Line data Source code
1 : //!
2 : //! Postgres wrapper (`compute_ctl`) is intended to be run as a Docker entrypoint or as a `systemd`
3 : //! `ExecStart` option. It will handle all the `Neon` specifics during compute node
4 : //! initialization:
5 : //! - `compute_ctl` accepts cluster (compute node) specification as a JSON file.
6 : //! - Every start is a fresh start, so the data directory is removed and
7 : //! initialized again on each run.
8 : //! - If remote_extension_config is provided, it will be used to fetch extensions list
9 : //! and download `shared_preload_libraries` from the remote storage.
10 : //! - Next it will put configuration files into the `PGDATA` directory.
11 : //! - Sync safekeepers and get commit LSN.
12 : //! - Get `basebackup` from pageserver using the returned on the previous step LSN.
13 : //! - Try to start `postgres` and wait until it is ready to accept connections.
14 : //! - Check and alter/drop/create roles and databases.
15 : //! - Hang waiting on the `postmaster` process to exit.
16 : //!
17 : //! Also `compute_ctl` spawns two separate service threads:
18 : //! - `compute-monitor` checks the last Postgres activity timestamp and saves it
19 : //! into the shared `ComputeNode`;
20 : //! - `http-endpoint` runs a Hyper HTTP API server, which serves readiness and the
21 : //! last activity requests.
22 : //!
23 : //! If `AUTOSCALING` environment variable is set, `compute_ctl` will start the
24 : //! `vm-monitor` located in [`neon/libs/vm_monitor`]. For VM compute nodes,
25 : //! `vm-monitor` communicates with the VM autoscaling system. It coordinates
26 : //! downscaling and requests immediate upscaling under resource pressure.
27 : //!
28 : //! Usage example:
29 : //! ```sh
30 : //! compute_ctl -D /var/db/postgres/compute \
31 : //! -C 'postgresql://cloud_admin@localhost/postgres' \
32 : //! -S /var/db/postgres/specs/current.json \
33 : //! -b /usr/local/bin/postgres \
34 : //! -r http://pg-ext-s3-gateway \
35 : //! ```
36 : use std::collections::HashMap;
37 : use std::fs::File;
38 : use std::path::Path;
39 : use std::process::exit;
40 : use std::sync::atomic::Ordering;
41 : use std::sync::{mpsc, Arc, Condvar, Mutex, RwLock};
42 : use std::{thread, time::Duration};
43 :
44 : use anyhow::{Context, Result};
45 : use chrono::Utc;
46 : use clap::Arg;
47 : use signal_hook::consts::{SIGQUIT, SIGTERM};
48 : use signal_hook::{consts::SIGINT, iterator::Signals};
49 : use tracing::{error, info, warn};
50 : use url::Url;
51 :
52 : use compute_api::responses::ComputeStatus;
53 : use compute_api::spec::ComputeSpec;
54 :
55 : use compute_tools::compute::{
56 : forward_termination_signal, ComputeNode, ComputeState, ParsedSpec, PG_PID,
57 : };
58 : use compute_tools::configurator::launch_configurator;
59 : use compute_tools::extension_server::get_pg_version;
60 : use compute_tools::http::api::launch_http_server;
61 : use compute_tools::logger::*;
62 : use compute_tools::monitor::launch_monitor;
63 : use compute_tools::params::*;
64 : use compute_tools::spec::*;
65 : use compute_tools::swap::resize_swap;
66 : use rlimit::{setrlimit, Resource};
67 :
68 : // this is an arbitrary build tag. Fine as a default / for testing purposes
69 : // in-case of not-set environment var
70 : const BUILD_TAG_DEFAULT: &str = "latest";
71 :
72 0 : fn main() -> Result<()> {
73 0 : let (build_tag, clap_args) = init()?;
74 :
75 : // enable core dumping for all child processes
76 0 : setrlimit(Resource::CORE, rlimit::INFINITY, rlimit::INFINITY)?;
77 :
78 0 : let (pg_handle, start_pg_result) = {
79 : // Enter startup tracing context
80 0 : let _startup_context_guard = startup_context_from_env();
81 :
82 0 : let cli_args = process_cli(&clap_args)?;
83 :
84 0 : let cli_spec = try_spec_from_cli(&clap_args, &cli_args)?;
85 :
86 0 : let wait_spec_result = wait_spec(build_tag, cli_args, cli_spec)?;
87 :
88 0 : start_postgres(&clap_args, wait_spec_result)?
89 :
90 : // Startup is finished, exit the startup tracing span
91 : };
92 :
93 : // PostgreSQL is now running, if startup was successful. Wait until it exits.
94 0 : let wait_pg_result = wait_postgres(pg_handle)?;
95 :
96 0 : let delay_exit = cleanup_after_postgres_exit(start_pg_result)?;
97 :
98 0 : maybe_delay_exit(delay_exit);
99 0 :
100 0 : deinit_and_exit(wait_pg_result);
101 0 : }
102 :
103 0 : fn init() -> Result<(String, clap::ArgMatches)> {
104 0 : init_tracing_and_logging(DEFAULT_LOG_LEVEL)?;
105 :
106 0 : let mut signals = Signals::new([SIGINT, SIGTERM, SIGQUIT])?;
107 0 : thread::spawn(move || {
108 0 : for sig in signals.forever() {
109 0 : handle_exit_signal(sig);
110 0 : }
111 0 : });
112 0 :
113 0 : let build_tag = option_env!("BUILD_TAG")
114 0 : .unwrap_or(BUILD_TAG_DEFAULT)
115 0 : .to_string();
116 0 : info!("build_tag: {build_tag}");
117 :
118 0 : Ok((build_tag, cli().get_matches()))
119 0 : }
120 :
121 0 : fn process_cli(matches: &clap::ArgMatches) -> Result<ProcessCliResult> {
122 0 : let pgbin_default = "postgres";
123 0 : let pgbin = matches
124 0 : .get_one::<String>("pgbin")
125 0 : .map(|s| s.as_str())
126 0 : .unwrap_or(pgbin_default);
127 0 :
128 0 : let ext_remote_storage = matches
129 0 : .get_one::<String>("remote-ext-config")
130 0 : // Compatibility hack: if the control plane specified any remote-ext-config
131 0 : // use the default value for extension storage proxy gateway.
132 0 : // Remove this once the control plane is updated to pass the gateway URL
133 0 : .map(|conf| {
134 0 : if conf.starts_with("http") {
135 0 : conf.trim_end_matches('/')
136 : } else {
137 0 : "http://pg-ext-s3-gateway"
138 : }
139 0 : });
140 0 :
141 0 : let http_port = *matches
142 0 : .get_one::<u16>("http-port")
143 0 : .expect("http-port is required");
144 0 : let pgdata = matches
145 0 : .get_one::<String>("pgdata")
146 0 : .expect("PGDATA path is required");
147 0 : let connstr = matches
148 0 : .get_one::<String>("connstr")
149 0 : .expect("Postgres connection string is required");
150 0 : let spec_json = matches.get_one::<String>("spec");
151 0 : let spec_path = matches.get_one::<String>("spec-path");
152 0 : let resize_swap_on_bind = matches.get_flag("resize-swap-on-bind");
153 0 :
154 0 : Ok(ProcessCliResult {
155 0 : connstr,
156 0 : pgdata,
157 0 : pgbin,
158 0 : ext_remote_storage,
159 0 : http_port,
160 0 : spec_json,
161 0 : spec_path,
162 0 : resize_swap_on_bind,
163 0 : })
164 0 : }
165 :
166 : struct ProcessCliResult<'clap> {
167 : connstr: &'clap str,
168 : pgdata: &'clap str,
169 : pgbin: &'clap str,
170 : ext_remote_storage: Option<&'clap str>,
171 : http_port: u16,
172 : spec_json: Option<&'clap String>,
173 : spec_path: Option<&'clap String>,
174 : resize_swap_on_bind: bool,
175 : }
176 :
177 0 : fn startup_context_from_env() -> Option<opentelemetry::ContextGuard> {
178 0 : // Extract OpenTelemetry context for the startup actions from the
179 0 : // TRACEPARENT and TRACESTATE env variables, and attach it to the current
180 0 : // tracing context.
181 0 : //
182 0 : // This is used to propagate the context for the 'start_compute' operation
183 0 : // from the neon control plane. This allows linking together the wider
184 0 : // 'start_compute' operation that creates the compute container, with the
185 0 : // startup actions here within the container.
186 0 : //
187 0 : // There is no standard for passing context in env variables, but a lot of
188 0 : // tools use TRACEPARENT/TRACESTATE, so we use that convention too. See
189 0 : // https://github.com/open-telemetry/opentelemetry-specification/issues/740
190 0 : //
191 0 : // Switch to the startup context here, and exit it once the startup has
192 0 : // completed and Postgres is up and running.
193 0 : //
194 0 : // If this pod is pre-created without binding it to any particular endpoint
195 0 : // yet, this isn't the right place to enter the startup context. In that
196 0 : // case, the control plane should pass the tracing context as part of the
197 0 : // /configure API call.
198 0 : //
199 0 : // NOTE: This is supposed to only cover the *startup* actions. Once
200 0 : // postgres is configured and up-and-running, we exit this span. Any other
201 0 : // actions that are performed on incoming HTTP requests, for example, are
202 0 : // performed in separate spans.
203 0 : //
204 0 : // XXX: If the pod is restarted, we perform the startup actions in the same
205 0 : // context as the original startup actions, which probably doesn't make
206 0 : // sense.
207 0 : let mut startup_tracing_carrier: HashMap<String, String> = HashMap::new();
208 0 : if let Ok(val) = std::env::var("TRACEPARENT") {
209 0 : startup_tracing_carrier.insert("traceparent".to_string(), val);
210 0 : }
211 0 : if let Ok(val) = std::env::var("TRACESTATE") {
212 0 : startup_tracing_carrier.insert("tracestate".to_string(), val);
213 0 : }
214 0 : if !startup_tracing_carrier.is_empty() {
215 : use opentelemetry::propagation::TextMapPropagator;
216 : use opentelemetry::sdk::propagation::TraceContextPropagator;
217 0 : let guard = TraceContextPropagator::new()
218 0 : .extract(&startup_tracing_carrier)
219 0 : .attach();
220 0 : info!("startup tracing context attached");
221 0 : Some(guard)
222 : } else {
223 0 : None
224 : }
225 0 : }
226 :
227 0 : fn try_spec_from_cli(
228 0 : matches: &clap::ArgMatches,
229 0 : ProcessCliResult {
230 0 : spec_json,
231 0 : spec_path,
232 0 : ..
233 0 : }: &ProcessCliResult,
234 0 : ) -> Result<CliSpecParams> {
235 0 : let compute_id = matches.get_one::<String>("compute-id");
236 0 : let control_plane_uri = matches.get_one::<String>("control-plane-uri");
237 0 :
238 0 : let spec;
239 0 : let mut live_config_allowed = false;
240 0 : match spec_json {
241 : // First, try to get cluster spec from the cli argument
242 0 : Some(json) => {
243 0 : info!("got spec from cli argument {}", json);
244 0 : spec = Some(serde_json::from_str(json)?);
245 : }
246 : None => {
247 : // Second, try to read it from the file if path is provided
248 0 : if let Some(sp) = spec_path {
249 0 : let path = Path::new(sp);
250 0 : let file = File::open(path)?;
251 0 : spec = Some(serde_json::from_reader(file)?);
252 0 : live_config_allowed = true;
253 0 : } else if let Some(id) = compute_id {
254 0 : if let Some(cp_base) = control_plane_uri {
255 0 : live_config_allowed = true;
256 0 : spec = match get_spec_from_control_plane(cp_base, id) {
257 0 : Ok(s) => s,
258 0 : Err(e) => {
259 0 : error!("cannot get response from control plane: {}", e);
260 0 : panic!("neither spec nor confirmation that compute is in the Empty state was received");
261 : }
262 : };
263 : } else {
264 0 : panic!("must specify both --control-plane-uri and --compute-id or none");
265 : }
266 : } else {
267 0 : panic!(
268 0 : "compute spec should be provided by one of the following ways: \
269 0 : --spec OR --spec-path OR --control-plane-uri and --compute-id"
270 0 : );
271 : }
272 : }
273 : };
274 :
275 0 : Ok(CliSpecParams {
276 0 : spec,
277 0 : live_config_allowed,
278 0 : })
279 0 : }
280 :
281 : struct CliSpecParams {
282 : /// If a spec was provided via CLI or file, the [`ComputeSpec`]
283 : spec: Option<ComputeSpec>,
284 : live_config_allowed: bool,
285 : }
286 :
287 0 : fn wait_spec(
288 0 : build_tag: String,
289 0 : ProcessCliResult {
290 0 : connstr,
291 0 : pgdata,
292 0 : pgbin,
293 0 : ext_remote_storage,
294 0 : resize_swap_on_bind,
295 0 : http_port,
296 0 : ..
297 0 : }: ProcessCliResult,
298 0 : CliSpecParams {
299 0 : spec,
300 0 : live_config_allowed,
301 0 : }: CliSpecParams,
302 0 : ) -> Result<WaitSpecResult> {
303 0 : let mut new_state = ComputeState::new();
304 : let spec_set;
305 :
306 0 : if let Some(spec) = spec {
307 0 : let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
308 0 : info!("new pspec.spec: {:?}", pspec.spec);
309 0 : new_state.pspec = Some(pspec);
310 0 : spec_set = true;
311 0 : } else {
312 0 : spec_set = false;
313 0 : }
314 0 : let compute_node = ComputeNode {
315 0 : connstr: Url::parse(connstr).context("cannot parse connstr as a URL")?,
316 0 : pgdata: pgdata.to_string(),
317 0 : pgbin: pgbin.to_string(),
318 0 : pgversion: get_pg_version(pgbin),
319 0 : live_config_allowed,
320 0 : state: Mutex::new(new_state),
321 0 : state_changed: Condvar::new(),
322 0 : ext_remote_storage: ext_remote_storage.map(|s| s.to_string()),
323 0 : ext_download_progress: RwLock::new(HashMap::new()),
324 0 : build_tag,
325 0 : };
326 0 : let compute = Arc::new(compute_node);
327 0 :
328 0 : // If this is a pooled VM, prewarm before starting HTTP server and becoming
329 0 : // available for binding. Prewarming helps Postgres start quicker later,
330 0 : // because QEMU will already have its memory allocated from the host, and
331 0 : // the necessary binaries will already be cached.
332 0 : if !spec_set {
333 0 : compute.prewarm_postgres()?;
334 0 : }
335 :
336 : // Launch http service first, so that we can serve control-plane requests
337 : // while configuration is still in progress.
338 0 : let _http_handle =
339 0 : launch_http_server(http_port, &compute).expect("cannot launch http endpoint thread");
340 0 :
341 0 : if !spec_set {
342 : // No spec provided, hang waiting for it.
343 0 : info!("no compute spec provided, waiting");
344 :
345 0 : let mut state = compute.state.lock().unwrap();
346 0 : while state.status != ComputeStatus::ConfigurationPending {
347 0 : state = compute.state_changed.wait(state).unwrap();
348 0 :
349 0 : if state.status == ComputeStatus::ConfigurationPending {
350 0 : info!("got spec, continue configuration");
351 : // Spec is already set by the http server handler.
352 0 : break;
353 0 : }
354 : }
355 :
356 : // Record for how long we slept waiting for the spec.
357 0 : let now = Utc::now();
358 0 : state.metrics.wait_for_spec_ms = now
359 0 : .signed_duration_since(state.start_time)
360 0 : .to_std()
361 0 : .unwrap()
362 0 : .as_millis() as u64;
363 0 :
364 0 : // Reset start time, so that the total startup time that is calculated later will
365 0 : // not include the time that we waited for the spec.
366 0 : state.start_time = now;
367 0 : }
368 :
369 0 : Ok(WaitSpecResult {
370 0 : compute,
371 0 : http_port,
372 0 : resize_swap_on_bind,
373 0 : })
374 0 : }
375 :
376 : struct WaitSpecResult {
377 : compute: Arc<ComputeNode>,
378 : // passed through from ProcessCliResult
379 : http_port: u16,
380 : resize_swap_on_bind: bool,
381 : }
382 :
383 0 : fn start_postgres(
384 0 : // need to allow unused because `matches` is only used if target_os = "linux"
385 0 : #[allow(unused_variables)] matches: &clap::ArgMatches,
386 0 : WaitSpecResult {
387 0 : compute,
388 0 : http_port,
389 0 : resize_swap_on_bind,
390 0 : }: WaitSpecResult,
391 0 : ) -> Result<(Option<PostgresHandle>, StartPostgresResult)> {
392 0 : // We got all we need, update the state.
393 0 : let mut state = compute.state.lock().unwrap();
394 0 : state.status = ComputeStatus::Init;
395 0 : compute.state_changed.notify_all();
396 0 :
397 0 : info!(
398 0 : "running compute with features: {:?}",
399 0 : state.pspec.as_ref().unwrap().spec.features
400 : );
401 : // before we release the mutex, fetch the swap size (if any) for later.
402 0 : let swap_size_bytes = state.pspec.as_ref().unwrap().spec.swap_size_bytes;
403 0 : drop(state);
404 0 :
405 0 : // Launch remaining service threads
406 0 : let _monitor_handle = launch_monitor(&compute);
407 0 : let _configurator_handle = launch_configurator(&compute);
408 0 :
409 0 : let mut prestartup_failed = false;
410 0 : let mut delay_exit = false;
411 :
412 : // Resize swap to the desired size if the compute spec says so
413 0 : if let (Some(size_bytes), true) = (swap_size_bytes, resize_swap_on_bind) {
414 : // To avoid 'swapoff' hitting postgres startup, we need to run resize-swap to completion
415 : // *before* starting postgres.
416 : //
417 : // In theory, we could do this asynchronously if SkipSwapon was enabled for VMs, but this
418 : // carries a risk of introducing hard-to-debug issues - e.g. if postgres sometimes gets
419 : // OOM-killed during startup because swap wasn't available yet.
420 0 : match resize_swap(size_bytes) {
421 : Ok(()) => {
422 0 : let size_gib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
423 0 : info!(%size_bytes, %size_gib, "resized swap");
424 : }
425 0 : Err(err) => {
426 0 : let err = err.context("failed to resize swap");
427 0 : error!("{err:#}");
428 :
429 : // Mark compute startup as failed; don't try to start postgres, and report this
430 : // error to the control plane when it next asks.
431 0 : prestartup_failed = true;
432 0 : let mut state = compute.state.lock().unwrap();
433 0 : state.error = Some(format!("{err:?}"));
434 0 : state.status = ComputeStatus::Failed;
435 0 : compute.state_changed.notify_all();
436 0 : delay_exit = true;
437 : }
438 : }
439 0 : }
440 :
441 0 : let extension_server_port: u16 = http_port;
442 0 :
443 0 : // Start Postgres
444 0 : let mut pg = None;
445 0 : if !prestartup_failed {
446 0 : pg = match compute.start_compute(extension_server_port) {
447 0 : Ok(pg) => Some(pg),
448 0 : Err(err) => {
449 0 : error!("could not start the compute node: {:#}", err);
450 0 : let mut state = compute.state.lock().unwrap();
451 0 : state.error = Some(format!("{:?}", err));
452 0 : state.status = ComputeStatus::Failed;
453 0 : // Notify others that Postgres failed to start. In case of configuring the
454 0 : // empty compute, it's likely that API handler is still waiting for compute
455 0 : // state change. With this we will notify it that compute is in Failed state,
456 0 : // so control plane will know about it earlier and record proper error instead
457 0 : // of timeout.
458 0 : compute.state_changed.notify_all();
459 0 : drop(state); // unlock
460 0 : delay_exit = true;
461 0 : None
462 : }
463 : };
464 : } else {
465 0 : warn!("skipping postgres startup because pre-startup step failed");
466 : }
467 :
468 : // Start the vm-monitor if directed to. The vm-monitor only runs on linux
469 : // because it requires cgroups.
470 : cfg_if::cfg_if! {
471 : if #[cfg(target_os = "linux")] {
472 : use std::env;
473 : use tokio_util::sync::CancellationToken;
474 0 : let vm_monitor_addr = matches
475 0 : .get_one::<String>("vm-monitor-addr")
476 0 : .expect("--vm-monitor-addr should always be set because it has a default arg");
477 0 : let file_cache_connstr = matches.get_one::<String>("filecache-connstr");
478 0 : let cgroup = matches.get_one::<String>("cgroup");
479 :
480 : // Only make a runtime if we need to.
481 : // Note: it seems like you can make a runtime in an inner scope and
482 : // if you start a task in it it won't be dropped. However, make it
483 : // in the outermost scope just to be safe.
484 0 : let rt = if env::var_os("AUTOSCALING").is_some() {
485 0 : Some(
486 0 : tokio::runtime::Builder::new_multi_thread()
487 0 : .worker_threads(4)
488 0 : .enable_all()
489 0 : .build()
490 0 : .expect("failed to create tokio runtime for monitor")
491 0 : )
492 : } else {
493 0 : None
494 : };
495 :
496 : // This token is used internally by the monitor to clean up all threads
497 0 : let token = CancellationToken::new();
498 0 :
499 0 : let vm_monitor = rt.as_ref().map(|rt| {
500 0 : rt.spawn(vm_monitor::start(
501 0 : Box::leak(Box::new(vm_monitor::Args {
502 0 : cgroup: cgroup.cloned(),
503 0 : pgconnstr: file_cache_connstr.cloned(),
504 0 : addr: vm_monitor_addr.clone(),
505 0 : })),
506 0 : token.clone(),
507 0 : ))
508 0 : });
509 0 : }
510 0 : }
511 0 :
512 0 : Ok((
513 0 : pg,
514 0 : StartPostgresResult {
515 0 : delay_exit,
516 0 : compute,
517 0 : #[cfg(target_os = "linux")]
518 0 : rt,
519 0 : #[cfg(target_os = "linux")]
520 0 : token,
521 0 : #[cfg(target_os = "linux")]
522 0 : vm_monitor,
523 0 : },
524 0 : ))
525 0 : }
526 :
527 : type PostgresHandle = (std::process::Child, std::thread::JoinHandle<()>);
528 :
529 : struct StartPostgresResult {
530 : delay_exit: bool,
531 : // passed through from WaitSpecResult
532 : compute: Arc<ComputeNode>,
533 :
534 : #[cfg(target_os = "linux")]
535 : rt: Option<tokio::runtime::Runtime>,
536 : #[cfg(target_os = "linux")]
537 : token: tokio_util::sync::CancellationToken,
538 : #[cfg(target_os = "linux")]
539 : vm_monitor: Option<tokio::task::JoinHandle<Result<()>>>,
540 : }
541 :
542 0 : fn wait_postgres(pg: Option<PostgresHandle>) -> Result<WaitPostgresResult> {
543 0 : // Wait for the child Postgres process forever. In this state Ctrl+C will
544 0 : // propagate to Postgres and it will be shut down as well.
545 0 : let mut exit_code = None;
546 0 : if let Some((mut pg, logs_handle)) = pg {
547 0 : let ecode = pg
548 0 : .wait()
549 0 : .expect("failed to start waiting on Postgres process");
550 0 : PG_PID.store(0, Ordering::SeqCst);
551 0 :
552 0 : // Process has exited, so we can join the logs thread.
553 0 : let _ = logs_handle
554 0 : .join()
555 0 : .map_err(|e| tracing::error!("log thread panicked: {:?}", e));
556 0 :
557 0 : info!("Postgres exited with code {}, shutting down", ecode);
558 0 : exit_code = ecode.code()
559 0 : }
560 :
561 0 : Ok(WaitPostgresResult { exit_code })
562 0 : }
563 :
564 : struct WaitPostgresResult {
565 : exit_code: Option<i32>,
566 : }
567 :
568 0 : fn cleanup_after_postgres_exit(
569 0 : StartPostgresResult {
570 0 : mut delay_exit,
571 0 : compute,
572 0 : #[cfg(target_os = "linux")]
573 0 : vm_monitor,
574 0 : #[cfg(target_os = "linux")]
575 0 : token,
576 0 : #[cfg(target_os = "linux")]
577 0 : rt,
578 0 : }: StartPostgresResult,
579 0 : ) -> Result<bool> {
580 : // Terminate the vm_monitor so it releases the file watcher on
581 : // /sys/fs/cgroup/neon-postgres.
582 : // Note: the vm-monitor only runs on linux because it requires cgroups.
583 : cfg_if::cfg_if! {
584 : if #[cfg(target_os = "linux")] {
585 0 : if let Some(handle) = vm_monitor {
586 0 : // Kills all threads spawned by the monitor
587 0 : token.cancel();
588 0 : // Kills the actual task running the monitor
589 0 : handle.abort();
590 0 :
591 0 : // If handle is some, rt must have been used to produce it, and
592 0 : // hence is also some
593 0 : rt.unwrap().shutdown_timeout(Duration::from_secs(2));
594 0 : }
595 : }
596 : }
597 :
598 : // Maybe sync safekeepers again, to speed up next startup
599 0 : let compute_state = compute.state.lock().unwrap().clone();
600 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
601 0 : if matches!(pspec.spec.mode, compute_api::spec::ComputeMode::Primary) {
602 0 : info!("syncing safekeepers on shutdown");
603 0 : let storage_auth_token = pspec.storage_auth_token.clone();
604 0 : let lsn = compute.sync_safekeepers(storage_auth_token)?;
605 0 : info!("synced safekeepers at lsn {lsn}");
606 0 : }
607 :
608 0 : let mut state = compute.state.lock().unwrap();
609 0 : if state.status == ComputeStatus::TerminationPending {
610 0 : state.status = ComputeStatus::Terminated;
611 0 : compute.state_changed.notify_all();
612 0 : // we were asked to terminate gracefully, don't exit to avoid restart
613 0 : delay_exit = true
614 0 : }
615 0 : drop(state);
616 :
617 0 : if let Err(err) = compute.check_for_core_dumps() {
618 0 : error!("error while checking for core dumps: {err:?}");
619 0 : }
620 :
621 0 : Ok(delay_exit)
622 0 : }
623 :
624 0 : fn maybe_delay_exit(delay_exit: bool) {
625 0 : // If launch failed, keep serving HTTP requests for a while, so the cloud
626 0 : // control plane can get the actual error.
627 0 : if delay_exit {
628 0 : info!("giving control plane 30s to collect the error before shutdown");
629 0 : thread::sleep(Duration::from_secs(30));
630 0 : }
631 0 : }
632 :
633 0 : fn deinit_and_exit(WaitPostgresResult { exit_code }: WaitPostgresResult) -> ! {
634 0 : // Shutdown trace pipeline gracefully, so that it has a chance to send any
635 0 : // pending traces before we exit. Shutting down OTEL tracing provider may
636 0 : // hang for quite some time, see, for example:
637 0 : // - https://github.com/open-telemetry/opentelemetry-rust/issues/868
638 0 : // - and our problems with staging https://github.com/neondatabase/cloud/issues/3707#issuecomment-1493983636
639 0 : //
640 0 : // Yet, we want computes to shut down fast enough, as we may need a new one
641 0 : // for the same timeline ASAP. So wait no longer than 2s for the shutdown to
642 0 : // complete, then just error out and exit the main thread.
643 0 : info!("shutting down tracing");
644 0 : let (sender, receiver) = mpsc::channel();
645 0 : let _ = thread::spawn(move || {
646 0 : tracing_utils::shutdown_tracing();
647 0 : sender.send(()).ok()
648 0 : });
649 0 : let shutdown_res = receiver.recv_timeout(Duration::from_millis(2000));
650 0 : if shutdown_res.is_err() {
651 0 : error!("timed out while shutting down tracing, exiting anyway");
652 0 : }
653 :
654 0 : info!("shutting down");
655 0 : exit(exit_code.unwrap_or(1))
656 : }
657 :
658 2 : fn cli() -> clap::Command {
659 2 : // Env variable is set by `cargo`
660 2 : let version = option_env!("CARGO_PKG_VERSION").unwrap_or("unknown");
661 2 : clap::Command::new("compute_ctl")
662 2 : .version(version)
663 2 : .arg(
664 2 : Arg::new("http-port")
665 2 : .long("http-port")
666 2 : .value_name("HTTP_PORT")
667 2 : .default_value("3080")
668 2 : .value_parser(clap::value_parser!(u16))
669 2 : .required(false),
670 2 : )
671 2 : .arg(
672 2 : Arg::new("connstr")
673 2 : .short('C')
674 2 : .long("connstr")
675 2 : .value_name("DATABASE_URL")
676 2 : .required(true),
677 2 : )
678 2 : .arg(
679 2 : Arg::new("pgdata")
680 2 : .short('D')
681 2 : .long("pgdata")
682 2 : .value_name("DATADIR")
683 2 : .required(true),
684 2 : )
685 2 : .arg(
686 2 : Arg::new("pgbin")
687 2 : .short('b')
688 2 : .long("pgbin")
689 2 : .default_value("postgres")
690 2 : .value_name("POSTGRES_PATH"),
691 2 : )
692 2 : .arg(
693 2 : Arg::new("spec")
694 2 : .short('s')
695 2 : .long("spec")
696 2 : .value_name("SPEC_JSON"),
697 2 : )
698 2 : .arg(
699 2 : Arg::new("spec-path")
700 2 : .short('S')
701 2 : .long("spec-path")
702 2 : .value_name("SPEC_PATH"),
703 2 : )
704 2 : .arg(
705 2 : Arg::new("compute-id")
706 2 : .short('i')
707 2 : .long("compute-id")
708 2 : .value_name("COMPUTE_ID"),
709 2 : )
710 2 : .arg(
711 2 : Arg::new("control-plane-uri")
712 2 : .short('p')
713 2 : .long("control-plane-uri")
714 2 : .value_name("CONTROL_PLANE_API_BASE_URI"),
715 2 : )
716 2 : .arg(
717 2 : Arg::new("remote-ext-config")
718 2 : .short('r')
719 2 : .long("remote-ext-config")
720 2 : .value_name("REMOTE_EXT_CONFIG"),
721 2 : )
722 2 : // TODO(fprasx): we currently have default arguments because the cloud PR
723 2 : // to pass them in hasn't been merged yet. We should get rid of them once
724 2 : // the PR is merged.
725 2 : .arg(
726 2 : Arg::new("vm-monitor-addr")
727 2 : .long("vm-monitor-addr")
728 2 : .default_value("0.0.0.0:10301")
729 2 : .value_name("VM_MONITOR_ADDR"),
730 2 : )
731 2 : .arg(
732 2 : Arg::new("cgroup")
733 2 : .long("cgroup")
734 2 : .default_value("neon-postgres")
735 2 : .value_name("CGROUP"),
736 2 : )
737 2 : .arg(
738 2 : Arg::new("filecache-connstr")
739 2 : .long("filecache-connstr")
740 2 : .default_value(
741 2 : "host=localhost port=5432 dbname=postgres user=cloud_admin sslmode=disable application_name=vm-monitor",
742 2 : )
743 2 : .value_name("FILECACHE_CONNSTR"),
744 2 : )
745 2 : .arg(
746 2 : Arg::new("resize-swap-on-bind")
747 2 : .long("resize-swap-on-bind")
748 2 : .action(clap::ArgAction::SetTrue),
749 2 : )
750 2 : }
751 :
752 : /// When compute_ctl is killed, send also termination signal to sync-safekeepers
753 : /// to prevent leakage. TODO: it is better to convert compute_ctl to async and
754 : /// wait for termination which would be easy then.
755 0 : fn handle_exit_signal(sig: i32) {
756 0 : info!("received {sig} termination signal");
757 0 : forward_termination_signal();
758 0 : exit(1);
759 : }
760 :
761 : #[test]
762 2 : fn verify_cli() {
763 2 : cli().debug_assert()
764 2 : }
|
