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::{handle_cancel_messages, 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::connection_with_credentials_provider::ConnectionWithCredentialsProvider;
22 : use proxy::redis::kv_ops::RedisKVClient;
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::task::JoinSet;
32 : use tokio_util::sync::CancellationToken;
33 : use tracing::{info, warn, Instrument};
34 : use utils::sentry_init::init_sentry;
35 : use utils::{project_build_tag, project_git_version};
36 :
37 : project_git_version!(GIT_VERSION);
38 : project_build_tag!(BUILD_TAG);
39 :
40 : use clap::{Parser, ValueEnum};
41 :
42 : #[global_allocator]
43 : static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
44 :
45 : #[derive(Clone, Debug, ValueEnum)]
46 : enum AuthBackendType {
47 : #[value(name("cplane-v1"), alias("control-plane"))]
48 : ControlPlaneV1,
49 :
50 : #[value(name("link"), alias("control-redirect"))]
51 : ConsoleRedirect,
52 :
53 : #[cfg(feature = "testing")]
54 : Postgres,
55 : }
56 :
57 : /// Neon proxy/router
58 : #[derive(Parser)]
59 : #[command(version = GIT_VERSION, about)]
60 : struct ProxyCliArgs {
61 : /// Name of the region this proxy is deployed in
62 1 : #[clap(long, default_value_t = String::new())]
63 0 : region: String,
64 : /// listen for incoming client connections on ip:port
65 : #[clap(short, long, default_value = "127.0.0.1:4432")]
66 0 : proxy: String,
67 1 : #[clap(value_enum, long, default_value_t = AuthBackendType::ConsoleRedirect)]
68 0 : auth_backend: AuthBackendType,
69 : /// listen for management callback connection on ip:port
70 : #[clap(short, long, default_value = "127.0.0.1:7000")]
71 0 : mgmt: String,
72 : /// listen for incoming http connections (metrics, etc) on ip:port
73 : #[clap(long, default_value = "127.0.0.1:7001")]
74 0 : http: String,
75 : /// listen for incoming wss connections on ip:port
76 : #[clap(long)]
77 : wss: Option<String>,
78 : /// redirect unauthenticated users to the given uri in case of console redirect auth
79 : #[clap(short, long, default_value = "http://localhost:3000/psql_session/")]
80 0 : uri: String,
81 : /// cloud API endpoint for authenticating users
82 : #[clap(
83 : short,
84 : long,
85 : default_value = "http://localhost:3000/authenticate_proxy_request/"
86 : )]
87 0 : auth_endpoint: String,
88 : /// JWT used to connect to control plane.
89 : #[clap(
90 : long,
91 : value_name = "JWT",
92 : default_value = "",
93 : env = "NEON_PROXY_TO_CONTROLPLANE_TOKEN"
94 : )]
95 0 : control_plane_token: Arc<str>,
96 : /// if this is not local proxy, this toggles whether we accept jwt or passwords for http
97 1 : #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
98 0 : is_auth_broker: bool,
99 : /// path to TLS key for client postgres connections
100 : ///
101 : /// tls-key and tls-cert are for backwards compatibility, we can put all certs in one dir
102 : #[clap(short = 'k', long, alias = "ssl-key")]
103 : tls_key: Option<String>,
104 : /// path to TLS cert for client postgres connections
105 : ///
106 : /// tls-key and tls-cert are for backwards compatibility, we can put all certs in one dir
107 : #[clap(short = 'c', long, alias = "ssl-cert")]
108 : tls_cert: Option<String>,
109 : /// Allow writing TLS session keys to the given file pointed to by the environment variable `SSLKEYLOGFILE`.
110 : #[clap(long, alias = "allow-ssl-keylogfile")]
111 0 : allow_tls_keylogfile: bool,
112 : /// path to directory with TLS certificates for client postgres connections
113 : #[clap(long)]
114 : certs_dir: Option<String>,
115 : /// timeout for the TLS handshake
116 : #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
117 0 : handshake_timeout: tokio::time::Duration,
118 : /// http endpoint to receive periodic metric updates
119 : #[clap(long)]
120 : metric_collection_endpoint: Option<String>,
121 : /// how often metrics should be sent to a collection endpoint
122 : #[clap(long)]
123 : metric_collection_interval: Option<String>,
124 : /// cache for `wake_compute` api method (use `size=0` to disable)
125 : #[clap(long, default_value = config::CacheOptions::CACHE_DEFAULT_OPTIONS)]
126 0 : wake_compute_cache: String,
127 : /// lock for `wake_compute` api method. example: "shards=32,permits=4,epoch=10m,timeout=1s". (use `permits=0` to disable).
128 : #[clap(long, default_value = config::ConcurrencyLockOptions::DEFAULT_OPTIONS_WAKE_COMPUTE_LOCK)]
129 0 : wake_compute_lock: String,
130 : /// lock for `connect_compute` api method. example: "shards=32,permits=4,epoch=10m,timeout=1s". (use `permits=0` to disable).
131 : #[clap(long, default_value = config::ConcurrencyLockOptions::DEFAULT_OPTIONS_CONNECT_COMPUTE_LOCK)]
132 0 : connect_compute_lock: String,
133 : #[clap(flatten)]
134 : sql_over_http: SqlOverHttpArgs,
135 : /// timeout for scram authentication protocol
136 : #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
137 0 : scram_protocol_timeout: tokio::time::Duration,
138 : /// size of the threadpool for password hashing
139 1 : #[clap(long, default_value_t = 4)]
140 0 : scram_thread_pool_size: u8,
141 : /// Endpoint rate limiter max number of requests per second.
142 : ///
143 : /// Provided in the form `<Requests Per Second>@<Bucket Duration Size>`.
144 : /// Can be given multiple times for different bucket sizes.
145 4 : #[clap(long, default_values_t = RateBucketInfo::DEFAULT_ENDPOINT_SET)]
146 1 : endpoint_rps_limit: Vec<RateBucketInfo>,
147 : /// Wake compute rate limiter max number of requests per second.
148 4 : #[clap(long, default_values_t = RateBucketInfo::DEFAULT_SET)]
149 1 : wake_compute_limit: Vec<RateBucketInfo>,
150 : /// Whether the auth rate limiter actually takes effect (for testing)
151 1 : #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
152 0 : auth_rate_limit_enabled: bool,
153 : /// Authentication rate limiter max number of hashes per second.
154 4 : #[clap(long, default_values_t = RateBucketInfo::DEFAULT_AUTH_SET)]
155 1 : auth_rate_limit: Vec<RateBucketInfo>,
156 : /// The IP subnet to use when considering whether two IP addresses are considered the same.
157 1 : #[clap(long, default_value_t = 64)]
158 0 : auth_rate_limit_ip_subnet: u8,
159 : /// Redis rate limiter max number of requests per second.
160 3 : #[clap(long, default_values_t = RateBucketInfo::DEFAULT_REDIS_SET)]
161 1 : redis_rps_limit: Vec<RateBucketInfo>,
162 : /// Cancellation channel size (max queue size for redis kv client)
163 : #[clap(long, default_value = "1024")]
164 0 : cancellation_ch_size: usize,
165 : /// cache for `allowed_ips` (use `size=0` to disable)
166 : #[clap(long, default_value = config::CacheOptions::CACHE_DEFAULT_OPTIONS)]
167 0 : allowed_ips_cache: String,
168 : /// cache for `role_secret` (use `size=0` to disable)
169 : #[clap(long, default_value = config::CacheOptions::CACHE_DEFAULT_OPTIONS)]
170 0 : role_secret_cache: String,
171 : /// redis url for notifications (if empty, redis_host:port will be used for both notifications and streaming connections)
172 : #[clap(long)]
173 : redis_notifications: Option<String>,
174 : /// what from the available authentications type to use for the regional redis we have. Supported are "irsa" and "plain".
175 : #[clap(long, default_value = "irsa")]
176 0 : redis_auth_type: String,
177 : /// redis host for streaming connections (might be different from the notifications host)
178 : #[clap(long)]
179 : redis_host: Option<String>,
180 : /// redis port for streaming connections (might be different from the notifications host)
181 : #[clap(long)]
182 : redis_port: Option<u16>,
183 : /// redis cluster name, used in aws elasticache
184 : #[clap(long)]
185 : redis_cluster_name: Option<String>,
186 : /// redis user_id, used in aws elasticache
187 : #[clap(long)]
188 : redis_user_id: Option<String>,
189 : /// aws region to retrieve credentials
190 1 : #[clap(long, default_value_t = String::new())]
191 0 : aws_region: String,
192 : /// cache for `project_info` (use `size=0` to disable)
193 : #[clap(long, default_value = config::ProjectInfoCacheOptions::CACHE_DEFAULT_OPTIONS)]
194 0 : project_info_cache: String,
195 : /// cache for all valid endpoints
196 : #[clap(long, default_value = config::EndpointCacheConfig::CACHE_DEFAULT_OPTIONS)]
197 0 : endpoint_cache_config: String,
198 : #[clap(flatten)]
199 : parquet_upload: ParquetUploadArgs,
200 :
201 : /// interval for backup metric collection
202 : #[clap(long, default_value = "10m", value_parser = humantime::parse_duration)]
203 0 : metric_backup_collection_interval: std::time::Duration,
204 : /// remote storage configuration for backup metric collection
205 : /// Encoded as toml (same format as pageservers), eg
206 : /// `{bucket_name='the-bucket',bucket_region='us-east-1',prefix_in_bucket='proxy',endpoint='http://minio:9000'}`
207 : #[clap(long, value_parser = remote_storage_from_toml)]
208 : metric_backup_collection_remote_storage: Option<RemoteStorageConfig>,
209 : /// chunk size for backup metric collection
210 : /// Size of each event is no more than 400 bytes, so 2**22 is about 200MB before the compression.
211 : #[clap(long, default_value = "4194304")]
212 0 : metric_backup_collection_chunk_size: usize,
213 : /// Whether to retry the connection to the compute node
214 : #[clap(long, default_value = config::RetryConfig::CONNECT_TO_COMPUTE_DEFAULT_VALUES)]
215 0 : connect_to_compute_retry: String,
216 : /// Whether to retry the wake_compute request
217 : #[clap(long, default_value = config::RetryConfig::WAKE_COMPUTE_DEFAULT_VALUES)]
218 0 : wake_compute_retry: String,
219 :
220 : /// Configure if this is a private access proxy for the POC: In that case the proxy will ignore the IP allowlist
221 1 : #[clap(long, default_value_t = false, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
222 0 : is_private_access_proxy: bool,
223 :
224 : /// Configure whether all incoming requests have a Proxy Protocol V2 packet.
225 : // TODO(conradludgate): switch default to rejected or required once we've updated all deployments
226 1 : #[clap(value_enum, long, default_value_t = ProxyProtocolV2::Supported)]
227 0 : proxy_protocol_v2: ProxyProtocolV2,
228 :
229 : /// Time the proxy waits for the webauth session to be confirmed by the control plane.
230 : // TODO: rename to `console_redirect_confirmation_timeout`.
231 : #[clap(long, default_value = "2m", value_parser = humantime::parse_duration)]
232 0 : webauth_confirmation_timeout: std::time::Duration,
233 : }
234 :
235 : #[derive(clap::Args, Clone, Copy, Debug)]
236 : struct SqlOverHttpArgs {
237 : /// timeout for http connection requests
238 : #[clap(long, default_value = "15s", value_parser = humantime::parse_duration)]
239 0 : sql_over_http_timeout: tokio::time::Duration,
240 :
241 : /// Whether the SQL over http pool is opt-in
242 1 : #[clap(long, default_value_t = true, value_parser = clap::builder::BoolishValueParser::new(), action = clap::ArgAction::Set)]
243 0 : sql_over_http_pool_opt_in: bool,
244 :
245 : /// How many connections to pool for each endpoint. Excess connections are discarded
246 1 : #[clap(long, default_value_t = 20)]
247 0 : sql_over_http_pool_max_conns_per_endpoint: usize,
248 :
249 : /// How many connections to pool for each endpoint. Excess connections are discarded
250 1 : #[clap(long, default_value_t = 20000)]
251 0 : sql_over_http_pool_max_total_conns: usize,
252 :
253 : /// How long pooled connections should remain idle for before closing
254 : #[clap(long, default_value = "5m", value_parser = humantime::parse_duration)]
255 0 : sql_over_http_idle_timeout: tokio::time::Duration,
256 :
257 : /// Duration each shard will wait on average before a GC sweep.
258 : /// A longer time will causes sweeps to take longer but will interfere less frequently.
259 : #[clap(long, default_value = "10m", value_parser = humantime::parse_duration)]
260 0 : sql_over_http_pool_gc_epoch: tokio::time::Duration,
261 :
262 : /// How many shards should the global pool have. Must be a power of two.
263 : /// More shards will introduce less contention for pool operations, but can
264 : /// increase memory used by the pool
265 1 : #[clap(long, default_value_t = 128)]
266 0 : sql_over_http_pool_shards: usize,
267 :
268 1 : #[clap(long, default_value_t = 10000)]
269 0 : sql_over_http_client_conn_threshold: u64,
270 :
271 1 : #[clap(long, default_value_t = 64)]
272 0 : sql_over_http_cancel_set_shards: usize,
273 :
274 1 : #[clap(long, default_value_t = 10 * 1024 * 1024)] // 10 MiB
275 0 : sql_over_http_max_request_size_bytes: usize,
276 :
277 1 : #[clap(long, default_value_t = 10 * 1024 * 1024)] // 10 MiB
278 0 : sql_over_http_max_response_size_bytes: usize,
279 : }
280 :
281 : #[tokio::main]
282 0 : async fn main() -> anyhow::Result<()> {
283 0 : let _logging_guard = proxy::logging::init().await?;
284 0 : let _panic_hook_guard = utils::logging::replace_panic_hook_with_tracing_panic_hook();
285 0 : let _sentry_guard = init_sentry(Some(GIT_VERSION.into()), &[]);
286 0 :
287 0 : // TODO: refactor these to use labels
288 0 : info!("Version: {GIT_VERSION}");
289 0 : info!("Build_tag: {BUILD_TAG}");
290 0 : let neon_metrics = ::metrics::NeonMetrics::new(::metrics::BuildInfo {
291 0 : revision: GIT_VERSION,
292 0 : build_tag: BUILD_TAG,
293 0 : });
294 0 :
295 0 : let jemalloc = match proxy::jemalloc::MetricRecorder::new() {
296 0 : Ok(t) => Some(t),
297 0 : Err(e) => {
298 0 : tracing::error!(error = ?e, "could not start jemalloc metrics loop");
299 0 : None
300 0 : }
301 0 : };
302 0 :
303 0 : let args = ProxyCliArgs::parse();
304 0 : let config = build_config(&args)?;
305 0 : let auth_backend = build_auth_backend(&args)?;
306 0 :
307 0 : match auth_backend {
308 0 : Either::Left(auth_backend) => info!("Authentication backend: {auth_backend}"),
309 0 : Either::Right(auth_backend) => info!("Authentication backend: {auth_backend:?}"),
310 0 : };
311 0 : info!("Using region: {}", args.aws_region);
312 0 :
313 0 : // TODO: untangle the config args
314 0 : let regional_redis_client = match (args.redis_auth_type.as_str(), &args.redis_notifications) {
315 0 : ("plain", redis_url) => match redis_url {
316 0 : None => {
317 0 : bail!("plain auth requires redis_notifications to be set");
318 0 : }
319 0 : Some(url) => Some(
320 0 : ConnectionWithCredentialsProvider::new_with_static_credentials(url.to_string()),
321 0 : ),
322 0 : },
323 0 : ("irsa", _) => match (&args.redis_host, args.redis_port) {
324 0 : (Some(host), Some(port)) => Some(
325 0 : ConnectionWithCredentialsProvider::new_with_credentials_provider(
326 0 : host.to_string(),
327 0 : port,
328 0 : elasticache::CredentialsProvider::new(
329 0 : args.aws_region,
330 0 : args.redis_cluster_name,
331 0 : args.redis_user_id,
332 0 : )
333 0 : .await,
334 0 : ),
335 0 : ),
336 0 : (None, None) => {
337 0 : warn!("irsa auth requires redis-host and redis-port to be set, continuing without regional_redis_client");
338 0 : None
339 0 : }
340 0 : _ => {
341 0 : bail!("redis-host and redis-port must be specified together");
342 0 : }
343 0 : },
344 0 : _ => {
345 0 : bail!("unknown auth type given");
346 0 : }
347 0 : };
348 0 :
349 0 : let redis_notifications_client = if let Some(url) = args.redis_notifications {
350 0 : Some(ConnectionWithCredentialsProvider::new_with_static_credentials(url.to_string()))
351 0 : } else {
352 0 : regional_redis_client.clone()
353 0 : };
354 0 :
355 0 : // Check that we can bind to address before further initialization
356 0 : let http_address: SocketAddr = args.http.parse()?;
357 0 : info!("Starting http on {http_address}");
358 0 : let http_listener = TcpListener::bind(http_address).await?.into_std()?;
359 0 :
360 0 : let mgmt_address: SocketAddr = args.mgmt.parse()?;
361 0 : info!("Starting mgmt on {mgmt_address}");
362 0 : let mgmt_listener = TcpListener::bind(mgmt_address).await?;
363 0 :
364 0 : let proxy_listener = if !args.is_auth_broker {
365 0 : let proxy_address: SocketAddr = args.proxy.parse()?;
366 0 : info!("Starting proxy on {proxy_address}");
367 0 :
368 0 : Some(TcpListener::bind(proxy_address).await?)
369 0 : } else {
370 0 : None
371 0 : };
372 0 :
373 0 : // TODO: rename the argument to something like serverless.
374 0 : // It now covers more than just websockets, it also covers SQL over HTTP.
375 0 : let serverless_listener = if let Some(serverless_address) = args.wss {
376 0 : let serverless_address: SocketAddr = serverless_address.parse()?;
377 0 : info!("Starting wss on {serverless_address}");
378 0 : Some(TcpListener::bind(serverless_address).await?)
379 0 : } else if args.is_auth_broker {
380 0 : bail!("wss arg must be present for auth-broker")
381 0 : } else {
382 0 : None
383 0 : };
384 0 :
385 0 : let cancellation_token = CancellationToken::new();
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_kv_client = regional_redis_client
391 0 : .as_ref()
392 0 : .map(|redis_publisher| RedisKVClient::new(redis_publisher.clone(), redis_rps_limit));
393 0 :
394 0 : // channel size should be higher than redis client limit to avoid blocking
395 0 : let cancel_ch_size = args.cancellation_ch_size;
396 0 : let (tx_cancel, rx_cancel) = tokio::sync::mpsc::channel(cancel_ch_size);
397 0 : let cancellation_handler = Arc::new(CancellationHandler::new(
398 0 : &config.connect_to_compute,
399 0 : Some(tx_cancel),
400 0 : ));
401 0 :
402 0 : // bit of a hack - find the min rps and max rps supported and turn it into
403 0 : // leaky bucket config instead
404 0 : let max = args
405 0 : .endpoint_rps_limit
406 0 : .iter()
407 0 : .map(|x| x.rps())
408 0 : .max_by(f64::total_cmp)
409 0 : .unwrap_or(EndpointRateLimiter::DEFAULT.max);
410 0 : let rps = args
411 0 : .endpoint_rps_limit
412 0 : .iter()
413 0 : .map(|x| x.rps())
414 0 : .min_by(f64::total_cmp)
415 0 : .unwrap_or(EndpointRateLimiter::DEFAULT.rps);
416 0 : let endpoint_rate_limiter = Arc::new(EndpointRateLimiter::new_with_shards(
417 0 : LeakyBucketConfig { rps, max },
418 0 : 64,
419 0 : ));
420 0 :
421 0 : // client facing tasks. these will exit on error or on cancellation
422 0 : // cancellation returns Ok(())
423 0 : let mut client_tasks = JoinSet::new();
424 0 : match auth_backend {
425 0 : Either::Left(auth_backend) => {
426 0 : if let Some(proxy_listener) = proxy_listener {
427 0 : client_tasks.spawn(proxy::proxy::task_main(
428 0 : config,
429 0 : auth_backend,
430 0 : proxy_listener,
431 0 : cancellation_token.clone(),
432 0 : cancellation_handler.clone(),
433 0 : endpoint_rate_limiter.clone(),
434 0 : ));
435 0 : }
436 0 :
437 0 : if let Some(serverless_listener) = serverless_listener {
438 0 : client_tasks.spawn(serverless::task_main(
439 0 : config,
440 0 : auth_backend,
441 0 : serverless_listener,
442 0 : cancellation_token.clone(),
443 0 : cancellation_handler.clone(),
444 0 : endpoint_rate_limiter.clone(),
445 0 : ));
446 0 : }
447 0 : }
448 0 : Either::Right(auth_backend) => {
449 0 : if let Some(proxy_listener) = proxy_listener {
450 0 : client_tasks.spawn(proxy::console_redirect_proxy::task_main(
451 0 : config,
452 0 : auth_backend,
453 0 : proxy_listener,
454 0 : cancellation_token.clone(),
455 0 : cancellation_handler.clone(),
456 0 : ));
457 0 : }
458 0 : }
459 0 : }
460 0 :
461 0 : client_tasks.spawn(proxy::context::parquet::worker(
462 0 : cancellation_token.clone(),
463 0 : args.parquet_upload,
464 0 : ));
465 0 :
466 0 : // maintenance tasks. these never return unless there's an error
467 0 : let mut maintenance_tasks = JoinSet::new();
468 0 : maintenance_tasks.spawn(proxy::signals::handle(cancellation_token.clone(), || {}));
469 0 : maintenance_tasks.spawn(http::health_server::task_main(
470 0 : http_listener,
471 0 : AppMetrics {
472 0 : jemalloc,
473 0 : neon_metrics,
474 0 : proxy: proxy::metrics::Metrics::get(),
475 0 : },
476 0 : ));
477 0 : maintenance_tasks.spawn(control_plane::mgmt::task_main(mgmt_listener));
478 0 :
479 0 : if let Some(metrics_config) = &config.metric_collection {
480 0 : // TODO: Add gc regardles of the metric collection being enabled.
481 0 : maintenance_tasks.spawn(usage_metrics::task_main(metrics_config));
482 0 : }
483 0 :
484 0 : if let Either::Left(auth::Backend::ControlPlane(api, _)) = &auth_backend {
485 0 : if let proxy::control_plane::client::ControlPlaneClient::ProxyV1(api) = &**api {
486 0 : match (redis_notifications_client, regional_redis_client.clone()) {
487 0 : (None, None) => {}
488 0 : (client1, client2) => {
489 0 : let cache = api.caches.project_info.clone();
490 0 : if let Some(client) = client1 {
491 0 : maintenance_tasks.spawn(notifications::task_main(
492 0 : client,
493 0 : cache.clone(),
494 0 : args.region.clone(),
495 0 : ));
496 0 : }
497 0 : if let Some(client) = client2 {
498 0 : maintenance_tasks.spawn(notifications::task_main(
499 0 : client,
500 0 : cache.clone(),
501 0 : args.region.clone(),
502 0 : ));
503 0 : }
504 0 : maintenance_tasks.spawn(async move { cache.clone().gc_worker().await });
505 0 : }
506 0 : }
507 0 :
508 0 : if let Some(mut redis_kv_client) = redis_kv_client {
509 0 : maintenance_tasks.spawn(async move {
510 0 : redis_kv_client.try_connect().await?;
511 0 : handle_cancel_messages(&mut redis_kv_client, rx_cancel).await
512 0 : });
513 0 : }
514 0 :
515 0 : if let Some(regional_redis_client) = regional_redis_client {
516 0 : let cache = api.caches.endpoints_cache.clone();
517 0 : let con = regional_redis_client;
518 0 : let span = tracing::info_span!("endpoints_cache");
519 0 : maintenance_tasks.spawn(
520 0 : async move { cache.do_read(con, cancellation_token.clone()).await }
521 0 : .instrument(span),
522 0 : );
523 0 : }
524 0 : }
525 0 : }
526 0 :
527 0 : let maintenance = loop {
528 0 : // get one complete task
529 0 : match futures::future::select(
530 0 : pin!(maintenance_tasks.join_next()),
531 0 : pin!(client_tasks.join_next()),
532 0 : )
533 0 : .await
534 0 : {
535 0 : // exit immediately on maintenance task completion
536 0 : Either::Left((Some(res), _)) => break proxy::error::flatten_err(res)?,
537 0 : // exit with error immediately if all maintenance tasks have ceased (should be caught by branch above)
538 0 : Either::Left((None, _)) => bail!("no maintenance tasks running. invalid state"),
539 0 : // exit immediately on client task error
540 0 : Either::Right((Some(res), _)) => proxy::error::flatten_err(res)?,
541 0 : // exit if all our client tasks have shutdown gracefully
542 0 : Either::Right((None, _)) => return Ok(()),
543 0 : }
544 0 : };
545 0 :
546 0 : // maintenance tasks return Infallible success values, this is an impossible value
547 0 : // so this match statically ensures that there are no possibilities for that value
548 0 : match maintenance {}
549 0 : }
550 :
551 : /// ProxyConfig is created at proxy startup, and lives forever.
552 0 : fn build_config(args: &ProxyCliArgs) -> anyhow::Result<&'static ProxyConfig> {
553 0 : let thread_pool = ThreadPool::new(args.scram_thread_pool_size);
554 0 : Metrics::install(thread_pool.metrics.clone());
555 :
556 0 : let tls_config = match (&args.tls_key, &args.tls_cert) {
557 0 : (Some(key_path), Some(cert_path)) => Some(config::configure_tls(
558 0 : key_path,
559 0 : cert_path,
560 0 : args.certs_dir.as_ref(),
561 0 : args.allow_tls_keylogfile,
562 0 : )?),
563 0 : (None, None) => None,
564 0 : _ => bail!("either both or neither tls-key and tls-cert must be specified"),
565 : };
566 :
567 0 : let backup_metric_collection_config = config::MetricBackupCollectionConfig {
568 0 : interval: args.metric_backup_collection_interval,
569 0 : remote_storage_config: args.metric_backup_collection_remote_storage.clone(),
570 0 : chunk_size: args.metric_backup_collection_chunk_size,
571 0 : };
572 :
573 0 : let metric_collection = match (
574 0 : &args.metric_collection_endpoint,
575 0 : &args.metric_collection_interval,
576 : ) {
577 0 : (Some(endpoint), Some(interval)) => Some(config::MetricCollectionConfig {
578 0 : endpoint: endpoint.parse()?,
579 0 : interval: humantime::parse_duration(interval)?,
580 0 : backup_metric_collection_config,
581 : }),
582 0 : (None, None) => None,
583 0 : _ => bail!(
584 0 : "either both or neither metric-collection-endpoint \
585 0 : and metric-collection-interval must be specified"
586 0 : ),
587 : };
588 :
589 : let config::ConcurrencyLockOptions {
590 0 : shards,
591 0 : limiter,
592 0 : epoch,
593 0 : timeout,
594 0 : } = args.connect_compute_lock.parse()?;
595 0 : info!(
596 : ?limiter,
597 : shards,
598 : ?epoch,
599 0 : "Using NodeLocks (connect_compute)"
600 : );
601 0 : let connect_compute_locks = control_plane::locks::ApiLocks::new(
602 0 : "connect_compute_lock",
603 0 : limiter,
604 0 : shards,
605 0 : timeout,
606 0 : epoch,
607 0 : &Metrics::get().proxy.connect_compute_lock,
608 0 : )?;
609 :
610 0 : let http_config = HttpConfig {
611 0 : accept_websockets: !args.is_auth_broker,
612 0 : pool_options: GlobalConnPoolOptions {
613 0 : max_conns_per_endpoint: args.sql_over_http.sql_over_http_pool_max_conns_per_endpoint,
614 0 : gc_epoch: args.sql_over_http.sql_over_http_pool_gc_epoch,
615 0 : pool_shards: args.sql_over_http.sql_over_http_pool_shards,
616 0 : idle_timeout: args.sql_over_http.sql_over_http_idle_timeout,
617 0 : opt_in: args.sql_over_http.sql_over_http_pool_opt_in,
618 0 : max_total_conns: args.sql_over_http.sql_over_http_pool_max_total_conns,
619 0 : },
620 0 : cancel_set: CancelSet::new(args.sql_over_http.sql_over_http_cancel_set_shards),
621 0 : client_conn_threshold: args.sql_over_http.sql_over_http_client_conn_threshold,
622 0 : max_request_size_bytes: args.sql_over_http.sql_over_http_max_request_size_bytes,
623 0 : max_response_size_bytes: args.sql_over_http.sql_over_http_max_response_size_bytes,
624 0 : };
625 0 : let authentication_config = AuthenticationConfig {
626 0 : jwks_cache: JwkCache::default(),
627 0 : thread_pool,
628 0 : scram_protocol_timeout: args.scram_protocol_timeout,
629 0 : rate_limiter_enabled: args.auth_rate_limit_enabled,
630 0 : rate_limiter: AuthRateLimiter::new(args.auth_rate_limit.clone()),
631 0 : rate_limit_ip_subnet: args.auth_rate_limit_ip_subnet,
632 0 : ip_allowlist_check_enabled: !args.is_private_access_proxy,
633 0 : is_auth_broker: args.is_auth_broker,
634 0 : accept_jwts: args.is_auth_broker,
635 0 : console_redirect_confirmation_timeout: args.webauth_confirmation_timeout,
636 0 : };
637 :
638 0 : let compute_config = ComputeConfig {
639 0 : retry: config::RetryConfig::parse(&args.connect_to_compute_retry)?,
640 0 : tls: Arc::new(compute_client_config_with_root_certs()?),
641 0 : timeout: Duration::from_secs(2),
642 : };
643 :
644 0 : let config = ProxyConfig {
645 0 : tls_config,
646 0 : metric_collection,
647 0 : http_config,
648 0 : authentication_config,
649 0 : proxy_protocol_v2: args.proxy_protocol_v2,
650 0 : handshake_timeout: args.handshake_timeout,
651 0 : region: args.region.clone(),
652 0 : wake_compute_retry_config: config::RetryConfig::parse(&args.wake_compute_retry)?,
653 0 : connect_compute_locks,
654 0 : connect_to_compute: compute_config,
655 0 : };
656 0 :
657 0 : let config = Box::leak(Box::new(config));
658 0 :
659 0 : tokio::spawn(config.connect_compute_locks.garbage_collect_worker());
660 0 :
661 0 : Ok(config)
662 0 : }
663 :
664 : /// auth::Backend is created at proxy startup, and lives forever.
665 0 : fn build_auth_backend(
666 0 : args: &ProxyCliArgs,
667 0 : ) -> anyhow::Result<Either<&'static auth::Backend<'static, ()>, &'static ConsoleRedirectBackend>> {
668 0 : match &args.auth_backend {
669 : AuthBackendType::ControlPlaneV1 => {
670 0 : let wake_compute_cache_config: CacheOptions = args.wake_compute_cache.parse()?;
671 0 : let project_info_cache_config: ProjectInfoCacheOptions =
672 0 : args.project_info_cache.parse()?;
673 0 : let endpoint_cache_config: config::EndpointCacheConfig =
674 0 : args.endpoint_cache_config.parse()?;
675 :
676 0 : info!("Using NodeInfoCache (wake_compute) with options={wake_compute_cache_config:?}");
677 0 : info!(
678 0 : "Using AllowedIpsCache (wake_compute) with options={project_info_cache_config:?}"
679 : );
680 0 : info!("Using EndpointCacheConfig with options={endpoint_cache_config:?}");
681 0 : let caches = Box::leak(Box::new(control_plane::caches::ApiCaches::new(
682 0 : wake_compute_cache_config,
683 0 : project_info_cache_config,
684 0 : endpoint_cache_config,
685 0 : )));
686 :
687 : let config::ConcurrencyLockOptions {
688 0 : shards,
689 0 : limiter,
690 0 : epoch,
691 0 : timeout,
692 0 : } = args.wake_compute_lock.parse()?;
693 0 : info!(?limiter, shards, ?epoch, "Using NodeLocks (wake_compute)");
694 0 : let locks = Box::leak(Box::new(control_plane::locks::ApiLocks::new(
695 0 : "wake_compute_lock",
696 0 : limiter,
697 0 : shards,
698 0 : timeout,
699 0 : epoch,
700 0 : &Metrics::get().wake_compute_lock,
701 0 : )?));
702 0 : tokio::spawn(locks.garbage_collect_worker());
703 :
704 0 : let url: proxy::url::ApiUrl = args.auth_endpoint.parse()?;
705 :
706 0 : let endpoint = http::Endpoint::new(url, http::new_client());
707 0 :
708 0 : let mut wake_compute_rps_limit = args.wake_compute_limit.clone();
709 0 : RateBucketInfo::validate(&mut wake_compute_rps_limit)?;
710 0 : let wake_compute_endpoint_rate_limiter =
711 0 : Arc::new(WakeComputeRateLimiter::new(wake_compute_rps_limit));
712 0 :
713 0 : let api = control_plane::client::cplane_proxy_v1::NeonControlPlaneClient::new(
714 0 : endpoint,
715 0 : args.control_plane_token.clone(),
716 0 : caches,
717 0 : locks,
718 0 : wake_compute_endpoint_rate_limiter,
719 0 : );
720 0 :
721 0 : let api = control_plane::client::ControlPlaneClient::ProxyV1(api);
722 0 : let auth_backend = auth::Backend::ControlPlane(MaybeOwned::Owned(api), ());
723 0 : let config = Box::leak(Box::new(auth_backend));
724 0 :
725 0 : Ok(Either::Left(config))
726 : }
727 :
728 : #[cfg(feature = "testing")]
729 : AuthBackendType::Postgres => {
730 0 : let url = args.auth_endpoint.parse()?;
731 0 : let api = control_plane::client::mock::MockControlPlane::new(
732 0 : url,
733 0 : !args.is_private_access_proxy,
734 0 : );
735 0 : let api = control_plane::client::ControlPlaneClient::PostgresMock(api);
736 0 :
737 0 : let auth_backend = auth::Backend::ControlPlane(MaybeOwned::Owned(api), ());
738 0 :
739 0 : let config = Box::leak(Box::new(auth_backend));
740 0 :
741 0 : Ok(Either::Left(config))
742 : }
743 :
744 : AuthBackendType::ConsoleRedirect => {
745 0 : let wake_compute_cache_config: CacheOptions = args.wake_compute_cache.parse()?;
746 0 : let project_info_cache_config: ProjectInfoCacheOptions =
747 0 : args.project_info_cache.parse()?;
748 0 : let endpoint_cache_config: config::EndpointCacheConfig =
749 0 : args.endpoint_cache_config.parse()?;
750 :
751 0 : info!("Using NodeInfoCache (wake_compute) with options={wake_compute_cache_config:?}");
752 0 : info!(
753 0 : "Using AllowedIpsCache (wake_compute) with options={project_info_cache_config:?}"
754 : );
755 0 : info!("Using EndpointCacheConfig with options={endpoint_cache_config:?}");
756 0 : let caches = Box::leak(Box::new(control_plane::caches::ApiCaches::new(
757 0 : wake_compute_cache_config,
758 0 : project_info_cache_config,
759 0 : endpoint_cache_config,
760 0 : )));
761 :
762 : let config::ConcurrencyLockOptions {
763 0 : shards,
764 0 : limiter,
765 0 : epoch,
766 0 : timeout,
767 0 : } = args.wake_compute_lock.parse()?;
768 0 : info!(?limiter, shards, ?epoch, "Using NodeLocks (wake_compute)");
769 0 : let locks = Box::leak(Box::new(control_plane::locks::ApiLocks::new(
770 0 : "wake_compute_lock",
771 0 : limiter,
772 0 : shards,
773 0 : timeout,
774 0 : epoch,
775 0 : &Metrics::get().wake_compute_lock,
776 0 : )?));
777 :
778 0 : let url = args.uri.clone().parse()?;
779 0 : let ep_url: proxy::url::ApiUrl = args.auth_endpoint.parse()?;
780 0 : let endpoint = http::Endpoint::new(ep_url, http::new_client());
781 0 : let mut wake_compute_rps_limit = args.wake_compute_limit.clone();
782 0 : RateBucketInfo::validate(&mut wake_compute_rps_limit)?;
783 0 : let wake_compute_endpoint_rate_limiter =
784 0 : Arc::new(WakeComputeRateLimiter::new(wake_compute_rps_limit));
785 0 :
786 0 : // Since we use only get_allowed_ips_and_secret() wake_compute_endpoint_rate_limiter
787 0 : // and locks are not used in ConsoleRedirectBackend,
788 0 : // but they are required by the NeonControlPlaneClient
789 0 : let api = control_plane::client::cplane_proxy_v1::NeonControlPlaneClient::new(
790 0 : endpoint,
791 0 : args.control_plane_token.clone(),
792 0 : caches,
793 0 : locks,
794 0 : wake_compute_endpoint_rate_limiter,
795 0 : );
796 0 :
797 0 : let backend = ConsoleRedirectBackend::new(url, api);
798 0 : let config = Box::leak(Box::new(backend));
799 0 :
800 0 : Ok(Either::Right(config))
801 : }
802 : }
803 0 : }
804 :
805 : #[cfg(test)]
806 : mod tests {
807 : use std::time::Duration;
808 :
809 : use clap::Parser;
810 : use proxy::rate_limiter::RateBucketInfo;
811 :
812 : #[test]
813 1 : fn parse_endpoint_rps_limit() {
814 1 : let config = super::ProxyCliArgs::parse_from([
815 1 : "proxy",
816 1 : "--endpoint-rps-limit",
817 1 : "100@1s",
818 1 : "--endpoint-rps-limit",
819 1 : "20@30s",
820 1 : ]);
821 1 :
822 1 : assert_eq!(
823 1 : config.endpoint_rps_limit,
824 1 : vec![
825 1 : RateBucketInfo::new(100, Duration::from_secs(1)),
826 1 : RateBucketInfo::new(20, Duration::from_secs(30)),
827 1 : ]
828 1 : );
829 1 : }
830 : }
|
