Line data Source code
1 : use anyhow::{Context, Result};
2 : use chrono::{DateTime, Utc};
3 : use compute_api::privilege::Privilege;
4 : use compute_api::responses::{
5 : ComputeConfig, ComputeCtlConfig, ComputeMetrics, ComputeStatus, LfcOffloadState,
6 : LfcPrewarmState, PromoteState, TlsConfig,
7 : };
8 : use compute_api::spec::{
9 : ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, ExtVersion, GenericOption,
10 : PageserverConnectionInfo, PageserverProtocol, PgIdent, Role,
11 : };
12 : use futures::StreamExt;
13 : use futures::future::join_all;
14 : use futures::stream::FuturesUnordered;
15 : use itertools::Itertools;
16 : use nix::sys::signal::{Signal, kill};
17 : use nix::unistd::Pid;
18 : use once_cell::sync::Lazy;
19 : use pageserver_page_api::{self as page_api, BaseBackupCompression};
20 : use postgres;
21 : use postgres::NoTls;
22 : use postgres::error::SqlState;
23 : use remote_storage::{DownloadError, RemotePath};
24 : use std::collections::{HashMap, HashSet};
25 : use std::ffi::OsString;
26 : use std::os::unix::fs::{PermissionsExt, symlink};
27 : use std::path::Path;
28 : use std::process::{Command, Stdio};
29 : use std::str::FromStr;
30 : use std::sync::atomic::{AtomicU32, AtomicU64, Ordering};
31 : use std::sync::{Arc, Condvar, Mutex, RwLock};
32 : use std::time::{Duration, Instant};
33 : use std::{env, fs};
34 : use tokio::{spawn, sync::watch, task::JoinHandle, time};
35 : use tracing::{Instrument, debug, error, info, instrument, warn};
36 : use url::Url;
37 : use utils::id::{TenantId, TimelineId};
38 : use utils::lsn::Lsn;
39 : use utils::measured_stream::MeasuredReader;
40 : use utils::pid_file;
41 : use utils::shard::{ShardIndex, ShardNumber, ShardStripeSize};
42 :
43 : use crate::configurator::launch_configurator;
44 : use crate::disk_quota::set_disk_quota;
45 : use crate::hadron_metrics::COMPUTE_ATTACHED;
46 : use crate::installed_extensions::get_installed_extensions;
47 : use crate::logger::{self, startup_context_from_env};
48 : use crate::lsn_lease::launch_lsn_lease_bg_task_for_static;
49 : use crate::metrics::COMPUTE_CTL_UP;
50 : use crate::monitor::launch_monitor;
51 : use crate::pg_helpers::*;
52 : use crate::pgbouncer::*;
53 : use crate::rsyslog::{
54 : PostgresLogsRsyslogConfig, configure_audit_rsyslog, configure_postgres_logs_export,
55 : launch_pgaudit_gc,
56 : };
57 : use crate::spec::*;
58 : use crate::swap::resize_swap;
59 : use crate::sync_sk::{check_if_synced, ping_safekeeper};
60 : use crate::tls::watch_cert_for_changes;
61 : use crate::{config, extension_server, local_proxy};
62 :
63 : pub static SYNC_SAFEKEEPERS_PID: AtomicU32 = AtomicU32::new(0);
64 : pub static PG_PID: AtomicU32 = AtomicU32::new(0);
65 : // This is an arbitrary build tag. Fine as a default / for testing purposes
66 : // in-case of not-set environment var
67 : const BUILD_TAG_DEFAULT: &str = "latest";
68 : /// Build tag/version of the compute node binaries/image. It's tricky and ugly
69 : /// to pass it everywhere as a part of `ComputeNodeParams`, so we use a
70 : /// global static variable.
71 0 : pub static BUILD_TAG: Lazy<String> = Lazy::new(|| {
72 0 : option_env!("BUILD_TAG")
73 0 : .unwrap_or(BUILD_TAG_DEFAULT)
74 0 : .to_string()
75 0 : });
76 : const DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL: u64 = 3600;
77 :
78 : /// Static configuration params that don't change after startup. These mostly
79 : /// come from the CLI args, or are derived from them.
80 : #[derive(Clone, Debug)]
81 : pub struct ComputeNodeParams {
82 : /// The ID of the compute
83 : pub compute_id: String,
84 :
85 : /// Url type maintains proper escaping
86 : pub connstr: url::Url,
87 :
88 : /// The name of the 'weak' superuser role, which we give to the users.
89 : /// It follows the allow list approach, i.e., we take a standard role
90 : /// and grant it extra permissions with explicit GRANTs here and there,
91 : /// and core patches.
92 : pub privileged_role_name: String,
93 :
94 : pub resize_swap_on_bind: bool,
95 : pub set_disk_quota_for_fs: Option<String>,
96 :
97 : // VM monitor parameters
98 : #[cfg(target_os = "linux")]
99 : pub filecache_connstr: String,
100 : #[cfg(target_os = "linux")]
101 : pub cgroup: String,
102 : #[cfg(target_os = "linux")]
103 : pub vm_monitor_addr: String,
104 :
105 : pub pgdata: String,
106 : pub pgbin: String,
107 : pub pgversion: String,
108 :
109 : /// The port that the compute's external HTTP server listens on
110 : pub external_http_port: u16,
111 : /// The port that the compute's internal HTTP server listens on
112 : pub internal_http_port: u16,
113 :
114 : /// the address of extension storage proxy gateway
115 : pub remote_ext_base_url: Option<Url>,
116 :
117 : /// Interval for installed extensions collection
118 : pub installed_extensions_collection_interval: Arc<AtomicU64>,
119 : /// Hadron instance ID of the compute node.
120 : pub instance_id: Option<String>,
121 : /// Timeout of PG compute startup in the Init state.
122 : pub pg_init_timeout: Option<Duration>,
123 : // Path to the `pg_isready` binary.
124 : pub pg_isready_bin: String,
125 : pub lakebase_mode: bool,
126 :
127 : pub build_tag: String,
128 : pub control_plane_uri: Option<String>,
129 : pub config_path_test_only: Option<OsString>,
130 : }
131 :
132 : type TaskHandle = Mutex<Option<JoinHandle<()>>>;
133 :
134 : /// Compute node info shared across several `compute_ctl` threads.
135 : pub struct ComputeNode {
136 : pub params: ComputeNodeParams,
137 :
138 : // We connect to Postgres from many different places, so build configs once
139 : // and reuse them where needed. These are derived from 'params.connstr'
140 : pub conn_conf: postgres::config::Config,
141 : pub tokio_conn_conf: tokio_postgres::config::Config,
142 :
143 : /// Volatile part of the `ComputeNode`, which should be used under `Mutex`.
144 : /// To allow HTTP API server to serving status requests, while configuration
145 : /// is in progress, lock should be held only for short periods of time to do
146 : /// read/write, not the whole configuration process.
147 : pub state: Mutex<ComputeState>,
148 : /// `Condvar` to allow notifying waiters about state changes.
149 : pub state_changed: Condvar,
150 :
151 : // key: ext_archive_name, value: started download time, download_completed?
152 : pub ext_download_progress: RwLock<HashMap<String, (DateTime<Utc>, bool)>>,
153 : pub compute_ctl_config: ComputeCtlConfig,
154 :
155 : /// Handle to the extension stats collection task
156 : extension_stats_task: TaskHandle,
157 : lfc_offload_task: TaskHandle,
158 : }
159 :
160 : // store some metrics about download size that might impact startup time
161 : #[derive(Clone, Debug)]
162 : pub struct RemoteExtensionMetrics {
163 : num_ext_downloaded: u64,
164 : largest_ext_size: u64,
165 : total_ext_download_size: u64,
166 : }
167 :
168 : #[derive(Clone, Debug)]
169 : pub struct ComputeState {
170 : pub start_time: DateTime<Utc>,
171 : pub pg_start_time: Option<DateTime<Utc>>,
172 : pub status: ComputeStatus,
173 : /// Timestamp of the last Postgres activity. It could be `None` if
174 : /// compute wasn't used since start.
175 : pub last_active: Option<DateTime<Utc>>,
176 : pub error: Option<String>,
177 :
178 : /// Compute spec. This can be received from the CLI or - more likely -
179 : /// passed by the control plane with a /configure HTTP request.
180 : pub pspec: Option<ParsedSpec>,
181 :
182 : /// If the spec is passed by a /configure request, 'startup_span' is the
183 : /// /configure request's tracing span. The main thread enters it when it
184 : /// processes the compute startup, so that the compute startup is considered
185 : /// to be part of the /configure request for tracing purposes.
186 : ///
187 : /// If the request handling thread/task called startup_compute() directly,
188 : /// it would automatically be a child of the request handling span, and we
189 : /// wouldn't need this. But because we use the main thread to perform the
190 : /// startup, and the /configure task just waits for it to finish, we need to
191 : /// set up the span relationship ourselves.
192 : pub startup_span: Option<tracing::span::Span>,
193 :
194 : pub lfc_prewarm_state: LfcPrewarmState,
195 : pub lfc_offload_state: LfcOffloadState,
196 :
197 : /// WAL flush LSN that is set after terminating Postgres and syncing safekeepers if
198 : /// mode == ComputeMode::Primary. None otherwise
199 : pub terminate_flush_lsn: Option<Lsn>,
200 : pub promote_state: Option<watch::Receiver<PromoteState>>,
201 :
202 : pub metrics: ComputeMetrics,
203 : }
204 :
205 : impl ComputeState {
206 0 : pub fn new() -> Self {
207 0 : Self {
208 0 : start_time: Utc::now(),
209 0 : pg_start_time: None,
210 0 : status: ComputeStatus::Empty,
211 0 : last_active: None,
212 0 : error: None,
213 0 : pspec: None,
214 0 : startup_span: None,
215 0 : metrics: ComputeMetrics::default(),
216 0 : lfc_prewarm_state: LfcPrewarmState::default(),
217 0 : lfc_offload_state: LfcOffloadState::default(),
218 0 : terminate_flush_lsn: None,
219 0 : promote_state: None,
220 0 : }
221 0 : }
222 :
223 0 : pub fn set_status(&mut self, status: ComputeStatus, state_changed: &Condvar) {
224 0 : let prev = self.status;
225 0 : info!("Changing compute status from {} to {}", prev, status);
226 0 : self.status = status;
227 0 : state_changed.notify_all();
228 :
229 0 : COMPUTE_CTL_UP.reset();
230 0 : COMPUTE_CTL_UP
231 0 : .with_label_values(&[&BUILD_TAG, status.to_string().as_str()])
232 0 : .set(1);
233 0 : }
234 :
235 0 : pub fn set_failed_status(&mut self, err: anyhow::Error, state_changed: &Condvar) {
236 0 : self.error = Some(format!("{err:?}"));
237 0 : self.set_status(ComputeStatus::Failed, state_changed);
238 0 : }
239 : }
240 :
241 : impl Default for ComputeState {
242 0 : fn default() -> Self {
243 0 : Self::new()
244 0 : }
245 : }
246 :
247 : #[derive(Clone, Debug)]
248 : pub struct ParsedSpec {
249 : pub spec: ComputeSpec,
250 : pub tenant_id: TenantId,
251 : pub timeline_id: TimelineId,
252 : pub pageserver_conninfo: PageserverConnectionInfo,
253 : pub safekeeper_connstrings: Vec<String>,
254 : pub storage_auth_token: Option<String>,
255 : /// k8s dns name and port
256 : pub endpoint_storage_addr: Option<String>,
257 : pub endpoint_storage_token: Option<String>,
258 : }
259 :
260 : impl ParsedSpec {
261 1 : pub fn validate(&self) -> Result<(), String> {
262 : // Only Primary nodes are using safekeeper_connstrings, and at the moment
263 : // this method only validates that part of the specs.
264 1 : if self.spec.mode != ComputeMode::Primary {
265 0 : return Ok(());
266 1 : }
267 :
268 : // While it seems like a good idea to check for an odd number of entries in
269 : // the safekeepers connection string, changes to the list of safekeepers might
270 : // incur appending a new server to a list of 3, in which case a list of 4
271 : // entries is okay in production.
272 : //
273 : // Still we want unique entries, and at least one entry in the vector
274 1 : if self.safekeeper_connstrings.is_empty() {
275 0 : return Err(String::from("safekeeper_connstrings is empty"));
276 1 : }
277 :
278 : // check for uniqueness of the connection strings in the set
279 1 : let mut connstrings = self.safekeeper_connstrings.clone();
280 :
281 1 : connstrings.sort();
282 1 : let mut previous = &connstrings[0];
283 :
284 2 : for current in connstrings.iter().skip(1) {
285 : // duplicate entry?
286 2 : if current == previous {
287 1 : return Err(format!(
288 1 : "duplicate entry in safekeeper_connstrings: {current}!",
289 1 : ));
290 1 : }
291 :
292 1 : previous = current;
293 : }
294 :
295 0 : Ok(())
296 1 : }
297 : }
298 :
299 : impl TryFrom<ComputeSpec> for ParsedSpec {
300 : type Error = anyhow::Error;
301 1 : fn try_from(spec: ComputeSpec) -> Result<Self, anyhow::Error> {
302 : // Extract the options from the spec file that are needed to connect to
303 : // the storage system.
304 : //
305 : // In compute specs generated by old control plane versions, the spec file might
306 : // be missing the `pageserver_connection_info` field. In that case, we need to dig
307 : // the pageserver connection info from the `pageserver_connstr` field instead, or
308 : // if that's missing too, from the GUC in the cluster.settings field.
309 1 : let mut pageserver_conninfo = spec.pageserver_connection_info.clone();
310 1 : if pageserver_conninfo.is_none() {
311 1 : if let Some(pageserver_connstr_field) = &spec.pageserver_connstring {
312 0 : pageserver_conninfo = Some(PageserverConnectionInfo::from_connstr(
313 0 : pageserver_connstr_field,
314 0 : spec.shard_stripe_size,
315 0 : )?);
316 1 : }
317 0 : }
318 1 : if pageserver_conninfo.is_none() {
319 1 : if let Some(guc) = spec.cluster.settings.find("neon.pageserver_connstring") {
320 1 : let stripe_size = if let Some(guc) = spec.cluster.settings.find("neon.stripe_size")
321 : {
322 0 : Some(ShardStripeSize(u32::from_str(&guc)?))
323 : } else {
324 1 : None
325 : };
326 1 : pageserver_conninfo =
327 1 : Some(PageserverConnectionInfo::from_connstr(&guc, stripe_size)?);
328 0 : }
329 0 : }
330 1 : let pageserver_conninfo = pageserver_conninfo.ok_or(anyhow::anyhow!(
331 1 : "pageserver connection information should be provided"
332 0 : ))?;
333 :
334 : // Similarly for safekeeper connection strings
335 1 : let safekeeper_connstrings = if spec.safekeeper_connstrings.is_empty() {
336 1 : if matches!(spec.mode, ComputeMode::Primary) {
337 1 : spec.cluster
338 1 : .settings
339 1 : .find("neon.safekeepers")
340 1 : .ok_or(anyhow::anyhow!("safekeeper connstrings should be provided"))?
341 1 : .split(',')
342 4 : .map(|str| str.to_string())
343 1 : .collect()
344 : } else {
345 0 : vec![]
346 : }
347 : } else {
348 0 : spec.safekeeper_connstrings.clone()
349 : };
350 :
351 1 : let storage_auth_token = spec.storage_auth_token.clone();
352 1 : let tenant_id: TenantId = if let Some(tenant_id) = spec.tenant_id {
353 0 : tenant_id
354 : } else {
355 1 : let guc = spec
356 1 : .cluster
357 1 : .settings
358 1 : .find("neon.tenant_id")
359 1 : .ok_or(anyhow::anyhow!("tenant id should be provided"))?;
360 1 : TenantId::from_str(&guc).context("invalid tenant id")?
361 : };
362 1 : let timeline_id: TimelineId = if let Some(timeline_id) = spec.timeline_id {
363 0 : timeline_id
364 : } else {
365 1 : let guc = spec
366 1 : .cluster
367 1 : .settings
368 1 : .find("neon.timeline_id")
369 1 : .ok_or(anyhow::anyhow!("timeline id should be provided"))?;
370 1 : TimelineId::from_str(&guc).context(anyhow::anyhow!("invalid timeline id"))?
371 : };
372 :
373 1 : let endpoint_storage_addr: Option<String> = spec
374 1 : .endpoint_storage_addr
375 1 : .clone()
376 1 : .or_else(|| spec.cluster.settings.find("neon.endpoint_storage_addr"));
377 1 : let endpoint_storage_token = spec
378 1 : .endpoint_storage_token
379 1 : .clone()
380 1 : .or_else(|| spec.cluster.settings.find("neon.endpoint_storage_token"));
381 :
382 1 : let res = ParsedSpec {
383 1 : spec,
384 1 : pageserver_conninfo,
385 1 : safekeeper_connstrings,
386 1 : storage_auth_token,
387 1 : tenant_id,
388 1 : timeline_id,
389 1 : endpoint_storage_addr,
390 1 : endpoint_storage_token,
391 1 : };
392 :
393 : // Now check validity of the parsed specification
394 1 : res.validate().map_err(anyhow::Error::msg)?;
395 0 : Ok(res)
396 1 : }
397 : }
398 :
399 : /// If we are a VM, returns a [`Command`] that will run in the `neon-postgres`
400 : /// cgroup. Otherwise returns the default `Command::new(cmd)`
401 : ///
402 : /// This function should be used to start postgres, as it will start it in the
403 : /// neon-postgres cgroup if we are a VM. This allows autoscaling to control
404 : /// postgres' resource usage. The cgroup will exist in VMs because vm-builder
405 : /// creates it during the sysinit phase of its inittab.
406 0 : fn maybe_cgexec(cmd: &str) -> Command {
407 : // The cplane sets this env var for autoscaling computes.
408 : // use `var_os` so we don't have to worry about the variable being valid
409 : // unicode. Should never be an concern . . . but just in case
410 0 : if env::var_os("AUTOSCALING").is_some() {
411 0 : let mut command = Command::new("cgexec");
412 0 : command.args(["-g", "memory:neon-postgres"]);
413 0 : command.arg(cmd);
414 0 : command
415 : } else {
416 0 : Command::new(cmd)
417 : }
418 0 : }
419 :
420 : struct PostgresHandle {
421 : postgres: std::process::Child,
422 : log_collector: JoinHandle<Result<()>>,
423 : }
424 :
425 : impl PostgresHandle {
426 : /// Return PID of the postgres (postmaster) process
427 0 : fn pid(&self) -> Pid {
428 0 : Pid::from_raw(self.postgres.id() as i32)
429 0 : }
430 : }
431 :
432 : struct StartVmMonitorResult {
433 : #[cfg(target_os = "linux")]
434 : token: tokio_util::sync::CancellationToken,
435 : #[cfg(target_os = "linux")]
436 : vm_monitor: Option<JoinHandle<Result<()>>>,
437 : }
438 :
439 : // BEGIN_HADRON
440 : /// This function creates roles that are used by Databricks.
441 : /// These roles are not needs to be botostrapped at PG Compute provisioning time.
442 : /// The auth method for these roles are configured in databricks_pg_hba.conf in universe repository.
443 0 : pub(crate) fn create_databricks_roles() -> Vec<String> {
444 0 : let roles = vec![
445 : // Role for prometheus_stats_exporter
446 0 : Role {
447 0 : name: "databricks_monitor".to_string(),
448 0 : // This uses "local" connection and auth method for that is "trust", so no password is needed.
449 0 : encrypted_password: None,
450 0 : options: Some(vec![GenericOption {
451 0 : name: "IN ROLE pg_monitor".to_string(),
452 0 : value: None,
453 0 : vartype: "string".to_string(),
454 0 : }]),
455 0 : },
456 : // Role for brickstore control plane
457 0 : Role {
458 0 : name: "databricks_control_plane".to_string(),
459 0 : // Certificate user does not need password.
460 0 : encrypted_password: None,
461 0 : options: Some(vec![GenericOption {
462 0 : name: "SUPERUSER".to_string(),
463 0 : value: None,
464 0 : vartype: "string".to_string(),
465 0 : }]),
466 0 : },
467 : // Role for brickstore httpgateway.
468 0 : Role {
469 0 : name: "databricks_gateway".to_string(),
470 0 : // Certificate user does not need password.
471 0 : encrypted_password: None,
472 0 : options: None,
473 0 : },
474 : ];
475 :
476 0 : roles
477 0 : .into_iter()
478 0 : .map(|role| {
479 0 : let query = format!(
480 0 : r#"
481 0 : DO $$
482 0 : BEGIN
483 0 : IF NOT EXISTS (
484 0 : SELECT FROM pg_catalog.pg_roles WHERE rolname = '{}')
485 0 : THEN
486 0 : CREATE ROLE {} {};
487 0 : END IF;
488 0 : END
489 0 : $$;"#,
490 : role.name,
491 0 : role.name.pg_quote(),
492 0 : role.to_pg_options(),
493 : );
494 0 : query
495 0 : })
496 0 : .collect()
497 0 : }
498 :
499 : /// Databricks-specific environment variables to be passed to the `postgres` sub-process.
500 : pub struct DatabricksEnvVars {
501 : /// The Databricks "endpoint ID" of the compute instance. Used by `postgres` to check
502 : /// the token scopes of internal auth tokens.
503 : pub endpoint_id: String,
504 : /// Hostname of the Databricks workspace URL this compute instance belongs to.
505 : /// Used by postgres to verify Databricks PAT tokens.
506 : pub workspace_host: String,
507 :
508 : pub lakebase_mode: bool,
509 : }
510 :
511 : impl DatabricksEnvVars {
512 0 : pub fn new(
513 0 : compute_spec: &ComputeSpec,
514 0 : compute_id: Option<&String>,
515 0 : instance_id: Option<String>,
516 0 : lakebase_mode: bool,
517 0 : ) -> Self {
518 0 : let endpoint_id = if let Some(instance_id) = instance_id {
519 : // Use instance_id as endpoint_id if it is set. This code path is for PuPr model.
520 0 : instance_id
521 : } else {
522 : // Use compute_id as endpoint_id if instance_id is not set. The code path is for PrPr model.
523 : // compute_id is a string format of "{endpoint_id}/{compute_idx}"
524 : // endpoint_id is a uuid. We only need to pass down endpoint_id to postgres.
525 : // Panics if compute_id is not set or not in the expected format.
526 0 : compute_id.unwrap().split('/').next().unwrap().to_string()
527 : };
528 0 : let workspace_host = compute_spec
529 0 : .databricks_settings
530 0 : .as_ref()
531 0 : .map(|s| s.databricks_workspace_host.clone())
532 0 : .unwrap_or("".to_string());
533 0 : Self {
534 0 : endpoint_id,
535 0 : workspace_host,
536 0 : lakebase_mode,
537 0 : }
538 0 : }
539 :
540 : /// Constants for the names of Databricks-specific postgres environment variables.
541 : const DATABRICKS_ENDPOINT_ID_ENVVAR: &'static str = "DATABRICKS_ENDPOINT_ID";
542 : const DATABRICKS_WORKSPACE_HOST_ENVVAR: &'static str = "DATABRICKS_WORKSPACE_HOST";
543 :
544 : /// Convert DatabricksEnvVars to a list of string pairs that can be passed as env vars. Consumes `self`.
545 0 : pub fn to_env_var_list(self) -> Vec<(String, String)> {
546 0 : if !self.lakebase_mode {
547 : // In neon env, we don't need to pass down the env vars to postgres.
548 0 : return vec![];
549 0 : }
550 0 : vec![
551 0 : (
552 0 : Self::DATABRICKS_ENDPOINT_ID_ENVVAR.to_string(),
553 0 : self.endpoint_id.clone(),
554 0 : ),
555 0 : (
556 0 : Self::DATABRICKS_WORKSPACE_HOST_ENVVAR.to_string(),
557 0 : self.workspace_host.clone(),
558 0 : ),
559 : ]
560 0 : }
561 : }
562 :
563 : impl ComputeNode {
564 0 : pub fn new(params: ComputeNodeParams, config: ComputeConfig) -> Result<Self> {
565 0 : let connstr = params.connstr.as_str();
566 0 : let mut conn_conf = postgres::config::Config::from_str(connstr)
567 0 : .context("cannot build postgres config from connstr")?;
568 0 : let mut tokio_conn_conf = tokio_postgres::config::Config::from_str(connstr)
569 0 : .context("cannot build tokio postgres config from connstr")?;
570 :
571 : // Users can set some configuration parameters per database with
572 : // ALTER DATABASE ... SET ...
573 : //
574 : // There are at least these parameters:
575 : //
576 : // - role=some_other_role
577 : // - default_transaction_read_only=on
578 : // - statement_timeout=1, i.e., 1ms, which will cause most of the queries to fail
579 : // - search_path=non_public_schema, this should be actually safe because
580 : // we don't call any functions in user databases, but better to always reset
581 : // it to public.
582 : //
583 : // that can affect `compute_ctl` and prevent it from properly configuring the database schema.
584 : // Unset them via connection string options before connecting to the database.
585 : // N.B. keep it in sync with `ZENITH_OPTIONS` in `get_maintenance_client()`.
586 : const EXTRA_OPTIONS: &str = "-c role=cloud_admin -c default_transaction_read_only=off -c search_path=public -c statement_timeout=0 -c pgaudit.log=none";
587 0 : let options = match conn_conf.get_options() {
588 : // Allow the control plane to override any options set by the
589 : // compute
590 0 : Some(options) => format!("{EXTRA_OPTIONS} {options}"),
591 0 : None => EXTRA_OPTIONS.to_string(),
592 : };
593 0 : conn_conf.options(&options);
594 0 : tokio_conn_conf.options(&options);
595 :
596 0 : let mut new_state = ComputeState::new();
597 0 : if let Some(spec) = config.spec {
598 0 : let pspec = ParsedSpec::try_from(spec).map_err(|msg| anyhow::anyhow!(msg))?;
599 0 : if params.lakebase_mode {
600 0 : ComputeNode::set_spec(¶ms, &mut new_state, pspec);
601 0 : } else {
602 0 : new_state.pspec = Some(pspec);
603 0 : }
604 0 : }
605 :
606 0 : Ok(ComputeNode {
607 0 : params,
608 0 : conn_conf,
609 0 : tokio_conn_conf,
610 0 : state: Mutex::new(new_state),
611 0 : state_changed: Condvar::new(),
612 0 : ext_download_progress: RwLock::new(HashMap::new()),
613 0 : compute_ctl_config: config.compute_ctl_config,
614 0 : extension_stats_task: Mutex::new(None),
615 0 : lfc_offload_task: Mutex::new(None),
616 0 : })
617 0 : }
618 :
619 : /// Top-level control flow of compute_ctl. Returns a process exit code we should
620 : /// exit with.
621 0 : pub fn run(self) -> Result<Option<i32>> {
622 0 : let this = Arc::new(self);
623 :
624 0 : let cli_spec = this.state.lock().unwrap().pspec.clone();
625 :
626 : // If this is a pooled VM, prewarm before starting HTTP server and becoming
627 : // available for binding. Prewarming helps Postgres start quicker later,
628 : // because QEMU will already have its memory allocated from the host, and
629 : // the necessary binaries will already be cached.
630 0 : if cli_spec.is_none() {
631 0 : this.prewarm_postgres_vm_memory()?;
632 0 : }
633 :
634 : // Set the up metric with Empty status before starting the HTTP server.
635 : // That way on the first metric scrape, an external observer will see us
636 : // as 'up' and 'empty' (unless the compute was started with a spec or
637 : // already configured by control plane).
638 0 : COMPUTE_CTL_UP
639 0 : .with_label_values(&[&BUILD_TAG, ComputeStatus::Empty.to_string().as_str()])
640 0 : .set(1);
641 :
642 : // Launch the external HTTP server first, so that we can serve control plane
643 : // requests while configuration is still in progress.
644 0 : crate::http::server::Server::External {
645 0 : port: this.params.external_http_port,
646 0 : config: this.compute_ctl_config.clone(),
647 0 : compute_id: this.params.compute_id.clone(),
648 0 : instance_id: this.params.instance_id.clone(),
649 0 : }
650 0 : .launch(&this);
651 :
652 : // The internal HTTP server could be launched later, but there isn't much
653 : // sense in waiting.
654 0 : crate::http::server::Server::Internal {
655 0 : port: this.params.internal_http_port,
656 0 : }
657 0 : .launch(&this);
658 :
659 : // If we got a spec from the CLI already, use that. Otherwise wait for the
660 : // control plane to pass it to us with a /configure HTTP request
661 0 : let pspec = if let Some(cli_spec) = cli_spec {
662 0 : cli_spec
663 : } else {
664 0 : this.wait_spec()?
665 : };
666 :
667 0 : launch_lsn_lease_bg_task_for_static(&this);
668 :
669 : // We have a spec, start the compute
670 0 : let mut delay_exit = false;
671 0 : let mut vm_monitor = None;
672 0 : let mut pg_process: Option<PostgresHandle> = None;
673 :
674 0 : match this.start_compute(&mut pg_process) {
675 0 : Ok(()) => {
676 0 : // Success! Launch remaining services (just vm-monitor currently)
677 0 : vm_monitor =
678 0 : Some(this.start_vm_monitor(pspec.spec.disable_lfc_resizing.unwrap_or(false)));
679 0 : }
680 0 : Err(err) => {
681 : // Something went wrong with the startup. Log it and expose the error to
682 : // HTTP status requests.
683 0 : error!("could not start the compute node: {:#}", err);
684 0 : this.set_failed_status(err);
685 0 : delay_exit = true;
686 :
687 : // If the error happened after starting PostgreSQL, kill it
688 0 : if let Some(ref pg_process) = pg_process {
689 0 : kill(pg_process.pid(), Signal::SIGQUIT).ok();
690 0 : }
691 : }
692 : }
693 :
694 : // If startup was successful, or it failed in the late stages,
695 : // PostgreSQL is now running. Wait until it exits.
696 0 : let exit_code = if let Some(pg_handle) = pg_process {
697 0 : let exit_status = this.wait_postgres(pg_handle);
698 0 : info!("Postgres exited with code {}, shutting down", exit_status);
699 0 : exit_status.code()
700 : } else {
701 0 : None
702 : };
703 :
704 0 : this.terminate_extension_stats_task();
705 0 : this.terminate_lfc_offload_task();
706 :
707 : // Terminate the vm_monitor so it releases the file watcher on
708 : // /sys/fs/cgroup/neon-postgres.
709 : // Note: the vm-monitor only runs on linux because it requires cgroups.
710 0 : if let Some(vm_monitor) = vm_monitor {
711 : cfg_if::cfg_if! {
712 : if #[cfg(target_os = "linux")] {
713 : // Kills all threads spawned by the monitor
714 0 : vm_monitor.token.cancel();
715 0 : if let Some(handle) = vm_monitor.vm_monitor {
716 0 : // Kills the actual task running the monitor
717 0 : handle.abort();
718 0 : }
719 : } else {
720 : _ = vm_monitor; // appease unused lint on macOS
721 : }
722 : }
723 0 : }
724 :
725 : // Reap the postgres process
726 0 : delay_exit |= this.cleanup_after_postgres_exit()?;
727 :
728 : // /terminate returns LSN. If we don't sleep at all, connection will break and we
729 : // won't get result. If we sleep too much, tests will take significantly longer
730 : // and Github Action run will error out
731 0 : let sleep_duration = if delay_exit {
732 0 : Duration::from_secs(30)
733 : } else {
734 0 : Duration::from_millis(300)
735 : };
736 :
737 : // If launch failed, keep serving HTTP requests for a while, so the cloud
738 : // control plane can get the actual error.
739 0 : if delay_exit {
740 0 : info!("giving control plane 30s to collect the error before shutdown");
741 0 : }
742 0 : std::thread::sleep(sleep_duration);
743 0 : Ok(exit_code)
744 0 : }
745 :
746 0 : pub fn wait_spec(&self) -> Result<ParsedSpec> {
747 0 : info!("no compute spec provided, waiting");
748 0 : let mut state = self.state.lock().unwrap();
749 0 : while state.status != ComputeStatus::ConfigurationPending {
750 0 : state = self.state_changed.wait(state).unwrap();
751 0 : }
752 :
753 0 : info!("got spec, continue configuration");
754 0 : let spec = state.pspec.as_ref().unwrap().clone();
755 :
756 : // Record for how long we slept waiting for the spec.
757 0 : let now = Utc::now();
758 0 : state.metrics.wait_for_spec_ms = now
759 0 : .signed_duration_since(state.start_time)
760 0 : .to_std()
761 0 : .unwrap()
762 0 : .as_millis() as u64;
763 :
764 : // Reset start time, so that the total startup time that is calculated later will
765 : // not include the time that we waited for the spec.
766 0 : state.start_time = now;
767 :
768 0 : Ok(spec)
769 0 : }
770 :
771 : /// Start compute.
772 : ///
773 : /// Prerequisites:
774 : /// - the compute spec has been placed in self.state.pspec
775 : ///
776 : /// On success:
777 : /// - status is set to ComputeStatus::Running
778 : /// - self.running_postgres is set
779 : ///
780 : /// On error:
781 : /// - status is left in ComputeStatus::Init. The caller is responsible for setting it to Failed
782 : /// - if Postgres was started before the fatal error happened, self.running_postgres is
783 : /// set. The caller is responsible for killing it.
784 : ///
785 : /// Note that this is in the critical path of a compute cold start. Keep this fast.
786 : /// Try to do things concurrently, to hide the latencies.
787 0 : fn start_compute(self: &Arc<Self>, pg_handle: &mut Option<PostgresHandle>) -> Result<()> {
788 : let compute_state: ComputeState;
789 :
790 : let start_compute_span;
791 : let _this_entered;
792 : {
793 0 : let mut state_guard = self.state.lock().unwrap();
794 :
795 : // Create a tracing span for the startup operation.
796 : //
797 : // We could otherwise just annotate the function with #[instrument], but if
798 : // we're being configured from a /configure HTTP request, we want the
799 : // startup to be considered part of the /configure request.
800 : //
801 : // Similarly, if a trace ID was passed in env variables, attach it to the span.
802 0 : start_compute_span = {
803 : // Temporarily enter the parent span, so that the new span becomes its child.
804 0 : if let Some(p) = state_guard.startup_span.take() {
805 0 : let _parent_entered = p.entered();
806 0 : tracing::info_span!("start_compute")
807 0 : } else if let Some(otel_context) = startup_context_from_env() {
808 : use tracing_opentelemetry::OpenTelemetrySpanExt;
809 0 : let span = tracing::info_span!("start_compute");
810 0 : span.set_parent(otel_context);
811 0 : span
812 : } else {
813 0 : tracing::info_span!("start_compute")
814 : }
815 : };
816 0 : _this_entered = start_compute_span.enter();
817 :
818 : // Hadron: Record postgres start time (used to enforce pg_init_timeout).
819 0 : state_guard.pg_start_time.replace(Utc::now());
820 :
821 0 : state_guard.set_status(ComputeStatus::Init, &self.state_changed);
822 0 : compute_state = state_guard.clone()
823 : }
824 :
825 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
826 0 : info!(
827 0 : "starting compute for project {}, operation {}, tenant {}, timeline {}, project {}, branch {}, endpoint {}, features {:?}, spec.remote_extensions {:?}",
828 0 : pspec.spec.cluster.cluster_id.as_deref().unwrap_or("None"),
829 0 : pspec.spec.operation_uuid.as_deref().unwrap_or("None"),
830 : pspec.tenant_id,
831 : pspec.timeline_id,
832 0 : pspec.spec.project_id.as_deref().unwrap_or("None"),
833 0 : pspec.spec.branch_id.as_deref().unwrap_or("None"),
834 0 : pspec.spec.endpoint_id.as_deref().unwrap_or("None"),
835 : pspec.spec.features,
836 : pspec.spec.remote_extensions,
837 : );
838 :
839 : ////// PRE-STARTUP PHASE: things that need to be finished before we start the Postgres process
840 :
841 : // Collect all the tasks that must finish here
842 0 : let mut pre_tasks = tokio::task::JoinSet::new();
843 :
844 : // Make sure TLS certificates are properly loaded and in the right place.
845 0 : if self.compute_ctl_config.tls.is_some() {
846 0 : let this = self.clone();
847 0 : pre_tasks.spawn(async move {
848 0 : this.watch_cert_for_changes().await;
849 :
850 0 : Ok::<(), anyhow::Error>(())
851 0 : });
852 0 : }
853 :
854 0 : let tls_config = self.tls_config(&pspec.spec);
855 :
856 : // If there are any remote extensions in shared_preload_libraries, start downloading them
857 0 : if pspec.spec.remote_extensions.is_some() {
858 0 : let (this, spec) = (self.clone(), pspec.spec.clone());
859 0 : pre_tasks.spawn(async move {
860 0 : this.download_preload_extensions(&spec)
861 0 : .in_current_span()
862 0 : .await
863 0 : });
864 0 : }
865 :
866 : // Prepare pgdata directory. This downloads the basebackup, among other things.
867 : {
868 0 : let (this, cs) = (self.clone(), compute_state.clone());
869 0 : pre_tasks.spawn_blocking_child(move || this.prepare_pgdata(&cs));
870 : }
871 :
872 : // Resize swap to the desired size if the compute spec says so
873 0 : if let (Some(size_bytes), true) =
874 0 : (pspec.spec.swap_size_bytes, self.params.resize_swap_on_bind)
875 : {
876 0 : pre_tasks.spawn_blocking_child(move || {
877 : // To avoid 'swapoff' hitting postgres startup, we need to run resize-swap to completion
878 : // *before* starting postgres.
879 : //
880 : // In theory, we could do this asynchronously if SkipSwapon was enabled for VMs, but this
881 : // carries a risk of introducing hard-to-debug issues - e.g. if postgres sometimes gets
882 : // OOM-killed during startup because swap wasn't available yet.
883 0 : resize_swap(size_bytes).context("failed to resize swap")?;
884 0 : let size_mib = size_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
885 0 : info!(%size_bytes, %size_mib, "resized swap");
886 :
887 0 : Ok::<(), anyhow::Error>(())
888 0 : });
889 0 : }
890 :
891 : // Set disk quota if the compute spec says so
892 0 : if let (Some(disk_quota_bytes), Some(disk_quota_fs_mountpoint)) = (
893 0 : pspec.spec.disk_quota_bytes,
894 0 : self.params.set_disk_quota_for_fs.as_ref(),
895 : ) {
896 0 : let disk_quota_fs_mountpoint = disk_quota_fs_mountpoint.clone();
897 0 : pre_tasks.spawn_blocking_child(move || {
898 0 : set_disk_quota(disk_quota_bytes, &disk_quota_fs_mountpoint)
899 0 : .context("failed to set disk quota")?;
900 0 : let size_mib = disk_quota_bytes as f32 / (1 << 20) as f32; // just for more coherent display.
901 0 : info!(%disk_quota_bytes, %size_mib, "set disk quota");
902 :
903 0 : Ok::<(), anyhow::Error>(())
904 0 : });
905 0 : }
906 :
907 : // tune pgbouncer
908 0 : if let Some(pgbouncer_settings) = &pspec.spec.pgbouncer_settings {
909 0 : info!("tuning pgbouncer");
910 :
911 0 : let pgbouncer_settings = pgbouncer_settings.clone();
912 0 : let tls_config = tls_config.clone();
913 :
914 : // Spawn a background task to do the tuning,
915 : // so that we don't block the main thread that starts Postgres.
916 0 : let _handle = tokio::spawn(async move {
917 0 : let res = tune_pgbouncer(pgbouncer_settings, tls_config).await;
918 0 : if let Err(err) = res {
919 0 : error!("error while tuning pgbouncer: {err:?}");
920 : // Continue with the startup anyway
921 0 : }
922 0 : });
923 0 : }
924 :
925 : // configure local_proxy
926 0 : if let Some(local_proxy) = &pspec.spec.local_proxy_config {
927 0 : info!("configuring local_proxy");
928 :
929 : // Spawn a background task to do the configuration,
930 : // so that we don't block the main thread that starts Postgres.
931 :
932 0 : let mut local_proxy = local_proxy.clone();
933 0 : local_proxy.tls = tls_config.clone();
934 :
935 0 : let _handle = tokio::spawn(async move {
936 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
937 0 : error!("error while configuring local_proxy: {err:?}");
938 : // Continue with the startup anyway
939 0 : }
940 0 : });
941 0 : }
942 :
943 : // Configure and start rsyslog for compliance audit logging
944 0 : match pspec.spec.audit_log_level {
945 : ComputeAudit::Hipaa | ComputeAudit::Extended | ComputeAudit::Full => {
946 0 : let remote_tls_endpoint =
947 0 : std::env::var("AUDIT_LOGGING_TLS_ENDPOINT").unwrap_or("".to_string());
948 0 : let remote_plain_endpoint =
949 0 : std::env::var("AUDIT_LOGGING_ENDPOINT").unwrap_or("".to_string());
950 :
951 0 : if remote_plain_endpoint.is_empty() && remote_tls_endpoint.is_empty() {
952 0 : anyhow::bail!(
953 0 : "AUDIT_LOGGING_ENDPOINT and AUDIT_LOGGING_TLS_ENDPOINT are both empty"
954 : );
955 0 : }
956 :
957 0 : let log_directory_path = Path::new(&self.params.pgdata).join("log");
958 0 : let log_directory_path = log_directory_path.to_string_lossy().to_string();
959 :
960 : // Add project_id,endpoint_id to identify the logs.
961 : //
962 : // These ids are passed from cplane,
963 0 : let endpoint_id = pspec.spec.endpoint_id.as_deref().unwrap_or("");
964 0 : let project_id = pspec.spec.project_id.as_deref().unwrap_or("");
965 :
966 0 : configure_audit_rsyslog(
967 0 : log_directory_path.clone(),
968 0 : endpoint_id,
969 0 : project_id,
970 0 : &remote_plain_endpoint,
971 0 : &remote_tls_endpoint,
972 0 : )?;
973 :
974 : // Launch a background task to clean up the audit logs
975 0 : launch_pgaudit_gc(log_directory_path);
976 : }
977 0 : _ => {}
978 : }
979 :
980 : // Configure and start rsyslog for Postgres logs export
981 0 : let conf = PostgresLogsRsyslogConfig::new(pspec.spec.logs_export_host.as_deref());
982 0 : configure_postgres_logs_export(conf)?;
983 :
984 : // Launch remaining service threads
985 0 : let _monitor_handle = launch_monitor(self);
986 0 : let _configurator_handle = launch_configurator(self);
987 :
988 : // Wait for all the pre-tasks to finish before starting postgres
989 0 : let rt = tokio::runtime::Handle::current();
990 0 : while let Some(res) = rt.block_on(pre_tasks.join_next()) {
991 0 : res??;
992 : }
993 :
994 : ////// START POSTGRES
995 0 : let start_time = Utc::now();
996 0 : let pg_process = self.start_postgres(pspec.storage_auth_token.clone())?;
997 0 : let postmaster_pid = pg_process.pid();
998 0 : *pg_handle = Some(pg_process);
999 :
1000 : // If this is a primary endpoint, perform some post-startup configuration before
1001 : // opening it up for the world.
1002 0 : let config_time = Utc::now();
1003 0 : if pspec.spec.mode == ComputeMode::Primary {
1004 0 : self.configure_as_primary(&compute_state)?;
1005 :
1006 0 : let conf = self.get_tokio_conn_conf(None);
1007 0 : tokio::task::spawn(async {
1008 0 : let _ = installed_extensions(conf).await;
1009 0 : });
1010 0 : }
1011 :
1012 : // All done!
1013 0 : let startup_end_time = Utc::now();
1014 0 : let metrics = {
1015 0 : let mut state = self.state.lock().unwrap();
1016 0 : state.metrics.start_postgres_ms = config_time
1017 0 : .signed_duration_since(start_time)
1018 0 : .to_std()
1019 0 : .unwrap()
1020 0 : .as_millis() as u64;
1021 0 : state.metrics.config_ms = startup_end_time
1022 0 : .signed_duration_since(config_time)
1023 0 : .to_std()
1024 0 : .unwrap()
1025 0 : .as_millis() as u64;
1026 0 : state.metrics.total_startup_ms = startup_end_time
1027 0 : .signed_duration_since(compute_state.start_time)
1028 0 : .to_std()
1029 0 : .unwrap()
1030 0 : .as_millis() as u64;
1031 0 : state.metrics.clone()
1032 : };
1033 0 : self.set_status(ComputeStatus::Running);
1034 :
1035 : // Log metrics so that we can search for slow operations in logs
1036 0 : info!(?metrics, postmaster_pid = %postmaster_pid, "compute start finished");
1037 :
1038 0 : self.spawn_extension_stats_task();
1039 :
1040 0 : if pspec.spec.autoprewarm {
1041 0 : info!("autoprewarming on startup as requested");
1042 0 : self.prewarm_lfc(None);
1043 0 : }
1044 0 : if let Some(seconds) = pspec.spec.offload_lfc_interval_seconds {
1045 0 : self.spawn_lfc_offload_task(Duration::from_secs(seconds.into()));
1046 0 : };
1047 0 : Ok(())
1048 0 : }
1049 :
1050 : #[instrument(skip_all)]
1051 : async fn download_preload_extensions(&self, spec: &ComputeSpec) -> Result<()> {
1052 : let remote_extensions = if let Some(remote_extensions) = &spec.remote_extensions {
1053 : remote_extensions
1054 : } else {
1055 : return Ok(());
1056 : };
1057 :
1058 : // First, create control files for all available extensions
1059 : extension_server::create_control_files(remote_extensions, &self.params.pgbin);
1060 :
1061 : let library_load_start_time = Utc::now();
1062 : let remote_ext_metrics = self.prepare_preload_libraries(spec).await?;
1063 :
1064 : let library_load_time = Utc::now()
1065 : .signed_duration_since(library_load_start_time)
1066 : .to_std()
1067 : .unwrap()
1068 : .as_millis() as u64;
1069 : let mut state = self.state.lock().unwrap();
1070 : state.metrics.load_ext_ms = library_load_time;
1071 : state.metrics.num_ext_downloaded = remote_ext_metrics.num_ext_downloaded;
1072 : state.metrics.largest_ext_size = remote_ext_metrics.largest_ext_size;
1073 : state.metrics.total_ext_download_size = remote_ext_metrics.total_ext_download_size;
1074 : info!(
1075 : "Loading shared_preload_libraries took {:?}ms",
1076 : library_load_time
1077 : );
1078 : info!("{:?}", remote_ext_metrics);
1079 :
1080 : Ok(())
1081 : }
1082 :
1083 : /// Start the vm-monitor if directed to. The vm-monitor only runs on linux
1084 : /// because it requires cgroups.
1085 0 : fn start_vm_monitor(&self, disable_lfc_resizing: bool) -> StartVmMonitorResult {
1086 : cfg_if::cfg_if! {
1087 : if #[cfg(target_os = "linux")] {
1088 : use std::env;
1089 : use tokio_util::sync::CancellationToken;
1090 :
1091 : // This token is used internally by the monitor to clean up all threads
1092 0 : let token = CancellationToken::new();
1093 :
1094 : // don't pass postgres connection string to vm-monitor if we don't want it to resize LFC
1095 0 : let pgconnstr = if disable_lfc_resizing {
1096 0 : None
1097 : } else {
1098 0 : Some(self.params.filecache_connstr.clone())
1099 : };
1100 :
1101 0 : let vm_monitor = if env::var_os("AUTOSCALING").is_some() {
1102 0 : let vm_monitor = tokio::spawn(vm_monitor::start(
1103 0 : Box::leak(Box::new(vm_monitor::Args {
1104 0 : cgroup: Some(self.params.cgroup.clone()),
1105 0 : pgconnstr,
1106 0 : addr: self.params.vm_monitor_addr.clone(),
1107 0 : })),
1108 0 : token.clone(),
1109 : ));
1110 0 : Some(vm_monitor)
1111 : } else {
1112 0 : None
1113 : };
1114 0 : StartVmMonitorResult { token, vm_monitor }
1115 : } else {
1116 : _ = disable_lfc_resizing; // appease unused lint on macOS
1117 : StartVmMonitorResult { }
1118 : }
1119 : }
1120 0 : }
1121 :
1122 0 : fn cleanup_after_postgres_exit(&self) -> Result<bool> {
1123 : // Maybe sync safekeepers again, to speed up next startup
1124 0 : let compute_state = self.state.lock().unwrap().clone();
1125 0 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1126 0 : let lsn = if matches!(pspec.spec.mode, compute_api::spec::ComputeMode::Primary) {
1127 0 : info!("syncing safekeepers on shutdown");
1128 0 : let storage_auth_token = pspec.storage_auth_token.clone();
1129 0 : let lsn = self.sync_safekeepers(storage_auth_token)?;
1130 0 : info!(%lsn, "synced safekeepers");
1131 0 : Some(lsn)
1132 : } else {
1133 0 : info!("not primary, not syncing safekeepers");
1134 0 : None
1135 : };
1136 :
1137 0 : let mut state = self.state.lock().unwrap();
1138 0 : state.terminate_flush_lsn = lsn;
1139 :
1140 0 : let delay_exit = state.status == ComputeStatus::TerminationPendingFast;
1141 0 : if state.status == ComputeStatus::TerminationPendingFast
1142 0 : || state.status == ComputeStatus::TerminationPendingImmediate
1143 : {
1144 0 : info!(
1145 0 : "Changing compute status from {} to {}",
1146 0 : state.status,
1147 : ComputeStatus::Terminated
1148 : );
1149 0 : state.status = ComputeStatus::Terminated;
1150 0 : self.state_changed.notify_all();
1151 0 : }
1152 0 : drop(state);
1153 :
1154 0 : if let Err(err) = self.check_for_core_dumps() {
1155 0 : error!("error while checking for core dumps: {err:?}");
1156 0 : }
1157 :
1158 0 : Ok(delay_exit)
1159 0 : }
1160 :
1161 : /// Check that compute node has corresponding feature enabled.
1162 0 : pub fn has_feature(&self, feature: ComputeFeature) -> bool {
1163 0 : let state = self.state.lock().unwrap();
1164 :
1165 0 : if let Some(s) = state.pspec.as_ref() {
1166 0 : s.spec.features.contains(&feature)
1167 : } else {
1168 0 : false
1169 : }
1170 0 : }
1171 :
1172 0 : pub fn set_status(&self, status: ComputeStatus) {
1173 0 : let mut state = self.state.lock().unwrap();
1174 0 : state.set_status(status, &self.state_changed);
1175 0 : }
1176 :
1177 0 : pub fn set_failed_status(&self, err: anyhow::Error) {
1178 0 : let mut state = self.state.lock().unwrap();
1179 0 : state.set_failed_status(err, &self.state_changed);
1180 0 : }
1181 :
1182 0 : pub fn get_status(&self) -> ComputeStatus {
1183 0 : self.state.lock().unwrap().status
1184 0 : }
1185 :
1186 0 : pub fn get_timeline_id(&self) -> Option<TimelineId> {
1187 0 : self.state
1188 0 : .lock()
1189 0 : .unwrap()
1190 0 : .pspec
1191 0 : .as_ref()
1192 0 : .map(|s| s.timeline_id)
1193 0 : }
1194 :
1195 : // Remove `pgdata` directory and create it again with right permissions.
1196 0 : fn create_pgdata(&self) -> Result<()> {
1197 : // Ignore removal error, likely it is a 'No such file or directory (os error 2)'.
1198 : // If it is something different then create_dir() will error out anyway.
1199 0 : let pgdata = &self.params.pgdata;
1200 0 : let _ok = fs::remove_dir_all(pgdata);
1201 0 : if self.params.lakebase_mode {
1202 0 : // Ignore creation errors if the directory already exists (e.g. mounting it ahead of time).
1203 0 : // If it is something different then PG startup will error out anyway.
1204 0 : let _ok = fs::create_dir(pgdata);
1205 0 : } else {
1206 0 : fs::create_dir(pgdata)?;
1207 : }
1208 :
1209 0 : fs::set_permissions(pgdata, fs::Permissions::from_mode(0o700))?;
1210 :
1211 0 : Ok(())
1212 0 : }
1213 :
1214 : /// Fetches a basebackup from the Pageserver using the compute state's Pageserver connstring and
1215 : /// unarchives it to `pgdata` directory, replacing any existing contents.
1216 : #[instrument(skip_all, fields(%lsn))]
1217 : fn try_get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
1218 : let spec = compute_state.pspec.as_ref().expect("spec must be set");
1219 :
1220 : let started = Instant::now();
1221 : let (connected, size) = match spec.pageserver_conninfo.prefer_protocol {
1222 : PageserverProtocol::Grpc => self.try_get_basebackup_grpc(spec, lsn)?,
1223 : PageserverProtocol::Libpq => self.try_get_basebackup_libpq(spec, lsn)?,
1224 : };
1225 :
1226 : self.fix_zenith_signal_neon_signal()?;
1227 :
1228 : let mut state = self.state.lock().unwrap();
1229 : state.metrics.pageserver_connect_micros =
1230 : connected.duration_since(started).as_micros() as u64;
1231 : state.metrics.basebackup_bytes = size as u64;
1232 : state.metrics.basebackup_ms = started.elapsed().as_millis() as u64;
1233 :
1234 : Ok(())
1235 : }
1236 :
1237 : /// Move the Zenith signal file to Neon signal file location.
1238 : /// This makes Compute compatible with older PageServers that don't yet
1239 : /// know about the Zenith->Neon rename.
1240 0 : fn fix_zenith_signal_neon_signal(&self) -> Result<()> {
1241 0 : let datadir = Path::new(&self.params.pgdata);
1242 :
1243 0 : let neonsig = datadir.join("neon.signal");
1244 :
1245 0 : if neonsig.is_file() {
1246 0 : return Ok(());
1247 0 : }
1248 :
1249 0 : let zenithsig = datadir.join("zenith.signal");
1250 :
1251 0 : if zenithsig.is_file() {
1252 0 : fs::copy(zenithsig, neonsig)?;
1253 0 : }
1254 :
1255 0 : Ok(())
1256 0 : }
1257 :
1258 : /// Fetches a basebackup via gRPC. The connstring must use grpc://. Returns the timestamp when
1259 : /// the connection was established, and the (compressed) size of the basebackup.
1260 0 : fn try_get_basebackup_grpc(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
1261 0 : let shard0_index = ShardIndex {
1262 0 : shard_number: ShardNumber(0),
1263 0 : shard_count: spec.pageserver_conninfo.shard_count,
1264 0 : };
1265 0 : let shard0_url = spec
1266 0 : .pageserver_conninfo
1267 0 : .shard_url(ShardNumber(0), PageserverProtocol::Grpc)?
1268 0 : .to_owned();
1269 0 : let (reader, connected) = tokio::runtime::Handle::current().block_on(async move {
1270 0 : let mut client = page_api::Client::connect(
1271 0 : shard0_url,
1272 0 : spec.tenant_id,
1273 0 : spec.timeline_id,
1274 0 : shard0_index,
1275 0 : spec.storage_auth_token.clone(),
1276 0 : None, // NB: base backups use payload compression
1277 0 : )
1278 0 : .await?;
1279 0 : let connected = Instant::now();
1280 0 : let reader = client
1281 0 : .get_base_backup(page_api::GetBaseBackupRequest {
1282 0 : lsn: (lsn != Lsn(0)).then_some(lsn),
1283 0 : compression: BaseBackupCompression::Gzip,
1284 0 : replica: spec.spec.mode != ComputeMode::Primary,
1285 0 : full: false,
1286 0 : })
1287 0 : .await?;
1288 0 : anyhow::Ok((reader, connected))
1289 0 : })?;
1290 :
1291 0 : let mut reader = MeasuredReader::new(tokio_util::io::SyncIoBridge::new(reader));
1292 :
1293 : // Set `ignore_zeros` so that unpack() reads the entire stream and doesn't just stop at the
1294 : // end-of-archive marker. If the server errors, the tar::Builder drop handler will write an
1295 : // end-of-archive marker before the error is emitted, and we would not see the error.
1296 0 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut reader));
1297 0 : ar.set_ignore_zeros(true);
1298 0 : ar.unpack(&self.params.pgdata)?;
1299 :
1300 0 : Ok((connected, reader.get_byte_count()))
1301 0 : }
1302 :
1303 : /// Fetches a basebackup via libpq. The connstring must use postgresql://. Returns the timestamp
1304 : /// when the connection was established, and the (compressed) size of the basebackup.
1305 0 : fn try_get_basebackup_libpq(&self, spec: &ParsedSpec, lsn: Lsn) -> Result<(Instant, usize)> {
1306 0 : let shard0_connstr = spec
1307 0 : .pageserver_conninfo
1308 0 : .shard_url(ShardNumber(0), PageserverProtocol::Libpq)?;
1309 0 : let mut config = postgres::Config::from_str(shard0_connstr)?;
1310 :
1311 : // Use the storage auth token from the config file, if given.
1312 : // Note: this overrides any password set in the connection string.
1313 0 : if let Some(storage_auth_token) = &spec.storage_auth_token {
1314 0 : info!("Got storage auth token from spec file");
1315 0 : config.password(storage_auth_token);
1316 : } else {
1317 0 : info!("Storage auth token not set");
1318 : }
1319 :
1320 0 : config.application_name("compute_ctl");
1321 0 : config.options(&format!(
1322 0 : "-c neon.compute_mode={}",
1323 0 : spec.spec.mode.to_type_str()
1324 0 : ));
1325 :
1326 : // Connect to pageserver
1327 0 : let mut client = config.connect(NoTls)?;
1328 0 : let connected = Instant::now();
1329 :
1330 0 : let basebackup_cmd = match lsn {
1331 : Lsn(0) => {
1332 0 : if spec.spec.mode != ComputeMode::Primary {
1333 0 : format!(
1334 0 : "basebackup {} {} --gzip --replica",
1335 : spec.tenant_id, spec.timeline_id
1336 : )
1337 : } else {
1338 0 : format!("basebackup {} {} --gzip", spec.tenant_id, spec.timeline_id)
1339 : }
1340 : }
1341 : _ => {
1342 0 : if spec.spec.mode != ComputeMode::Primary {
1343 0 : format!(
1344 0 : "basebackup {} {} {} --gzip --replica",
1345 : spec.tenant_id, spec.timeline_id, lsn
1346 : )
1347 : } else {
1348 0 : format!(
1349 0 : "basebackup {} {} {} --gzip",
1350 : spec.tenant_id, spec.timeline_id, lsn
1351 : )
1352 : }
1353 : }
1354 : };
1355 :
1356 0 : let copyreader = client.copy_out(basebackup_cmd.as_str())?;
1357 0 : let mut measured_reader = MeasuredReader::new(copyreader);
1358 0 : let mut bufreader = std::io::BufReader::new(&mut measured_reader);
1359 :
1360 : // Read the archive directly from the `CopyOutReader`
1361 : //
1362 : // Set `ignore_zeros` so that unpack() reads all the Copy data and
1363 : // doesn't stop at the end-of-archive marker. Otherwise, if the server
1364 : // sends an Error after finishing the tarball, we will not notice it.
1365 : // The tar::Builder drop handler will write an end-of-archive marker
1366 : // before emitting the error, and we would not see it otherwise.
1367 0 : let mut ar = tar::Archive::new(flate2::read::GzDecoder::new(&mut bufreader));
1368 0 : ar.set_ignore_zeros(true);
1369 0 : ar.unpack(&self.params.pgdata)?;
1370 :
1371 0 : Ok((connected, measured_reader.get_byte_count()))
1372 0 : }
1373 :
1374 : // Gets the basebackup in a retry loop
1375 : #[instrument(skip_all, fields(%lsn))]
1376 : pub fn get_basebackup(&self, compute_state: &ComputeState, lsn: Lsn) -> Result<()> {
1377 : let mut retry_period_ms = 500.0;
1378 : let mut attempts = 0;
1379 : const DEFAULT_ATTEMPTS: u16 = 10;
1380 : #[cfg(feature = "testing")]
1381 : let max_attempts = if let Ok(v) = env::var("NEON_COMPUTE_TESTING_BASEBACKUP_RETRIES") {
1382 : u16::from_str(&v).unwrap()
1383 : } else {
1384 : DEFAULT_ATTEMPTS
1385 : };
1386 : #[cfg(not(feature = "testing"))]
1387 : let max_attempts = DEFAULT_ATTEMPTS;
1388 : loop {
1389 : let result = self.try_get_basebackup(compute_state, lsn);
1390 : match result {
1391 : Ok(_) => {
1392 : return result;
1393 : }
1394 : Err(ref e) if attempts < max_attempts => {
1395 : warn!("Failed to get basebackup: {e:?} (attempt {attempts}/{max_attempts})");
1396 : std::thread::sleep(std::time::Duration::from_millis(retry_period_ms as u64));
1397 : retry_period_ms *= 1.5;
1398 : }
1399 : Err(_) => {
1400 : return result;
1401 : }
1402 : }
1403 : attempts += 1;
1404 : }
1405 : }
1406 :
1407 0 : pub async fn check_safekeepers_synced_async(
1408 0 : &self,
1409 0 : compute_state: &ComputeState,
1410 0 : ) -> Result<Option<Lsn>> {
1411 : // Construct a connection config for each safekeeper
1412 0 : let pspec: ParsedSpec = compute_state
1413 0 : .pspec
1414 0 : .as_ref()
1415 0 : .expect("spec must be set")
1416 0 : .clone();
1417 0 : let sk_connstrs: Vec<String> = pspec.safekeeper_connstrings.clone();
1418 0 : let sk_configs = sk_connstrs.into_iter().map(|connstr| {
1419 : // Format connstr
1420 0 : let id = connstr.clone();
1421 0 : let connstr = format!("postgresql://no_user@{connstr}");
1422 0 : let options = format!(
1423 0 : "-c timeline_id={} tenant_id={}",
1424 : pspec.timeline_id, pspec.tenant_id
1425 : );
1426 :
1427 : // Construct client
1428 0 : let mut config = tokio_postgres::Config::from_str(&connstr).unwrap();
1429 0 : config.options(&options);
1430 0 : if let Some(storage_auth_token) = pspec.storage_auth_token.clone() {
1431 0 : config.password(storage_auth_token);
1432 0 : }
1433 :
1434 0 : (id, config)
1435 0 : });
1436 :
1437 : // Create task set to query all safekeepers
1438 0 : let mut tasks = FuturesUnordered::new();
1439 0 : let quorum = sk_configs.len() / 2 + 1;
1440 0 : for (id, config) in sk_configs {
1441 0 : let timeout = tokio::time::Duration::from_millis(100);
1442 0 : let task = tokio::time::timeout(timeout, ping_safekeeper(id, config));
1443 0 : tasks.push(tokio::spawn(task));
1444 0 : }
1445 :
1446 : // Get a quorum of responses or errors
1447 0 : let mut responses = Vec::new();
1448 0 : let mut join_errors = Vec::new();
1449 0 : let mut task_errors = Vec::new();
1450 0 : let mut timeout_errors = Vec::new();
1451 0 : while let Some(response) = tasks.next().await {
1452 0 : match response {
1453 0 : Ok(Ok(Ok(r))) => responses.push(r),
1454 0 : Ok(Ok(Err(e))) => task_errors.push(e),
1455 0 : Ok(Err(e)) => timeout_errors.push(e),
1456 0 : Err(e) => join_errors.push(e),
1457 : };
1458 0 : if responses.len() >= quorum {
1459 0 : break;
1460 0 : }
1461 0 : if join_errors.len() + task_errors.len() + timeout_errors.len() >= quorum {
1462 0 : break;
1463 0 : }
1464 : }
1465 :
1466 : // In case of error, log and fail the check, but don't crash.
1467 : // We're playing it safe because these errors could be transient
1468 : // and we don't yet retry.
1469 0 : if responses.len() < quorum {
1470 0 : error!(
1471 0 : "failed sync safekeepers check {:?} {:?} {:?}",
1472 : join_errors, task_errors, timeout_errors
1473 : );
1474 0 : return Ok(None);
1475 0 : }
1476 :
1477 0 : Ok(check_if_synced(responses))
1478 0 : }
1479 :
1480 : // Fast path for sync_safekeepers. If they're already synced we get the lsn
1481 : // in one roundtrip. If not, we should do a full sync_safekeepers.
1482 : #[instrument(skip_all)]
1483 : pub fn check_safekeepers_synced(&self, compute_state: &ComputeState) -> Result<Option<Lsn>> {
1484 : let start_time = Utc::now();
1485 :
1486 : let rt = tokio::runtime::Handle::current();
1487 : let result = rt.block_on(self.check_safekeepers_synced_async(compute_state));
1488 :
1489 : // Record runtime
1490 : self.state.lock().unwrap().metrics.sync_sk_check_ms = Utc::now()
1491 : .signed_duration_since(start_time)
1492 : .to_std()
1493 : .unwrap()
1494 : .as_millis() as u64;
1495 : result
1496 : }
1497 :
1498 : // Run `postgres` in a special mode with `--sync-safekeepers` argument
1499 : // and return the reported LSN back to the caller.
1500 : #[instrument(skip_all)]
1501 : pub fn sync_safekeepers(&self, storage_auth_token: Option<String>) -> Result<Lsn> {
1502 : let start_time = Utc::now();
1503 :
1504 : let mut sync_handle = maybe_cgexec(&self.params.pgbin)
1505 : .args(["--sync-safekeepers"])
1506 : .env("PGDATA", &self.params.pgdata) // we cannot use -D in this mode
1507 : .envs(if let Some(storage_auth_token) = &storage_auth_token {
1508 : vec![("NEON_AUTH_TOKEN", storage_auth_token)]
1509 : } else {
1510 : vec![]
1511 : })
1512 : .stdout(Stdio::piped())
1513 : .stderr(Stdio::piped())
1514 : .spawn()
1515 : .expect("postgres --sync-safekeepers failed to start");
1516 : SYNC_SAFEKEEPERS_PID.store(sync_handle.id(), Ordering::SeqCst);
1517 :
1518 : // `postgres --sync-safekeepers` will print all log output to stderr and
1519 : // final LSN to stdout. So we leave stdout to collect LSN, while stderr logs
1520 : // will be collected in a child thread.
1521 : let stderr = sync_handle
1522 : .stderr
1523 : .take()
1524 : .expect("stderr should be captured");
1525 : let logs_handle = handle_postgres_logs(stderr);
1526 :
1527 : let sync_output = sync_handle
1528 : .wait_with_output()
1529 : .expect("postgres --sync-safekeepers failed");
1530 : SYNC_SAFEKEEPERS_PID.store(0, Ordering::SeqCst);
1531 :
1532 : // Process has exited, so we can join the logs thread.
1533 : let _ = tokio::runtime::Handle::current()
1534 : .block_on(logs_handle)
1535 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
1536 :
1537 : if !sync_output.status.success() {
1538 : anyhow::bail!(
1539 : "postgres --sync-safekeepers exited with non-zero status: {}. stdout: {}",
1540 : sync_output.status,
1541 : String::from_utf8(sync_output.stdout)
1542 : .expect("postgres --sync-safekeepers exited, and stdout is not utf-8"),
1543 : );
1544 : }
1545 :
1546 : self.state.lock().unwrap().metrics.sync_safekeepers_ms = Utc::now()
1547 : .signed_duration_since(start_time)
1548 : .to_std()
1549 : .unwrap()
1550 : .as_millis() as u64;
1551 :
1552 : let lsn = Lsn::from_str(String::from_utf8(sync_output.stdout)?.trim())?;
1553 :
1554 : Ok(lsn)
1555 : }
1556 :
1557 : /// Do all the preparations like PGDATA directory creation, configuration,
1558 : /// safekeepers sync, basebackup, etc.
1559 : #[instrument(skip_all)]
1560 : pub fn prepare_pgdata(&self, compute_state: &ComputeState) -> Result<()> {
1561 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
1562 : let spec = &pspec.spec;
1563 : let pgdata_path = Path::new(&self.params.pgdata);
1564 :
1565 : let tls_config = self.tls_config(&pspec.spec);
1566 : let databricks_settings = spec.databricks_settings.as_ref();
1567 : let postgres_port = self.params.connstr.port();
1568 :
1569 : // Remove/create an empty pgdata directory and put configuration there.
1570 : self.create_pgdata()?;
1571 : config::write_postgres_conf(
1572 : pgdata_path,
1573 : &self.params,
1574 : &pspec.spec,
1575 : postgres_port,
1576 : self.params.internal_http_port,
1577 : tls_config,
1578 : databricks_settings,
1579 : self.params.lakebase_mode,
1580 : )?;
1581 :
1582 : // Syncing safekeepers is only safe with primary nodes: if a primary
1583 : // is already connected it will be kicked out, so a secondary (standby)
1584 : // cannot sync safekeepers.
1585 : let lsn = match spec.mode {
1586 : ComputeMode::Primary => {
1587 : info!("checking if safekeepers are synced");
1588 : let lsn = if let Ok(Some(lsn)) = self.check_safekeepers_synced(compute_state) {
1589 : lsn
1590 : } else {
1591 : info!("starting safekeepers syncing");
1592 : self.sync_safekeepers(pspec.storage_auth_token.clone())
1593 : .with_context(|| "failed to sync safekeepers")?
1594 : };
1595 : info!("safekeepers synced at LSN {}", lsn);
1596 : lsn
1597 : }
1598 : ComputeMode::Static(lsn) => {
1599 : info!("Starting read-only node at static LSN {}", lsn);
1600 : lsn
1601 : }
1602 : ComputeMode::Replica => {
1603 : info!("Initializing standby from latest Pageserver LSN");
1604 : Lsn(0)
1605 : }
1606 : };
1607 :
1608 : self.get_basebackup(compute_state, lsn)
1609 0 : .with_context(|| format!("failed to get basebackup@{lsn}"))?;
1610 :
1611 : if let Some(settings) = databricks_settings {
1612 : copy_tls_certificates(
1613 : &settings.pg_compute_tls_settings.key_file,
1614 : &settings.pg_compute_tls_settings.cert_file,
1615 : pgdata_path,
1616 : )?;
1617 :
1618 : // Update pg_hba.conf received with basebackup including additional databricks settings.
1619 : update_pg_hba(pgdata_path, Some(&settings.databricks_pg_hba))?;
1620 : update_pg_ident(pgdata_path, Some(&settings.databricks_pg_ident))?;
1621 : } else {
1622 : // Update pg_hba.conf received with basebackup.
1623 : update_pg_hba(pgdata_path, None)?;
1624 : }
1625 :
1626 : if let Some(databricks_settings) = spec.databricks_settings.as_ref() {
1627 : copy_tls_certificates(
1628 : &databricks_settings.pg_compute_tls_settings.key_file,
1629 : &databricks_settings.pg_compute_tls_settings.cert_file,
1630 : pgdata_path,
1631 : )?;
1632 : }
1633 :
1634 : // Place pg_dynshmem under /dev/shm. This allows us to use
1635 : // 'dynamic_shared_memory_type = mmap' so that the files are placed in
1636 : // /dev/shm, similar to how 'dynamic_shared_memory_type = posix' works.
1637 : //
1638 : // Why on earth don't we just stick to the 'posix' default, you might
1639 : // ask. It turns out that making large allocations with 'posix' doesn't
1640 : // work very well with autoscaling. The behavior we want is that:
1641 : //
1642 : // 1. You can make large DSM allocations, larger than the current RAM
1643 : // size of the VM, without errors
1644 : //
1645 : // 2. If the allocated memory is really used, the VM is scaled up
1646 : // automatically to accommodate that
1647 : //
1648 : // We try to make that possible by having swap in the VM. But with the
1649 : // default 'posix' DSM implementation, we fail step 1, even when there's
1650 : // plenty of swap available. PostgreSQL uses posix_fallocate() to create
1651 : // the shmem segment, which is really just a file in /dev/shm in Linux,
1652 : // but posix_fallocate() on tmpfs returns ENOMEM if the size is larger
1653 : // than available RAM.
1654 : //
1655 : // Using 'dynamic_shared_memory_type = mmap' works around that, because
1656 : // the Postgres 'mmap' DSM implementation doesn't use
1657 : // posix_fallocate(). Instead, it uses repeated calls to write(2) to
1658 : // fill the file with zeros. It's weird that that differs between
1659 : // 'posix' and 'mmap', but we take advantage of it. When the file is
1660 : // filled slowly with write(2), the kernel allows it to grow larger, as
1661 : // long as there's swap available.
1662 : //
1663 : // In short, using 'dynamic_shared_memory_type = mmap' allows us one DSM
1664 : // segment to be larger than currently available RAM. But because we
1665 : // don't want to store it on a real file, which the kernel would try to
1666 : // flush to disk, so symlink pg_dynshm to /dev/shm.
1667 : //
1668 : // We don't set 'dynamic_shared_memory_type = mmap' here, we let the
1669 : // control plane control that option. If 'mmap' is not used, this
1670 : // symlink doesn't affect anything.
1671 : //
1672 : // See https://github.com/neondatabase/autoscaling/issues/800
1673 : std::fs::remove_dir_all(pgdata_path.join("pg_dynshmem"))?;
1674 : symlink("/dev/shm/", pgdata_path.join("pg_dynshmem"))?;
1675 :
1676 : match spec.mode {
1677 : ComputeMode::Primary => {}
1678 : ComputeMode::Replica | ComputeMode::Static(..) => {
1679 : add_standby_signal(pgdata_path)?;
1680 : }
1681 : }
1682 :
1683 : Ok(())
1684 : }
1685 :
1686 : /// Start and stop a postgres process to warm up the VM for startup.
1687 0 : pub fn prewarm_postgres_vm_memory(&self) -> Result<()> {
1688 0 : if self.params.lakebase_mode {
1689 : // We are running in Hadron mode. Disabling this prewarming step for now as it could run
1690 : // into dblet port conflicts and also doesn't add much value with our current infra.
1691 0 : info!("Skipping postgres prewarming in Hadron mode");
1692 0 : return Ok(());
1693 0 : }
1694 0 : info!("prewarming VM memory");
1695 :
1696 : // Create pgdata
1697 0 : let pgdata = &format!("{}.warmup", self.params.pgdata);
1698 0 : create_pgdata(pgdata)?;
1699 :
1700 : // Run initdb to completion
1701 0 : info!("running initdb");
1702 0 : let initdb_bin = Path::new(&self.params.pgbin)
1703 0 : .parent()
1704 0 : .unwrap()
1705 0 : .join("initdb");
1706 0 : Command::new(initdb_bin)
1707 0 : .args(["--pgdata", pgdata])
1708 0 : .output()
1709 0 : .expect("cannot start initdb process");
1710 :
1711 : // Write conf
1712 : use std::io::Write;
1713 0 : let conf_path = Path::new(pgdata).join("postgresql.conf");
1714 0 : let mut file = std::fs::File::create(conf_path)?;
1715 0 : writeln!(file, "shared_buffers=65536")?;
1716 0 : writeln!(file, "port=51055")?; // Nobody should be connecting
1717 0 : writeln!(file, "shared_preload_libraries = 'neon'")?;
1718 :
1719 : // Start postgres
1720 0 : info!("starting postgres");
1721 0 : let mut pg = maybe_cgexec(&self.params.pgbin)
1722 0 : .args(["-D", pgdata])
1723 0 : .spawn()
1724 0 : .expect("cannot start postgres process");
1725 :
1726 : // Stop it when it's ready
1727 0 : info!("waiting for postgres");
1728 0 : wait_for_postgres(&mut pg, Path::new(pgdata))?;
1729 : // SIGQUIT orders postgres to exit immediately. We don't want to SIGKILL
1730 : // it to avoid orphaned processes prowling around while datadir is
1731 : // wiped.
1732 0 : let pm_pid = Pid::from_raw(pg.id() as i32);
1733 0 : kill(pm_pid, Signal::SIGQUIT)?;
1734 0 : info!("sent SIGQUIT signal");
1735 0 : pg.wait()?;
1736 0 : info!("done prewarming vm memory");
1737 :
1738 : // clean up
1739 0 : let _ok = fs::remove_dir_all(pgdata);
1740 0 : Ok(())
1741 0 : }
1742 :
1743 : /// Start Postgres as a child process and wait for it to start accepting
1744 : /// connections.
1745 : ///
1746 : /// Returns a handle to the child process and a handle to the logs thread.
1747 : #[instrument(skip_all)]
1748 : pub fn start_postgres(&self, storage_auth_token: Option<String>) -> Result<PostgresHandle> {
1749 : let pgdata_path = Path::new(&self.params.pgdata);
1750 :
1751 : let env_vars: Vec<(String, String)> = if self.params.lakebase_mode {
1752 : let databricks_env_vars = {
1753 : let state = self.state.lock().unwrap();
1754 : let spec = &state.pspec.as_ref().unwrap().spec;
1755 : DatabricksEnvVars::new(
1756 : spec,
1757 : Some(&self.params.compute_id),
1758 : self.params.instance_id.clone(),
1759 : self.params.lakebase_mode,
1760 : )
1761 : };
1762 :
1763 : info!(
1764 : "Starting Postgres for databricks endpoint id: {}",
1765 : &databricks_env_vars.endpoint_id
1766 : );
1767 :
1768 : let mut env_vars = databricks_env_vars.to_env_var_list();
1769 0 : env_vars.extend(storage_auth_token.map(|t| ("NEON_AUTH_TOKEN".to_string(), t)));
1770 : env_vars
1771 : } else if let Some(storage_auth_token) = &storage_auth_token {
1772 : vec![("NEON_AUTH_TOKEN".to_owned(), storage_auth_token.to_owned())]
1773 : } else {
1774 : vec![]
1775 : };
1776 :
1777 : // Run postgres as a child process.
1778 : let mut pg = maybe_cgexec(&self.params.pgbin)
1779 : .args(["-D", &self.params.pgdata])
1780 : .envs(env_vars)
1781 : .stderr(Stdio::piped())
1782 : .spawn()
1783 : .expect("cannot start postgres process");
1784 : PG_PID.store(pg.id(), Ordering::SeqCst);
1785 :
1786 : // Start a task to collect logs from stderr.
1787 : let stderr = pg.stderr.take().expect("stderr should be captured");
1788 : let logs_handle = handle_postgres_logs(stderr);
1789 :
1790 : wait_for_postgres(&mut pg, pgdata_path)?;
1791 :
1792 : Ok(PostgresHandle {
1793 : postgres: pg,
1794 : log_collector: logs_handle,
1795 : })
1796 : }
1797 :
1798 : /// Wait for the child Postgres process forever. In this state Ctrl+C will
1799 : /// propagate to Postgres and it will be shut down as well.
1800 0 : fn wait_postgres(&self, mut pg_handle: PostgresHandle) -> std::process::ExitStatus {
1801 0 : info!(postmaster_pid = %pg_handle.postgres.id(), "Waiting for Postgres to exit");
1802 :
1803 0 : let ecode = pg_handle
1804 0 : .postgres
1805 0 : .wait()
1806 0 : .expect("failed to start waiting on Postgres process");
1807 0 : PG_PID.store(0, Ordering::SeqCst);
1808 :
1809 : // Process has exited. Wait for the log collecting task to finish.
1810 0 : let _ = tokio::runtime::Handle::current()
1811 0 : .block_on(pg_handle.log_collector)
1812 0 : .map_err(|e| tracing::error!("log task panicked: {:?}", e));
1813 :
1814 0 : ecode
1815 0 : }
1816 :
1817 : /// Do post configuration of the already started Postgres. This function spawns a background task to
1818 : /// configure the database after applying the compute spec. Currently, it upgrades the neon extension
1819 : /// version. In the future, it may upgrade all 3rd-party extensions.
1820 : #[instrument(skip_all)]
1821 : pub fn post_apply_config(&self) -> Result<()> {
1822 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:post_apply_config"));
1823 0 : tokio::spawn(async move {
1824 0 : let res = async {
1825 0 : let (mut client, connection) = conf.connect(NoTls).await?;
1826 0 : tokio::spawn(async move {
1827 0 : if let Err(e) = connection.await {
1828 0 : eprintln!("connection error: {e}");
1829 0 : }
1830 0 : });
1831 :
1832 0 : handle_neon_extension_upgrade(&mut client)
1833 0 : .await
1834 0 : .context("handle_neon_extension_upgrade")?;
1835 0 : Ok::<_, anyhow::Error>(())
1836 0 : }
1837 0 : .await;
1838 0 : if let Err(err) = res {
1839 0 : error!("error while post_apply_config: {err:#}");
1840 0 : }
1841 0 : });
1842 : Ok(())
1843 : }
1844 :
1845 0 : pub fn get_conn_conf(&self, application_name: Option<&str>) -> postgres::Config {
1846 0 : let mut conf = self.conn_conf.clone();
1847 0 : if let Some(application_name) = application_name {
1848 0 : conf.application_name(application_name);
1849 0 : }
1850 0 : conf
1851 0 : }
1852 :
1853 0 : pub fn get_tokio_conn_conf(&self, application_name: Option<&str>) -> tokio_postgres::Config {
1854 0 : let mut conf = self.tokio_conn_conf.clone();
1855 0 : if let Some(application_name) = application_name {
1856 0 : conf.application_name(application_name);
1857 0 : }
1858 0 : conf
1859 0 : }
1860 :
1861 0 : pub async fn get_maintenance_client(
1862 0 : conf: &tokio_postgres::Config,
1863 0 : ) -> Result<tokio_postgres::Client> {
1864 0 : let mut conf = conf.clone();
1865 0 : conf.application_name("compute_ctl:apply_config");
1866 :
1867 0 : let (client, conn) = match conf.connect(NoTls).await {
1868 : // If connection fails, it may be the old node with `zenith_admin` superuser.
1869 : //
1870 : // In this case we need to connect with old `zenith_admin` name
1871 : // and create new user. We cannot simply rename connected user,
1872 : // but we can create a new one and grant it all privileges.
1873 0 : Err(e) => match e.code() {
1874 : Some(&SqlState::INVALID_PASSWORD)
1875 : | Some(&SqlState::INVALID_AUTHORIZATION_SPECIFICATION) => {
1876 : // Connect with `zenith_admin` if `cloud_admin` could not authenticate
1877 0 : info!(
1878 0 : "cannot connect to Postgres: {}, retrying with 'zenith_admin' username",
1879 : e
1880 : );
1881 0 : let mut zenith_admin_conf = postgres::config::Config::from(conf.clone());
1882 0 : zenith_admin_conf.application_name("compute_ctl:apply_config");
1883 0 : zenith_admin_conf.user("zenith_admin");
1884 :
1885 : // It doesn't matter what were the options before, here we just want
1886 : // to connect and create a new superuser role.
1887 : const ZENITH_OPTIONS: &str = "-c role=zenith_admin -c default_transaction_read_only=off -c search_path=public -c statement_timeout=0";
1888 0 : zenith_admin_conf.options(ZENITH_OPTIONS);
1889 :
1890 0 : let mut client =
1891 0 : zenith_admin_conf.connect(NoTls)
1892 0 : .context("broken cloud_admin credential: tried connecting with cloud_admin but could not authenticate, and zenith_admin does not work either")?;
1893 :
1894 : // Disable forwarding so that users don't get a cloud_admin role
1895 0 : let mut func = || {
1896 0 : client.simple_query("SET neon.forward_ddl = false")?;
1897 0 : client.simple_query("CREATE USER cloud_admin WITH SUPERUSER")?;
1898 0 : client.simple_query("GRANT zenith_admin TO cloud_admin")?;
1899 0 : Ok::<_, anyhow::Error>(())
1900 0 : };
1901 0 : func().context("apply_config setup cloud_admin")?;
1902 :
1903 0 : drop(client);
1904 :
1905 : // Reconnect with connstring with expected name
1906 0 : conf.connect(NoTls).await?
1907 : }
1908 0 : _ => return Err(e.into()),
1909 : },
1910 0 : Ok((client, conn)) => (client, conn),
1911 : };
1912 :
1913 0 : spawn(async move {
1914 0 : if let Err(e) = conn.await {
1915 0 : error!("maintenance client connection error: {}", e);
1916 0 : }
1917 0 : });
1918 :
1919 : // Disable DDL forwarding because control plane already knows about the roles/databases
1920 : // we're about to modify.
1921 0 : client
1922 0 : .simple_query("SET neon.forward_ddl = false")
1923 0 : .await
1924 0 : .context("apply_config SET neon.forward_ddl = false")?;
1925 :
1926 0 : Ok(client)
1927 0 : }
1928 :
1929 : /// Do initial configuration of the already started Postgres.
1930 : #[instrument(skip_all)]
1931 : pub fn apply_config(&self, compute_state: &ComputeState) -> Result<()> {
1932 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:apply_config"));
1933 :
1934 : if self.params.lakebase_mode {
1935 : // Set a 2-minute statement_timeout for the session applying config. The individual SQL statements
1936 : // used in apply_spec_sql() should not take long (they are just creating users and installing
1937 : // extensions). If any of them are stuck for an extended period of time it usually indicates a
1938 : // pageserver connectivity problem and we should bail out.
1939 : conf.options("-c statement_timeout=2min");
1940 : }
1941 :
1942 : let conf = Arc::new(conf);
1943 : let spec = Arc::new(
1944 : compute_state
1945 : .pspec
1946 : .as_ref()
1947 : .expect("spec must be set")
1948 : .spec
1949 : .clone(),
1950 : );
1951 :
1952 : let mut tls_config = None::<TlsConfig>;
1953 : if spec.features.contains(&ComputeFeature::TlsExperimental) {
1954 : tls_config = self.compute_ctl_config.tls.clone();
1955 : }
1956 :
1957 : self.update_installed_extensions_collection_interval(&spec);
1958 :
1959 : let max_concurrent_connections = self.max_service_connections(compute_state, &spec);
1960 :
1961 : // Merge-apply spec & changes to PostgreSQL state.
1962 : self.apply_spec_sql(spec.clone(), conf.clone(), max_concurrent_connections)?;
1963 :
1964 : if let Some(local_proxy) = &spec.clone().local_proxy_config {
1965 : let mut local_proxy = local_proxy.clone();
1966 : local_proxy.tls = tls_config.clone();
1967 :
1968 : info!("configuring local_proxy");
1969 : local_proxy::configure(&local_proxy).context("apply_config local_proxy")?;
1970 : }
1971 :
1972 : // Run migrations separately to not hold up cold starts
1973 : let lakebase_mode = self.params.lakebase_mode;
1974 : let params = self.params.clone();
1975 0 : tokio::spawn(async move {
1976 0 : let mut conf = conf.as_ref().clone();
1977 0 : conf.application_name("compute_ctl:migrations");
1978 :
1979 0 : match conf.connect(NoTls).await {
1980 0 : Ok((mut client, connection)) => {
1981 0 : tokio::spawn(async move {
1982 0 : if let Err(e) = connection.await {
1983 0 : eprintln!("connection error: {e}");
1984 0 : }
1985 0 : });
1986 0 : if let Err(e) = handle_migrations(params, &mut client, lakebase_mode).await {
1987 0 : error!("Failed to run migrations: {}", e);
1988 0 : }
1989 : }
1990 0 : Err(e) => {
1991 0 : error!(
1992 0 : "Failed to connect to the compute for running migrations: {}",
1993 : e
1994 : );
1995 : }
1996 : };
1997 0 : });
1998 :
1999 : Ok::<(), anyhow::Error>(())
2000 : }
2001 :
2002 : // Signal to the configurator to refresh the configuration by pulling a new spec from the HCC.
2003 : // Note that this merely triggers a notification on a condition variable the configurator thread
2004 : // waits on. The configurator thread (in configurator.rs) pulls the new spec from the HCC and
2005 : // applies it.
2006 0 : pub async fn signal_refresh_configuration(&self) -> Result<()> {
2007 0 : let states_allowing_configuration_refresh = [
2008 0 : ComputeStatus::Running,
2009 0 : ComputeStatus::Failed,
2010 0 : ComputeStatus::RefreshConfigurationPending,
2011 0 : ];
2012 :
2013 0 : let mut state = self.state.lock().expect("state lock poisoned");
2014 0 : if states_allowing_configuration_refresh.contains(&state.status) {
2015 0 : state.status = ComputeStatus::RefreshConfigurationPending;
2016 0 : self.state_changed.notify_all();
2017 0 : Ok(())
2018 0 : } else if state.status == ComputeStatus::Init {
2019 : // If the compute is in Init state, we can't refresh the configuration immediately,
2020 : // but we should be able to do that soon.
2021 0 : Ok(())
2022 : } else {
2023 0 : Err(anyhow::anyhow!(
2024 0 : "Cannot refresh compute configuration in state {:?}",
2025 0 : state.status
2026 0 : ))
2027 : }
2028 0 : }
2029 :
2030 : // Wrapped this around `pg_ctl reload`, but right now we don't use
2031 : // `pg_ctl` for start / stop.
2032 : #[instrument(skip_all)]
2033 : fn pg_reload_conf(&self) -> Result<()> {
2034 : let pgctl_bin = Path::new(&self.params.pgbin)
2035 : .parent()
2036 : .unwrap()
2037 : .join("pg_ctl");
2038 : Command::new(pgctl_bin)
2039 : .args(["reload", "-D", &self.params.pgdata])
2040 : .output()
2041 : .expect("cannot run pg_ctl process");
2042 : Ok(())
2043 : }
2044 :
2045 : /// Similar to `apply_config()`, but does a bit different sequence of operations,
2046 : /// as it's used to reconfigure a previously started and configured Postgres node.
2047 : #[instrument(skip_all)]
2048 : pub fn reconfigure(&self) -> Result<()> {
2049 : let spec = self.state.lock().unwrap().pspec.clone().unwrap().spec;
2050 :
2051 : let tls_config = self.tls_config(&spec);
2052 :
2053 : self.update_installed_extensions_collection_interval(&spec);
2054 :
2055 : if let Some(ref pgbouncer_settings) = spec.pgbouncer_settings {
2056 : info!("tuning pgbouncer");
2057 :
2058 : let pgbouncer_settings = pgbouncer_settings.clone();
2059 : let tls_config = tls_config.clone();
2060 :
2061 : // Spawn a background task to do the tuning,
2062 : // so that we don't block the main thread that starts Postgres.
2063 0 : tokio::spawn(async move {
2064 0 : let res = tune_pgbouncer(pgbouncer_settings, tls_config).await;
2065 0 : if let Err(err) = res {
2066 0 : error!("error while tuning pgbouncer: {err:?}");
2067 0 : }
2068 0 : });
2069 : }
2070 :
2071 : if let Some(ref local_proxy) = spec.local_proxy_config {
2072 : info!("configuring local_proxy");
2073 :
2074 : // Spawn a background task to do the configuration,
2075 : // so that we don't block the main thread that starts Postgres.
2076 : let mut local_proxy = local_proxy.clone();
2077 : local_proxy.tls = tls_config.clone();
2078 0 : tokio::spawn(async move {
2079 0 : if let Err(err) = local_proxy::configure(&local_proxy) {
2080 0 : error!("error while configuring local_proxy: {err:?}");
2081 0 : }
2082 0 : });
2083 : }
2084 :
2085 : // Reconfigure rsyslog for Postgres logs export
2086 : let conf = PostgresLogsRsyslogConfig::new(spec.logs_export_host.as_deref());
2087 : configure_postgres_logs_export(conf)?;
2088 :
2089 : // Write new config
2090 : let pgdata_path = Path::new(&self.params.pgdata);
2091 : let postgres_port = self.params.connstr.port();
2092 : config::write_postgres_conf(
2093 : pgdata_path,
2094 : &self.params,
2095 : &spec,
2096 : postgres_port,
2097 : self.params.internal_http_port,
2098 : tls_config,
2099 : spec.databricks_settings.as_ref(),
2100 : self.params.lakebase_mode,
2101 : )?;
2102 :
2103 : self.pg_reload_conf()?;
2104 :
2105 : if !spec.skip_pg_catalog_updates {
2106 : let max_concurrent_connections = spec.reconfigure_concurrency;
2107 : // Temporarily reset max_cluster_size in config
2108 : // to avoid the possibility of hitting the limit, while we are reconfiguring:
2109 : // creating new extensions, roles, etc.
2110 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
2111 0 : self.pg_reload_conf()?;
2112 :
2113 0 : if spec.mode == ComputeMode::Primary {
2114 0 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:reconfigure"));
2115 0 : let conf = Arc::new(conf);
2116 :
2117 0 : let spec = Arc::new(spec.clone());
2118 :
2119 0 : self.apply_spec_sql(spec, conf, max_concurrent_connections)?;
2120 0 : }
2121 :
2122 0 : Ok(())
2123 0 : })?;
2124 : self.pg_reload_conf()?;
2125 : }
2126 :
2127 : let unknown_op = "unknown".to_string();
2128 : let op_id = spec.operation_uuid.as_ref().unwrap_or(&unknown_op);
2129 : info!(
2130 : "finished reconfiguration of compute node for operation {}",
2131 : op_id
2132 : );
2133 :
2134 : Ok(())
2135 : }
2136 :
2137 : #[instrument(skip_all)]
2138 : pub fn configure_as_primary(&self, compute_state: &ComputeState) -> Result<()> {
2139 : let pspec = compute_state.pspec.as_ref().expect("spec must be set");
2140 :
2141 : assert!(pspec.spec.mode == ComputeMode::Primary);
2142 : if !pspec.spec.skip_pg_catalog_updates {
2143 : let pgdata_path = Path::new(&self.params.pgdata);
2144 : // temporarily reset max_cluster_size in config
2145 : // to avoid the possibility of hitting the limit, while we are applying config:
2146 : // creating new extensions, roles, etc...
2147 0 : config::with_compute_ctl_tmp_override(pgdata_path, "neon.max_cluster_size=-1", || {
2148 0 : self.pg_reload_conf()?;
2149 :
2150 0 : self.apply_config(compute_state)?;
2151 :
2152 0 : Ok(())
2153 0 : })?;
2154 :
2155 : let postgresql_conf_path = pgdata_path.join("postgresql.conf");
2156 : if config::line_in_file(
2157 : &postgresql_conf_path,
2158 : "neon.disable_logical_replication_subscribers=false",
2159 : )? {
2160 : info!(
2161 : "updated postgresql.conf to set neon.disable_logical_replication_subscribers=false"
2162 : );
2163 : }
2164 : self.pg_reload_conf()?;
2165 : }
2166 : self.post_apply_config()?;
2167 :
2168 : Ok(())
2169 : }
2170 :
2171 0 : pub async fn watch_cert_for_changes(self: Arc<Self>) {
2172 : // update status on cert renewal
2173 0 : if let Some(tls_config) = &self.compute_ctl_config.tls {
2174 0 : let tls_config = tls_config.clone();
2175 :
2176 : // wait until the cert exists.
2177 0 : let mut cert_watch = watch_cert_for_changes(tls_config.cert_path.clone()).await;
2178 :
2179 0 : tokio::task::spawn_blocking(move || {
2180 0 : let handle = tokio::runtime::Handle::current();
2181 : 'cert_update: loop {
2182 : // let postgres/pgbouncer/local_proxy know the new cert/key exists.
2183 : // we need to wait until it's configurable first.
2184 :
2185 0 : let mut state = self.state.lock().unwrap();
2186 : 'status_update: loop {
2187 0 : match state.status {
2188 : // let's update the state to config pending
2189 : ComputeStatus::ConfigurationPending | ComputeStatus::Running => {
2190 0 : state.set_status(
2191 0 : ComputeStatus::ConfigurationPending,
2192 0 : &self.state_changed,
2193 0 : );
2194 0 : break 'status_update;
2195 : }
2196 :
2197 : // exit loop
2198 : ComputeStatus::Failed
2199 : | ComputeStatus::TerminationPendingFast
2200 : | ComputeStatus::TerminationPendingImmediate
2201 0 : | ComputeStatus::Terminated => break 'cert_update,
2202 :
2203 : // wait
2204 : ComputeStatus::Init
2205 : | ComputeStatus::Configuration
2206 : | ComputeStatus::RefreshConfiguration
2207 : | ComputeStatus::RefreshConfigurationPending
2208 0 : | ComputeStatus::Empty => {
2209 0 : state = self.state_changed.wait(state).unwrap();
2210 0 : }
2211 : }
2212 : }
2213 0 : drop(state);
2214 :
2215 : // wait for a new certificate update
2216 0 : if handle.block_on(cert_watch.changed()).is_err() {
2217 0 : break;
2218 0 : }
2219 : }
2220 0 : });
2221 0 : }
2222 0 : }
2223 :
2224 0 : pub fn tls_config(&self, spec: &ComputeSpec) -> &Option<TlsConfig> {
2225 0 : if spec.features.contains(&ComputeFeature::TlsExperimental) {
2226 0 : &self.compute_ctl_config.tls
2227 : } else {
2228 0 : &None::<TlsConfig>
2229 : }
2230 0 : }
2231 :
2232 : /// Update the `last_active` in the shared state, but ensure that it's a more recent one.
2233 0 : pub fn update_last_active(&self, last_active: Option<DateTime<Utc>>) {
2234 0 : let mut state = self.state.lock().unwrap();
2235 : // NB: `Some(<DateTime>)` is always greater than `None`.
2236 0 : if last_active > state.last_active {
2237 0 : state.last_active = last_active;
2238 0 : debug!("set the last compute activity time to: {:?}", last_active);
2239 0 : }
2240 0 : }
2241 :
2242 : // Look for core dumps and collect backtraces.
2243 : //
2244 : // EKS worker nodes have following core dump settings:
2245 : // /proc/sys/kernel/core_pattern -> core
2246 : // /proc/sys/kernel/core_uses_pid -> 1
2247 : // ulimit -c -> unlimited
2248 : // which results in core dumps being written to postgres data directory as core.<pid>.
2249 : //
2250 : // Use that as a default location and pattern, except macos where core dumps are written
2251 : // to /cores/ directory by default.
2252 : //
2253 : // With default Linux settings, the core dump file is called just "core", so check for
2254 : // that too.
2255 0 : pub fn check_for_core_dumps(&self) -> Result<()> {
2256 0 : let core_dump_dir = match std::env::consts::OS {
2257 0 : "macos" => Path::new("/cores/"),
2258 : // BEGIN HADRON
2259 : // NB: Read core dump files from a fixed location outside of
2260 : // the data directory since `compute_ctl` wipes the data directory
2261 : // across container restarts.
2262 : _ => {
2263 0 : if self.params.lakebase_mode {
2264 0 : Path::new("/databricks/logs/brickstore")
2265 : } else {
2266 0 : Path::new(&self.params.pgdata)
2267 : }
2268 : } // END HADRON
2269 : };
2270 :
2271 : // Collect core dump paths if any
2272 0 : info!("checking for core dumps in {}", core_dump_dir.display());
2273 0 : let files = fs::read_dir(core_dump_dir)?;
2274 0 : let cores = files.filter_map(|entry| {
2275 0 : let entry = entry.ok()?;
2276 :
2277 0 : let is_core_dump = match entry.file_name().to_str()? {
2278 0 : n if n.starts_with("core.") => true,
2279 0 : "core" => true,
2280 0 : _ => false,
2281 : };
2282 0 : if is_core_dump {
2283 0 : Some(entry.path())
2284 : } else {
2285 0 : None
2286 : }
2287 0 : });
2288 :
2289 : // Print backtrace for each core dump
2290 0 : for core_path in cores {
2291 0 : warn!(
2292 0 : "core dump found: {}, collecting backtrace",
2293 0 : core_path.display()
2294 : );
2295 :
2296 : // Try first with gdb
2297 0 : let backtrace = Command::new("gdb")
2298 0 : .args(["--batch", "-q", "-ex", "bt", &self.params.pgbin])
2299 0 : .arg(&core_path)
2300 0 : .output();
2301 :
2302 : // Try lldb if no gdb is found -- that is handy for local testing on macOS
2303 0 : let backtrace = match backtrace {
2304 0 : Err(ref e) if e.kind() == std::io::ErrorKind::NotFound => {
2305 0 : warn!("cannot find gdb, trying lldb");
2306 0 : Command::new("lldb")
2307 0 : .arg("-c")
2308 0 : .arg(&core_path)
2309 0 : .args(["--batch", "-o", "bt all", "-o", "quit"])
2310 0 : .output()
2311 : }
2312 0 : _ => backtrace,
2313 0 : }?;
2314 :
2315 0 : warn!(
2316 0 : "core dump backtrace: {}",
2317 0 : String::from_utf8_lossy(&backtrace.stdout)
2318 : );
2319 0 : warn!(
2320 0 : "debugger stderr: {}",
2321 0 : String::from_utf8_lossy(&backtrace.stderr)
2322 : );
2323 : }
2324 :
2325 0 : Ok(())
2326 0 : }
2327 :
2328 : /// Select `pg_stat_statements` data and return it as a stringified JSON
2329 0 : pub async fn collect_insights(&self) -> String {
2330 0 : let mut result_rows: Vec<String> = Vec::new();
2331 0 : let conf = self.get_tokio_conn_conf(Some("compute_ctl:collect_insights"));
2332 0 : let connect_result = conf.connect(NoTls).await;
2333 0 : let (client, connection) = connect_result.unwrap();
2334 0 : tokio::spawn(async move {
2335 0 : if let Err(e) = connection.await {
2336 0 : eprintln!("connection error: {e}");
2337 0 : }
2338 0 : });
2339 0 : let result = client
2340 0 : .simple_query(
2341 0 : "SELECT
2342 0 : row_to_json(pg_stat_statements)
2343 0 : FROM
2344 0 : pg_stat_statements
2345 0 : WHERE
2346 0 : userid != 'cloud_admin'::regrole::oid
2347 0 : ORDER BY
2348 0 : (mean_exec_time + mean_plan_time) DESC
2349 0 : LIMIT 100",
2350 0 : )
2351 0 : .await;
2352 :
2353 0 : if let Ok(raw_rows) = result {
2354 0 : for message in raw_rows.iter() {
2355 0 : if let postgres::SimpleQueryMessage::Row(row) = message {
2356 0 : if let Some(json) = row.get(0) {
2357 0 : result_rows.push(json.to_string());
2358 0 : }
2359 0 : }
2360 : }
2361 :
2362 0 : format!("{{\"pg_stat_statements\": [{}]}}", result_rows.join(","))
2363 : } else {
2364 0 : "{{\"pg_stat_statements\": []}}".to_string()
2365 : }
2366 0 : }
2367 :
2368 : // download an archive, unzip and place files in correct locations
2369 0 : pub async fn download_extension(
2370 0 : &self,
2371 0 : real_ext_name: String,
2372 0 : ext_path: RemotePath,
2373 0 : ) -> Result<u64, DownloadError> {
2374 0 : let remote_ext_base_url =
2375 0 : self.params
2376 0 : .remote_ext_base_url
2377 0 : .as_ref()
2378 0 : .ok_or(DownloadError::BadInput(anyhow::anyhow!(
2379 0 : "Remote extensions storage is not configured",
2380 0 : )))?;
2381 :
2382 0 : let ext_archive_name = ext_path.object_name().expect("bad path");
2383 :
2384 0 : let mut first_try = false;
2385 0 : if !self
2386 0 : .ext_download_progress
2387 0 : .read()
2388 0 : .expect("lock err")
2389 0 : .contains_key(ext_archive_name)
2390 0 : {
2391 0 : self.ext_download_progress
2392 0 : .write()
2393 0 : .expect("lock err")
2394 0 : .insert(ext_archive_name.to_string(), (Utc::now(), false));
2395 0 : first_try = true;
2396 0 : }
2397 0 : let (download_start, download_completed) =
2398 0 : self.ext_download_progress.read().expect("lock err")[ext_archive_name];
2399 0 : let start_time_delta = Utc::now()
2400 0 : .signed_duration_since(download_start)
2401 0 : .to_std()
2402 0 : .unwrap()
2403 0 : .as_millis() as u64;
2404 :
2405 : // how long to wait for extension download if it was started by another process
2406 : const HANG_TIMEOUT: u64 = 3000; // milliseconds
2407 :
2408 0 : if download_completed {
2409 0 : info!("extension already downloaded, skipping re-download");
2410 0 : return Ok(0);
2411 0 : } else if start_time_delta < HANG_TIMEOUT && !first_try {
2412 0 : info!(
2413 0 : "download {ext_archive_name} already started by another process, hanging untill completion or timeout"
2414 : );
2415 0 : let mut interval = tokio::time::interval(tokio::time::Duration::from_millis(500));
2416 : loop {
2417 0 : info!("waiting for download");
2418 0 : interval.tick().await;
2419 0 : let (_, download_completed_now) =
2420 0 : self.ext_download_progress.read().expect("lock")[ext_archive_name];
2421 0 : if download_completed_now {
2422 0 : info!("download finished by whoever else downloaded it");
2423 0 : return Ok(0);
2424 0 : }
2425 : }
2426 : // NOTE: the above loop will get terminated
2427 : // based on the timeout of the download function
2428 0 : }
2429 :
2430 : // if extension hasn't been downloaded before or the previous
2431 : // attempt to download was at least HANG_TIMEOUT ms ago
2432 : // then we try to download it here
2433 0 : info!("downloading new extension {ext_archive_name}");
2434 :
2435 0 : let download_size = extension_server::download_extension(
2436 0 : &real_ext_name,
2437 0 : &ext_path,
2438 0 : remote_ext_base_url,
2439 0 : &self.params.pgbin,
2440 0 : )
2441 0 : .await
2442 0 : .map_err(DownloadError::Other);
2443 :
2444 0 : if download_size.is_ok() {
2445 0 : self.ext_download_progress
2446 0 : .write()
2447 0 : .expect("bad lock")
2448 0 : .insert(ext_archive_name.to_string(), (download_start, true));
2449 0 : }
2450 :
2451 0 : download_size
2452 0 : }
2453 :
2454 0 : pub async fn set_role_grants(
2455 0 : &self,
2456 0 : db_name: &PgIdent,
2457 0 : schema_name: &PgIdent,
2458 0 : privileges: &[Privilege],
2459 0 : role_name: &PgIdent,
2460 0 : ) -> Result<()> {
2461 : use tokio_postgres::NoTls;
2462 :
2463 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:set_role_grants"));
2464 0 : conf.dbname(db_name);
2465 :
2466 0 : let (db_client, conn) = conf
2467 0 : .connect(NoTls)
2468 0 : .await
2469 0 : .context("Failed to connect to the database")?;
2470 0 : tokio::spawn(conn);
2471 :
2472 : // TODO: support other types of grants apart from schemas?
2473 :
2474 : // check the role grants first - to gracefully handle read-replicas.
2475 0 : let select = "SELECT privilege_type
2476 0 : FROM pg_namespace
2477 0 : JOIN LATERAL (SELECT * FROM aclexplode(nspacl) AS x) acl ON true
2478 0 : JOIN pg_user users ON acl.grantee = users.usesysid
2479 0 : WHERE users.usename = $1
2480 0 : AND nspname = $2";
2481 0 : let rows = db_client
2482 0 : .query(select, &[role_name, schema_name])
2483 0 : .await
2484 0 : .with_context(|| format!("Failed to execute query: {select}"))?;
2485 :
2486 0 : let already_granted: HashSet<String> = rows.into_iter().map(|row| row.get(0)).collect();
2487 :
2488 0 : let grants = privileges
2489 0 : .iter()
2490 0 : .filter(|p| !already_granted.contains(p.as_str()))
2491 : // should not be quoted as it's part of the command.
2492 : // is already sanitized so it's ok
2493 0 : .map(|p| p.as_str())
2494 0 : .join(", ");
2495 :
2496 0 : if !grants.is_empty() {
2497 : // quote the schema and role name as identifiers to sanitize them.
2498 0 : let schema_name = schema_name.pg_quote();
2499 0 : let role_name = role_name.pg_quote();
2500 :
2501 0 : let query = format!("GRANT {grants} ON SCHEMA {schema_name} TO {role_name}",);
2502 0 : db_client
2503 0 : .simple_query(&query)
2504 0 : .await
2505 0 : .with_context(|| format!("Failed to execute query: {query}"))?;
2506 0 : }
2507 :
2508 0 : Ok(())
2509 0 : }
2510 :
2511 0 : pub async fn install_extension(
2512 0 : &self,
2513 0 : ext_name: &PgIdent,
2514 0 : db_name: &PgIdent,
2515 0 : ext_version: ExtVersion,
2516 0 : ) -> Result<ExtVersion> {
2517 : use tokio_postgres::NoTls;
2518 :
2519 0 : let mut conf = self.get_tokio_conn_conf(Some("compute_ctl:install_extension"));
2520 0 : conf.dbname(db_name);
2521 :
2522 0 : let (db_client, conn) = conf
2523 0 : .connect(NoTls)
2524 0 : .await
2525 0 : .context("Failed to connect to the database")?;
2526 0 : tokio::spawn(conn);
2527 :
2528 0 : let version_query = "SELECT extversion FROM pg_extension WHERE extname = $1";
2529 0 : let version: Option<ExtVersion> = db_client
2530 0 : .query_opt(version_query, &[&ext_name])
2531 0 : .await
2532 0 : .with_context(|| format!("Failed to execute query: {version_query}"))?
2533 0 : .map(|row| row.get(0));
2534 :
2535 : // sanitize the inputs as postgres idents.
2536 0 : let ext_name: String = ext_name.pg_quote();
2537 0 : let quoted_version: String = ext_version.pg_quote();
2538 :
2539 0 : if let Some(installed_version) = version {
2540 0 : if installed_version == ext_version {
2541 0 : return Ok(installed_version);
2542 0 : }
2543 0 : let query = format!("ALTER EXTENSION {ext_name} UPDATE TO {quoted_version}");
2544 0 : db_client
2545 0 : .simple_query(&query)
2546 0 : .await
2547 0 : .with_context(|| format!("Failed to execute query: {query}"))?;
2548 : } else {
2549 0 : let query =
2550 0 : format!("CREATE EXTENSION IF NOT EXISTS {ext_name} WITH VERSION {quoted_version}");
2551 0 : db_client
2552 0 : .simple_query(&query)
2553 0 : .await
2554 0 : .with_context(|| format!("Failed to execute query: {query}"))?;
2555 : }
2556 :
2557 0 : Ok(ext_version)
2558 0 : }
2559 :
2560 0 : pub async fn prepare_preload_libraries(
2561 0 : &self,
2562 0 : spec: &ComputeSpec,
2563 0 : ) -> Result<RemoteExtensionMetrics> {
2564 0 : if self.params.remote_ext_base_url.is_none() {
2565 0 : return Ok(RemoteExtensionMetrics {
2566 0 : num_ext_downloaded: 0,
2567 0 : largest_ext_size: 0,
2568 0 : total_ext_download_size: 0,
2569 0 : });
2570 0 : }
2571 0 : let remote_extensions = spec
2572 0 : .remote_extensions
2573 0 : .as_ref()
2574 0 : .ok_or(anyhow::anyhow!("Remote extensions are not configured"))?;
2575 :
2576 0 : info!("parse shared_preload_libraries from spec.cluster.settings");
2577 0 : let mut libs_vec = Vec::new();
2578 0 : if let Some(libs) = spec.cluster.settings.find("shared_preload_libraries") {
2579 0 : libs_vec = libs
2580 0 : .split(&[',', '\'', ' '])
2581 0 : .filter(|s| *s != "neon" && *s != "databricks_auth" && !s.is_empty())
2582 0 : .map(str::to_string)
2583 0 : .collect();
2584 0 : }
2585 0 : info!("parse shared_preload_libraries from provided postgresql.conf");
2586 :
2587 : // that is used in neon_local and python tests
2588 0 : if let Some(conf) = &spec.cluster.postgresql_conf {
2589 0 : let conf_lines = conf.split('\n').collect::<Vec<&str>>();
2590 0 : let mut shared_preload_libraries_line = "";
2591 0 : for line in conf_lines {
2592 0 : if line.starts_with("shared_preload_libraries") {
2593 0 : shared_preload_libraries_line = line;
2594 0 : }
2595 : }
2596 0 : let mut preload_libs_vec = Vec::new();
2597 0 : if let Some(libs) = shared_preload_libraries_line.split("='").nth(1) {
2598 0 : preload_libs_vec = libs
2599 0 : .split(&[',', '\'', ' '])
2600 0 : .filter(|s| *s != "neon" && *s != "databricks_auth" && !s.is_empty())
2601 0 : .map(str::to_string)
2602 0 : .collect();
2603 0 : }
2604 0 : libs_vec.extend(preload_libs_vec);
2605 0 : }
2606 :
2607 : // Don't try to download libraries that are not in the index.
2608 : // Assume that they are already present locally.
2609 0 : libs_vec.retain(|lib| remote_extensions.library_index.contains_key(lib));
2610 :
2611 0 : info!("Downloading to shared preload libraries: {:?}", &libs_vec);
2612 :
2613 0 : let mut download_tasks = Vec::new();
2614 0 : for library in &libs_vec {
2615 0 : let (ext_name, ext_path) =
2616 0 : remote_extensions.get_ext(library, true, &BUILD_TAG, &self.params.pgversion)?;
2617 0 : download_tasks.push(self.download_extension(ext_name, ext_path));
2618 : }
2619 0 : let results = join_all(download_tasks).await;
2620 :
2621 0 : let mut remote_ext_metrics = RemoteExtensionMetrics {
2622 0 : num_ext_downloaded: 0,
2623 0 : largest_ext_size: 0,
2624 0 : total_ext_download_size: 0,
2625 0 : };
2626 0 : for result in results {
2627 0 : let download_size = match result {
2628 0 : Ok(res) => {
2629 0 : remote_ext_metrics.num_ext_downloaded += 1;
2630 0 : res
2631 : }
2632 0 : Err(err) => {
2633 : // if we failed to download an extension, we don't want to fail the whole
2634 : // process, but we do want to log the error
2635 0 : error!("Failed to download extension: {}", err);
2636 0 : 0
2637 : }
2638 : };
2639 :
2640 0 : remote_ext_metrics.largest_ext_size =
2641 0 : std::cmp::max(remote_ext_metrics.largest_ext_size, download_size);
2642 0 : remote_ext_metrics.total_ext_download_size += download_size;
2643 : }
2644 0 : Ok(remote_ext_metrics)
2645 0 : }
2646 :
2647 : /// Waits until current thread receives a state changed notification and
2648 : /// the pageserver connection strings has changed.
2649 : ///
2650 : /// The operation will time out after a specified duration.
2651 0 : pub fn wait_timeout_while_pageserver_connstr_unchanged(&self, duration: Duration) {
2652 0 : let state = self.state.lock().unwrap();
2653 0 : let old_pageserver_conninfo = state
2654 0 : .pspec
2655 0 : .as_ref()
2656 0 : .expect("spec must be set")
2657 0 : .pageserver_conninfo
2658 0 : .clone();
2659 0 : let mut unchanged = true;
2660 0 : let _ = self
2661 0 : .state_changed
2662 0 : .wait_timeout_while(state, duration, |s| {
2663 0 : let pageserver_conninfo = &s
2664 0 : .pspec
2665 0 : .as_ref()
2666 0 : .expect("spec must be set")
2667 0 : .pageserver_conninfo;
2668 0 : unchanged = pageserver_conninfo == &old_pageserver_conninfo;
2669 0 : unchanged
2670 0 : })
2671 0 : .unwrap();
2672 0 : if !unchanged {
2673 0 : info!("Pageserver config changed");
2674 0 : }
2675 0 : }
2676 :
2677 0 : pub fn spawn_extension_stats_task(&self) {
2678 0 : self.terminate_extension_stats_task();
2679 :
2680 0 : let conf = self.tokio_conn_conf.clone();
2681 0 : let atomic_interval = self.params.installed_extensions_collection_interval.clone();
2682 0 : let mut installed_extensions_collection_interval =
2683 0 : 2 * atomic_interval.load(std::sync::atomic::Ordering::SeqCst);
2684 0 : info!(
2685 0 : "[NEON_EXT_SPAWN] Spawning background installed extensions worker with Timeout: {}",
2686 : installed_extensions_collection_interval
2687 : );
2688 0 : let handle = tokio::spawn(async move {
2689 : loop {
2690 0 : info!(
2691 0 : "[NEON_EXT_INT_SLEEP]: Interval: {}",
2692 : installed_extensions_collection_interval
2693 : );
2694 : // Sleep at the start of the loop to ensure that two collections don't happen at the same time.
2695 : // The first collection happens during compute startup.
2696 0 : tokio::time::sleep(tokio::time::Duration::from_secs(
2697 0 : installed_extensions_collection_interval,
2698 0 : ))
2699 0 : .await;
2700 0 : let _ = installed_extensions(conf.clone()).await;
2701 : // Acquire a read lock on the compute spec and then update the interval if necessary
2702 0 : installed_extensions_collection_interval = std::cmp::max(
2703 0 : installed_extensions_collection_interval,
2704 0 : 2 * atomic_interval.load(std::sync::atomic::Ordering::SeqCst),
2705 0 : );
2706 : }
2707 : });
2708 :
2709 : // Store the new task handle
2710 0 : *self.extension_stats_task.lock().unwrap() = Some(handle);
2711 0 : }
2712 :
2713 0 : fn terminate_extension_stats_task(&self) {
2714 0 : if let Some(h) = self.extension_stats_task.lock().unwrap().take() {
2715 0 : h.abort()
2716 0 : }
2717 0 : }
2718 :
2719 0 : pub fn spawn_lfc_offload_task(self: &Arc<Self>, interval: Duration) {
2720 0 : self.terminate_lfc_offload_task();
2721 0 : let secs = interval.as_secs();
2722 0 : let this = self.clone();
2723 :
2724 0 : info!("spawning LFC offload worker with {secs}s interval");
2725 0 : let handle = spawn(async move {
2726 0 : let mut interval = time::interval(interval);
2727 0 : interval.tick().await; // returns immediately
2728 : loop {
2729 0 : interval.tick().await;
2730 :
2731 0 : let prewarm_state = this.state.lock().unwrap().lfc_prewarm_state.clone();
2732 : // Do not offload LFC state if we are currently prewarming or any issue occurred.
2733 : // If we'd do that, we might override the LFC state in endpoint storage with some
2734 : // incomplete state. Imagine a situation:
2735 : // 1. Endpoint started with `autoprewarm: true`
2736 : // 2. While prewarming is not completed, we upload the new incomplete state
2737 : // 3. Compute gets interrupted and restarts
2738 : // 4. We start again and try to prewarm with the state from 2. instead of the previous complete state
2739 0 : if matches!(
2740 0 : prewarm_state,
2741 : LfcPrewarmState::Completed
2742 : | LfcPrewarmState::NotPrewarmed
2743 : | LfcPrewarmState::Skipped
2744 : ) {
2745 0 : this.offload_lfc_async().await;
2746 0 : }
2747 : }
2748 : });
2749 0 : *self.lfc_offload_task.lock().unwrap() = Some(handle);
2750 0 : }
2751 :
2752 0 : fn terminate_lfc_offload_task(&self) {
2753 0 : if let Some(h) = self.lfc_offload_task.lock().unwrap().take() {
2754 0 : h.abort()
2755 0 : }
2756 0 : }
2757 :
2758 0 : fn update_installed_extensions_collection_interval(&self, spec: &ComputeSpec) {
2759 : // Update the interval for collecting installed extensions statistics
2760 : // If the value is -1, we never suspend so set the value to default collection.
2761 : // If the value is 0, it means default, we will just continue to use the default.
2762 0 : if spec.suspend_timeout_seconds == -1 || spec.suspend_timeout_seconds == 0 {
2763 0 : self.params.installed_extensions_collection_interval.store(
2764 0 : DEFAULT_INSTALLED_EXTENSIONS_COLLECTION_INTERVAL,
2765 0 : std::sync::atomic::Ordering::SeqCst,
2766 0 : );
2767 0 : } else {
2768 0 : self.params.installed_extensions_collection_interval.store(
2769 0 : spec.suspend_timeout_seconds as u64,
2770 0 : std::sync::atomic::Ordering::SeqCst,
2771 0 : );
2772 0 : }
2773 0 : }
2774 :
2775 : /// Set the compute spec and update related metrics.
2776 : /// This is the central place where pspec is updated.
2777 0 : pub fn set_spec(params: &ComputeNodeParams, state: &mut ComputeState, pspec: ParsedSpec) {
2778 0 : state.pspec = Some(pspec);
2779 0 : ComputeNode::update_attached_metric(params, state);
2780 0 : let _ = logger::update_ids(¶ms.instance_id, &Some(params.compute_id.clone()));
2781 0 : }
2782 :
2783 0 : pub fn update_attached_metric(params: &ComputeNodeParams, state: &mut ComputeState) {
2784 : // Update the pg_cctl_attached gauge when all identifiers are available.
2785 0 : if let Some(instance_id) = ¶ms.instance_id {
2786 0 : if let Some(pspec) = &state.pspec {
2787 0 : // Clear all values in the metric
2788 0 : COMPUTE_ATTACHED.reset();
2789 0 :
2790 0 : // Set new metric value
2791 0 : COMPUTE_ATTACHED
2792 0 : .with_label_values(&[
2793 0 : ¶ms.compute_id,
2794 0 : instance_id,
2795 0 : &pspec.tenant_id.to_string(),
2796 0 : &pspec.timeline_id.to_string(),
2797 0 : ])
2798 0 : .set(1);
2799 0 : }
2800 0 : }
2801 0 : }
2802 : }
2803 :
2804 0 : pub async fn installed_extensions(conf: tokio_postgres::Config) -> Result<()> {
2805 0 : let res = get_installed_extensions(conf).await;
2806 0 : match res {
2807 0 : Ok(extensions) => {
2808 0 : info!(
2809 0 : "[NEON_EXT_STAT] {}",
2810 0 : serde_json::to_string(&extensions).expect("failed to serialize extensions list")
2811 : );
2812 : }
2813 0 : Err(err) => error!("could not get installed extensions: {err}"),
2814 : }
2815 0 : Ok(())
2816 0 : }
2817 :
2818 0 : pub fn forward_termination_signal(dev_mode: bool) {
2819 0 : let ss_pid = SYNC_SAFEKEEPERS_PID.load(Ordering::SeqCst);
2820 0 : if ss_pid != 0 {
2821 0 : let ss_pid = nix::unistd::Pid::from_raw(ss_pid as i32);
2822 0 : kill(ss_pid, Signal::SIGTERM).ok();
2823 0 : }
2824 :
2825 0 : if !dev_mode {
2826 : // Terminate pgbouncer with SIGKILL
2827 0 : match pid_file::read(PGBOUNCER_PIDFILE.into()) {
2828 0 : Ok(pid_file::PidFileRead::LockedByOtherProcess(pid)) => {
2829 0 : info!("sending SIGKILL to pgbouncer process pid: {}", pid);
2830 0 : if let Err(e) = kill(pid, Signal::SIGKILL) {
2831 0 : error!("failed to terminate pgbouncer: {}", e);
2832 0 : }
2833 : }
2834 : // pgbouncer does not lock the pid file, so we read and kill the process directly
2835 : Ok(pid_file::PidFileRead::NotHeldByAnyProcess(_)) => {
2836 0 : if let Ok(pid_str) = std::fs::read_to_string(PGBOUNCER_PIDFILE) {
2837 0 : if let Ok(pid) = pid_str.trim().parse::<i32>() {
2838 0 : info!(
2839 0 : "sending SIGKILL to pgbouncer process pid: {} (from unlocked pid file)",
2840 : pid
2841 : );
2842 0 : if let Err(e) = kill(Pid::from_raw(pid), Signal::SIGKILL) {
2843 0 : error!("failed to terminate pgbouncer: {}", e);
2844 0 : }
2845 0 : }
2846 : } else {
2847 0 : info!("pgbouncer pid file exists but process not running");
2848 : }
2849 : }
2850 : Ok(pid_file::PidFileRead::NotExist) => {
2851 0 : info!("pgbouncer pid file not found, process may not be running");
2852 : }
2853 0 : Err(e) => {
2854 0 : error!("error reading pgbouncer pid file: {}", e);
2855 : }
2856 : }
2857 :
2858 : // Terminate local_proxy
2859 0 : match pid_file::read("/etc/local_proxy/pid".into()) {
2860 0 : Ok(pid_file::PidFileRead::LockedByOtherProcess(pid)) => {
2861 0 : info!("sending SIGTERM to local_proxy process pid: {}", pid);
2862 0 : if let Err(e) = kill(pid, Signal::SIGTERM) {
2863 0 : error!("failed to terminate local_proxy: {}", e);
2864 0 : }
2865 : }
2866 : Ok(pid_file::PidFileRead::NotHeldByAnyProcess(_)) => {
2867 0 : info!("local_proxy PID file exists but process not running");
2868 : }
2869 : Ok(pid_file::PidFileRead::NotExist) => {
2870 0 : info!("local_proxy PID file not found, process may not be running");
2871 : }
2872 0 : Err(e) => {
2873 0 : error!("error reading local_proxy PID file: {}", e);
2874 : }
2875 : }
2876 : } else {
2877 0 : info!("Skipping pgbouncer and local_proxy termination because in dev mode");
2878 : }
2879 :
2880 0 : let pg_pid = PG_PID.load(Ordering::SeqCst);
2881 0 : if pg_pid != 0 {
2882 0 : let pg_pid = nix::unistd::Pid::from_raw(pg_pid as i32);
2883 0 : // Use 'fast' shutdown (SIGINT) because it also creates a shutdown checkpoint, which is important for
2884 0 : // ROs to get a list of running xacts faster instead of going through the CLOG.
2885 0 : // See https://www.postgresql.org/docs/current/server-shutdown.html for the list of modes and signals.
2886 0 : kill(pg_pid, Signal::SIGINT).ok();
2887 0 : }
2888 0 : }
2889 :
2890 : // helper trait to call JoinSet::spawn_blocking(f), but propagates the current
2891 : // tracing span to the thread.
2892 : trait JoinSetExt<T> {
2893 : fn spawn_blocking_child<F>(&mut self, f: F) -> tokio::task::AbortHandle
2894 : where
2895 : F: FnOnce() -> T + Send + 'static,
2896 : T: Send;
2897 : }
2898 :
2899 : impl<T: 'static> JoinSetExt<T> for tokio::task::JoinSet<T> {
2900 0 : fn spawn_blocking_child<F>(&mut self, f: F) -> tokio::task::AbortHandle
2901 0 : where
2902 0 : F: FnOnce() -> T + Send + 'static,
2903 0 : T: Send,
2904 : {
2905 0 : let sp = tracing::Span::current();
2906 0 : self.spawn_blocking(move || {
2907 0 : let _e = sp.enter();
2908 0 : f()
2909 0 : })
2910 0 : }
2911 : }
2912 :
2913 : #[cfg(test)]
2914 : mod tests {
2915 : use std::fs::File;
2916 :
2917 : use super::*;
2918 :
2919 : #[test]
2920 1 : fn duplicate_safekeeper_connstring() {
2921 1 : let file = File::open("tests/cluster_spec.json").unwrap();
2922 1 : let spec: ComputeSpec = serde_json::from_reader(file).unwrap();
2923 :
2924 1 : match ParsedSpec::try_from(spec.clone()) {
2925 0 : Ok(_p) => panic!("Failed to detect duplicate entry"),
2926 1 : Err(e) => assert!(
2927 1 : e.to_string()
2928 1 : .starts_with("duplicate entry in safekeeper_connstrings:")
2929 : ),
2930 : };
2931 1 : }
2932 : }
|
