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,
21 : DEFAULT_GLOBAL_DISK_CHECK_INTERVAL, DEFAULT_HEARTBEAT_TIMEOUT, DEFAULT_HTTP_LISTEN_ADDR,
22 : DEFAULT_MAX_GLOBAL_DISK_USAGE_RATIO, DEFAULT_MAX_OFFLOADER_LAG_BYTES,
23 : DEFAULT_MAX_REELECT_OFFLOADER_LAG_BYTES, DEFAULT_MAX_TIMELINE_DISK_USAGE_BYTES,
24 : DEFAULT_PARTIAL_BACKUP_CONCURRENCY, DEFAULT_PARTIAL_BACKUP_TIMEOUT, DEFAULT_PG_LISTEN_ADDR,
25 : DEFAULT_SSL_CERT_FILE, DEFAULT_SSL_CERT_RELOAD_PERIOD, DEFAULT_SSL_KEY_FILE,
26 : };
27 : use safekeeper::hadron;
28 : use safekeeper::wal_backup::WalBackup;
29 : use safekeeper::{
30 : BACKGROUND_RUNTIME, BROKER_RUNTIME, GlobalTimelines, HTTP_RUNTIME, SafeKeeperConf,
31 : WAL_SERVICE_RUNTIME, broker, control_file, http, wal_service,
32 : };
33 : use sd_notify::NotifyState;
34 : use storage_broker::{DEFAULT_ENDPOINT, Uri};
35 : use tokio::runtime::Handle;
36 : use tokio::signal::unix::{SignalKind, signal};
37 : use tokio::task::JoinError;
38 : use tracing::*;
39 : use utils::auth::{JwtAuth, Scope, SwappableJwtAuth};
40 : use utils::id::NodeId;
41 : use utils::logging::{self, LogFormat, SecretString};
42 : use utils::metrics_collector::{METRICS_COLLECTION_INTERVAL, METRICS_COLLECTOR};
43 : use utils::sentry_init::init_sentry;
44 : use utils::{pid_file, project_build_tag, project_git_version, tcp_listener};
45 :
46 : use safekeeper::hadron::{
47 : GLOBAL_DISK_LIMIT_EXCEEDED, get_filesystem_capacity, get_filesystem_usage,
48 : };
49 : use safekeeper::metrics::GLOBAL_DISK_UTIL_CHECK_SECONDS;
50 : use std::sync::atomic::Ordering;
51 :
52 : #[global_allocator]
53 : static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
54 :
55 : /// Configure jemalloc to profile heap allocations by sampling stack traces every 2 MB (1 << 21).
56 : /// This adds roughly 3% overhead for allocations on average, which is acceptable considering
57 : /// performance-sensitive code will avoid allocations as far as possible anyway.
58 : #[allow(non_upper_case_globals)]
59 : #[unsafe(export_name = "malloc_conf")]
60 : pub static malloc_conf: &[u8] = b"prof:true,prof_active:true,lg_prof_sample:21\0";
61 :
62 : const PID_FILE_NAME: &str = "safekeeper.pid";
63 : const ID_FILE_NAME: &str = "safekeeper.id";
64 :
65 : project_git_version!(GIT_VERSION);
66 : project_build_tag!(BUILD_TAG);
67 :
68 : const FEATURES: &[&str] = &[
69 : #[cfg(feature = "testing")]
70 : "testing",
71 : ];
72 :
73 0 : fn version() -> String {
74 0 : format!(
75 0 : "{GIT_VERSION} failpoints: {}, features: {:?}",
76 0 : fail::has_failpoints(),
77 : FEATURES,
78 : )
79 0 : }
80 :
81 : const ABOUT: &str = r#"
82 : A fleet of safekeepers is responsible for reliably storing WAL received from
83 : compute, passing it through consensus (mitigating potential computes brain
84 : split), and serving the hardened part further downstream to pageserver(s).
85 : "#;
86 :
87 : #[derive(Parser)]
88 : #[command(name = "Neon safekeeper", version = GIT_VERSION, about = ABOUT, long_about = None)]
89 : struct Args {
90 : /// Path to the safekeeper data directory.
91 : #[arg(short = 'D', long, default_value = "./")]
92 : datadir: Utf8PathBuf,
93 : /// Safekeeper node id.
94 : #[arg(long)]
95 : id: Option<u64>,
96 : /// Initialize safekeeper with given id and exit.
97 : #[arg(long)]
98 : init: bool,
99 : /// Listen endpoint for receiving/sending WAL in the form host:port.
100 : #[arg(short, long, default_value = DEFAULT_PG_LISTEN_ADDR)]
101 : listen_pg: String,
102 : /// Listen endpoint for receiving/sending WAL in the form host:port allowing
103 : /// only tenant scoped auth tokens. Pointless if auth is disabled.
104 : #[arg(long, default_value = None, verbatim_doc_comment)]
105 : listen_pg_tenant_only: Option<String>,
106 : /// Listen http endpoint for management and metrics in the form host:port.
107 : #[arg(long, default_value = DEFAULT_HTTP_LISTEN_ADDR)]
108 : listen_http: String,
109 : /// Listen https endpoint for management and metrics in the form host:port.
110 : #[arg(long, default_value = None)]
111 : listen_https: Option<String>,
112 : /// Advertised endpoint for receiving/sending WAL in the form host:port. If not
113 : /// specified, listen_pg is used to advertise instead.
114 : #[arg(long, default_value = None)]
115 : advertise_pg: Option<String>,
116 : /// Availability zone of the safekeeper.
117 : #[arg(long)]
118 : availability_zone: Option<String>,
119 : /// Do not wait for changes to be written safely to disk. Unsafe.
120 : #[arg(short, long)]
121 : no_sync: bool,
122 : /// Dump control file at path specified by this argument and exit.
123 : #[arg(long)]
124 : dump_control_file: Option<Utf8PathBuf>,
125 : /// Broker endpoint for storage nodes coordination in the form
126 : /// http[s]://host:port. In case of https schema TLS is connection is
127 : /// established; plaintext otherwise.
128 : #[arg(long, default_value = DEFAULT_ENDPOINT, verbatim_doc_comment)]
129 : broker_endpoint: Uri,
130 : /// Broker keepalive interval.
131 : #[arg(long, value_parser= humantime::parse_duration, default_value = storage_broker::DEFAULT_KEEPALIVE_INTERVAL)]
132 : broker_keepalive_interval: Duration,
133 : /// Peer safekeeper is considered dead after not receiving heartbeats from
134 : /// it during this period passed as a human readable duration.
135 : #[arg(long, value_parser= humantime::parse_duration, default_value = DEFAULT_HEARTBEAT_TIMEOUT, verbatim_doc_comment)]
136 : heartbeat_timeout: Duration,
137 : /// Enable/disable peer recovery.
138 : #[arg(long, default_value = "false", action=ArgAction::Set)]
139 : peer_recovery: bool,
140 : /// Remote storage configuration for WAL backup (offloading to s3) as TOML
141 : /// inline table, e.g.
142 : /// {max_concurrent_syncs = 17, max_sync_errors = 13, bucket_name = "<BUCKETNAME>", bucket_region = "<REGION>", concurrency_limit = 119}
143 : /// Safekeeper offloads WAL to
144 : /// [prefix_in_bucket/]<tenant_id>/<timeline_id>/<segment_file>, mirroring
145 : /// structure on the file system.
146 : #[arg(long, value_parser = parse_remote_storage, verbatim_doc_comment)]
147 : remote_storage: Option<RemoteStorageConfig>,
148 : /// Safekeeper won't be elected for WAL offloading if it is lagging for more than this value in bytes
149 : #[arg(long, default_value_t = DEFAULT_MAX_OFFLOADER_LAG_BYTES)]
150 : max_offloader_lag: u64,
151 : /* BEGIN_HADRON */
152 : /// Safekeeper will re-elect a new offloader if the current backup lagging for more than this value in bytes
153 : #[arg(long, default_value_t = DEFAULT_MAX_REELECT_OFFLOADER_LAG_BYTES)]
154 : max_reelect_offloader_lag_bytes: u64,
155 : /// Safekeeper will stop accepting new WALs if the timeline disk usage exceeds this value in bytes.
156 : /// Setting this value to 0 disables the limit.
157 : #[arg(long, default_value_t = DEFAULT_MAX_TIMELINE_DISK_USAGE_BYTES)]
158 : max_timeline_disk_usage_bytes: u64,
159 : /* END_HADRON */
160 : /// Number of max parallel WAL segments to be offloaded to remote storage.
161 : #[arg(long, default_value = "5")]
162 : wal_backup_parallel_jobs: usize,
163 : /// Disable WAL backup to s3. When disabled, safekeeper removes WAL ignoring
164 : /// WAL backup horizon.
165 : #[arg(long)]
166 : disable_wal_backup: bool,
167 : /// If given, enables auth on incoming connections to WAL service endpoint
168 : /// (--listen-pg). Value specifies path to a .pem public key used for
169 : /// validations of JWT tokens. Empty string is allowed and means disabling
170 : /// auth.
171 : #[arg(long, verbatim_doc_comment, value_parser = opt_pathbuf_parser)]
172 : pg_auth_public_key_path: Option<Utf8PathBuf>,
173 : /// If given, enables auth on incoming connections to tenant only WAL
174 : /// service endpoint (--listen-pg-tenant-only). Value specifies path to a
175 : /// .pem public key used for validations of JWT tokens. Empty string is
176 : /// allowed and means disabling auth.
177 : #[arg(long, verbatim_doc_comment, value_parser = opt_pathbuf_parser)]
178 : pg_tenant_only_auth_public_key_path: Option<Utf8PathBuf>,
179 : /// If given, enables auth on incoming connections to http management
180 : /// service endpoint (--listen-http). Value specifies path to a .pem public
181 : /// key used for validations of JWT tokens. Empty string is allowed and
182 : /// means disabling auth.
183 : #[arg(long, verbatim_doc_comment, value_parser = opt_pathbuf_parser)]
184 : http_auth_public_key_path: Option<Utf8PathBuf>,
185 : /// Format for logging, either 'plain' or 'json'.
186 : #[arg(long, default_value = "plain")]
187 : log_format: String,
188 : /// Run everything in single threaded current thread runtime, might be
189 : /// useful for debugging.
190 : #[arg(long)]
191 : current_thread_runtime: bool,
192 : /// Keep horizon for walsenders, i.e. don't remove WAL segments that are
193 : /// still needed for existing replication connection.
194 : #[arg(long)]
195 : walsenders_keep_horizon: bool,
196 : /// Controls how long backup will wait until uploading the partial segment.
197 : #[arg(long, value_parser = humantime::parse_duration, default_value = DEFAULT_PARTIAL_BACKUP_TIMEOUT, verbatim_doc_comment)]
198 : partial_backup_timeout: Duration,
199 : /// Disable task to push messages to broker every second. Supposed to
200 : /// be used in tests.
201 : #[arg(long)]
202 : disable_periodic_broker_push: bool,
203 : /// Enable automatic switching to offloaded state.
204 : #[arg(long)]
205 : enable_offload: bool,
206 : /// Delete local WAL files after offloading. When disabled, they will be left on disk.
207 : #[arg(long)]
208 : delete_offloaded_wal: bool,
209 : /// Pending updates to control file will be automatically saved after this interval.
210 : #[arg(long, value_parser = humantime::parse_duration, default_value = DEFAULT_CONTROL_FILE_SAVE_INTERVAL)]
211 : control_file_save_interval: Duration,
212 : /// Number of allowed concurrent uploads of partial segments to remote storage.
213 : #[arg(long, default_value = DEFAULT_PARTIAL_BACKUP_CONCURRENCY)]
214 : partial_backup_concurrency: usize,
215 : /// How long a timeline must be resident before it is eligible for eviction.
216 : /// Usually, timeline eviction has to wait for `partial_backup_timeout` before being eligible for eviction,
217 : /// but if a timeline is un-evicted and then _not_ written to, it would immediately flap to evicting again,
218 : /// if it weren't for `eviction_min_resident` preventing that.
219 : ///
220 : /// Also defines interval for eviction retries.
221 : #[arg(long, value_parser = humantime::parse_duration, default_value = DEFAULT_EVICTION_MIN_RESIDENT)]
222 : eviction_min_resident: Duration,
223 : /// Enable fanning out WAL to different shards from the same reader
224 : #[arg(long)]
225 : wal_reader_fanout: bool,
226 : /// Only fan out the WAL reader if the absoulte delta between the new requested position
227 : /// and the current position of the reader is smaller than this value.
228 : #[arg(long)]
229 : max_delta_for_fanout: Option<u64>,
230 : /// Path to a file with certificate's private key for https API.
231 : #[arg(long, default_value = DEFAULT_SSL_KEY_FILE)]
232 : ssl_key_file: Utf8PathBuf,
233 : /// Path to a file with a X509 certificate for https API.
234 : #[arg(long, default_value = DEFAULT_SSL_CERT_FILE)]
235 : ssl_cert_file: Utf8PathBuf,
236 : /// Period to reload certificate and private key from files.
237 : #[arg(long, value_parser = humantime::parse_duration, default_value = DEFAULT_SSL_CERT_RELOAD_PERIOD)]
238 : ssl_cert_reload_period: Duration,
239 : /// Trusted root CA certificates to use in https APIs.
240 : #[arg(long)]
241 : ssl_ca_file: Option<Utf8PathBuf>,
242 : /// Flag to use https for requests to peer's safekeeper API.
243 : #[arg(long)]
244 : use_https_safekeeper_api: bool,
245 : /// Path to the JWT auth token used to authenticate with other safekeepers.
246 : #[arg(long)]
247 : auth_token_path: Option<Utf8PathBuf>,
248 :
249 : /// Enable TLS in WAL service API.
250 : /// Does not force TLS: the client negotiates TLS usage during the handshake.
251 : /// Uses key and certificate from ssl_key_file/ssl_cert_file.
252 : #[arg(long)]
253 : enable_tls_wal_service_api: bool,
254 :
255 : /// Controls whether to collect all metrics on each scrape or to return potentially stale
256 : /// results.
257 : #[arg(long, default_value_t = true)]
258 : force_metric_collection_on_scrape: bool,
259 :
260 : /// Run in development mode (disables security checks)
261 : #[arg(long, help = "Run in development mode (disables security checks)")]
262 : dev: bool,
263 : /* BEGIN_HADRON */
264 : #[arg(long)]
265 : enable_pull_timeline_on_startup: bool,
266 : /// How often to scan entire data-dir for total disk usage
267 : #[arg(long, value_parser=humantime::parse_duration, default_value = DEFAULT_GLOBAL_DISK_CHECK_INTERVAL)]
268 : global_disk_check_interval: Duration,
269 : /// The portion of the filesystem capacity that can be used by all timelines.
270 : /// A circuit breaker will trip and reject all WAL writes if the total usage
271 : /// exceeds this ratio.
272 : /// Set to 0 to disable the global disk usage limit.
273 : #[arg(long, default_value_t = DEFAULT_MAX_GLOBAL_DISK_USAGE_RATIO)]
274 : max_global_disk_usage_ratio: f64,
275 : /* END_HADRON */
276 : }
277 :
278 : // Like PathBufValueParser, but allows empty string.
279 0 : fn opt_pathbuf_parser(s: &str) -> Result<Utf8PathBuf, String> {
280 0 : Ok(Utf8PathBuf::from_str(s).unwrap())
281 0 : }
282 :
283 : #[tokio::main(flavor = "current_thread")]
284 0 : async fn main() -> anyhow::Result<()> {
285 : // We want to allow multiple occurences of the same arg (taking the last) so
286 : // that neon_local could generate command with defaults + overrides without
287 : // getting 'argument cannot be used multiple times' error. This seems to be
288 : // impossible with pure Derive API, so convert struct to Command, modify it,
289 : // parse arguments, and then fill the struct back.
290 0 : let cmd = <Args as clap::CommandFactory>::command()
291 0 : .args_override_self(true)
292 0 : .version(version());
293 0 : let mut matches = cmd.get_matches();
294 0 : let mut args = <Args as clap::FromArgMatches>::from_arg_matches_mut(&mut matches)?;
295 :
296 : // I failed to modify opt_pathbuf_parser to return Option<PathBuf> in
297 : // reasonable time, so turn empty string into option post factum.
298 0 : if let Some(pb) = &args.pg_auth_public_key_path {
299 0 : if pb.as_os_str().is_empty() {
300 0 : args.pg_auth_public_key_path = None;
301 0 : }
302 0 : }
303 0 : if let Some(pb) = &args.pg_tenant_only_auth_public_key_path {
304 0 : if pb.as_os_str().is_empty() {
305 0 : args.pg_tenant_only_auth_public_key_path = None;
306 0 : }
307 0 : }
308 0 : if let Some(pb) = &args.http_auth_public_key_path {
309 0 : if pb.as_os_str().is_empty() {
310 0 : args.http_auth_public_key_path = None;
311 0 : }
312 0 : }
313 :
314 0 : if let Some(addr) = args.dump_control_file {
315 0 : let state = control_file::FileStorage::load_control_file(addr)?;
316 0 : let json = serde_json::to_string(&state)?;
317 0 : print!("{json}");
318 0 : return Ok(());
319 0 : }
320 :
321 : // important to keep the order of:
322 : // 1. init logging
323 : // 2. tracing panic hook
324 : // 3. sentry
325 0 : logging::init(
326 0 : LogFormat::from_config(&args.log_format)?,
327 0 : logging::TracingErrorLayerEnablement::Disabled,
328 0 : logging::Output::Stdout,
329 0 : )?;
330 0 : logging::replace_panic_hook_with_tracing_panic_hook().forget();
331 0 : info!("version: {GIT_VERSION}");
332 0 : info!("buld_tag: {BUILD_TAG}");
333 :
334 0 : let args_workdir = &args.datadir;
335 0 : let workdir = args_workdir.canonicalize_utf8().with_context(|| {
336 0 : format!("Failed to get the absolute path for input workdir {args_workdir:?}")
337 0 : })?;
338 :
339 : // Change into the data directory.
340 0 : std::env::set_current_dir(&workdir)?;
341 :
342 : // Prevent running multiple safekeepers on the same directory
343 0 : let lock_file_path = workdir.join(PID_FILE_NAME);
344 0 : let lock_file =
345 0 : pid_file::claim_for_current_process(&lock_file_path).context("claim pid file")?;
346 0 : info!("claimed pid file at {lock_file_path:?}");
347 : // ensure that the lock file is held even if the main thread of the process is panics
348 : // we need to release the lock file only when the current process is gone
349 0 : std::mem::forget(lock_file);
350 :
351 : // Set or read our ID.
352 0 : let id = set_id(&workdir, args.id.map(NodeId))?;
353 0 : if args.init {
354 0 : return Ok(());
355 0 : }
356 :
357 0 : let pg_auth = match args.pg_auth_public_key_path.as_ref() {
358 : None => {
359 0 : info!("pg auth is disabled");
360 0 : None
361 : }
362 0 : Some(path) => {
363 0 : info!("loading pg auth JWT key from {path}");
364 0 : Some(Arc::new(
365 0 : JwtAuth::from_key_path(path).context("failed to load the auth key")?,
366 : ))
367 : }
368 : };
369 0 : let pg_tenant_only_auth = match args.pg_tenant_only_auth_public_key_path.as_ref() {
370 : None => {
371 0 : info!("pg tenant only auth is disabled");
372 0 : None
373 : }
374 0 : Some(path) => {
375 0 : info!("loading pg tenant only auth JWT key from {path}");
376 0 : Some(Arc::new(
377 0 : JwtAuth::from_key_path(path).context("failed to load the auth key")?,
378 : ))
379 : }
380 : };
381 0 : let http_auth = match args.http_auth_public_key_path.as_ref() {
382 : None => {
383 0 : info!("http auth is disabled");
384 0 : None
385 : }
386 0 : Some(path) => {
387 0 : info!("loading http auth JWT key(s) from {path}");
388 0 : let jwt_auth = JwtAuth::from_key_path(path).context("failed to load the auth key")?;
389 0 : Some(Arc::new(SwappableJwtAuth::new(jwt_auth)))
390 : }
391 : };
392 :
393 : // Load JWT auth token to connect to other safekeepers for pull_timeline.
394 0 : let sk_auth_token = if let Some(auth_token_path) = args.auth_token_path.as_ref() {
395 0 : info!("loading JWT token for authentication with safekeepers from {auth_token_path}");
396 0 : let auth_token = tokio::fs::read_to_string(auth_token_path).await?;
397 0 : Some(SecretString::from(auth_token.trim().to_owned()))
398 : } else {
399 0 : info!("no JWT token for authentication with safekeepers detected");
400 0 : None
401 : };
402 :
403 0 : let ssl_ca_certs = match args.ssl_ca_file.as_ref() {
404 0 : Some(ssl_ca_file) => {
405 0 : tracing::info!("Using ssl root CA file: {ssl_ca_file:?}");
406 0 : let buf = tokio::fs::read(ssl_ca_file).await?;
407 0 : pem::parse_many(&buf)?
408 0 : .into_iter()
409 0 : .filter(|pem| pem.tag() == "CERTIFICATE")
410 0 : .collect()
411 : }
412 0 : None => Vec::new(),
413 : };
414 :
415 0 : let conf = Arc::new(SafeKeeperConf {
416 0 : workdir,
417 0 : my_id: id,
418 0 : listen_pg_addr: args.listen_pg,
419 0 : listen_pg_addr_tenant_only: args.listen_pg_tenant_only,
420 0 : listen_http_addr: args.listen_http,
421 0 : listen_https_addr: args.listen_https,
422 0 : advertise_pg_addr: args.advertise_pg,
423 0 : availability_zone: args.availability_zone,
424 0 : no_sync: args.no_sync,
425 0 : broker_endpoint: args.broker_endpoint,
426 0 : broker_keepalive_interval: args.broker_keepalive_interval,
427 0 : heartbeat_timeout: args.heartbeat_timeout,
428 0 : peer_recovery_enabled: args.peer_recovery,
429 0 : remote_storage: args.remote_storage,
430 0 : max_offloader_lag_bytes: args.max_offloader_lag,
431 0 : /* BEGIN_HADRON */
432 0 : max_reelect_offloader_lag_bytes: args.max_reelect_offloader_lag_bytes,
433 0 : max_timeline_disk_usage_bytes: args.max_timeline_disk_usage_bytes,
434 0 : /* END_HADRON */
435 0 : wal_backup_enabled: !args.disable_wal_backup,
436 0 : backup_parallel_jobs: args.wal_backup_parallel_jobs,
437 0 : pg_auth,
438 0 : pg_tenant_only_auth,
439 0 : http_auth,
440 0 : sk_auth_token,
441 0 : current_thread_runtime: args.current_thread_runtime,
442 0 : walsenders_keep_horizon: args.walsenders_keep_horizon,
443 0 : partial_backup_timeout: args.partial_backup_timeout,
444 0 : disable_periodic_broker_push: args.disable_periodic_broker_push,
445 0 : enable_offload: args.enable_offload,
446 0 : delete_offloaded_wal: args.delete_offloaded_wal,
447 0 : control_file_save_interval: args.control_file_save_interval,
448 0 : partial_backup_concurrency: args.partial_backup_concurrency,
449 0 : eviction_min_resident: args.eviction_min_resident,
450 0 : wal_reader_fanout: args.wal_reader_fanout,
451 0 : max_delta_for_fanout: args.max_delta_for_fanout,
452 0 : ssl_key_file: args.ssl_key_file,
453 0 : ssl_cert_file: args.ssl_cert_file,
454 0 : ssl_cert_reload_period: args.ssl_cert_reload_period,
455 0 : ssl_ca_certs,
456 0 : use_https_safekeeper_api: args.use_https_safekeeper_api,
457 0 : enable_tls_wal_service_api: args.enable_tls_wal_service_api,
458 0 : force_metric_collection_on_scrape: args.force_metric_collection_on_scrape,
459 0 : /* BEGIN_HADRON */
460 0 : advertise_pg_addr_tenant_only: None,
461 0 : enable_pull_timeline_on_startup: args.enable_pull_timeline_on_startup,
462 0 : hcc_base_url: None,
463 0 : global_disk_check_interval: args.global_disk_check_interval,
464 0 : max_global_disk_usage_ratio: args.max_global_disk_usage_ratio,
465 0 : /* END_HADRON */
466 0 : });
467 :
468 : // initialize sentry if SENTRY_DSN is provided
469 0 : let _sentry_guard = init_sentry(
470 0 : Some(GIT_VERSION.into()),
471 0 : &[("node_id", &conf.my_id.to_string())],
472 : );
473 0 : start_safekeeper(conf).await
474 0 : }
475 :
476 : /// Result of joining any of main tasks: upper error means task failed to
477 : /// complete, e.g. panicked, inner is error produced by task itself.
478 : type JoinTaskRes = Result<anyhow::Result<()>, JoinError>;
479 :
480 0 : async fn start_safekeeper(conf: Arc<SafeKeeperConf>) -> Result<()> {
481 : // fsync the datadir to make sure we have a consistent state on disk.
482 0 : if !conf.no_sync {
483 0 : let dfd = File::open(&conf.workdir).context("open datadir for syncfs")?;
484 0 : let started = Instant::now();
485 0 : utils::crashsafe::syncfs(dfd)?;
486 0 : let elapsed = started.elapsed();
487 0 : info!(
488 0 : elapsed_ms = elapsed.as_millis(),
489 0 : "syncfs data directory done"
490 : );
491 0 : }
492 :
493 0 : info!("starting safekeeper WAL service on {}", conf.listen_pg_addr);
494 0 : let pg_listener = tcp_listener::bind(conf.listen_pg_addr.clone()).map_err(|e| {
495 0 : error!("failed to bind to address {}: {}", conf.listen_pg_addr, e);
496 0 : e
497 0 : })?;
498 :
499 0 : let pg_listener_tenant_only =
500 0 : if let Some(listen_pg_addr_tenant_only) = &conf.listen_pg_addr_tenant_only {
501 0 : info!(
502 0 : "starting safekeeper tenant scoped WAL service on {}",
503 : listen_pg_addr_tenant_only
504 : );
505 0 : let listener = tcp_listener::bind(listen_pg_addr_tenant_only.clone()).map_err(|e| {
506 0 : error!(
507 0 : "failed to bind to address {}: {}",
508 : listen_pg_addr_tenant_only, e
509 : );
510 0 : e
511 0 : })?;
512 0 : Some(listener)
513 : } else {
514 0 : None
515 : };
516 :
517 0 : info!(
518 0 : "starting safekeeper HTTP service on {}",
519 0 : conf.listen_http_addr
520 : );
521 0 : let http_listener = tcp_listener::bind(conf.listen_http_addr.clone()).map_err(|e| {
522 0 : error!("failed to bind to address {}: {}", conf.listen_http_addr, e);
523 0 : e
524 0 : })?;
525 :
526 0 : let https_listener = match conf.listen_https_addr.as_ref() {
527 0 : Some(listen_https_addr) => {
528 0 : info!("starting safekeeper HTTPS service on {}", listen_https_addr);
529 0 : Some(tcp_listener::bind(listen_https_addr).map_err(|e| {
530 0 : error!("failed to bind to address {}: {}", listen_https_addr, e);
531 0 : e
532 0 : })?)
533 : }
534 0 : None => None,
535 : };
536 :
537 0 : let wal_backup = Arc::new(WalBackup::new(&conf).await?);
538 :
539 0 : let global_timelines = Arc::new(GlobalTimelines::new(conf.clone(), wal_backup.clone()));
540 :
541 : // Register metrics collector for active timelines. It's important to do this
542 : // after daemonizing, otherwise process collector will be upset.
543 0 : let timeline_collector = safekeeper::metrics::TimelineCollector::new(global_timelines.clone());
544 0 : metrics::register_internal(Box::new(timeline_collector))?;
545 :
546 : // Keep handles to main tasks to die if any of them disappears.
547 0 : let mut tasks_handles: FuturesUnordered<BoxFuture<(String, JoinTaskRes)>> =
548 0 : FuturesUnordered::new();
549 :
550 : // Start wal backup launcher before loading timelines as we'll notify it
551 : // through the channel about timelines which need offloading, not draining
552 : // the channel would cause deadlock.
553 0 : let current_thread_rt = conf
554 0 : .current_thread_runtime
555 0 : .then(|| Handle::try_current().expect("no runtime in main"));
556 :
557 : // Load all timelines from disk to memory.
558 0 : global_timelines.init().await?;
559 :
560 : /* BEGIN_HADRON */
561 0 : if conf.enable_pull_timeline_on_startup && global_timelines.timelines_count() == 0 {
562 0 : match hadron::hcc_pull_timelines(&conf, global_timelines.clone()).await {
563 : Ok(_) => {
564 0 : info!("Successfully pulled all timelines from peer safekeepers");
565 : }
566 0 : Err(e) => {
567 0 : error!("Failed to pull timelines from peer safekeepers: {:?}", e);
568 0 : return Err(e);
569 : }
570 : }
571 0 : }
572 : /* END_HADRON */
573 :
574 : // Run everything in current thread rt, if asked.
575 0 : if conf.current_thread_runtime {
576 0 : info!("running in current thread runtime");
577 0 : }
578 :
579 0 : let tls_server_config = if conf.listen_https_addr.is_some() || conf.enable_tls_wal_service_api {
580 0 : let ssl_key_file = conf.ssl_key_file.clone();
581 0 : let ssl_cert_file = conf.ssl_cert_file.clone();
582 0 : let ssl_cert_reload_period = conf.ssl_cert_reload_period;
583 :
584 : // Create resolver in BACKGROUND_RUNTIME, so the background certificate reloading
585 : // task is run in this runtime.
586 0 : let cert_resolver = current_thread_rt
587 0 : .as_ref()
588 0 : .unwrap_or_else(|| BACKGROUND_RUNTIME.handle())
589 0 : .spawn(async move {
590 0 : ReloadingCertificateResolver::new(
591 0 : "main",
592 0 : &ssl_key_file,
593 0 : &ssl_cert_file,
594 0 : ssl_cert_reload_period,
595 0 : )
596 0 : .await
597 0 : })
598 0 : .await??;
599 :
600 0 : let config = rustls::ServerConfig::builder()
601 0 : .with_no_client_auth()
602 0 : .with_cert_resolver(cert_resolver);
603 :
604 0 : Some(Arc::new(config))
605 : } else {
606 0 : None
607 : };
608 :
609 0 : let wal_service_handle = current_thread_rt
610 0 : .as_ref()
611 0 : .unwrap_or_else(|| WAL_SERVICE_RUNTIME.handle())
612 0 : .spawn(wal_service::task_main(
613 0 : conf.clone(),
614 0 : pg_listener,
615 0 : Scope::SafekeeperData,
616 0 : conf.enable_tls_wal_service_api
617 0 : .then(|| tls_server_config.clone())
618 0 : .flatten(),
619 0 : global_timelines.clone(),
620 : ))
621 : // wrap with task name for error reporting
622 0 : .map(|res| ("WAL service main".to_owned(), res));
623 0 : tasks_handles.push(Box::pin(wal_service_handle));
624 :
625 0 : let global_timelines_ = global_timelines.clone();
626 0 : let timeline_housekeeping_handle = current_thread_rt
627 0 : .as_ref()
628 0 : .unwrap_or_else(|| WAL_SERVICE_RUNTIME.handle())
629 0 : .spawn(async move {
630 : const TOMBSTONE_TTL: Duration = Duration::from_secs(3600 * 24);
631 : loop {
632 0 : tokio::time::sleep(TOMBSTONE_TTL).await;
633 0 : global_timelines_.housekeeping(&TOMBSTONE_TTL);
634 : }
635 : })
636 0 : .map(|res| ("Timeline map housekeeping".to_owned(), res));
637 0 : tasks_handles.push(Box::pin(timeline_housekeeping_handle));
638 :
639 : /* BEGIN_HADRON */
640 : // Spawn global disk usage watcher task, if a global disk usage limit is specified.
641 0 : let interval = conf.global_disk_check_interval;
642 0 : let data_dir = conf.workdir.clone();
643 : // Use the safekeeper data directory to compute filesystem capacity. This only runs once on startup, so
644 : // there is little point to continue if we can't have the proper protections in place.
645 0 : let fs_capacity_bytes = get_filesystem_capacity(data_dir.as_std_path())
646 0 : .expect("Failed to get filesystem capacity for data directory");
647 0 : let limit: u64 = (conf.max_global_disk_usage_ratio * fs_capacity_bytes as f64) as u64;
648 0 : if limit > 0 {
649 0 : let disk_usage_watch_handle = BACKGROUND_RUNTIME
650 0 : .handle()
651 0 : .spawn(async move {
652 : // Use Tokio interval to preserve fixed cadence between filesystem utilization checks
653 0 : let mut ticker = tokio::time::interval(interval);
654 0 : ticker.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
655 :
656 : loop {
657 0 : ticker.tick().await;
658 0 : let data_dir_clone = data_dir.clone();
659 0 : let check_start = Instant::now();
660 :
661 0 : let usage = tokio::task::spawn_blocking(move || {
662 0 : get_filesystem_usage(data_dir_clone.as_std_path())
663 0 : })
664 0 : .await
665 0 : .unwrap_or(0);
666 :
667 0 : let elapsed = check_start.elapsed().as_secs_f64();
668 0 : GLOBAL_DISK_UTIL_CHECK_SECONDS.observe(elapsed);
669 0 : if usage > limit {
670 0 : warn!(
671 0 : "Global disk usage exceeded limit. Usage: {} bytes, limit: {} bytes",
672 : usage, limit
673 : );
674 0 : }
675 0 : GLOBAL_DISK_LIMIT_EXCEEDED.store(usage > limit, Ordering::Relaxed);
676 : }
677 : })
678 0 : .map(|res| ("Global disk usage watcher".to_string(), res));
679 0 : tasks_handles.push(Box::pin(disk_usage_watch_handle));
680 0 : }
681 : /* END_HADRON */
682 0 : if let Some(pg_listener_tenant_only) = pg_listener_tenant_only {
683 0 : let wal_service_handle = current_thread_rt
684 0 : .as_ref()
685 0 : .unwrap_or_else(|| WAL_SERVICE_RUNTIME.handle())
686 0 : .spawn(wal_service::task_main(
687 0 : conf.clone(),
688 0 : pg_listener_tenant_only,
689 0 : Scope::Tenant,
690 0 : conf.enable_tls_wal_service_api
691 0 : .then(|| tls_server_config.clone())
692 0 : .flatten(),
693 0 : global_timelines.clone(),
694 : ))
695 : // wrap with task name for error reporting
696 0 : .map(|res| ("WAL service tenant only main".to_owned(), res));
697 0 : tasks_handles.push(Box::pin(wal_service_handle));
698 0 : }
699 :
700 0 : let http_handle = current_thread_rt
701 0 : .as_ref()
702 0 : .unwrap_or_else(|| HTTP_RUNTIME.handle())
703 0 : .spawn(http::task_main_http(
704 0 : conf.clone(),
705 0 : http_listener,
706 0 : global_timelines.clone(),
707 : ))
708 0 : .map(|res| ("HTTP service main".to_owned(), res));
709 0 : tasks_handles.push(Box::pin(http_handle));
710 :
711 0 : if let Some(https_listener) = https_listener {
712 0 : let https_handle = current_thread_rt
713 0 : .as_ref()
714 0 : .unwrap_or_else(|| HTTP_RUNTIME.handle())
715 0 : .spawn(http::task_main_https(
716 0 : conf.clone(),
717 0 : https_listener,
718 0 : tls_server_config.expect("tls_server_config is set earlier if https is enabled"),
719 0 : global_timelines.clone(),
720 : ))
721 0 : .map(|res| ("HTTPS service main".to_owned(), res));
722 0 : tasks_handles.push(Box::pin(https_handle));
723 0 : }
724 :
725 0 : let broker_task_handle = current_thread_rt
726 0 : .as_ref()
727 0 : .unwrap_or_else(|| BROKER_RUNTIME.handle())
728 0 : .spawn(
729 0 : broker::task_main(conf.clone(), global_timelines.clone())
730 0 : .instrument(info_span!("broker")),
731 : )
732 0 : .map(|res| ("broker main".to_owned(), res));
733 0 : tasks_handles.push(Box::pin(broker_task_handle));
734 :
735 : /* BEGIN_HADRON */
736 0 : if conf.force_metric_collection_on_scrape {
737 0 : let metrics_handle = current_thread_rt
738 0 : .as_ref()
739 0 : .unwrap_or_else(|| BACKGROUND_RUNTIME.handle())
740 0 : .spawn(async move {
741 0 : let mut interval: tokio::time::Interval =
742 0 : tokio::time::interval(METRICS_COLLECTION_INTERVAL);
743 : loop {
744 0 : interval.tick().await;
745 0 : tokio::task::spawn_blocking(|| {
746 0 : METRICS_COLLECTOR.run_once(true);
747 0 : });
748 : }
749 : })
750 0 : .map(|res| ("broker main".to_owned(), res));
751 0 : tasks_handles.push(Box::pin(metrics_handle));
752 0 : }
753 : /* END_HADRON */
754 :
755 0 : set_build_info_metric(GIT_VERSION, BUILD_TAG);
756 :
757 : // TODO: update tokio-stream, convert to real async Stream with
758 : // SignalStream, map it to obtain missing signal name, combine streams into
759 : // single stream we can easily sit on.
760 0 : let mut sigquit_stream = signal(SignalKind::quit())?;
761 0 : let mut sigint_stream = signal(SignalKind::interrupt())?;
762 0 : let mut sigterm_stream = signal(SignalKind::terminate())?;
763 :
764 : // Notify systemd that we are ready. This is important as currently loading
765 : // timelines takes significant time (~30s in busy regions).
766 0 : if let Err(e) = sd_notify::notify(true, &[NotifyState::Ready]) {
767 0 : warn!("systemd notify failed: {:?}", e);
768 0 : }
769 :
770 0 : tokio::select! {
771 0 : Some((task_name, res)) = tasks_handles.next()=> {
772 0 : error!("{} task failed: {:?}, exiting", task_name, res);
773 0 : std::process::exit(1);
774 : }
775 : // On any shutdown signal, log receival and exit. Additionally, handling
776 : // SIGQUIT prevents coredump.
777 0 : _ = sigquit_stream.recv() => info!("received SIGQUIT, terminating"),
778 0 : _ = sigint_stream.recv() => info!("received SIGINT, terminating"),
779 0 : _ = sigterm_stream.recv() => info!("received SIGTERM, terminating")
780 :
781 : };
782 0 : std::process::exit(0);
783 0 : }
784 :
785 : /// Determine safekeeper id.
786 0 : fn set_id(workdir: &Utf8Path, given_id: Option<NodeId>) -> Result<NodeId> {
787 0 : let id_file_path = workdir.join(ID_FILE_NAME);
788 :
789 : let my_id: NodeId;
790 : // If file with ID exists, read it in; otherwise set one passed.
791 0 : match fs::read(&id_file_path) {
792 0 : Ok(id_serialized) => {
793 : my_id = NodeId(
794 0 : std::str::from_utf8(&id_serialized)
795 0 : .context("failed to parse safekeeper id")?
796 0 : .parse()
797 0 : .context("failed to parse safekeeper id")?,
798 : );
799 0 : if let Some(given_id) = given_id {
800 0 : if given_id != my_id {
801 0 : bail!(
802 0 : "safekeeper already initialized with id {}, can't set {}",
803 : my_id,
804 : given_id
805 : );
806 0 : }
807 0 : }
808 0 : info!("safekeeper ID {}", my_id);
809 : }
810 0 : Err(error) => match error.kind() {
811 : ErrorKind::NotFound => {
812 0 : my_id = if let Some(given_id) = given_id {
813 0 : given_id
814 : } else {
815 0 : bail!("safekeeper id is not specified");
816 : };
817 0 : let mut f = File::create(&id_file_path)
818 0 : .with_context(|| format!("Failed to create id file at {id_file_path:?}"))?;
819 0 : f.write_all(my_id.to_string().as_bytes())?;
820 0 : f.sync_all()?;
821 0 : info!("initialized safekeeper id {}", my_id);
822 : }
823 : _ => {
824 0 : return Err(error.into());
825 : }
826 : },
827 : }
828 0 : Ok(my_id)
829 0 : }
830 :
831 0 : fn parse_remote_storage(storage_conf: &str) -> anyhow::Result<RemoteStorageConfig> {
832 0 : RemoteStorageConfig::from_toml(&storage_conf.parse()?)
833 0 : }
834 :
835 : #[test]
836 1 : fn verify_cli() {
837 : use clap::CommandFactory;
838 1 : Args::command().debug_assert()
839 1 : }
|
