ferrox-server 0.10.0

OpenAI-compatible HTTP server for the Ferrox inference engine
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244

//! Optional bearer-token auth and a simple rate limiter for
//! ferrox-server. Both are off by default -- unset `FERROX_API_KEY` /
//! `FERROX_RATE_LIMIT_PER_MINUTE` and the server behaves exactly as
//! before -- following the same opt-in convention llama.cpp's server
//! uses for `--api-key`, so existing deployments and tests aren't
//! affected unless explicitly configured.

use axum::{
    body::Body,
    extract::State,
    http::{
        header::{AUTHORIZATION, RETRY_AFTER},
        HeaderValue, Request, StatusCode,
    },
    middleware::Next,
    response::{IntoResponse, Response},
    Json,
};
use std::sync::{Arc, Mutex};
use std::time::Instant;

#[derive(Clone)]
pub struct AuthConfig {
    pub api_key: Arc<String>,
}

/// Rejects any request whose `Authorization: Bearer <key>` header
/// doesn't match the configured key. Only ever added to the router
/// when `FERROX_API_KEY` is set (see `main.rs`) -- there is no
/// "disabled" state to check here, keeping the common unauthenticated
/// path free of any per-request branch.
pub async fn require_api_key(
    State(config): State<AuthConfig>,
    req: Request<Body>,
    next: Next,
) -> Response {
    let provided = req
        .headers()
        .get(AUTHORIZATION)
        .and_then(|v| v.to_str().ok())
        .and_then(|v| v.strip_prefix("Bearer "));

    if provided == Some(config.api_key.as_str()) {
        next.run(req).await
    } else {
        (
            StatusCode::UNAUTHORIZED,
            Json(serde_json::json!({"error": {"message": "invalid or missing API key"}})),
        )
            .into_response()
    }
}

struct Bucket {
    tokens: f64,
    capacity: f64,
    refill_per_sec: f64,
    last_refill: Instant,
}

/// A simple token-bucket rate limiter, shared across all requests
/// (global, not per-client): `capacity` tokens, continuously refilled
/// at `refill_per_sec` tokens/second, one token consumed per request
/// that passes through it. A reasonable minimum for a single-model
/// self-hosted deployment; a real multi-tenant deployment would want
/// per-API-key buckets instead of one global bucket, a possible
/// follow-on.
pub struct RateLimiter {
    inner: Mutex<Bucket>,
}

impl RateLimiter {
    pub fn new(capacity: u32, refill_per_sec: f64) -> Self {
        RateLimiter {
            inner: Mutex::new(Bucket {
                tokens: capacity as f64,
                capacity: capacity as f64,
                refill_per_sec,
                last_refill: Instant::now(),
            }),
        }
    }

    pub fn per_minute(requests_per_minute: u32) -> Self {
        // Bucket capacity equals a full minute's allowance, so a burst
        // after idle time isn't punished, refilled continuously at the
        // equivalent per-second rate.
        Self::new(requests_per_minute, requests_per_minute as f64 / 60.0)
    }

    /// Attempts to consume one token. Returns `false` (and consumes
    /// nothing) if none are available right now.
    pub fn try_acquire(&self) -> bool {
        let mut b = self.inner.lock().unwrap_or_else(|p| p.into_inner());
        let now = Instant::now();
        let elapsed = now.duration_since(b.last_refill).as_secs_f64();
        b.tokens = (b.tokens + elapsed * b.refill_per_sec).min(b.capacity);
        b.last_refill = now;
        if b.tokens >= 1.0 {
            b.tokens -= 1.0;
            true
        } else {
            false
        }
    }
}

pub async fn rate_limit(
    State(limiter): State<Arc<RateLimiter>>,
    req: Request<Body>,
    next: Next,
) -> Response {
    if limiter.try_acquire() {
        next.run(req).await
    } else {
        (
            StatusCode::TOO_MANY_REQUESTS,
            Json(serde_json::json!({"error": {"message": "rate limit exceeded"}})),
        )
            .into_response()
    }
}

/// Seconds a client is told to wait before retrying a 503.
///
/// A fixed short hint rather than a computed one: the honest estimate
/// would be "queue depth divided by throughput", and this server does
/// not know a caller's throughput. One second is short enough that a
/// legitimate client recovers quickly and long enough that a retry
/// storm loses its coordination.
pub const RETRY_AFTER_SECONDS: u64 = 1;

/// Stamps `Retry-After` on any 503 that does not already carry one.
///
/// Lives in a layer rather than at each rejection site because the
/// header is a property of the *status*, not of any one handler: 503
/// means "temporarily unavailable" (RFC 9110), so every 503 this server
/// emits -- queue full, KV pool exhausted, no model loaded -- is by
/// definition worth retrying, and a client that has to guess how long
/// to wait will guess "immediately". An existing header is never
/// overwritten, so a handler that knows better keeps its own value.
pub async fn retry_after(req: Request<Body>, next: Next) -> Response {
    let mut response = next.run(req).await;
    if response.status() == StatusCode::SERVICE_UNAVAILABLE
        && !response.headers().contains_key(RETRY_AFTER)
    {
        response
            .headers_mut()
            .insert(RETRY_AFTER, HeaderValue::from(RETRY_AFTER_SECONDS));
    }
    response
}

#[cfg(test)]
mod tests {
    use super::*;
    use axum::{routing::get, Router};
    use tower::ServiceExt;

    async fn call(app: Router, path: &str) -> Response {
        app.oneshot(
            Request::builder()
                .uri(path)
                .body(Body::empty())
                .expect("request"),
        )
        .await
        .expect("response")
    }

    fn retry_after_router() -> Router {
        Router::new()
            .route("/busy", get(|| async { StatusCode::SERVICE_UNAVAILABLE }))
            .route("/fine", get(|| async { StatusCode::OK }))
            .route(
                "/busy-with-hint",
                get(|| async { ([(RETRY_AFTER, "30")], StatusCode::SERVICE_UNAVAILABLE) }),
            )
            .layer(axum::middleware::from_fn(retry_after))
    }

    /// A 503 without a retry hint tells a client nothing except "try
    /// again", which clients read as "try again now" -- the retry storm
    /// the queue cap exists to survive.
    #[tokio::test]
    async fn a_503_gets_a_retry_after_header() {
        let response = call(retry_after_router(), "/busy").await;
        assert_eq!(response.status(), StatusCode::SERVICE_UNAVAILABLE);
        assert_eq!(
            response.headers().get(RETRY_AFTER).expect("header present"),
            &HeaderValue::from(RETRY_AFTER_SECONDS)
        );
    }

    #[tokio::test]
    async fn a_successful_response_is_left_alone() {
        let response = call(retry_after_router(), "/fine").await;
        assert_eq!(response.status(), StatusCode::OK);
        assert!(response.headers().get(RETRY_AFTER).is_none());
    }

    #[tokio::test]
    async fn an_existing_retry_after_is_not_overwritten() {
        let response = call(retry_after_router(), "/busy-with-hint").await;
        assert_eq!(
            response.headers().get(RETRY_AFTER).expect("header present"),
            "30",
            "a handler that knows a better value keeps it"
        );
    }

    #[test]
    fn rate_limiter_allows_up_to_capacity_then_blocks() {
        let limiter = RateLimiter::new(3, 0.0); // no refill, isolate the capacity check
        assert!(limiter.try_acquire());
        assert!(limiter.try_acquire());
        assert!(limiter.try_acquire());
        assert!(
            !limiter.try_acquire(),
            "a 4th request within the same instant must be rejected"
        );
    }

    #[test]
    fn rate_limiter_refills_over_time() {
        let limiter = RateLimiter::new(1, 1000.0); // fast refill for a quick test
        assert!(limiter.try_acquire());
        assert!(!limiter.try_acquire());
        std::thread::sleep(std::time::Duration::from_millis(5));
        assert!(
            limiter.try_acquire(),
            "tokens must refill over time, not stay exhausted forever"
        );
    }

    #[test]
    fn per_minute_constructor_sets_a_full_minute_of_burst_capacity() {
        let limiter = RateLimiter::per_minute(60);
        for _ in 0..60 {
            assert!(limiter.try_acquire());
        }
        assert!(!limiter.try_acquire());
    }
}