Line data Source code
1 : use std::net::SocketAddr;
2 : use std::pin::pin;
3 : use std::sync::Arc;
4 : use std::time::Duration;
5 :
6 : use anyhow::bail;
7 : use futures::future::Either;
8 : use proxy::auth::backend::jwt::JwkCache;
9 : use proxy::auth::backend::{AuthRateLimiter, ConsoleRedirectBackend, MaybeOwned};
10 : use proxy::cancellation::{CancelMap, CancellationHandler};
11 : use proxy::config::{
12 : self, remote_storage_from_toml, AuthenticationConfig, CacheOptions, ComputeConfig, HttpConfig,
13 : ProjectInfoCacheOptions, ProxyConfig, ProxyProtocolV2,
14 : };
15 : use proxy::context::parquet::ParquetUploadArgs;
16 : use proxy::http::health_server::AppMetrics;
17 : use proxy::metrics::Metrics;
18 : use proxy::rate_limiter::{
19 : EndpointRateLimiter, LeakyBucketConfig, RateBucketInfo, WakeComputeRateLimiter,
20 : };
21 : use proxy::redis::cancellation_publisher::RedisPublisherClient;
22 : use proxy::redis::connection_with_credentials_provider::ConnectionWithCredentialsProvider;
23 : use proxy::redis::{elasticache, notifications};
24 : use proxy::scram::threadpool::ThreadPool;
25 : use proxy::serverless::cancel_set::CancelSet;
26 : use proxy::serverless::GlobalConnPoolOptions;
27 : use proxy::tls::client_config::compute_client_config_with_root_certs;
28 : use proxy::{auth, control_plane, http, serverless, usage_metrics};
29 : use remote_storage::RemoteStorageConfig;
30 : use tokio::net::TcpListener;
31 : use tokio::sync::Mutex;
32 : use tokio::task::JoinSet;
33 : use tokio_util::sync::CancellationToken;
34 : use tracing::{info, warn, Instrument};
35 : use utils::sentry_init::init_sentry;
36 : use utils::{project_build_tag, project_git_version};
37 :
38 : project_git_version!(GIT_VERSION);
39 : project_build_tag!(BUILD_TAG);
40 :
41 : use clap::{Parser, ValueEnum};
42 :
43 : #[global_allocator]
44 : static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
45 :
46 : #[derive(Clone, Debug, ValueEnum)]
47 : enum AuthBackendType {
48 : #[value(name("cplane-v1"), alias("control-plane"))]
49 : ControlPlaneV1,
50 :
51 : #[value(name("link"), alias("control-redirect"))]
52 : ConsoleRedirect,
53 :
54 : #[cfg(feature = "testing")]
55 : Postgres,
56 : }
57 :
58 : /// Neon proxy/router
59 : #[derive(Parser)]
60 : #[command(version = GIT_VERSION, about)]
61 : struct ProxyCliArgs {
62 : /// Name of the region this proxy is deployed in
63 1 : #[clap(long, default_value_t = String::new())]
64 0 : region: String,
65 : /// listen for incoming client connections on ip:port
66 : #[clap(short, long, default_value = "127.0.0.1:4432")]
67 0 : proxy: String,
68 1 : #[clap(value_enum, long, default_value_t = AuthBackendType::ConsoleRedirect)]
69 0 : auth_backend: AuthBackendType,
70 : /// listen for management callback connection on ip:port
71 : #[clap(short, long, default_value = "127.0.0.1:7000")]
72 0 : mgmt: String,
73 : /// listen for incoming http connections (metrics, etc) on ip:port
74 : #[clap(long, default_value = "127.0.0.1:7001")]
75 0 : http: String,
76 : /// listen for incoming wss connections on ip:port
77 : #[clap(long)]
78 : wss: Option<String>,
79 : /// redirect unauthenticated users to the given uri in case of console redirect auth
80 : #[clap(short, long, default_value = "http://localhost:3000/psql_session/")]
81 0 : uri: String,
82 : /// cloud API endpoint for authenticating users
83 : #[clap(
84 : short,
85 : long,
86 : default_value = "http://localhost:3000/authenticate_proxy_request/"
87 : )]
88 0 : auth_endpoint: String,
89 : /// JWT used to connect to control plane.
90 : #[clap(
91 : long,
92 : value_name = "JWT",
93 : default_value = "",
94 : env = "NEON_PROXY_TO_CONTROLPLANE_TOKEN"
95 : )]
96 0 : control_plane_token: Arc<str>,
97 : /// if this is not local proxy, this toggles whether we accept jwt or passwords for http
98 1 : #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
99 0 : is_auth_broker: bool,
100 : /// path to TLS key for client postgres connections
101 : ///
102 : /// tls-key and tls-cert are for backwards compatibility, we can put all certs in one dir
103 : #[clap(short = 'k', long, alias = "ssl-key")]
104 : tls_key: Option<String>,
105 : /// path to TLS cert for client postgres connections
106 : ///
107 : /// tls-key and tls-cert are for backwards compatibility, we can put all certs in one dir
108 : #[clap(short = 'c', long, alias = "ssl-cert")]
109 : tls_cert: Option<String>,
110 : /// Allow writing TLS session keys to the given file pointed to by the environment variable `SSLKEYLOGFILE`.
111 : #[clap(long, alias = "allow-ssl-keylogfile")]
112 0 : allow_tls_keylogfile: bool,
113 : /// path to directory with TLS certificates for client postgres connections
114 : #[clap(long)]
115 : certs_dir: Option<String>,
116 : /// timeout for the TLS handshake
117 : #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
118 0 : handshake_timeout: tokio::time::Duration,
119 : /// http endpoint to receive periodic metric updates
120 : #[clap(long)]
121 : metric_collection_endpoint: Option<String>,
122 : /// how often metrics should be sent to a collection endpoint
123 : #[clap(long)]
124 : metric_collection_interval: Option<String>,
125 : /// cache for `wake_compute` api method (use `size=0` to disable)
126 : #[clap(long, default_value = config::CacheOptions::CACHE_DEFAULT_OPTIONS)]
127 0 : wake_compute_cache: String,
128 : /// lock for `wake_compute` api method. example: "shards=32,permits=4,epoch=10m,timeout=1s". (use `permits=0` to disable).
129 : #[clap(long, default_value = config::ConcurrencyLockOptions::DEFAULT_OPTIONS_WAKE_COMPUTE_LOCK)]
130 0 : wake_compute_lock: String,
131 : /// lock for `connect_compute` api method. example: "shards=32,permits=4,epoch=10m,timeout=1s". (use `permits=0` to disable).
132 : #[clap(long, default_value = config::ConcurrencyLockOptions::DEFAULT_OPTIONS_CONNECT_COMPUTE_LOCK)]
133 0 : connect_compute_lock: String,
134 : #[clap(flatten)]
135 : sql_over_http: SqlOverHttpArgs,
136 : /// timeout for scram authentication protocol
137 : #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
138 0 : scram_protocol_timeout: tokio::time::Duration,
139 : /// size of the threadpool for password hashing
140 1 : #[clap(long, default_value_t = 4)]
141 0 : scram_thread_pool_size: u8,
142 : /// Endpoint rate limiter max number of requests per second.
143 : ///
144 : /// Provided in the form `<Requests Per Second>@<Bucket Duration Size>`.
145 : /// Can be given multiple times for different bucket sizes.
146 4 : #[clap(long, default_values_t = RateBucketInfo::DEFAULT_ENDPOINT_SET)]
147 1 : endpoint_rps_limit: Vec<RateBucketInfo>,
148 : /// Wake compute rate limiter max number of requests per second.
149 4 : #[clap(long, default_values_t = RateBucketInfo::DEFAULT_SET)]
150 1 : wake_compute_limit: Vec<RateBucketInfo>,
151 : /// Whether the auth rate limiter actually takes effect (for testing)
152 1 : #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
153 0 : auth_rate_limit_enabled: bool,
154 : /// Authentication rate limiter max number of hashes per second.
155 4 : #[clap(long, default_values_t = RateBucketInfo::DEFAULT_AUTH_SET)]
156 1 : auth_rate_limit: Vec<RateBucketInfo>,
157 : /// The IP subnet to use when considering whether two IP addresses are considered the same.
158 1 : #[clap(long, default_value_t = 64)]
159 0 : auth_rate_limit_ip_subnet: u8,
160 : /// Redis rate limiter max number of requests per second.
161 4 : #[clap(long, default_values_t = RateBucketInfo::DEFAULT_SET)]
162 1 : redis_rps_limit: Vec<RateBucketInfo>,
163 : /// cache for `allowed_ips` (use `size=0` to disable)
164 : #[clap(long, default_value = config::CacheOptions::CACHE_DEFAULT_OPTIONS)]
165 0 : allowed_ips_cache: String,
166 : /// cache for `role_secret` (use `size=0` to disable)
167 : #[clap(long, default_value = config::CacheOptions::CACHE_DEFAULT_OPTIONS)]
168 0 : role_secret_cache: String,
169 : /// redis url for notifications (if empty, redis_host:port will be used for both notifications and streaming connections)
170 : #[clap(long)]
171 : redis_notifications: Option<String>,
172 : /// what from the available authentications type to use for the regional redis we have. Supported are "irsa" and "plain".
173 : #[clap(long, default_value = "irsa")]
174 0 : redis_auth_type: String,
175 : /// redis host for streaming connections (might be different from the notifications host)
176 : #[clap(long)]
177 : redis_host: Option<String>,
178 : /// redis port for streaming connections (might be different from the notifications host)
179 : #[clap(long)]
180 : redis_port: Option<u16>,
181 : /// redis cluster name, used in aws elasticache
182 : #[clap(long)]
183 : redis_cluster_name: Option<String>,
184 : /// redis user_id, used in aws elasticache
185 : #[clap(long)]
186 : redis_user_id: Option<String>,
187 : /// aws region to retrieve credentials
188 1 : #[clap(long, default_value_t = String::new())]
189 0 : aws_region: String,
190 : /// cache for `project_info` (use `size=0` to disable)
191 : #[clap(long, default_value = config::ProjectInfoCacheOptions::CACHE_DEFAULT_OPTIONS)]
192 0 : project_info_cache: String,
193 : /// cache for all valid endpoints
194 : #[clap(long, default_value = config::EndpointCacheConfig::CACHE_DEFAULT_OPTIONS)]
195 0 : endpoint_cache_config: String,
196 : #[clap(flatten)]
197 : parquet_upload: ParquetUploadArgs,
198 :
199 : /// interval for backup metric collection
200 : #[clap(long, default_value = "10m", value_parser = humantime::parse_duration)]
201 0 : metric_backup_collection_interval: std::time::Duration,
202 : /// remote storage configuration for backup metric collection
203 : /// Encoded as toml (same format as pageservers), eg
204 : /// `{bucket_name='the-bucket',bucket_region='us-east-1',prefix_in_bucket='proxy',endpoint='http://minio:9000'}`
205 : #[clap(long, value_parser = remote_storage_from_toml)]
206 : metric_backup_collection_remote_storage: Option<RemoteStorageConfig>,
207 : /// chunk size for backup metric collection
208 : /// Size of each event is no more than 400 bytes, so 2**22 is about 200MB before the compression.
209 : #[clap(long, default_value = "4194304")]
210 0 : metric_backup_collection_chunk_size: usize,
211 : /// Whether to retry the connection to the compute node
212 : #[clap(long, default_value = config::RetryConfig::CONNECT_TO_COMPUTE_DEFAULT_VALUES)]
213 0 : connect_to_compute_retry: String,
214 : /// Whether to retry the wake_compute request
215 : #[clap(long, default_value = config::RetryConfig::WAKE_COMPUTE_DEFAULT_VALUES)]
216 0 : wake_compute_retry: String,
217 :
218 : /// Configure if this is a private access proxy for the POC: In that case the proxy will ignore the IP allowlist
219 1 : #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
220 0 : is_private_access_proxy: bool,
221 :
222 : /// Configure whether all incoming requests have a Proxy Protocol V2 packet.
223 : // TODO(conradludgate): switch default to rejected or required once we've updated all deployments
224 1 : #[clap(value_enum, long, default_value_t = ProxyProtocolV2::Supported)]
225 0 : proxy_protocol_v2: ProxyProtocolV2,
226 :
227 : /// Time the proxy waits for the webauth session to be confirmed by the control plane.
228 : // TODO: rename to `console_redirect_confirmation_timeout`.
229 : #[clap(long, default_value = "2m", value_parser = humantime::parse_duration)]
230 0 : webauth_confirmation_timeout: std::time::Duration,
231 : }
232 :
233 : #[derive(clap::Args, Clone, Copy, Debug)]
234 : struct SqlOverHttpArgs {
235 : /// timeout for http connection requests
236 : #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
237 0 : sql_over_http_timeout: tokio::time::Duration,
238 :
239 : /// Whether the SQL over http pool is opt-in
240 1 : #[clap(long, default_value_t = true, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
241 0 : sql_over_http_pool_opt_in: bool,
242 :
243 : /// How many connections to pool for each endpoint. Excess connections are discarded
244 1 : #[clap(long, default_value_t = 20)]
245 0 : sql_over_http_pool_max_conns_per_endpoint: usize,
246 :
247 : /// How many connections to pool for each endpoint. Excess connections are discarded
248 1 : #[clap(long, default_value_t = 20000)]
249 0 : sql_over_http_pool_max_total_conns: usize,
250 :
251 : /// How long pooled connections should remain idle for before closing
252 : #[clap(long, default_value = "5m", value_parser = humantime::parse_duration)]
253 0 : sql_over_http_idle_timeout: tokio::time::Duration,
254 :
255 : /// Duration each shard will wait on average before a GC sweep.
256 : /// A longer time will causes sweeps to take longer but will interfere less frequently.
257 : #[clap(long, default_value = "10m", value_parser = humantime::parse_duration)]
258 0 : sql_over_http_pool_gc_epoch: tokio::time::Duration,
259 :
260 : /// How many shards should the global pool have. Must be a power of two.
261 : /// More shards will introduce less contention for pool operations, but can
262 : /// increase memory used by the pool
263 1 : #[clap(long, default_value_t = 128)]
264 0 : sql_over_http_pool_shards: usize,
265 :
266 1 : #[clap(long, default_value_t = 10000)]
267 0 : sql_over_http_client_conn_threshold: u64,
268 :
269 1 : #[clap(long, default_value_t = 64)]
270 0 : sql_over_http_cancel_set_shards: usize,
271 :
272 1 : #[clap(long, default_value_t = 10 * 1024 * 1024)] // 10 MiB
273 0 : sql_over_http_max_request_size_bytes: usize,
274 :
275 1 : #[clap(long, default_value_t = 10 * 1024 * 1024)] // 10 MiB
276 0 : sql_over_http_max_response_size_bytes: usize,
277 : }
278 :
279 : #[tokio::main]
280 0 : async fn main() -> anyhow::Result<()> {
281 0 : let _logging_guard = proxy::logging::init().await?;
282 0 : let _panic_hook_guard = utils::logging::replace_panic_hook_with_tracing_panic_hook();
283 0 : let _sentry_guard = init_sentry(Some(GIT_VERSION.into()), &[]);
284 0 :
285 0 : // TODO: refactor these to use labels
286 0 : info!("Version: {GIT_VERSION}");
287 0 : info!("Build_tag: {BUILD_TAG}");
288 0 : let neon_metrics = ::metrics::NeonMetrics::new(::metrics::BuildInfo {
289 0 : revision: GIT_VERSION,
290 0 : build_tag: BUILD_TAG,
291 0 : });
292 0 :
293 0 : let jemalloc = match proxy::jemalloc::MetricRecorder::new() {
294 0 : Ok(t) => Some(t),
295 0 : Err(e) => {
296 0 : tracing::error!(error = ?e, "could not start jemalloc metrics loop");
297 0 : None
298 0 : }
299 0 : };
300 0 :
301 0 : let args = ProxyCliArgs::parse();
302 0 : let config = build_config(&args)?;
303 0 : let auth_backend = build_auth_backend(&args)?;
304 0 :
305 0 : match auth_backend {
306 0 : Either::Left(auth_backend) => info!("Authentication backend: {auth_backend}"),
307 0 : Either::Right(auth_backend) => info!("Authentication backend: {auth_backend:?}"),
308 0 : };
309 0 : info!("Using region: {}", args.aws_region);
310 0 :
311 0 : // TODO: untangle the config args
312 0 : let regional_redis_client = match (args.redis_auth_type.as_str(), &args.redis_notifications) {
313 0 : ("plain", redis_url) => match redis_url {
314 0 : None => {
315 0 : bail!("plain auth requires redis_notifications to be set");
316 0 : }
317 0 : Some(url) => Some(
318 0 : ConnectionWithCredentialsProvider::new_with_static_credentials(url.to_string()),
319 0 : ),
320 0 : },
321 0 : ("irsa", _) => match (&args.redis_host, args.redis_port) {
322 0 : (Some(host), Some(port)) => Some(
323 0 : ConnectionWithCredentialsProvider::new_with_credentials_provider(
324 0 : host.to_string(),
325 0 : port,
326 0 : elasticache::CredentialsProvider::new(
327 0 : args.aws_region,
328 0 : args.redis_cluster_name,
329 0 : args.redis_user_id,
330 0 : )
331 0 : .await,
332 0 : ),
333 0 : ),
334 0 : (None, None) => {
335 0 : warn!("irsa auth requires redis-host and redis-port to be set, continuing without regional_redis_client");
336 0 : None
337 0 : }
338 0 : _ => {
339 0 : bail!("redis-host and redis-port must be specified together");
340 0 : }
341 0 : },
342 0 : _ => {
343 0 : bail!("unknown auth type given");
344 0 : }
345 0 : };
346 0 :
347 0 : let redis_notifications_client = if let Some(url) = args.redis_notifications {
348 0 : Some(ConnectionWithCredentialsProvider::new_with_static_credentials(url.to_string()))
349 0 : } else {
350 0 : regional_redis_client.clone()
351 0 : };
352 0 :
353 0 : // Check that we can bind to address before further initialization
354 0 : let http_address: SocketAddr = args.http.parse()?;
355 0 : info!("Starting http on {http_address}");
356 0 : let http_listener = TcpListener::bind(http_address).await?.into_std()?;
357 0 :
358 0 : let mgmt_address: SocketAddr = args.mgmt.parse()?;
359 0 : info!("Starting mgmt on {mgmt_address}");
360 0 : let mgmt_listener = TcpListener::bind(mgmt_address).await?;
361 0 :
362 0 : let proxy_listener = if !args.is_auth_broker {
363 0 : let proxy_address: SocketAddr = args.proxy.parse()?;
364 0 : info!("Starting proxy on {proxy_address}");
365 0 :
366 0 : Some(TcpListener::bind(proxy_address).await?)
367 0 : } else {
368 0 : None
369 0 : };
370 0 :
371 0 : // TODO: rename the argument to something like serverless.
372 0 : // It now covers more than just websockets, it also covers SQL over HTTP.
373 0 : let serverless_listener = if let Some(serverless_address) = args.wss {
374 0 : let serverless_address: SocketAddr = serverless_address.parse()?;
375 0 : info!("Starting wss on {serverless_address}");
376 0 : Some(TcpListener::bind(serverless_address).await?)
377 0 : } else if args.is_auth_broker {
378 0 : bail!("wss arg must be present for auth-broker")
379 0 : } else {
380 0 : None
381 0 : };
382 0 :
383 0 : let cancellation_token = CancellationToken::new();
384 0 :
385 0 : let cancel_map = CancelMap::default();
386 0 :
387 0 : let redis_rps_limit = Vec::leak(args.redis_rps_limit.clone());
388 0 : RateBucketInfo::validate(redis_rps_limit)?;
389 0 :
390 0 : let redis_publisher = match ®ional_redis_client {
391 0 : Some(redis_publisher) => Some(Arc::new(Mutex::new(RedisPublisherClient::new(
392 0 : redis_publisher.clone(),
393 0 : args.region.clone(),
394 0 : redis_rps_limit,
395 0 : )?))),
396 0 : None => None,
397 0 : };
398 0 :
399 0 : let cancellation_handler = Arc::new(CancellationHandler::<
400 0 : Option<Arc<Mutex<RedisPublisherClient>>>,
401 0 : >::new(
402 0 : &config.connect_to_compute,
403 0 : cancel_map.clone(),
404 0 : redis_publisher,
405 0 : proxy::metrics::CancellationSource::FromClient,
406 0 : ));
407 0 :
408 0 : // bit of a hack - find the min rps and max rps supported and turn it into
409 0 : // leaky bucket config instead
410 0 : let max = args
411 0 : .endpoint_rps_limit
412 0 : .iter()
413 0 : .map(|x| x.rps())
414 0 : .max_by(f64::total_cmp)
415 0 : .unwrap_or(EndpointRateLimiter::DEFAULT.max);
416 0 : let rps = args
417 0 : .endpoint_rps_limit
418 0 : .iter()
419 0 : .map(|x| x.rps())
420 0 : .min_by(f64::total_cmp)
421 0 : .unwrap_or(EndpointRateLimiter::DEFAULT.rps);
422 0 : let endpoint_rate_limiter = Arc::new(EndpointRateLimiter::new_with_shards(
423 0 : LeakyBucketConfig { rps, max },
424 0 : 64,
425 0 : ));
426 0 :
427 0 : // client facing tasks. these will exit on error or on cancellation
428 0 : // cancellation returns Ok(())
429 0 : let mut client_tasks = JoinSet::new();
430 0 : match auth_backend {
431 0 : Either::Left(auth_backend) => {
432 0 : if let Some(proxy_listener) = proxy_listener {
433 0 : client_tasks.spawn(proxy::proxy::task_main(
434 0 : config,
435 0 : auth_backend,
436 0 : proxy_listener,
437 0 : cancellation_token.clone(),
438 0 : cancellation_handler.clone(),
439 0 : endpoint_rate_limiter.clone(),
440 0 : ));
441 0 : }
442 0 :
443 0 : if let Some(serverless_listener) = serverless_listener {
444 0 : client_tasks.spawn(serverless::task_main(
445 0 : config,
446 0 : auth_backend,
447 0 : serverless_listener,
448 0 : cancellation_token.clone(),
449 0 : cancellation_handler.clone(),
450 0 : endpoint_rate_limiter.clone(),
451 0 : ));
452 0 : }
453 0 : }
454 0 : Either::Right(auth_backend) => {
455 0 : if let Some(proxy_listener) = proxy_listener {
456 0 : client_tasks.spawn(proxy::console_redirect_proxy::task_main(
457 0 : config,
458 0 : auth_backend,
459 0 : proxy_listener,
460 0 : cancellation_token.clone(),
461 0 : cancellation_handler.clone(),
462 0 : ));
463 0 : }
464 0 : }
465 0 : }
466 0 :
467 0 : client_tasks.spawn(proxy::context::parquet::worker(
468 0 : cancellation_token.clone(),
469 0 : args.parquet_upload,
470 0 : ));
471 0 :
472 0 : // maintenance tasks. these never return unless there's an error
473 0 : let mut maintenance_tasks = JoinSet::new();
474 0 : maintenance_tasks.spawn(proxy::signals::handle(cancellation_token.clone(), || {}));
475 0 : maintenance_tasks.spawn(http::health_server::task_main(
476 0 : http_listener,
477 0 : AppMetrics {
478 0 : jemalloc,
479 0 : neon_metrics,
480 0 : proxy: proxy::metrics::Metrics::get(),
481 0 : },
482 0 : ));
483 0 : maintenance_tasks.spawn(control_plane::mgmt::task_main(mgmt_listener));
484 0 :
485 0 : if let Some(metrics_config) = &config.metric_collection {
486 0 : // TODO: Add gc regardles of the metric collection being enabled.
487 0 : maintenance_tasks.spawn(usage_metrics::task_main(metrics_config));
488 0 : }
489 0 :
490 0 : if let Either::Left(auth::Backend::ControlPlane(api, _)) = &auth_backend {
491 0 : if let proxy::control_plane::client::ControlPlaneClient::ProxyV1(api) = &**api {
492 0 : match (redis_notifications_client, regional_redis_client.clone()) {
493 0 : (None, None) => {}
494 0 : (client1, client2) => {
495 0 : let cache = api.caches.project_info.clone();
496 0 : if let Some(client) = client1 {
497 0 : maintenance_tasks.spawn(notifications::task_main(
498 0 : config,
499 0 : client,
500 0 : cache.clone(),
501 0 : cancel_map.clone(),
502 0 : args.region.clone(),
503 0 : ));
504 0 : }
505 0 : if let Some(client) = client2 {
506 0 : maintenance_tasks.spawn(notifications::task_main(
507 0 : config,
508 0 : client,
509 0 : cache.clone(),
510 0 : cancel_map.clone(),
511 0 : args.region.clone(),
512 0 : ));
513 0 : }
514 0 : maintenance_tasks.spawn(async move { cache.clone().gc_worker().await });
515 0 : }
516 0 : }
517 0 : if let Some(regional_redis_client) = regional_redis_client {
518 0 : let cache = api.caches.endpoints_cache.clone();
519 0 : let con = regional_redis_client;
520 0 : let span = tracing::info_span!("endpoints_cache");
521 0 : maintenance_tasks.spawn(
522 0 : async move { cache.do_read(con, cancellation_token.clone()).await }
523 0 : .instrument(span),
524 0 : );
525 0 : }
526 0 : }
527 0 : }
528 0 :
529 0 : let maintenance = loop {
530 0 : // get one complete task
531 0 : match futures::future::select(
532 0 : pin!(maintenance_tasks.join_next()),
533 0 : pin!(client_tasks.join_next()),
534 0 : )
535 0 : .await
536 0 : {
537 0 : // exit immediately on maintenance task completion
538 0 : Either::Left((Some(res), _)) => break proxy::error::flatten_err(res)?,
539 0 : // exit with error immediately if all maintenance tasks have ceased (should be caught by branch above)
540 0 : Either::Left((None, _)) => bail!("no maintenance tasks running. invalid state"),
541 0 : // exit immediately on client task error
542 0 : Either::Right((Some(res), _)) => proxy::error::flatten_err(res)?,
543 0 : // exit if all our client tasks have shutdown gracefully
544 0 : Either::Right((None, _)) => return Ok(()),
545 0 : }
546 0 : };
547 0 :
548 0 : // maintenance tasks return Infallible success values, this is an impossible value
549 0 : // so this match statically ensures that there are no possibilities for that value
550 0 : match maintenance {}
551 0 : }
552 :
553 : /// ProxyConfig is created at proxy startup, and lives forever.
554 0 : fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
555 0 : let thread_pool = ThreadPool::new(args.scram_thread_pool_size);
556 0 : Metrics::install(thread_pool.metrics.clone());
557 :
558 0 : let tls_config = match (&args.tls_key, &args.tls_cert) {
559 0 : (Some(key_path), Some(cert_path)) => Some(config::configure_tls(
560 0 : key_path,
561 0 : cert_path,
562 0 : args.certs_dir.as_ref(),
563 0 : args.allow_tls_keylogfile,
564 0 : )?),
565 0 : (None, None) => None,
566 0 : _ => bail!("either both or neither tls-key and tls-cert must be specified"),
567 : };
568 :
569 0 : let backup_metric_collection_config = config::MetricBackupCollectionConfig {
570 0 : interval: args.metric_backup_collection_interval,
571 0 : remote_storage_config: args.metric_backup_collection_remote_storage.clone(),
572 0 : chunk_size: args.metric_backup_collection_chunk_size,
573 0 : };
574 :
575 0 : let metric_collection = match (
576 0 : &args.metric_collection_endpoint,
577 0 : &args.metric_collection_interval,
578 : ) {
579 0 : (Some(endpoint), Some(interval)) => Some(config::MetricCollectionConfig {
580 0 : endpoint: endpoint.parse()?,
581 0 : interval: humantime::parse_duration(interval)?,
582 0 : backup_metric_collection_config,
583 : }),
584 0 : (None, None) => None,
585 0 : _ => bail!(
586 0 : "either both or neither metric-collection-endpoint \
587 0 : and metric-collection-interval must be specified"
588 0 : ),
589 : };
590 :
591 : let config::ConcurrencyLockOptions {
592 0 : shards,
593 0 : limiter,
594 0 : epoch,
595 0 : timeout,
596 0 : } = args.connect_compute_lock.parse()?;
597 0 : info!(
598 : ?limiter,
599 : shards,
600 : ?epoch,
601 0 : "Using NodeLocks (connect_compute)"
602 : );
603 0 : let connect_compute_locks = control_plane::locks::ApiLocks::new(
604 0 : "connect_compute_lock",
605 0 : limiter,
606 0 : shards,
607 0 : timeout,
608 0 : epoch,
609 0 : &Metrics::get().proxy.connect_compute_lock,
610 0 : )?;
611 :
612 0 : let http_config = HttpConfig {
613 0 : accept_websockets: !args.is_auth_broker,
614 0 : pool_options: GlobalConnPoolOptions {
615 0 : max_conns_per_endpoint: args.sql_over_http.sql_over_http_pool_max_conns_per_endpoint,
616 0 : gc_epoch: args.sql_over_http.sql_over_http_pool_gc_epoch,
617 0 : pool_shards: args.sql_over_http.sql_over_http_pool_shards,
618 0 : idle_timeout: args.sql_over_http.sql_over_http_idle_timeout,
619 0 : opt_in: args.sql_over_http.sql_over_http_pool_opt_in,
620 0 : max_total_conns: args.sql_over_http.sql_over_http_pool_max_total_conns,
621 0 : },
622 0 : cancel_set: CancelSet::new(args.sql_over_http.sql_over_http_cancel_set_shards),
623 0 : client_conn_threshold: args.sql_over_http.sql_over_http_client_conn_threshold,
624 0 : max_request_size_bytes: args.sql_over_http.sql_over_http_max_request_size_bytes,
625 0 : max_response_size_bytes: args.sql_over_http.sql_over_http_max_response_size_bytes,
626 0 : };
627 0 : let authentication_config = AuthenticationConfig {
628 0 : jwks_cache: JwkCache::default(),
629 0 : thread_pool,
630 0 : scram_protocol_timeout: args.scram_protocol_timeout,
631 0 : rate_limiter_enabled: args.auth_rate_limit_enabled,
632 0 : rate_limiter: AuthRateLimiter::new(args.auth_rate_limit.clone()),
633 0 : rate_limit_ip_subnet: args.auth_rate_limit_ip_subnet,
634 0 : ip_allowlist_check_enabled: !args.is_private_access_proxy,
635 0 : is_auth_broker: args.is_auth_broker,
636 0 : accept_jwts: args.is_auth_broker,
637 0 : console_redirect_confirmation_timeout: args.webauth_confirmation_timeout,
638 0 : };
639 :
640 0 : let compute_config = ComputeConfig {
641 0 : retry: config::RetryConfig::parse(&args.connect_to_compute_retry)?,
642 0 : tls: Arc::new(compute_client_config_with_root_certs()?),
643 0 : timeout: Duration::from_secs(2),
644 : };
645 :
646 0 : let config = ProxyConfig {
647 0 : tls_config,
648 0 : metric_collection,
649 0 : http_config,
650 0 : authentication_config,
651 0 : proxy_protocol_v2: args.proxy_protocol_v2,
652 0 : handshake_timeout: args.handshake_timeout,
653 0 : region: args.region.clone(),
654 0 : wake_compute_retry_config: config::RetryConfig::parse(&args.wake_compute_retry)?,
655 0 : connect_compute_locks,
656 0 : connect_to_compute: compute_config,
657 0 : };
658 0 :
659 0 : let config = Box::leak(Box::new(config));
660 0 :
661 0 : tokio::spawn(config.connect_compute_locks.garbage_collect_worker());
662 0 :
663 0 : Ok(config)
664 0 : }
665 :
666 : /// auth::Backend is created at proxy startup, and lives forever.
667 0 : fn build_auth_backend(
668 0 : args: &ProxyCliArgs,
669 0 : ) -> anyhow::Result<Either<&'static auth::Backend<'static, ()>, &'static ConsoleRedirectBackend>> {
670 0 : match &args.auth_backend {
671 : AuthBackendType::ControlPlaneV1 => {
672 0 : let wake_compute_cache_config: CacheOptions = args.wake_compute_cache.parse()?;
673 0 : let project_info_cache_config: ProjectInfoCacheOptions =
674 0 : args.project_info_cache.parse()?;
675 0 : let endpoint_cache_config: config::EndpointCacheConfig =
676 0 : args.endpoint_cache_config.parse()?;
677 :
678 0 : info!("Using NodeInfoCache (wake_compute) with options={wake_compute_cache_config:?}");
679 0 : info!(
680 0 : "Using AllowedIpsCache (wake_compute) with options={project_info_cache_config:?}"
681 : );
682 0 : info!("Using EndpointCacheConfig with options={endpoint_cache_config:?}");
683 0 : let caches = Box::leak(Box::new(control_plane::caches::ApiCaches::new(
684 0 : wake_compute_cache_config,
685 0 : project_info_cache_config,
686 0 : endpoint_cache_config,
687 0 : )));
688 :
689 : let config::ConcurrencyLockOptions {
690 0 : shards,
691 0 : limiter,
692 0 : epoch,
693 0 : timeout,
694 0 : } = args.wake_compute_lock.parse()?;
695 0 : info!(?limiter, shards, ?epoch, "Using NodeLocks (wake_compute)");
696 0 : let locks = Box::leak(Box::new(control_plane::locks::ApiLocks::new(
697 0 : "wake_compute_lock",
698 0 : limiter,
699 0 : shards,
700 0 : timeout,
701 0 : epoch,
702 0 : &Metrics::get().wake_compute_lock,
703 0 : )?));
704 0 : tokio::spawn(locks.garbage_collect_worker());
705 :
706 0 : let url: proxy::url::ApiUrl = args.auth_endpoint.parse()?;
707 :
708 0 : let endpoint = http::Endpoint::new(url, http::new_client());
709 0 :
710 0 : let mut wake_compute_rps_limit = args.wake_compute_limit.clone();
711 0 : RateBucketInfo::validate(&mut wake_compute_rps_limit)?;
712 0 : let wake_compute_endpoint_rate_limiter =
713 0 : Arc::new(WakeComputeRateLimiter::new(wake_compute_rps_limit));
714 0 :
715 0 : let api = control_plane::client::cplane_proxy_v1::NeonControlPlaneClient::new(
716 0 : endpoint,
717 0 : args.control_plane_token.clone(),
718 0 : caches,
719 0 : locks,
720 0 : wake_compute_endpoint_rate_limiter,
721 0 : );
722 0 :
723 0 : let api = control_plane::client::ControlPlaneClient::ProxyV1(api);
724 0 : let auth_backend = auth::Backend::ControlPlane(MaybeOwned::Owned(api), ());
725 0 : let config = Box::leak(Box::new(auth_backend));
726 0 :
727 0 : Ok(Either::Left(config))
728 : }
729 :
730 : #[cfg(feature = "testing")]
731 : AuthBackendType::Postgres => {
732 0 : let url = args.auth_endpoint.parse()?;
733 0 : let api = control_plane::client::mock::MockControlPlane::new(
734 0 : url,
735 0 : !args.is_private_access_proxy,
736 0 : );
737 0 : let api = control_plane::client::ControlPlaneClient::PostgresMock(api);
738 0 :
739 0 : let auth_backend = auth::Backend::ControlPlane(MaybeOwned::Owned(api), ());
740 0 :
741 0 : let config = Box::leak(Box::new(auth_backend));
742 0 :
743 0 : Ok(Either::Left(config))
744 : }
745 :
746 : AuthBackendType::ConsoleRedirect => {
747 0 : let url = args.uri.parse()?;
748 0 : let backend = ConsoleRedirectBackend::new(url);
749 0 :
750 0 : let config = Box::leak(Box::new(backend));
751 0 :
752 0 : Ok(Either::Right(config))
753 : }
754 : }
755 0 : }
756 :
757 : #[cfg(test)]
758 : mod tests {
759 : use std::time::Duration;
760 :
761 : use clap::Parser;
762 : use proxy::rate_limiter::RateBucketInfo;
763 :
764 : #[test]
765 1 : fn parse_endpoint_rps_limit() {
766 1 : let config = super::ProxyCliArgs::parse_from([
767 1 : "proxy",
768 1 : "--endpoint-rps-limit",
769 1 : "100@1s",
770 1 : "--endpoint-rps-limit",
771 1 : "20@30s",
772 1 : ]);
773 1 :
774 1 : assert_eq!(
775 1 : config.endpoint_rps_limit,
776 1 : vec![
777 1 : RateBucketInfo::new(100, Duration::from_secs(1)),
778 1 : RateBucketInfo::new(20, Duration::from_secs(30)),
779 1 : ]
780 1 : );
781 1 : }
782 : }
|
