pagesift_api/
limits.rs

1//! 按 API key 的限流和配额：每分钟请求数（令牌桶）、同时进行的任务数、每天（UTC）的页面数。
2//! 超限时返回 Firecrawl 兼容的 429 错误体。没配置任何限额时什么都不检查。
3//!
4//! 当天已用的页数写进存储（有的话），重启不会把配额清零。
5//!
6//! 多服务器网关上（[`Limiter::with_ledger`]）每分钟请求数和每日页数改记在 `pagesift-cluster` 的账本里：
7//! 用量是追加的冻结 / 结算记录，崩溃重放、温备接管后一致；令牌桶在账本里，接管时由最近一分钟的
8//! 请求记录重建。单主机不经过账本，行为不变。
9
10use std::collections::HashMap;
11use std::sync::{Arc, Mutex};
12use std::time::{Instant, SystemTime, UNIX_EPOCH};
13
14use pagesift_store::Store;
15
16use crate::error::{ApiError, ErrorCode};
17
18/// 一个 API key 的限额，各项不设就是不限。
19#[derive(Debug, Clone, Default, PartialEq, Eq)]
20pub struct KeyLimits {
21    /// 日志、监控和配额记录里用的名字；不设时用 key 的指纹，key 本身不会写进日志或磁盘。
22    pub name: Option<String>,
23    pub requests_per_minute: Option<u32>,
24    pub max_concurrent_jobs: Option<usize>,
25    pub pages_per_day: Option<u64>,
26}
27
28/// 发起请求的调用方：鉴权中间件放进请求扩展里。
29#[derive(Debug, Clone)]
30pub(crate) struct Caller {
31    pub account: Arc<str>,
32    pub limits: Arc<KeyLimits>,
33}
34
35impl Caller {
36    pub(crate) fn new(key: Option<&str>, limits: Arc<KeyLimits>) -> Self {
37        let account = match (&limits.name, key) {
38            (Some(n), _) => n.clone(),
39            (None, Some(k)) => fingerprint(k),
40            (None, None) => "anonymous".into(),
41        };
42        Self {
43            account: account.into(),
44            limits,
45        }
46    }
47}
48
49impl Caller {
50    /// 恢复任务时按保存的名字找回调用方。
51    pub(crate) fn named(account: &str, limits: Arc<KeyLimits>) -> Self {
52        Self {
53            account: account.into(),
54            limits,
55        }
56    }
57}
58
59/// key 的指纹（FNV-1a 64 位），稳定、不可逆到能用的程度，用来区分调用方。
60fn fingerprint(key: &str) -> String {
61    let mut h: u64 = 0xcbf2_9ce4_8422_2325;
62    for b in key.bytes() {
63        h ^= u64::from(b);
64        h = h.wrapping_mul(0x0100_0000_01b3);
65    }
66    format!("key-{:08x}", h >> 32)
67}
68
69struct Bucket {
70    tokens: f64,
71    last: Instant,
72}
73
74#[derive(Clone, Copy)]
75struct Usage {
76    day: u64,
77    used: u64,
78}
79
80#[derive(Default)]
81pub(crate) struct Limiter {
82    buckets: Mutex<HashMap<Arc<str>, Bucket>>,
83    usage: Mutex<HashMap<Arc<str>, Usage>>,
84    store: Option<Store>,
85    ledger: Option<Arc<pagesift_cluster::Ledger>>,
86}
87
88fn rate_error(rpm: u32, wait: u64) -> ApiError {
89    ApiError::new(
90        ErrorCode::RateLimitExceeded,
91        format!(
92            "rate limit exceeded: {rpm} requests per minute for this API key; retry after {wait}s"
93        ),
94    )
95    .with_retry_after(wait)
96}
97
98fn ledger_error(e: impl std::fmt::Display) -> ApiError {
99    ApiError::new(ErrorCode::Internal, format!("quota ledger: {e}"))
100}
101
102fn today() -> u64 {
103    SystemTime::now()
104        .duration_since(UNIX_EPOCH)
105        .map_or(0, |d| d.as_secs() / 86_400)
106}
107
108fn secs_to_midnight() -> u64 {
109    let now = SystemTime::now()
110        .duration_since(UNIX_EPOCH)
111        .map_or(0, |d| d.as_secs());
112    86_400 - now % 86_400
113}
114
115fn usage_key(account: &str) -> String {
116    format!("quota/{account}")
117}
118
119impl Limiter {
120    pub(crate) fn new(store: Option<Store>) -> Self {
121        Self {
122            store,
123            ..Default::default()
124        }
125    }
126
127    /// 网关：限额记在账本里。`accounts` 是全部账户（名字 → 限额），事先登记。
128    pub(crate) fn with_ledger(
129        ledger: Arc<pagesift_cluster::Ledger>,
130        accounts: &HashMap<String, Arc<KeyLimits>>,
131    ) -> Self {
132        for (name, l) in accounts {
133            ledger.set_limits(
134                name,
135                pagesift_cluster::AccountLimits {
136                    pages_per_day: l.pages_per_day,
137                    requests_per_minute: l.requests_per_minute,
138                },
139            );
140        }
141        Self {
142            ledger: Some(ledger),
143            ..Default::default()
144        }
145    }
146
147    /// 每分钟请求数：容量等于每分钟请求数的令牌桶，匀速补充，允许短时突发。
148    pub(crate) fn check_rate(&self, c: &Caller) -> Result<(), ApiError> {
149        let Some(rpm) = c.limits.requests_per_minute else {
150            return Ok(());
151        };
152        if let Some(l) = &self.ledger {
153            use pagesift_cluster::Refusal;
154            return match l.check_rate(&c.account) {
155                Ok(()) => Ok(()),
156                Err(Refusal::Rate { retry_after, .. }) => Err(rate_error(
157                    rpm,
158                    retry_after.as_secs_f64().ceil().max(1.0) as u64,
159                )),
160                Err(Refusal::Quota { .. }) => Err(quota_error(c)),
161                Err(Refusal::Error(e)) => Err(ledger_error(e)),
162            };
163        }
164        let cap = f64::from(rpm.max(1));
165        let per_sec = cap / 60.0;
166        let now = Instant::now();
167        let mut buckets = self.buckets.lock().unwrap_or_else(|e| e.into_inner());
168        let b = buckets.entry(c.account.clone()).or_insert(Bucket {
169            tokens: cap,
170            last: now,
171        });
172        b.tokens = (b.tokens + now.duration_since(b.last).as_secs_f64() * per_sec).min(cap);
173        b.last = now;
174        if b.tokens >= 1.0 {
175            b.tokens -= 1.0;
176            return Ok(());
177        }
178        let wait = ((1.0 - b.tokens) / per_sec).ceil().max(1.0) as u64;
179        Err(rate_error(rpm, wait))
180    }
181
182    /// 同时进行的任务数是否还能再加一个。
183    pub(crate) fn check_jobs(&self, c: &Caller, active: usize) -> Result<(), ApiError> {
184        match c.limits.max_concurrent_jobs {
185            Some(max) if active >= max => Err(ApiError::new(
186                ErrorCode::ConcurrencyLimitExceeded,
187                format!("concurrency limit exceeded: at most {max} crawl or batch jobs may run at once for this API key"),
188            )
189            .with_retry_after(5)),
190            _ => Ok(()),
191        }
192    }
193
194    fn usage(&self, account: &Arc<str>) -> Usage {
195        let day = today();
196        let mut map = self.usage.lock().unwrap_or_else(|e| e.into_inner());
197        if let Some(u) = map.get(account).filter(|u| u.day == day) {
198            return *u;
199        }
200        // 第一次用到这个调用方（或者跨天了）：从存储里读当天的记录
201        let stored = self
202            .store
203            .as_ref()
204            .and_then(|s| s.tasks().kv_get(usage_key(account)).ok().flatten())
205            .and_then(|v| {
206                let text = String::from_utf8(v).ok()?;
207                let (d, n) = text.split_once(' ')?;
208                Some(Usage {
209                    day: d.parse().ok()?,
210                    used: n.parse().ok()?,
211                })
212            })
213            .filter(|u| u.day == day);
214        let u = stored.unwrap_or(Usage { day, used: 0 });
215        map.insert(account.clone(), u);
216        u
217    }
218
219    /// 当天还剩多少页；没设配额时是 `None`。
220    pub(crate) fn remaining_pages(&self, c: &Caller) -> Option<u64> {
221        let max = c.limits.pages_per_day?;
222        if let Some(l) = &self.ledger {
223            return l.balance(&c.account);
224        }
225        Some(max.saturating_sub(self.usage(&c.account).used))
226    }
227
228    pub(crate) fn check_quota(&self, c: &Caller) -> Result<(), ApiError> {
229        if self.remaining_pages(c) == Some(0) {
230            return Err(quota_error(c));
231        }
232        Ok(())
233    }
234
235    /// 记入当天用掉的页数。只有设了配额的调用方才计数和落盘。
236    pub(crate) async fn consume_pages(&self, c: &Caller, n: u64) {
237        if c.limits.pages_per_day.is_none() || n == 0 {
238            return;
239        }
240        if let Some(l) = &self.ledger {
241            // 冻结后立即按实际用量结算：页面已经抓完。余额不够冻结时（并发请求都通过了检查）
242            // 当天已经用满，不再记账，余额照样是 0
243            match l.reserve(&c.account, n) {
244                Ok(hold) => {
245                    if let Err(e) = l.settle(&hold, n) {
246                        tracing::warn!(error = %e, "settling page quota failed");
247                    }
248                }
249                Err(pagesift_cluster::Refusal::Error(e)) => {
250                    tracing::warn!(error = %e, "reserving page quota failed");
251                }
252                Err(_) => {}
253            }
254            return;
255        }
256        let u = self.usage(&c.account);
257        let u = {
258            let mut map = self.usage.lock().unwrap_or_else(|e| e.into_inner());
259            let e = map.entry(c.account.clone()).or_insert(u);
260            if e.day != u.day {
261                *e = u;
262            }
263            e.used += n;
264            *e
265        };
266        if let Some(store) = &self.store {
267            let key = usage_key(&c.account);
268            let value = format!("{} {}", u.day, u.used);
269            let r = store
270                .blocking(move |s| s.tasks().kv_put(key, value.as_bytes()))
271                .await;
272            if let Err(e) = r {
273                tracing::warn!(error = %e, "saving page quota usage failed");
274            }
275        }
276    }
277}
278
279pub(crate) fn quota_error(c: &Caller) -> ApiError {
280    ApiError::new(
281        ErrorCode::QuotaExceeded,
282        format!(
283            "daily page quota exceeded: {} pages per day for this API key",
284            c.limits.pages_per_day.unwrap_or(0)
285        ),
286    )
287    .with_retry_after(secs_to_midnight())
288}
289
290#[cfg(test)]
291mod tests {
292    use super::*;
293
294    fn caller(limits: KeyLimits) -> Caller {
295        Caller::new(Some("secret-key"), Arc::new(limits))
296    }
297
298    #[test]
299    fn token_bucket_allows_a_burst_then_429s() {
300        let l = Limiter::default();
301        let c = caller(KeyLimits {
302            requests_per_minute: Some(3),
303            ..Default::default()
304        });
305        for _ in 0..3 {
306            l.check_rate(&c).unwrap();
307        }
308        let e = l.check_rate(&c).unwrap_err();
309        assert_eq!(e.code, ErrorCode::RateLimitExceeded);
310        assert!(e.retry_after.unwrap() >= 1);
311        // 别的调用方有自己的桶
312        let other = Caller::new(Some("other"), c.limits.clone());
313        l.check_rate(&other).unwrap();
314        // 不限速的调用方不受影响
315        l.check_rate(&caller(KeyLimits::default())).unwrap();
316    }
317
318    #[test]
319    fn accounts_use_names_or_fingerprints() {
320        let named = Caller::new(
321            Some("k"),
322            Arc::new(KeyLimits {
323                name: Some("team-a".into()),
324                ..Default::default()
325            }),
326        );
327        assert_eq!(&*named.account, "team-a");
328        let anon = caller(KeyLimits::default());
329        assert!(anon.account.starts_with("key-") && !anon.account.contains("secret"));
330        assert_eq!(&*Caller::new(None, Arc::default()).account, "anonymous");
331    }
332
333    /// 网关：限额记在账本里，行为与单主机一致（桶先突发后 429，配额用完 QUOTA_EXCEEDED）。
334    #[tokio::test]
335    async fn ledger_backed_limits_match_the_local_ones() {
336        let ledger = Arc::new(
337            pagesift_cluster::Ledger::open(pagesift_cluster::LedgerConfig::new(
338                pagesift_cluster::MonoClock::new(),
339                false,
340            ))
341            .unwrap(),
342        );
343        let limits = Arc::new(KeyLimits {
344            requests_per_minute: Some(2),
345            pages_per_day: Some(2),
346            ..Default::default()
347        });
348        let c = Caller::new(Some("k"), limits.clone());
349        let accounts = HashMap::from([(c.account.to_string(), limits)]);
350        let l = Limiter::with_ledger(ledger, &accounts);
351        l.check_rate(&c).unwrap();
352        l.check_rate(&c).unwrap();
353        let e = l.check_rate(&c).unwrap_err();
354        assert_eq!(e.code, ErrorCode::RateLimitExceeded);
355        assert!(e.retry_after.unwrap() >= 1);
356        assert_eq!(l.remaining_pages(&c), Some(2));
357        l.consume_pages(&c, 1).await;
358        assert!(l.check_quota(&c).is_ok());
359        l.consume_pages(&c, 1).await;
360        assert_eq!(l.remaining_pages(&c), Some(0));
361        assert_eq!(
362            l.check_quota(&c).unwrap_err().code,
363            ErrorCode::QuotaExceeded
364        );
365        // 不限额的调用方不受影响
366        l.check_rate(&caller(KeyLimits::default())).unwrap();
367    }
368
369    #[tokio::test]
370    async fn page_quota_survives_a_restart() {
371        let dir = tempfile::tempdir().unwrap();
372        let c = caller(KeyLimits {
373            pages_per_day: Some(5),
374            ..Default::default()
375        });
376        {
377            let store = Store::open_dir(dir.path()).unwrap();
378            let l = Limiter::new(Some(store));
379            assert_eq!(l.remaining_pages(&c), Some(5));
380            l.consume_pages(&c, 4).await;
381            assert!(l.check_quota(&c).is_ok());
382            l.consume_pages(&c, 1).await;
383            assert_eq!(
384                l.check_quota(&c).unwrap_err().code,
385                ErrorCode::QuotaExceeded
386            );
387        }
388        let l = Limiter::new(Some(Store::open_dir(dir.path()).unwrap()));
389        assert_eq!(l.remaining_pages(&c), Some(0));
390        let jobs = caller(KeyLimits {
391            max_concurrent_jobs: Some(1),
392            ..Default::default()
393        });
394        assert!(l.check_jobs(&jobs, 0).is_ok());
395        assert_eq!(
396            l.check_jobs(&jobs, 1).unwrap_err().code,
397            ErrorCode::ConcurrencyLimitExceeded
398        );
399    }
400}
pagesift_api/limits.rs

pagesift_api/
limits.rs
