "type"].iter().any(|field| {
340        value
341            .pointer(&format!("/error/{field}"))
342            .and_then(serde_json::Value::as_str)
343            == Some(INSUFFICIENT_QUOTA)
344    })
345}
346
347/// Retries with a [`RetryPolicy`], and optionally reads the body of a `429` to tell a permanent error from a rate
348/// limit first.
349///
350/// `RetryLayer::new(policy)` retries exactly like `tower::retry::RetryLayer::new(policy)`. With
351/// [`RetryLayer::permanent_429`], the body of each `429` is read (up to [`RetryLayer::body_limit`]) and given to a
352/// function; when it returns `true`, the response is marked [`Permanent`] and returned at once. The body is put back,
353/// so the caller (and a generated client's error) still sees all of it.
354///
355/// ```no_run
356/// # #[cfg(feature = "reqwest")]
357/// # fn demo() {
358/// use timvw_rust_runtime::retry::{openai_insufficient_quota, RetryLayer, RetryPolicy};
359/// use timvw_rust_runtime::{Client, ReqwestService};
360///
361/// let client = Client::from_service(ReqwestService::default())
362///     .layer(RetryLayer::new(RetryPolicy::default()).permanent_429(openai_insufficient_quota));
363/// # }
364/// ```
365#[derive(Clone)]
366pub struct RetryLayer {
367    policy: RetryPolicy,
368    classify: Option<Classifier>,
369    body_limit: usize,
370}
371
372impl RetryLayer {
373    /// Retries with `policy`, without reading any body.
374    pub fn new(policy: RetryPolicy) -> Self {
375        Self {
376            policy,
377            classify: None,
378            body_limit: DEFAULT_BODY_LIMIT,
379        }
380    }
381
382    /// Reads the body of each `429` and does not retry it when `is_permanent` returns `true` for the body's bytes.
383    ///
384    /// A body longer than [`RetryLayer::body_limit`], or one that fails to read, is not given to `is_permanent`: the
385    /// response is retried as a rate limit, and the caller still reads the whole body (or its error).
386    pub fn permanent_429<F>(mut self, is_permanent: F) -> Self
387    where
388        F: Fn(&[u8]) -> bool + Send + Sync + 'static,
389    {
390        self.classify = Some(Arc::new(is_permanent));
391        self
392    }
393
394    /// How many bytes of a `429` body are read to classify it (default [`DEFAULT_BODY_LIMIT`], 64 KiB).
395    pub fn body_limit(mut self, limit: usize) -> Self {
396        self.body_limit = limit;
397        self
398    }
399
400    /// The retry policy.
401    pub fn policy(&self) -> &RetryPolicy {
402        &self.policy
403    }
404}
405
406/// Retries with [`RetryPolicy::default`].
407impl Default for RetryLayer {
408    fn default() -> Self {
409        Self::new(RetryPolicy::default())
410    }
411}
412
413impl fmt::Debug for RetryLayer {
414    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
415        f.debug_struct("RetryLayer")
416            .field("policy", &self.policy)
417            .field("permanent_429", &self.classify.is_some())
418            .field("body_limit", &self.body_limit)
419            .finish()
420    }
421}
422
423impl<S> Layer<S> for RetryLayer {
424    type Service = tower::retry::Retry<RetryPolicy, Classify429<S>>;
425
426    fn layer(&self, inner: S) -> Self::Service {
427        tower::retry::Retry::new(
428            self.policy.clone(),
429            Classify429 {
430                inner,
431                classify: self.classify.clone(),
432                body_limit: self.body_limit,
433            },
434        )
435    }
436}
437
438/// The service inside the retries of a [`RetryLayer`]: marks a `429` [`Permanent`] when its body says so.
439#[derive(Clone)]
440pub struct Classify429<S> {
441    inner: S,
442    classify: Option<Classifier>,
443    body_limit: usize,
444}
445
446impl<S: fmt::Debug> fmt::Debug for Classify429<S> {
447    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
448        f.debug_struct("Classify429")
449            .field("inner", &self.inner)
450            .field("permanent_429", &self.classify.is_some())
451            .field("body_limit", &self.body_limit)
452            .finish()
453    }
454}
455
456impl<S, R, B> Service<R> for Classify429<S>
457where
458    S: Service<R, Response = http::Response<B>>,
459    S::Future: MaybeSend + 'static,
460    B: http_body::Body + MaybeSend + MaybeSync + 'static,
461    B::Data: Into<Bytes>,
462    B::Error: Into<BoxError>,
463{
464    type Response = http::Response<Body>;
465    type Error = S::Error;
466    type Future = BoxFuture<'static, Result<Self::Response, Self::Error>>;
467
468    fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
469        self.inner.poll_ready(cx)
470    }
471
472    fn call(&mut self, request: R) -> Self::Future {
473        let response = self.inner.call(request);
474        let classify = self.classify.clone();
475        let body_limit = self.body_limit;
476        Box::pin(async move {
477            let mut response = response.await?.map(Body::wrap);
478            Ok(match classify {
479                Some(classify) if response.status() == StatusCode::TOO_MANY_REQUESTS => {
480                    response.extensions_mut().insert(ReceivedAt(Instant::now()));
481                    classify_429(response, &*classify, body_limit).await
482                }
483                _ => response,
484            })
485        })
486    }
487}
488
489/// Reads a `429` body, up to `limit` bytes, marks the response [`Permanent`] if `classify` says so, and puts the body
490/// back.
491async fn classify_429(
492    response: http::Response<Body>,
493    classify: &(dyn Fn(&[u8]) -> bool + Send + Sync),
494    limit: usize,
495) -> http::Response<Body> {
496    let (mut parts, mut body) = response.into_parts();
497    let mut read = BytesMut::new();
498    let body = loop {
499        match body.frame().await {
500            None => {
501                let read = read.freeze();
502                if classify(&read) {
503                    parts.extensions.insert(Permanent);
504                    #[cfg(feature = "tracing")]
505                    tracing::debug!(
506                        target: "timvw_rust_runtime::trace",
507                        status = parts.status.as_u16(),
508                        "not retrying: the body says the error is permanent"
509                    );
510                }
511                break Body::from(read);
512            }
513            // Trailers on a 429 are dropped.
514            Some(Ok(frame)) => {
515                let Ok(data) = frame.into_data() else {
516                    continue;
517                };
518                read.extend_from_slice(&data);
519                if read.len() > limit {
520                    let read = futures_util::stream::once(async move { Ok(read.freeze()) });
521                    break Body::wrap_stream(futures_util::StreamExt::chain(
522                        read,
523                        BodyDataStream::new(body),
524                    ));
525                }
526            }
527            Some(Err(error)) => {
528                break Body::wrap_stream(futures_util::stream::iter([
529                    Ok(read.freeze()),
530                    Err::<Bytes, BoxError>(error),
531                ]));
532            }
533        }
534    };
535    http::Response::from_parts(parts, body)
536}
537
538#[cfg(test)]
539mod tests {
540    use super::*;
541
542    fn response(status: u16, headers: &[(&str, &str)]) -> Result<http::Response<()>, BoxError> {
543        let mut builder = http::Response::builder().status(status);
544        for (name, value) in headers {
545            builder = builder.header(*name, *value);
546        }
547        Ok(builder.body(()).unwrap())
548    }
549
550    /// Whether `result` is retried, and after how long, measured from now.
551    fn wait(
552        result: &Result<http::Response<()>, BoxError>,
553        server_directed_429: bool,
554    ) -> Option<Option<Duration>> {
555        let now = Instant::now();
556        retryable(result, server_directed_429)
557            .map(|deadline| deadline.map(|deadline| deadline.saturating_duration_since(now)))
558    }
559
560    #[tokio::test(start_paused = true)]
561    async fn retries_transient_statuses_only() {
562        assert_eq!(wait(&response(429, &[]), false), Some(None));
563        assert_eq!(
564            wait(&response(503, &[("retry-after", "2")]), false),
565            Some(Some(Duration::from_secs(2)))
566        );
567        assert_eq!(wait(&response(408, &[]), false), Some(None));
568        assert_eq!(wait(&response(400, &[]), false), None);
569        assert_eq!(wait(&response(200, &[]), false), None);
570        assert_eq!(wait(&Err("boom".into()), false), None);
571    }
572
573    #[tokio::test(start_paused = true)]
574    async fn server_directed_429_needs_advice() {
575        assert_eq!(wait(&response(429, &[]), true), None);
576        assert_eq!(wait(&response(429, &[("retry-after", "soon")]), true), None);
577        assert_eq!(
578            wait(&response(429, &[("retry-after", "0")]), true),
579            Some(Some(Duration::ZERO))
580        );
581        assert_eq!(
582            wait(&response(429, &[("retry-after-ms", "300")]), true),
583            Some(Some(Duration::from_millis(300)))
584        );
585        let past = httpdate::fmt_http_date(std::time::SystemTime::now() - Duration::from_secs(60));
586        assert_eq!(
587            wait(&response(429, &[("retry-after", &past)]), true),
588            Some(Some(Duration::ZERO))
589        );
590        // Only 429s: other transient statuses back off without advice.
591        assert_eq!(wait(&response(503, &[]), true), Some(None));
592        assert_eq!(wait(&response(408, &[]), true), Some(None));
593    }
594
595    #[tokio::test(start_paused = true)]
596    async fn permanent_responses_are_not_retried() {
597        let mut result = response(429, &[("retry-after", "1")]);
598        result.as_mut().unwrap().extensions_mut().insert(Permanent);
599        assert_eq!(wait(&result, false), None);
600    }
601
602    fn clocks() -> Clocks {
603        let now = Instant::now();
604        Clocks {
605            received_at: now,
606            now,
607            wall: httpdate::parse_http_date("Wed, 21 Oct 2015 07:28:00 GMT")
608                .unwrap()
609                .duration_since(std::time::UNIX_EPOCH)
610                .unwrap(),
611        }
612    }
613
614    fn parse(headers: &[(&str, &str)], clocks: Clocks) -> Option<Duration> {
615        let mut map = HeaderMap::new();
616        for (name, value) in headers {
617            map.append(
618                http::HeaderName::from_bytes(name.as_bytes()).unwrap(),
619                value.parse().unwrap(),
620            );
621        }
622        retry_after(&map, clocks).map(|deadline| deadline.saturating_duration_since(clocks.now))
623    }
624
625    #[tokio::test(start_paused = true)]
626    async fn parses_retry_after_and_retry_after_ms() {
627        let wait = |headers: &[(&str, &str)]| parse(headers, clocks());
628        assert_eq!(
629            wait(&[("retry-after", " 20 ")]),
630            Some(Duration::from_secs(20))
631        );
632        assert_eq!(
633            wait(&[("retry-after", "1.5")]),
634            Some(Duration::from_millis(1500))
635        );
636        assert_eq!(
637            wait(&[("retry-after-ms", "250"), ("retry-after", "1")]),
638            Some(Duration::from_millis(250))
639        );
640        assert_eq!(
641            wait(&[("retry-after-ms", "12.5")]),
642            Some(Duration::from_micros(12_500))
643        );
644        assert_eq!(wait(&[("retry-after", "-1")]), None);
645        assert_eq!(wait(&[("retry-after", "")]), None);
646        assert_eq!(wait(&[("retry-after", "later")]), None);
647        assert_eq!(
648            wait(&[("retry-after-ms", "nan"), ("retry-after", "3")]),
649            Some(Duration::from_secs(3))
650        );
651        assert_eq!(wait(&[]), None);
652    }
653
654    #[tokio::test(start_paused = true)]
655    async fn parses_http_dates_against_the_wall_clock() {
656        let wait = |value: &str| parse(&[("retry-after", value)], clocks());
657        // The wall clock of `clocks()` is 07:28:00.
658        assert_eq!(
659            wait("Wed, 21 Oct 2015 07:28:30 GMT"),
660            Some(Duration::from_secs(30))
661        );
662        assert_eq!(wait("Wed, 21 Oct 2015 07:28:00 GMT"), Some(Duration::ZERO));
663        assert_eq!(wait("Wed, 21 Oct 2015 07:00:00 GMT"), Some(Duration::ZERO));
664        // The obsolete formats HTTP still accepts.
665        assert_eq!(
666            wait("Wednesday, 21-Oct-15 07:29:00 GMT"),
667            Some(Duration::from_secs(60))
668        );
669        assert_eq!(
670            wait("Wed Oct 21 07:28:05 2015"),
671            Some(Duration::from_secs(5))
672        );
673    }
674
675    #[tokio::test(start_paused = true)]
676    async fn seconds_count_from_receipt_and_dates_from_now() {
677        let mut clocks = clocks();
678        clocks.now = clocks.received_at + Duration::from_secs(2); // two seconds reading the body
679        assert_eq!(
680            parse(&[("retry-after", "5")], clocks),
681            Some(Duration::from_secs(3))
682        );
683        assert_eq!(
684            parse(&[("retry-after-ms", "1000")], clocks),
685            Some(Duration::ZERO)
686        );
687        assert_eq!(
688            parse(&[("retry-after", "Wed, 21 Oct 2015 07:28:05 GMT")], clocks),
689            Some(Duration::from_secs(5))
690        );
691    }
692
693    #[tokio::test(start_paused = true)]
694    async fn backs_off_and_caps() {
695        let mut policy = RetryPolicy::new(10)
696            .base_delay(Duration::from_millis(100))
697            .max_delay(Duration::from_secs(1))
698            .jitter(false);
699        let now = Instant::now();
700        assert_eq!(
701            policy.deadline(None, now),
702            Some(now + Duration::from_millis(100))
703        );
704        policy.attempts = 2;
705        assert_eq!(
706            policy.deadline(None, now),
707            Some(now + Duration::from_millis(400))
708        );
709        policy.attempts = 8;
710        assert_eq!(
711            policy.deadline(None, now),
712            Some(now + Duration::from_secs(1))
713        );
714    }
715
716    #[tokio::test(start_paused = true)]
717    async fn honours_retry_after_beyond_max_delay_up_to_max_retry_after() {
718        let policy = RetryPolicy::new(3).max_delay(Duration::from_secs(1));
719        let now = Instant::now();
720        let after = |secs| Some(now + Duration::from_secs(secs));
721        assert_eq!(policy.deadline(after(30), now), after(30));
722        assert_eq!(policy.deadline(after(60), now), after(60));
723        assert_eq!(policy.deadline(after(61), now), None);
724        // A deadline already passed is kept: retry now.
725        assert_eq!(
726            policy.deadline(Some(now), now + Duration::from_secs(1)),
727            Some(now)
728        );
729        let policy = policy.max_retry_after(Duration::from_secs(5));
730        assert_eq!(policy.deadline(after(6), now), None);
731    }
732
733    #[tokio::test(start_paused = true)]
734    async fn jitter_stays_between_half_and_all_of_the_backoff() {
735        let backoff = Duration::from_millis(800);
736        assert_eq!(jittered(backoff, 0.0), Duration::from_millis(400));
737        assert_eq!(jittered(backoff, 0.5), Duration::from_millis(600));
738        assert!(jittered(backoff, 0.999_999) <= backoff);
739
740        let mut policy = RetryPolicy::new(10)
741            .base_delay(Duration::from_millis(100))
742            .max_delay(Duration::from_secs(1));
743        policy.attempts = 3; // 800 ms before jitter
744        let delays: Vec<_> = (0..200).map(|_| policy.backoff()).collect();
745        for delay in &delays {
746            assert!(
747                (Duration::from_millis(400)..=backoff).contains(delay),
748                "{delay:?}"
749            );
750        }
751        assert!(delays.iter().any(|delay| *delay != delays[0]), "random");
752        // The waits a server asks for are not jittered.
753        let now = Instant::now();
754        let deadline = Some(now + Duration::from_secs(2));
755        assert_eq!(policy.deadline(deadline, now), deadline);
756    }
757
758    #[test]
759    fn random_unit_is_in_range() {
760        for _ in 0..1000 {
761            assert!((0.0..1.0).contains(&random_unit()));
762        }
763    }
764
765    #[test]
766    fn recognises_openai_insufficient_quota() {
767        assert!(openai_insufficient_quota(
768            br#"{"error":{"message":"You exceeded your current quota","type":"insufficient_quota","param":null,"code":"insufficient_quota"}}"#
769        ));
770        assert!(openai_insufficient_quota(
771            br#"{"error":{"type":"insufficient_quota"}}"#
772        ));
773        assert!(openai_insufficient_quota(
774            br#"{"error":{"type":"requests","code":"insufficient_quota"}}"#
775        ));
776        assert!(!openai_insufficient_quota(
777            br#"{"error":{"message":"Rate limit reached","type":"requests","code":"rate_limit_exceeded"}}"#
778        ));
779        assert!(!openai_insufficient_quota(b"insufficient_quota"));
780        assert!(!openai_insufficient_quota(b""));
781    }
782
783    #[test]
784    fn only_bodies_in_memory_are_cloned() {
785        let mut policy = RetryPolicy::default();
786        let mut clone_request = |request: &http::Request<Body>| {
787            Policy::<_, http::Response<()>, BoxError>::clone_request(&mut policy, request)
788        };
789        let request = http::Request::post("https://example.com")
790            .header("x", "1")
791            .body(Body::from("a"))
792            .unwrap();
793        let clone = clone_request(&request).unwrap();
794        assert_eq!(clone.uri(), request.uri());
795        assert_eq!(clone.headers()["x"], "1");
796        let chunks =
797            futures_util::stream::iter(vec![Ok::<_, std::io::Error>(bytes::Bytes::from("a"))]);
798        let streamed = http::Request::post("https://example.com")
799            .body(Body::wrap_stream(chunks))
800            .unwrap();
801        assert!(clone_request(&streamed).is_none());
802    }
803
804    fn body_429(chunks: Vec<Result<&'static str, &'static str>>) -> http::Response<Body> {
805        let stream = futures_util::stream::iter(
806            chunks
807                .into_iter()
808                .map(|chunk| chunk.map(Bytes::from).map_err(BoxError::from)),
809        );
810        http::Response::builder()
811            .status(429)
812            .body(Body::wrap_stream(stream))
813            .unwrap()
814    }
815
816    #[tokio::test]
817    async fn classifies_the_whole_body_and_puts_it_back() {
818        let response = body_429(vec![
819            Ok(r#"{"error":{"type":"#),
820            Ok(r#""insufficient_quota"}}"#),
821        ]);
822        let response = classify_429(response, &openai_insufficient_quota, 1024).await;
823        assert!(response.extensions().get::<Permanent>().is_some());
824        let body = response.into_body().collect_bytes().await.unwrap();
825        assert_eq!(body, r#"{"error":{"type":"insufficient_quota"}}"#);
826    }
827
828    #[tokio::test]
829    async fn a_body_over_the_limit_is_not_classified_but_kept_whole() {
830        let response = body_429(vec![Ok("0123"), Ok("4567"), Ok("89")]);
831        let response = classify_429(response, &|_: &[u8]| true, 5).await;
832        assert!(response.extensions().get::<Permanent>().is_none());
833        let body = response.into_body().collect_bytes().await.unwrap();
834        assert_eq!(body, "0123456789");
835    }
836
837    #[tokio::test]
838    async fn a_body_that_fails_is_not_classified_and_keeps_its_error() {
839        let response = body_429(vec![Ok("01"), Err("reset")]);
840        let response = classify_429(response, &|_: &[u8]| true, 1024).await;
841        assert!(response.extensions().get::<Permanent>().is_none());
842        let error = response.into_body().collect_bytes().await.unwrap_err();
843        assert_eq!(error.to_string(), "reset");
844    }
845}
timvw_rust_runtime/retry.rs

timvw_rust_runtime/
retry.rs
