Line data Source code
1 : //
2 : // Main entry point for the safekeeper executable
3 : //
4 : use std::fs::{self, File};
5 : use std::io::{ErrorKind, Write};
6 : use std::str::FromStr;
7 : use std::sync::Arc;
8 : use std::time::{Duration, Instant};
9 :
10 : use anyhow::{Context, Result, bail};
11 : use camino::{Utf8Path, Utf8PathBuf};
12 : use clap::{ArgAction, Parser};
13 : use futures::future::BoxFuture;
14 : use futures::stream::FuturesUnordered;
15 : use futures::{FutureExt, StreamExt};
16 : use http_utils::tls_certs::ReloadingCertificateResolver;
17 : use metrics::set_build_info_metric;
18 : use remote_storage::RemoteStorageConfig;
19 : use safekeeper::defaults::{
20 : DEFAULT_CONTROL_FILE_SAVE_INTERVAL, DEFAULT_EVICTION_MIN_RESIDENT, DEFAULT_HEARTBEAT_TIMEOUT,
21 : DEFAULT_HTTP_LISTEN_ADDR, DEFAULT_MAX_OFFLOADER_LAG_BYTES, DEFAULT_PARTIAL_BACKUP_CONCURRENCY,
22 : DEFAULT_PARTIAL_BACKUP_TIMEOUT, DEFAULT_PG_LISTEN_ADDR, DEFAULT_SSL_CERT_FILE,
23 : DEFAULT_SSL_CERT_RELOAD_PERIOD, DEFAULT_SSL_KEY_FILE,
24 : };
25 : use safekeeper::{
26 : BACKGROUND_RUNTIME, BROKER_RUNTIME, GlobalTimelines, HTTP_RUNTIME, SafeKeeperConf,
27 : WAL_SERVICE_RUNTIME, broker, control_file, http, wal_backup, wal_service,
28 : };
29 : use sd_notify::NotifyState;
30 : use storage_broker::{DEFAULT_ENDPOINT, Uri};
31 : use tokio::runtime::Handle;
32 : use tokio::signal::unix::{SignalKind, signal};
33 : use tokio::task::JoinError;
34 : use tracing::*;
35 : use utils::auth::{JwtAuth, Scope, SwappableJwtAuth};
36 : use utils::id::NodeId;
37 : use utils::logging::{self, LogFormat, SecretString};
38 : use utils::sentry_init::init_sentry;
39 : use utils::{pid_file, project_build_tag, project_git_version, tcp_listener};
40 :
41 : #[global_allocator]
42 : static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
43 :
44 : /// Configure jemalloc to profile heap allocations by sampling stack traces every 2 MB (1 << 21).
45 : /// This adds roughly 3% overhead for allocations on average, which is acceptable considering
46 : /// performance-sensitive code will avoid allocations as far as possible anyway.
47 : #[allow(non_upper_case_globals)]
48 : #[unsafe(export_name = "malloc_conf")]
49 : pub static malloc_conf: &[u8] = b"prof:true,prof_active:true,lg_prof_sample:21\0";
50 :
51 : const PID_FILE_NAME: &str = "safekeeper.pid";
52 : const ID_FILE_NAME: &str = "safekeeper.id";
53 :
54 : project_git_version!(GIT_VERSION);
55 : project_build_tag!(BUILD_TAG);
56 :
57 : const FEATURES: &[&str] = &[
58 : #[cfg(feature = "testing")]
59 : "testing",
60 : ];
61 :
62 0 : fn version() -> String {
63 0 : format!(
64 0 : "{GIT_VERSION} failpoints: {}, features: {:?}",
65 0 : fail::has_failpoints(),
66 0 : FEATURES,
67 0 : )
68 0 : }
69 :
70 : const ABOUT: &str = r#"
71 : A fleet of safekeepers is responsible for reliably storing WAL received from
72 : compute, passing it through consensus (mitigating potential computes brain
73 : split), and serving the hardened part further downstream to pageserver(s).
74 : "#;
75 :
76 : #[derive(Parser)]
77 : #[command(name = "Neon safekeeper", version = GIT_VERSION, about = ABOUT, long_about = None)]
78 : struct Args {
79 : /// Path to the safekeeper data directory.
80 : #[arg(short = 'D', long, default_value = "./")]
81 0 : datadir: Utf8PathBuf,
82 : /// Safekeeper node id.
83 : #[arg(long)]
84 : id: Option<u64>,
85 : /// Initialize safekeeper with given id and exit.
86 : #[arg(long)]
87 0 : init: bool,
88 : /// Listen endpoint for receiving/sending WAL in the form host:port.
89 : #[arg(short, long, default_value = DEFAULT_PG_LISTEN_ADDR)]
90 0 : listen_pg: String,
91 : /// Listen endpoint for receiving/sending WAL in the form host:port allowing
92 : /// only tenant scoped auth tokens. Pointless if auth is disabled.
93 : #[arg(long, default_value = None, verbatim_doc_comment)]
94 : listen_pg_tenant_only: Option<String>,
95 : /// Listen http endpoint for management and metrics in the form host:port.
96 : #[arg(long, default_value = DEFAULT_HTTP_LISTEN_ADDR)]
97 0 : listen_http: String,
98 : /// Listen https endpoint for management and metrics in the form host:port.
99 : #[arg(long, default_value = None)]
100 : listen_https: Option<String>,
101 : /// Advertised endpoint for receiving/sending WAL in the form host:port. If not
102 : /// specified, listen_pg is used to advertise instead.
103 : #[arg(long, default_value = None)]
104 : advertise_pg: Option<String>,
105 : /// Availability zone of the safekeeper.
106 : #[arg(long)]
107 : availability_zone: Option<String>,
108 : /// Do not wait for changes to be written safely to disk. Unsafe.
109 : #[arg(short, long)]
110 0 : no_sync: bool,
111 : /// Dump control file at path specified by this argument and exit.
112 : #[arg(long)]
113 : dump_control_file: Option<Utf8PathBuf>,
114 : /// Broker endpoint for storage nodes coordination in the form
115 : /// http[s]://host:port. In case of https schema TLS is connection is
116 : /// established; plaintext otherwise.
117 : #[arg(long, default_value = DEFAULT_ENDPOINT, verbatim_doc_comment)]
118 0 : broker_endpoint: Uri,
119 : /// Broker keepalive interval.
120 : #[arg(long, value_parser= humantime::parse_duration, default_value = storage_broker::DEFAULT_KEEPALIVE_INTERVAL)]
121 0 : broker_keepalive_interval: Duration,
122 : /// Peer safekeeper is considered dead after not receiving heartbeats from
123 : /// it during this period passed as a human readable duration.
124 : #[arg(long, value_parser= humantime::parse_duration, default_value = DEFAULT_HEARTBEAT_TIMEOUT, verbatim_doc_comment)]
125 0 : heartbeat_timeout: Duration,
126 : /// Enable/disable peer recovery.
127 : #[arg(long, default_value = "false", action=ArgAction::Set)]
128 0 : peer_recovery: bool,
129 : /// Remote storage configuration for WAL backup (offloading to s3) as TOML
130 : /// inline table, e.g.
131 : /// {max_concurrent_syncs = 17, max_sync_errors = 13, bucket_name = "<BUCKETNAME>", bucket_region = "<REGION>", concurrency_limit = 119}
132 : /// Safekeeper offloads WAL to
133 : /// [prefix_in_bucket/]<tenant_id>/<timeline_id>/<segment_file>, mirroring
134 : /// structure on the file system.
135 : #[arg(long, value_parser = parse_remote_storage, verbatim_doc_comment)]
136 : remote_storage: Option<RemoteStorageConfig>,
137 : /// Safekeeper won't be elected for WAL offloading if it is lagging for more than this value in bytes
138 1 : #[arg(long, default_value_t = DEFAULT_MAX_OFFLOADER_LAG_BYTES)]
139 0 : max_offloader_lag: u64,
140 : /// Number of max parallel WAL segments to be offloaded to remote storage.
141 : #[arg(long, default_value = "5")]
142 0 : wal_backup_parallel_jobs: usize,
143 : /// Disable WAL backup to s3. When disabled, safekeeper removes WAL ignoring
144 : /// WAL backup horizon.
145 : #[arg(long)]
146 0 : disable_wal_backup: bool,
147 : /// If given, enables auth on incoming connections to WAL service endpoint
148 : /// (--listen-pg). Value specifies path to a .pem public key used for
149 : /// validations of JWT tokens. Empty string is allowed and means disabling
150 : /// auth.
151 : #[arg(long, verbatim_doc_comment, value_parser = opt_pathbuf_parser)]
152 : pg_auth_public_key_path: Option<Utf8PathBuf>,
153 : /// If given, enables auth on incoming connections to tenant only WAL
154 : /// service endpoint (--listen-pg-tenant-only). Value specifies path to a
155 : /// .pem public key used for validations of JWT tokens. Empty string is
156 : /// allowed and means disabling auth.
157 : #[arg(long, verbatim_doc_comment, value_parser = opt_pathbuf_parser)]
158 : pg_tenant_only_auth_public_key_path: Option<Utf8PathBuf>,
159 : /// If given, enables auth on incoming connections to http management
160 : /// service endpoint (--listen-http). Value specifies path to a .pem public
161 : /// key used for validations of JWT tokens. Empty string is allowed and
162 : /// means disabling auth.
163 : #[arg(long, verbatim_doc_comment, value_parser = opt_pathbuf_parser)]
164 : http_auth_public_key_path: Option<Utf8PathBuf>,
165 : /// Format for logging, either 'plain' or 'json'.
166 : #[arg(long, default_value = "plain")]
167 0 : log_format: String,
168 : /// Run everything in single threaded current thread runtime, might be
169 : /// useful for debugging.
170 : #[arg(long)]
171 0 : current_thread_runtime: bool,
172 : /// Keep horizon for walsenders, i.e. don't remove WAL segments that are
173 : /// still needed for existing replication connection.
174 : #[arg(long)]
175 0 : walsenders_keep_horizon: bool,
176 : /// Controls how long backup will wait until uploading the partial segment.
177 : #[arg(long, value_parser = humantime::parse_duration, default_value = DEFAULT_PARTIAL_BACKUP_TIMEOUT, verbatim_doc_comment)]
178 0 : partial_backup_timeout: Duration,
179 : /// Disable task to push messages to broker every second. Supposed to
180 : /// be used in tests.
181 : #[arg(long)]
182 0 : disable_periodic_broker_push: bool,
183 : /// Enable automatic switching to offloaded state.
184 : #[arg(long)]
185 0 : enable_offload: bool,
186 : /// Delete local WAL files after offloading. When disabled, they will be left on disk.
187 : #[arg(long)]
188 0 : delete_offloaded_wal: bool,
189 : /// Pending updates to control file will be automatically saved after this interval.
190 : #[arg(long, value_parser = humantime::parse_duration, default_value = DEFAULT_CONTROL_FILE_SAVE_INTERVAL)]
191 0 : control_file_save_interval: Duration,
192 : /// Number of allowed concurrent uploads of partial segments to remote storage.
193 : #[arg(long, default_value = DEFAULT_PARTIAL_BACKUP_CONCURRENCY)]
194 0 : partial_backup_concurrency: usize,
195 : /// How long a timeline must be resident before it is eligible for eviction.
196 : /// Usually, timeline eviction has to wait for `partial_backup_timeout` before being eligible for eviction,
197 : /// but if a timeline is un-evicted and then _not_ written to, it would immediately flap to evicting again,
198 : /// if it weren't for `eviction_min_resident` preventing that.
199 : ///
200 : /// Also defines interval for eviction retries.
201 : #[arg(long, value_parser = humantime::parse_duration, default_value = DEFAULT_EVICTION_MIN_RESIDENT)]
202 0 : eviction_min_resident: Duration,
203 : /// Enable fanning out WAL to different shards from the same reader
204 : #[arg(long)]
205 0 : wal_reader_fanout: bool,
206 : /// Only fan out the WAL reader if the absoulte delta between the new requested position
207 : /// and the current position of the reader is smaller than this value.
208 : #[arg(long)]
209 : max_delta_for_fanout: Option<u64>,
210 : /// Path to a file with certificate's private key for https API.
211 : #[arg(long, default_value = DEFAULT_SSL_KEY_FILE)]
212 0 : ssl_key_file: Utf8PathBuf,
213 : /// Path to a file with a X509 certificate for https API.
214 : #[arg(long, default_value = DEFAULT_SSL_CERT_FILE)]
215 0 : ssl_cert_file: Utf8PathBuf,
216 : /// Period to reload certificate and private key from files.
217 : #[arg(long, value_parser = humantime::parse_duration, default_value = DEFAULT_SSL_CERT_RELOAD_PERIOD)]
218 0 : ssl_cert_reload_period: Duration,
219 : /// Trusted root CA certificates to use in https APIs.
220 : #[arg(long)]
221 : ssl_ca_file: Option<Utf8PathBuf>,
222 : /// Flag to use https for requests to peer's safekeeper API.
223 : #[arg(long)]
224 0 : use_https_safekeeper_api: bool,
225 : /// Path to the JWT auth token used to authenticate with other safekeepers.
226 : #[arg(long)]
227 : auth_token_path: Option<Utf8PathBuf>,
228 :
229 : /// Enable TLS in WAL service API.
230 : /// Does not force TLS: the client negotiates TLS usage during the handshake.
231 : /// Uses key and certificate from ssl_key_file/ssl_cert_file.
232 : #[arg(long)]
233 0 : enable_tls_wal_service_api: bool,
234 :
235 : /// Run in development mode (disables security checks)
236 : #[arg(long, help = "Run in development mode (disables security checks)")]
237 0 : dev: bool,
238 : }
239 :
240 : // Like PathBufValueParser, but allows empty string.
241 0 : fn opt_pathbuf_parser(s: &str) -> Result<Utf8PathBuf, String> {
242 0 : Ok(Utf8PathBuf::from_str(s).unwrap())
243 0 : }
244 :
245 : #[tokio::main(flavor = "current_thread")]
246 0 : async fn main() -> anyhow::Result<()> {
247 0 : // We want to allow multiple occurences of the same arg (taking the last) so
248 0 : // that neon_local could generate command with defaults + overrides without
249 0 : // getting 'argument cannot be used multiple times' error. This seems to be
250 0 : // impossible with pure Derive API, so convert struct to Command, modify it,
251 0 : // parse arguments, and then fill the struct back.
252 0 : let cmd = <Args as clap::CommandFactory>::command()
253 0 : .args_override_self(true)
254 0 : .version(version());
255 0 : let mut matches = cmd.get_matches();
256 0 : let mut args = <Args as clap::FromArgMatches>::from_arg_matches_mut(&mut matches)?;
257 0 :
258 0 : // I failed to modify opt_pathbuf_parser to return Option<PathBuf> in
259 0 : // reasonable time, so turn empty string into option post factum.
260 0 : if let Some(pb) = &args.pg_auth_public_key_path {
261 0 : if pb.as_os_str().is_empty() {
262 0 : args.pg_auth_public_key_path = None;
263 0 : }
264 0 : }
265 0 : if let Some(pb) = &args.pg_tenant_only_auth_public_key_path {
266 0 : if pb.as_os_str().is_empty() {
267 0 : args.pg_tenant_only_auth_public_key_path = None;
268 0 : }
269 0 : }
270 0 : if let Some(pb) = &args.http_auth_public_key_path {
271 0 : if pb.as_os_str().is_empty() {
272 0 : args.http_auth_public_key_path = None;
273 0 : }
274 0 : }
275 0 :
276 0 : if let Some(addr) = args.dump_control_file {
277 0 : let state = control_file::FileStorage::load_control_file(addr)?;
278 0 : let json = serde_json::to_string(&state)?;
279 0 : print!("{json}");
280 0 : return Ok(());
281 0 : }
282 0 :
283 0 : // important to keep the order of:
284 0 : // 1. init logging
285 0 : // 2. tracing panic hook
286 0 : // 3. sentry
287 0 : logging::init(
288 0 : LogFormat::from_config(&args.log_format)?,
289 0 : logging::TracingErrorLayerEnablement::Disabled,
290 0 : logging::Output::Stdout,
291 0 : )?;
292 0 : logging::replace_panic_hook_with_tracing_panic_hook().forget();
293 0 : info!("version: {GIT_VERSION}");
294 0 : info!("buld_tag: {BUILD_TAG}");
295 0 :
296 0 : let args_workdir = &args.datadir;
297 0 : let workdir = args_workdir.canonicalize_utf8().with_context(|| {
298 0 : format!("Failed to get the absolute path for input workdir {args_workdir:?}")
299 0 : })?;
300 0 :
301 0 : // Change into the data directory.
302 0 : std::env::set_current_dir(&workdir)?;
303 0 :
304 0 : // Prevent running multiple safekeepers on the same directory
305 0 : let lock_file_path = workdir.join(PID_FILE_NAME);
306 0 : let lock_file =
307 0 : pid_file::claim_for_current_process(&lock_file_path).context("claim pid file")?;
308 0 : info!("claimed pid file at {lock_file_path:?}");
309 0 : // ensure that the lock file is held even if the main thread of the process is panics
310 0 : // we need to release the lock file only when the current process is gone
311 0 : std::mem::forget(lock_file);
312 0 :
313 0 : // Set or read our ID.
314 0 : let id = set_id(&workdir, args.id.map(NodeId))?;
315 0 : if args.init {
316 0 : return Ok(());
317 0 : }
318 0 :
319 0 : let pg_auth = match args.pg_auth_public_key_path.as_ref() {
320 0 : None => {
321 0 : info!("pg auth is disabled");
322 0 : None
323 0 : }
324 0 : Some(path) => {
325 0 : info!("loading pg auth JWT key from {path}");
326 0 : Some(Arc::new(
327 0 : JwtAuth::from_key_path(path).context("failed to load the auth key")?,
328 0 : ))
329 0 : }
330 0 : };
331 0 : let pg_tenant_only_auth = match args.pg_tenant_only_auth_public_key_path.as_ref() {
332 0 : None => {
333 0 : info!("pg tenant only auth is disabled");
334 0 : None
335 0 : }
336 0 : Some(path) => {
337 0 : info!("loading pg tenant only auth JWT key from {path}");
338 0 : Some(Arc::new(
339 0 : JwtAuth::from_key_path(path).context("failed to load the auth key")?,
340 0 : ))
341 0 : }
342 0 : };
343 0 : let http_auth = match args.http_auth_public_key_path.as_ref() {
344 0 : None => {
345 0 : info!("http auth is disabled");
346 0 : None
347 0 : }
348 0 : Some(path) => {
349 0 : info!("loading http auth JWT key(s) from {path}");
350 0 : let jwt_auth = JwtAuth::from_key_path(path).context("failed to load the auth key")?;
351 0 : Some(Arc::new(SwappableJwtAuth::new(jwt_auth)))
352 0 : }
353 0 : };
354 0 :
355 0 : // Load JWT auth token to connect to other safekeepers for pull_timeline.
356 0 : let sk_auth_token = if let Some(auth_token_path) = args.auth_token_path.as_ref() {
357 0 : info!("loading JWT token for authentication with safekeepers from {auth_token_path}");
358 0 : let auth_token = tokio::fs::read_to_string(auth_token_path).await?;
359 0 : Some(SecretString::from(auth_token.trim().to_owned()))
360 0 : } else {
361 0 : info!("no JWT token for authentication with safekeepers detected");
362 0 : None
363 0 : };
364 0 :
365 0 : let ssl_ca_certs = match args.ssl_ca_file.as_ref() {
366 0 : Some(ssl_ca_file) => {
367 0 : tracing::info!("Using ssl root CA file: {ssl_ca_file:?}");
368 0 : let buf = tokio::fs::read(ssl_ca_file).await?;
369 0 : pem::parse_many(&buf)?
370 0 : .into_iter()
371 0 : .filter(|pem| pem.tag() == "CERTIFICATE")
372 0 : .collect()
373 0 : }
374 0 : None => Vec::new(),
375 0 : };
376 0 :
377 0 : let conf = Arc::new(SafeKeeperConf {
378 0 : workdir,
379 0 : my_id: id,
380 0 : listen_pg_addr: args.listen_pg,
381 0 : listen_pg_addr_tenant_only: args.listen_pg_tenant_only,
382 0 : listen_http_addr: args.listen_http,
383 0 : listen_https_addr: args.listen_https,
384 0 : advertise_pg_addr: args.advertise_pg,
385 0 : availability_zone: args.availability_zone,
386 0 : no_sync: args.no_sync,
387 0 : broker_endpoint: args.broker_endpoint,
388 0 : broker_keepalive_interval: args.broker_keepalive_interval,
389 0 : heartbeat_timeout: args.heartbeat_timeout,
390 0 : peer_recovery_enabled: args.peer_recovery,
391 0 : remote_storage: args.remote_storage,
392 0 : max_offloader_lag_bytes: args.max_offloader_lag,
393 0 : wal_backup_enabled: !args.disable_wal_backup,
394 0 : backup_parallel_jobs: args.wal_backup_parallel_jobs,
395 0 : pg_auth,
396 0 : pg_tenant_only_auth,
397 0 : http_auth,
398 0 : sk_auth_token,
399 0 : current_thread_runtime: args.current_thread_runtime,
400 0 : walsenders_keep_horizon: args.walsenders_keep_horizon,
401 0 : partial_backup_timeout: args.partial_backup_timeout,
402 0 : disable_periodic_broker_push: args.disable_periodic_broker_push,
403 0 : enable_offload: args.enable_offload,
404 0 : delete_offloaded_wal: args.delete_offloaded_wal,
405 0 : control_file_save_interval: args.control_file_save_interval,
406 0 : partial_backup_concurrency: args.partial_backup_concurrency,
407 0 : eviction_min_resident: args.eviction_min_resident,
408 0 : wal_reader_fanout: args.wal_reader_fanout,
409 0 : max_delta_for_fanout: args.max_delta_for_fanout,
410 0 : ssl_key_file: args.ssl_key_file,
411 0 : ssl_cert_file: args.ssl_cert_file,
412 0 : ssl_cert_reload_period: args.ssl_cert_reload_period,
413 0 : ssl_ca_certs,
414 0 : use_https_safekeeper_api: args.use_https_safekeeper_api,
415 0 : enable_tls_wal_service_api: args.enable_tls_wal_service_api,
416 0 : });
417 0 :
418 0 : // initialize sentry if SENTRY_DSN is provided
419 0 : let _sentry_guard = init_sentry(
420 0 : Some(GIT_VERSION.into()),
421 0 : &[("node_id", &conf.my_id.to_string())],
422 0 : );
423 0 : start_safekeeper(conf).await
424 0 : }
425 :
426 : /// Result of joining any of main tasks: upper error means task failed to
427 : /// complete, e.g. panicked, inner is error produced by task itself.
428 : type JoinTaskRes = Result<anyhow::Result<()>, JoinError>;
429 :
430 0 : async fn start_safekeeper(conf: Arc<SafeKeeperConf>) -> Result<()> {
431 0 : // fsync the datadir to make sure we have a consistent state on disk.
432 0 : if !conf.no_sync {
433 0 : let dfd = File::open(&conf.workdir).context("open datadir for syncfs")?;
434 0 : let started = Instant::now();
435 0 : utils::crashsafe::syncfs(dfd)?;
436 0 : let elapsed = started.elapsed();
437 0 : info!(
438 0 : elapsed_ms = elapsed.as_millis(),
439 0 : "syncfs data directory done"
440 : );
441 0 : }
442 :
443 0 : info!("starting safekeeper WAL service on {}", conf.listen_pg_addr);
444 0 : let pg_listener = tcp_listener::bind(conf.listen_pg_addr.clone()).map_err(|e| {
445 0 : error!("failed to bind to address {}: {}", conf.listen_pg_addr, e);
446 0 : e
447 0 : })?;
448 :
449 0 : let pg_listener_tenant_only =
450 0 : if let Some(listen_pg_addr_tenant_only) = &conf.listen_pg_addr_tenant_only {
451 0 : info!(
452 0 : "starting safekeeper tenant scoped WAL service on {}",
453 : listen_pg_addr_tenant_only
454 : );
455 0 : let listener = tcp_listener::bind(listen_pg_addr_tenant_only.clone()).map_err(|e| {
456 0 : error!(
457 0 : "failed to bind to address {}: {}",
458 : listen_pg_addr_tenant_only, e
459 : );
460 0 : e
461 0 : })?;
462 0 : Some(listener)
463 : } else {
464 0 : None
465 : };
466 :
467 0 : info!(
468 0 : "starting safekeeper HTTP service on {}",
469 0 : conf.listen_http_addr
470 : );
471 0 : let http_listener = tcp_listener::bind(conf.listen_http_addr.clone()).map_err(|e| {
472 0 : error!("failed to bind to address {}: {}", conf.listen_http_addr, e);
473 0 : e
474 0 : })?;
475 :
476 0 : let https_listener = match conf.listen_https_addr.as_ref() {
477 0 : Some(listen_https_addr) => {
478 0 : info!("starting safekeeper HTTPS service on {}", listen_https_addr);
479 0 : Some(tcp_listener::bind(listen_https_addr).map_err(|e| {
480 0 : error!("failed to bind to address {}: {}", listen_https_addr, e);
481 0 : e
482 0 : })?)
483 : }
484 0 : None => None,
485 : };
486 :
487 0 : let global_timelines = Arc::new(GlobalTimelines::new(conf.clone()));
488 0 :
489 0 : // Register metrics collector for active timelines. It's important to do this
490 0 : // after daemonizing, otherwise process collector will be upset.
491 0 : let timeline_collector = safekeeper::metrics::TimelineCollector::new(global_timelines.clone());
492 0 : metrics::register_internal(Box::new(timeline_collector))?;
493 :
494 0 : wal_backup::init_remote_storage(&conf).await;
495 :
496 : // Keep handles to main tasks to die if any of them disappears.
497 0 : let mut tasks_handles: FuturesUnordered<BoxFuture<(String, JoinTaskRes)>> =
498 0 : FuturesUnordered::new();
499 0 :
500 0 : // Start wal backup launcher before loading timelines as we'll notify it
501 0 : // through the channel about timelines which need offloading, not draining
502 0 : // the channel would cause deadlock.
503 0 : let current_thread_rt = conf
504 0 : .current_thread_runtime
505 0 : .then(|| Handle::try_current().expect("no runtime in main"));
506 0 :
507 0 : // Load all timelines from disk to memory.
508 0 : global_timelines.init().await?;
509 :
510 : // Run everything in current thread rt, if asked.
511 0 : if conf.current_thread_runtime {
512 0 : info!("running in current thread runtime");
513 0 : }
514 :
515 0 : let tls_server_config = if conf.listen_https_addr.is_some() || conf.enable_tls_wal_service_api {
516 0 : let ssl_key_file = conf.ssl_key_file.clone();
517 0 : let ssl_cert_file = conf.ssl_cert_file.clone();
518 0 : let ssl_cert_reload_period = conf.ssl_cert_reload_period;
519 :
520 : // Create resolver in BACKGROUND_RUNTIME, so the background certificate reloading
521 : // task is run in this runtime.
522 0 : let cert_resolver = current_thread_rt
523 0 : .as_ref()
524 0 : .unwrap_or_else(|| BACKGROUND_RUNTIME.handle())
525 0 : .spawn(async move {
526 0 : ReloadingCertificateResolver::new(
527 0 : "main",
528 0 : &ssl_key_file,
529 0 : &ssl_cert_file,
530 0 : ssl_cert_reload_period,
531 0 : )
532 0 : .await
533 0 : })
534 0 : .await??;
535 :
536 0 : let config = rustls::ServerConfig::builder()
537 0 : .with_no_client_auth()
538 0 : .with_cert_resolver(cert_resolver);
539 0 :
540 0 : Some(Arc::new(config))
541 : } else {
542 0 : None
543 : };
544 :
545 0 : let wal_service_handle = current_thread_rt
546 0 : .as_ref()
547 0 : .unwrap_or_else(|| WAL_SERVICE_RUNTIME.handle())
548 0 : .spawn(wal_service::task_main(
549 0 : conf.clone(),
550 0 : pg_listener,
551 0 : Scope::SafekeeperData,
552 0 : conf.enable_tls_wal_service_api
553 0 : .then(|| tls_server_config.clone())
554 0 : .flatten(),
555 0 : global_timelines.clone(),
556 0 : ))
557 0 : // wrap with task name for error reporting
558 0 : .map(|res| ("WAL service main".to_owned(), res));
559 0 : tasks_handles.push(Box::pin(wal_service_handle));
560 0 :
561 0 : let global_timelines_ = global_timelines.clone();
562 0 : let timeline_housekeeping_handle = current_thread_rt
563 0 : .as_ref()
564 0 : .unwrap_or_else(|| WAL_SERVICE_RUNTIME.handle())
565 0 : .spawn(async move {
566 : const TOMBSTONE_TTL: Duration = Duration::from_secs(3600 * 24);
567 : loop {
568 0 : tokio::time::sleep(TOMBSTONE_TTL).await;
569 0 : global_timelines_.housekeeping(&TOMBSTONE_TTL);
570 : }
571 0 : })
572 0 : .map(|res| ("Timeline map housekeeping".to_owned(), res));
573 0 : tasks_handles.push(Box::pin(timeline_housekeeping_handle));
574 :
575 0 : if let Some(pg_listener_tenant_only) = pg_listener_tenant_only {
576 0 : let wal_service_handle = current_thread_rt
577 0 : .as_ref()
578 0 : .unwrap_or_else(|| WAL_SERVICE_RUNTIME.handle())
579 0 : .spawn(wal_service::task_main(
580 0 : conf.clone(),
581 0 : pg_listener_tenant_only,
582 0 : Scope::Tenant,
583 0 : conf.enable_tls_wal_service_api
584 0 : .then(|| tls_server_config.clone())
585 0 : .flatten(),
586 0 : global_timelines.clone(),
587 0 : ))
588 0 : // wrap with task name for error reporting
589 0 : .map(|res| ("WAL service tenant only main".to_owned(), res));
590 0 : tasks_handles.push(Box::pin(wal_service_handle));
591 0 : }
592 :
593 0 : let http_handle = current_thread_rt
594 0 : .as_ref()
595 0 : .unwrap_or_else(|| HTTP_RUNTIME.handle())
596 0 : .spawn(http::task_main_http(
597 0 : conf.clone(),
598 0 : http_listener,
599 0 : global_timelines.clone(),
600 0 : ))
601 0 : .map(|res| ("HTTP service main".to_owned(), res));
602 0 : tasks_handles.push(Box::pin(http_handle));
603 :
604 0 : if let Some(https_listener) = https_listener {
605 0 : let https_handle = current_thread_rt
606 0 : .as_ref()
607 0 : .unwrap_or_else(|| HTTP_RUNTIME.handle())
608 0 : .spawn(http::task_main_https(
609 0 : conf.clone(),
610 0 : https_listener,
611 0 : tls_server_config.expect("tls_server_config is set earlier if https is enabled"),
612 0 : global_timelines.clone(),
613 0 : ))
614 0 : .map(|res| ("HTTPS service main".to_owned(), res));
615 0 : tasks_handles.push(Box::pin(https_handle));
616 0 : }
617 :
618 0 : let broker_task_handle = current_thread_rt
619 0 : .as_ref()
620 0 : .unwrap_or_else(|| BROKER_RUNTIME.handle())
621 0 : .spawn(
622 0 : broker::task_main(conf.clone(), global_timelines.clone())
623 0 : .instrument(info_span!("broker")),
624 : )
625 0 : .map(|res| ("broker main".to_owned(), res));
626 0 : tasks_handles.push(Box::pin(broker_task_handle));
627 0 :
628 0 : set_build_info_metric(GIT_VERSION, BUILD_TAG);
629 :
630 : // TODO: update tokio-stream, convert to real async Stream with
631 : // SignalStream, map it to obtain missing signal name, combine streams into
632 : // single stream we can easily sit on.
633 0 : let mut sigquit_stream = signal(SignalKind::quit())?;
634 0 : let mut sigint_stream = signal(SignalKind::interrupt())?;
635 0 : let mut sigterm_stream = signal(SignalKind::terminate())?;
636 :
637 : // Notify systemd that we are ready. This is important as currently loading
638 : // timelines takes significant time (~30s in busy regions).
639 0 : if let Err(e) = sd_notify::notify(true, &[NotifyState::Ready]) {
640 0 : warn!("systemd notify failed: {:?}", e);
641 0 : }
642 :
643 0 : tokio::select! {
644 0 : Some((task_name, res)) = tasks_handles.next()=> {
645 0 : error!("{} task failed: {:?}, exiting", task_name, res);
646 0 : std::process::exit(1);
647 : }
648 : // On any shutdown signal, log receival and exit. Additionally, handling
649 : // SIGQUIT prevents coredump.
650 0 : _ = sigquit_stream.recv() => info!("received SIGQUIT, terminating"),
651 0 : _ = sigint_stream.recv() => info!("received SIGINT, terminating"),
652 0 : _ = sigterm_stream.recv() => info!("received SIGTERM, terminating")
653 :
654 : };
655 0 : std::process::exit(0);
656 0 : }
657 :
658 : /// Determine safekeeper id.
659 0 : fn set_id(workdir: &Utf8Path, given_id: Option<NodeId>) -> Result<NodeId> {
660 0 : let id_file_path = workdir.join(ID_FILE_NAME);
661 0 :
662 0 : let my_id: NodeId;
663 0 : // If file with ID exists, read it in; otherwise set one passed.
664 0 : match fs::read(&id_file_path) {
665 0 : Ok(id_serialized) => {
666 0 : my_id = NodeId(
667 0 : std::str::from_utf8(&id_serialized)
668 0 : .context("failed to parse safekeeper id")?
669 0 : .parse()
670 0 : .context("failed to parse safekeeper id")?,
671 : );
672 0 : if let Some(given_id) = given_id {
673 0 : if given_id != my_id {
674 0 : bail!(
675 0 : "safekeeper already initialized with id {}, can't set {}",
676 0 : my_id,
677 0 : given_id
678 0 : );
679 0 : }
680 0 : }
681 0 : info!("safekeeper ID {}", my_id);
682 : }
683 0 : Err(error) => match error.kind() {
684 : ErrorKind::NotFound => {
685 0 : my_id = if let Some(given_id) = given_id {
686 0 : given_id
687 : } else {
688 0 : bail!("safekeeper id is not specified");
689 : };
690 0 : let mut f = File::create(&id_file_path)
691 0 : .with_context(|| format!("Failed to create id file at {id_file_path:?}"))?;
692 0 : f.write_all(my_id.to_string().as_bytes())?;
693 0 : f.sync_all()?;
694 0 : info!("initialized safekeeper id {}", my_id);
695 : }
696 : _ => {
697 0 : return Err(error.into());
698 : }
699 : },
700 : }
701 0 : Ok(my_id)
702 0 : }
703 :
704 0 : fn parse_remote_storage(storage_conf: &str) -> anyhow::Result<RemoteStorageConfig> {
705 0 : RemoteStorageConfig::from_toml(&storage_conf.parse()?)
706 0 : }
707 :
708 : #[test]
709 1 : fn verify_cli() {
710 : use clap::CommandFactory;
711 1 : Args::command().debug_assert()
712 1 : }
|
