Line data Source code
1 : use std::{
2 : borrow::Cow,
3 : collections::hash_map::RandomState,
4 : hash::{BuildHasher, Hash},
5 : sync::{
6 : atomic::{AtomicUsize, Ordering},
7 : Mutex,
8 : },
9 : };
10 :
11 : use anyhow::bail;
12 : use dashmap::DashMap;
13 : use itertools::Itertools;
14 : use rand::{rngs::StdRng, Rng, SeedableRng};
15 : use tokio::time::{Duration, Instant};
16 : use tracing::info;
17 :
18 : use crate::intern::EndpointIdInt;
19 :
20 : pub struct GlobalRateLimiter {
21 : data: Vec<RateBucket>,
22 : info: Vec<RateBucketInfo>,
23 : }
24 :
25 : impl GlobalRateLimiter {
26 0 : pub fn new(info: Vec<RateBucketInfo>) -> Self {
27 0 : Self {
28 0 : data: vec![
29 0 : RateBucket {
30 0 : start: Instant::now(),
31 0 : count: 0,
32 0 : };
33 0 : info.len()
34 0 : ],
35 0 : info,
36 0 : }
37 0 : }
38 :
39 : /// Check that number of connections is below `max_rps` rps.
40 0 : pub fn check(&mut self) -> bool {
41 0 : let now = Instant::now();
42 0 :
43 0 : let should_allow_request = self
44 0 : .data
45 0 : .iter_mut()
46 0 : .zip(&self.info)
47 0 : .all(|(bucket, info)| bucket.should_allow_request(info, now, 1));
48 0 :
49 0 : if should_allow_request {
50 0 : // only increment the bucket counts if the request will actually be accepted
51 0 : self.data.iter_mut().for_each(|b| b.inc(1));
52 0 : }
53 :
54 0 : should_allow_request
55 0 : }
56 : }
57 :
58 : // Simple per-endpoint rate limiter.
59 : //
60 : // Check that number of connections to the endpoint is below `max_rps` rps.
61 : // Purposefully ignore user name and database name as clients can reconnect
62 : // with different names, so we'll end up sending some http requests to
63 : // the control plane.
64 : pub type EndpointRateLimiter = BucketRateLimiter<EndpointIdInt, StdRng, RandomState>;
65 :
66 : pub struct BucketRateLimiter<Key, Rand = StdRng, Hasher = RandomState> {
67 : map: DashMap<Key, Vec<RateBucket>, Hasher>,
68 : info: Cow<'static, [RateBucketInfo]>,
69 : access_count: AtomicUsize,
70 : rand: Mutex<Rand>,
71 : }
72 :
73 : #[derive(Clone, Copy)]
74 : struct RateBucket {
75 : start: Instant,
76 : count: u32,
77 : }
78 :
79 : impl RateBucket {
80 6001830 : fn should_allow_request(&mut self, info: &RateBucketInfo, now: Instant, n: u32) -> bool {
81 6001830 : if now - self.start < info.interval {
82 6001814 : self.count + n <= info.max_rpi
83 : } else {
84 : // bucket expired, reset
85 16 : self.count = 0;
86 16 : self.start = now;
87 16 :
88 16 : true
89 : }
90 6001830 : }
91 :
92 6001818 : fn inc(&mut self, n: u32) {
93 6001818 : self.count += n;
94 6001818 : }
95 : }
96 :
97 : #[derive(Clone, Copy, PartialEq)]
98 : pub struct RateBucketInfo {
99 : pub interval: Duration,
100 : // requests per interval
101 : pub max_rpi: u32,
102 : }
103 :
104 : impl std::fmt::Display for RateBucketInfo {
105 32 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
106 32 : let rps = (self.max_rpi as u64) * 1000 / self.interval.as_millis() as u64;
107 32 : write!(f, "{rps}@{}", humantime::format_duration(self.interval))
108 32 : }
109 : }
110 :
111 : impl std::fmt::Debug for RateBucketInfo {
112 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
113 0 : write!(f, "{self}")
114 0 : }
115 : }
116 :
117 : impl std::str::FromStr for RateBucketInfo {
118 : type Err = anyhow::Error;
119 :
120 52 : fn from_str(s: &str) -> Result<Self, Self::Err> {
121 52 : let Some((max_rps, interval)) = s.split_once('@') else {
122 0 : bail!("invalid rate info")
123 : };
124 52 : let max_rps = max_rps.parse()?;
125 52 : let interval = humantime::parse_duration(interval)?;
126 52 : Ok(Self::new(max_rps, interval))
127 52 : }
128 : }
129 :
130 : impl RateBucketInfo {
131 : pub const DEFAULT_ENDPOINT_SET: [Self; 3] = [
132 : Self::new(300, Duration::from_secs(1)),
133 : Self::new(200, Duration::from_secs(60)),
134 : Self::new(100, Duration::from_secs(600)),
135 : ];
136 :
137 6 : pub fn validate(info: &mut [Self]) -> anyhow::Result<()> {
138 16 : info.sort_unstable_by_key(|info| info.interval);
139 6 : let invalid = info
140 6 : .iter()
141 6 : .tuple_windows()
142 8 : .find(|(a, b)| a.max_rpi > b.max_rpi);
143 6 : if let Some((a, b)) = invalid {
144 2 : bail!(
145 2 : "invalid bucket RPS limits. {b} allows fewer requests per bucket than {a} ({} vs {})",
146 2 : b.max_rpi,
147 2 : a.max_rpi,
148 2 : );
149 4 : }
150 4 :
151 4 : Ok(())
152 6 : }
153 :
154 60 : pub const fn new(max_rps: u32, interval: Duration) -> Self {
155 60 : Self {
156 60 : interval,
157 60 : max_rpi: ((max_rps as u64) * (interval.as_millis() as u64) / 1000) as u32,
158 60 : }
159 60 : }
160 : }
161 :
162 : impl<K: Hash + Eq> BucketRateLimiter<K> {
163 8 : pub fn new(info: impl Into<Cow<'static, [RateBucketInfo]>>) -> Self {
164 8 : Self::new_with_rand_and_hasher(info, StdRng::from_entropy(), RandomState::new())
165 8 : }
166 : }
167 :
168 : impl<K: Hash + Eq, R: Rng, S: BuildHasher + Clone> BucketRateLimiter<K, R, S> {
169 10 : fn new_with_rand_and_hasher(
170 10 : info: impl Into<Cow<'static, [RateBucketInfo]>>,
171 10 : rand: R,
172 10 : hasher: S,
173 10 : ) -> Self {
174 10 : let info = info.into();
175 10 : info!(buckets = ?info, "endpoint rate limiter");
176 10 : Self {
177 10 : info,
178 10 : map: DashMap::with_hasher_and_shard_amount(hasher, 64),
179 10 : access_count: AtomicUsize::new(1), // start from 1 to avoid GC on the first request
180 10 : rand: Mutex::new(rand),
181 10 : }
182 10 : }
183 :
184 : /// Check that number of connections to the endpoint is below `max_rps` rps.
185 2000914 : pub fn check(&self, key: K, n: u32) -> bool {
186 2000914 : // do a partial GC every 2k requests. This cleans up ~ 1/64th of the map.
187 2000914 : // worst case memory usage is about:
188 2000914 : // = 2 * 2048 * 64 * (48B + 72B)
189 2000914 : // = 30MB
190 2000914 : if self.access_count.fetch_add(1, Ordering::AcqRel) % 2048 == 0 {
191 976 : self.do_gc();
192 1999938 : }
193 :
194 2000914 : let now = Instant::now();
195 2000914 : let mut entry = self.map.entry(key).or_insert_with(|| {
196 2000008 : vec![
197 2000008 : RateBucket {
198 2000008 : start: now,
199 2000008 : count: 0,
200 2000008 : };
201 2000008 : self.info.len()
202 2000008 : ]
203 2000914 : });
204 2000914 :
205 2000914 : let should_allow_request = entry
206 2000914 : .iter_mut()
207 2000914 : .zip(&*self.info)
208 6001830 : .all(|(bucket, info)| bucket.should_allow_request(info, now, n));
209 2000914 :
210 2000914 : if should_allow_request {
211 2000906 : // only increment the bucket counts if the request will actually be accepted
212 6001818 : entry.iter_mut().for_each(|b| b.inc(n));
213 2000906 : }
214 :
215 2000914 : should_allow_request
216 2000914 : }
217 :
218 : /// Clean the map. Simple strategy: remove all entries in a random shard.
219 : /// At worst, we'll double the effective max_rps during the cleanup.
220 : /// But that way deletion does not aquire mutex on each entry access.
221 976 : pub fn do_gc(&self) {
222 976 : info!(
223 0 : "cleaning up bucket rate limiter, current size = {}",
224 0 : self.map.len()
225 0 : );
226 976 : let n = self.map.shards().len();
227 976 : // this lock is ok as the periodic cycle of do_gc makes this very unlikely to collide
228 976 : // (impossible, infact, unless we have 2048 threads)
229 976 : let shard = self.rand.lock().unwrap().gen_range(0..n);
230 976 : self.map.shards()[shard].write().clear();
231 976 : }
232 : }
233 :
234 : #[cfg(test)]
235 : mod tests {
236 : use std::{hash::BuildHasherDefault, time::Duration};
237 :
238 : use rand::SeedableRng;
239 : use rustc_hash::FxHasher;
240 : use tokio::time;
241 :
242 : use super::{BucketRateLimiter, EndpointRateLimiter};
243 : use crate::{intern::EndpointIdInt, rate_limiter::RateBucketInfo, EndpointId};
244 :
245 : #[test]
246 2 : fn rate_bucket_rpi() {
247 2 : let rate_bucket = RateBucketInfo::new(50, Duration::from_secs(5));
248 2 : assert_eq!(rate_bucket.max_rpi, 50 * 5);
249 :
250 2 : let rate_bucket = RateBucketInfo::new(50, Duration::from_millis(500));
251 2 : assert_eq!(rate_bucket.max_rpi, 50 / 2);
252 2 : }
253 :
254 : #[test]
255 2 : fn rate_bucket_parse() {
256 2 : let rate_bucket: RateBucketInfo = "100@10s".parse().unwrap();
257 2 : assert_eq!(rate_bucket.interval, Duration::from_secs(10));
258 2 : assert_eq!(rate_bucket.max_rpi, 100 * 10);
259 2 : assert_eq!(rate_bucket.to_string(), "100@10s");
260 :
261 2 : let rate_bucket: RateBucketInfo = "100@1m".parse().unwrap();
262 2 : assert_eq!(rate_bucket.interval, Duration::from_secs(60));
263 2 : assert_eq!(rate_bucket.max_rpi, 100 * 60);
264 2 : assert_eq!(rate_bucket.to_string(), "100@1m");
265 2 : }
266 :
267 : #[test]
268 2 : fn default_rate_buckets() {
269 2 : let mut defaults = RateBucketInfo::DEFAULT_ENDPOINT_SET;
270 2 : RateBucketInfo::validate(&mut defaults[..]).unwrap();
271 2 : }
272 :
273 : #[test]
274 : #[should_panic = "invalid bucket RPS limits. 10@10s allows fewer requests per bucket than 300@1s (100 vs 300)"]
275 2 : fn rate_buckets_validate() {
276 2 : let mut rates: Vec<RateBucketInfo> = ["300@1s", "10@10s"]
277 2 : .into_iter()
278 4 : .map(|s| s.parse().unwrap())
279 2 : .collect();
280 2 : RateBucketInfo::validate(&mut rates).unwrap();
281 2 : }
282 :
283 : #[tokio::test]
284 2 : async fn test_rate_limits() {
285 2 : let mut rates: Vec<RateBucketInfo> = ["100@1s", "20@30s"]
286 2 : .into_iter()
287 4 : .map(|s| s.parse().unwrap())
288 2 : .collect();
289 2 : RateBucketInfo::validate(&mut rates).unwrap();
290 2 : let limiter = EndpointRateLimiter::new(rates);
291 2 :
292 2 : let endpoint = EndpointId::from("ep-my-endpoint-1234");
293 2 : let endpoint = EndpointIdInt::from(endpoint);
294 2 :
295 2 : time::pause();
296 2 :
297 202 : for _ in 0..100 {
298 200 : assert!(limiter.check(endpoint, 1));
299 2 : }
300 2 : // more connections fail
301 2 : assert!(!limiter.check(endpoint, 1));
302 2 :
303 2 : // fail even after 500ms as it's in the same bucket
304 2 : time::advance(time::Duration::from_millis(500)).await;
305 2 : assert!(!limiter.check(endpoint, 1));
306 2 :
307 2 : // after a full 1s, 100 requests are allowed again
308 2 : time::advance(time::Duration::from_millis(500)).await;
309 12 : for _ in 1..6 {
310 510 : for _ in 0..50 {
311 500 : assert!(limiter.check(endpoint, 2));
312 2 : }
313 10 : time::advance(time::Duration::from_millis(1000)).await;
314 2 : }
315 2 :
316 2 : // more connections after 600 will exceed the 20rps@30s limit
317 2 : assert!(!limiter.check(endpoint, 1));
318 2 :
319 2 : // will still fail before the 30 second limit
320 2 : time::advance(time::Duration::from_millis(30_000 - 6_000 - 1)).await;
321 2 : assert!(!limiter.check(endpoint, 1));
322 2 :
323 2 : // after the full 30 seconds, 100 requests are allowed again
324 2 : time::advance(time::Duration::from_millis(1)).await;
325 202 : for _ in 0..100 {
326 200 : assert!(limiter.check(endpoint, 1));
327 2 : }
328 2 : }
329 :
330 : #[tokio::test]
331 2 : async fn test_rate_limits_gc() {
332 2 : // fixed seeded random/hasher to ensure that the test is not flaky
333 2 : let rand = rand::rngs::StdRng::from_seed([1; 32]);
334 2 : let hasher = BuildHasherDefault::<FxHasher>::default();
335 2 :
336 2 : let limiter = BucketRateLimiter::new_with_rand_and_hasher(
337 2 : &RateBucketInfo::DEFAULT_ENDPOINT_SET,
338 2 : rand,
339 2 : hasher,
340 2 : );
341 2000002 : for i in 0..1_000_000 {
342 2000000 : limiter.check(i, 1);
343 2000000 : }
344 2 : assert!(limiter.map.len() < 150_000);
345 2 : }
346 : }
|
