Line data Source code
1 : //! Code to manage compute endpoints
2 : //!
3 : //! In the local test environment, the data for each endpoint is stored in
4 : //!
5 : //! ```text
6 : //! .neon/endpoints/<endpoint id>
7 : //! ```
8 : //!
9 : //! Some basic information about the endpoint, like the tenant and timeline IDs,
10 : //! are stored in the `endpoint.json` file. The `endpoint.json` file is created
11 : //! when the endpoint is created, and doesn't change afterwards.
12 : //!
13 : //! The endpoint is managed by the `compute_ctl` binary. When an endpoint is
14 : //! started, we launch `compute_ctl` It synchronizes the safekeepers, downloads
15 : //! the basebackup from the pageserver to initialize the data directory, and
16 : //! finally launches the PostgreSQL process. It watches the PostgreSQL process
17 : //! until it exits.
18 : //!
19 : //! When an endpoint is created, a `postgresql.conf` file is also created in
20 : //! the endpoint's directory. The file can be modified before starting PostgreSQL.
21 : //! However, the `postgresql.conf` file in the endpoint directory is not used directly
22 : //! by PostgreSQL. It is passed to `compute_ctl`, and `compute_ctl` writes another
23 : //! copy of it in the data directory.
24 : //!
25 : //! Directory contents:
26 : //!
27 : //! ```text
28 : //! .neon/endpoints/main/
29 : //! compute.log - log output of `compute_ctl` and `postgres`
30 : //! endpoint.json - serialized `EndpointConf` struct
31 : //! postgresql.conf - postgresql settings
32 : //! config.json - passed to `compute_ctl`
33 : //! pgdata/
34 : //! postgresql.conf - copy of postgresql.conf created by `compute_ctl`
35 : //! zenith.signal
36 : //! <other PostgreSQL files>
37 : //! ```
38 : //!
39 : use std::collections::BTreeMap;
40 : use std::net::{IpAddr, Ipv4Addr, SocketAddr, TcpStream};
41 : use std::path::PathBuf;
42 : use std::process::Command;
43 : use std::str::FromStr;
44 : use std::sync::Arc;
45 : use std::time::{Duration, Instant};
46 :
47 : use anyhow::{Context, Result, anyhow, bail};
48 : use compute_api::requests::{
49 : COMPUTE_AUDIENCE, ComputeClaims, ComputeClaimsScope, ConfigurationRequest,
50 : };
51 : use compute_api::responses::{
52 : ComputeConfig, ComputeCtlConfig, ComputeStatus, ComputeStatusResponse, TlsConfig,
53 : };
54 : use compute_api::spec::{
55 : Cluster, ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, Database, PgIdent,
56 : RemoteExtSpec, Role,
57 : };
58 : use jsonwebtoken::jwk::{
59 : AlgorithmParameters, CommonParameters, EllipticCurve, Jwk, JwkSet, KeyAlgorithm, KeyOperations,
60 : OctetKeyPairParameters, OctetKeyPairType, PublicKeyUse,
61 : };
62 : use nix::sys::signal::{Signal, kill};
63 : use pageserver_api::shard::ShardStripeSize;
64 : use pem::Pem;
65 : use reqwest::header::CONTENT_TYPE;
66 : use safekeeper_api::membership::SafekeeperGeneration;
67 : use serde::{Deserialize, Serialize};
68 : use sha2::{Digest, Sha256};
69 : use spki::der::Decode;
70 : use spki::{SubjectPublicKeyInfo, SubjectPublicKeyInfoRef};
71 : use tracing::debug;
72 : use url::Host;
73 : use utils::id::{NodeId, TenantId, TimelineId};
74 :
75 : use crate::local_env::LocalEnv;
76 : use crate::postgresql_conf::PostgresConf;
77 : use crate::storage_controller::StorageController;
78 :
79 : // contents of a endpoint.json file
80 0 : #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
81 : pub struct EndpointConf {
82 : endpoint_id: String,
83 : tenant_id: TenantId,
84 : timeline_id: TimelineId,
85 : mode: ComputeMode,
86 : pg_port: u16,
87 : external_http_port: u16,
88 : internal_http_port: u16,
89 : pg_version: u32,
90 : skip_pg_catalog_updates: bool,
91 : reconfigure_concurrency: usize,
92 : drop_subscriptions_before_start: bool,
93 : features: Vec<ComputeFeature>,
94 : cluster: Option<Cluster>,
95 : compute_ctl_config: ComputeCtlConfig,
96 : }
97 :
98 : //
99 : // ComputeControlPlane
100 : //
101 : pub struct ComputeControlPlane {
102 : base_port: u16,
103 :
104 : // endpoint ID is the key
105 : pub endpoints: BTreeMap<String, Arc<Endpoint>>,
106 :
107 : env: LocalEnv,
108 : }
109 :
110 : impl ComputeControlPlane {
111 : // Load current endpoints from the endpoints/ subdirectories
112 0 : pub fn load(env: LocalEnv) -> Result<ComputeControlPlane> {
113 0 : let mut endpoints = BTreeMap::default();
114 0 : for endpoint_dir in std::fs::read_dir(env.endpoints_path())
115 0 : .with_context(|| format!("failed to list {}", env.endpoints_path().display()))?
116 : {
117 0 : let ep_res = Endpoint::from_dir_entry(endpoint_dir?, &env);
118 0 : let ep = match ep_res {
119 0 : Ok(ep) => ep,
120 0 : Err(e) => match e.downcast::<std::io::Error>() {
121 0 : Ok(e) => {
122 0 : // A parallel task could delete an endpoint while we have just scanned the directory
123 0 : if e.kind() == std::io::ErrorKind::NotFound {
124 0 : continue;
125 : } else {
126 0 : Err(e)?
127 : }
128 : }
129 0 : Err(e) => Err(e)?,
130 : },
131 : };
132 0 : endpoints.insert(ep.endpoint_id.clone(), Arc::new(ep));
133 : }
134 :
135 0 : Ok(ComputeControlPlane {
136 0 : base_port: 55431,
137 0 : endpoints,
138 0 : env,
139 0 : })
140 0 : }
141 :
142 0 : fn get_port(&mut self) -> u16 {
143 0 : 1 + self
144 0 : .endpoints
145 0 : .values()
146 0 : .map(|ep| std::cmp::max(ep.pg_address.port(), ep.external_http_address.port()))
147 0 : .max()
148 0 : .unwrap_or(self.base_port)
149 0 : }
150 :
151 : /// Create a JSON Web Key Set. This ideally matches the way we create a JWKS
152 : /// from the production control plane.
153 0 : fn create_jwks_from_pem(pem: &Pem) -> Result<JwkSet> {
154 0 : let spki: SubjectPublicKeyInfoRef = SubjectPublicKeyInfo::from_der(pem.contents())?;
155 0 : let public_key = spki.subject_public_key.raw_bytes();
156 0 :
157 0 : let mut hasher = Sha256::new();
158 0 : hasher.update(public_key);
159 0 : let key_hash = hasher.finalize();
160 0 :
161 0 : Ok(JwkSet {
162 0 : keys: vec![Jwk {
163 0 : common: CommonParameters {
164 0 : public_key_use: Some(PublicKeyUse::Signature),
165 0 : key_operations: Some(vec![KeyOperations::Verify]),
166 0 : key_algorithm: Some(KeyAlgorithm::EdDSA),
167 0 : key_id: Some(base64::encode_config(key_hash, base64::URL_SAFE_NO_PAD)),
168 0 : x509_url: None::<String>,
169 0 : x509_chain: None::<Vec<String>>,
170 0 : x509_sha1_fingerprint: None::<String>,
171 0 : x509_sha256_fingerprint: None::<String>,
172 0 : },
173 0 : algorithm: AlgorithmParameters::OctetKeyPair(OctetKeyPairParameters {
174 0 : key_type: OctetKeyPairType::OctetKeyPair,
175 0 : curve: EllipticCurve::Ed25519,
176 0 : x: base64::encode_config(public_key, base64::URL_SAFE_NO_PAD),
177 0 : }),
178 0 : }],
179 0 : })
180 0 : }
181 :
182 : #[allow(clippy::too_many_arguments)]
183 0 : pub fn new_endpoint(
184 0 : &mut self,
185 0 : endpoint_id: &str,
186 0 : tenant_id: TenantId,
187 0 : timeline_id: TimelineId,
188 0 : pg_port: Option<u16>,
189 0 : external_http_port: Option<u16>,
190 0 : internal_http_port: Option<u16>,
191 0 : pg_version: u32,
192 0 : mode: ComputeMode,
193 0 : skip_pg_catalog_updates: bool,
194 0 : drop_subscriptions_before_start: bool,
195 0 : ) -> Result<Arc<Endpoint>> {
196 0 : let pg_port = pg_port.unwrap_or_else(|| self.get_port());
197 0 : let external_http_port = external_http_port.unwrap_or_else(|| self.get_port() + 1);
198 0 : let internal_http_port = internal_http_port.unwrap_or_else(|| external_http_port + 1);
199 0 : let compute_ctl_config = ComputeCtlConfig {
200 0 : jwks: Self::create_jwks_from_pem(&self.env.read_public_key()?)?,
201 0 : tls: None::<TlsConfig>,
202 0 : };
203 0 : let ep = Arc::new(Endpoint {
204 0 : endpoint_id: endpoint_id.to_owned(),
205 0 : pg_address: SocketAddr::new(IpAddr::from(Ipv4Addr::LOCALHOST), pg_port),
206 0 : external_http_address: SocketAddr::new(
207 0 : IpAddr::from(Ipv4Addr::UNSPECIFIED),
208 0 : external_http_port,
209 0 : ),
210 0 : internal_http_address: SocketAddr::new(
211 0 : IpAddr::from(Ipv4Addr::LOCALHOST),
212 0 : internal_http_port,
213 0 : ),
214 0 : env: self.env.clone(),
215 0 : timeline_id,
216 0 : mode,
217 0 : tenant_id,
218 0 : pg_version,
219 0 : // We don't setup roles and databases in the spec locally, so we don't need to
220 0 : // do catalog updates. Catalog updates also include check availability
221 0 : // data creation. Yet, we have tests that check that size and db dump
222 0 : // before and after start are the same. So, skip catalog updates,
223 0 : // with this we basically test a case of waking up an idle compute, where
224 0 : // we also skip catalog updates in the cloud.
225 0 : skip_pg_catalog_updates,
226 0 : drop_subscriptions_before_start,
227 0 : reconfigure_concurrency: 1,
228 0 : features: vec![],
229 0 : cluster: None,
230 0 : compute_ctl_config: compute_ctl_config.clone(),
231 0 : });
232 0 :
233 0 : ep.create_endpoint_dir()?;
234 : std::fs::write(
235 0 : ep.endpoint_path().join("endpoint.json"),
236 0 : serde_json::to_string_pretty(&EndpointConf {
237 0 : endpoint_id: endpoint_id.to_string(),
238 0 : tenant_id,
239 0 : timeline_id,
240 0 : mode,
241 0 : external_http_port,
242 0 : internal_http_port,
243 0 : pg_port,
244 0 : pg_version,
245 0 : skip_pg_catalog_updates,
246 0 : drop_subscriptions_before_start,
247 0 : reconfigure_concurrency: 1,
248 0 : features: vec![],
249 0 : cluster: None,
250 0 : compute_ctl_config,
251 0 : })?,
252 0 : )?;
253 : std::fs::write(
254 0 : ep.endpoint_path().join("postgresql.conf"),
255 0 : ep.setup_pg_conf()?.to_string(),
256 0 : )?;
257 :
258 0 : self.endpoints
259 0 : .insert(ep.endpoint_id.clone(), Arc::clone(&ep));
260 0 :
261 0 : Ok(ep)
262 0 : }
263 :
264 0 : pub fn check_conflicting_endpoints(
265 0 : &self,
266 0 : mode: ComputeMode,
267 0 : tenant_id: TenantId,
268 0 : timeline_id: TimelineId,
269 0 : ) -> Result<()> {
270 0 : if matches!(mode, ComputeMode::Primary) {
271 : // this check is not complete, as you could have a concurrent attempt at
272 : // creating another primary, both reading the state before checking it here,
273 : // but it's better than nothing.
274 0 : let mut duplicates = self.endpoints.iter().filter(|(_k, v)| {
275 0 : v.tenant_id == tenant_id
276 0 : && v.timeline_id == timeline_id
277 0 : && v.mode == mode
278 0 : && v.status() != EndpointStatus::Stopped
279 0 : });
280 :
281 0 : if let Some((key, _)) = duplicates.next() {
282 0 : bail!(
283 0 : "attempting to create a duplicate primary endpoint on tenant {tenant_id}, timeline {timeline_id}: endpoint {key:?} exists already. please don't do this, it is not supported."
284 0 : );
285 0 : }
286 0 : }
287 0 : Ok(())
288 0 : }
289 : }
290 :
291 : ///////////////////////////////////////////////////////////////////////////////
292 :
293 : pub struct Endpoint {
294 : /// used as the directory name
295 : endpoint_id: String,
296 : pub tenant_id: TenantId,
297 : pub timeline_id: TimelineId,
298 : pub mode: ComputeMode,
299 :
300 : // port and address of the Postgres server and `compute_ctl`'s HTTP APIs
301 : pub pg_address: SocketAddr,
302 : pub external_http_address: SocketAddr,
303 : pub internal_http_address: SocketAddr,
304 :
305 : // postgres major version in the format: 14, 15, etc.
306 : pg_version: u32,
307 :
308 : // These are not part of the endpoint as such, but the environment
309 : // the endpoint runs in.
310 : pub env: LocalEnv,
311 :
312 : // Optimizations
313 : skip_pg_catalog_updates: bool,
314 :
315 : drop_subscriptions_before_start: bool,
316 : reconfigure_concurrency: usize,
317 : // Feature flags
318 : features: Vec<ComputeFeature>,
319 : // Cluster settings
320 : cluster: Option<Cluster>,
321 :
322 : /// The compute_ctl config for the endpoint's compute.
323 : compute_ctl_config: ComputeCtlConfig,
324 : }
325 :
326 : #[derive(PartialEq, Eq)]
327 : pub enum EndpointStatus {
328 : Running,
329 : Stopped,
330 : Crashed,
331 : RunningNoPidfile,
332 : }
333 :
334 : impl std::fmt::Display for EndpointStatus {
335 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
336 0 : let s = match self {
337 0 : Self::Running => "running",
338 0 : Self::Stopped => "stopped",
339 0 : Self::Crashed => "crashed",
340 0 : Self::RunningNoPidfile => "running, no pidfile",
341 : };
342 0 : write!(writer, "{}", s)
343 0 : }
344 : }
345 :
346 : impl Endpoint {
347 0 : fn from_dir_entry(entry: std::fs::DirEntry, env: &LocalEnv) -> Result<Endpoint> {
348 0 : if !entry.file_type()?.is_dir() {
349 0 : anyhow::bail!(
350 0 : "Endpoint::from_dir_entry failed: '{}' is not a directory",
351 0 : entry.path().display()
352 0 : );
353 0 : }
354 0 :
355 0 : // parse data directory name
356 0 : let fname = entry.file_name();
357 0 : let endpoint_id = fname.to_str().unwrap().to_string();
358 :
359 : // Read the endpoint.json file
360 0 : let conf: EndpointConf =
361 0 : serde_json::from_slice(&std::fs::read(entry.path().join("endpoint.json"))?)?;
362 :
363 0 : debug!("serialized endpoint conf: {:?}", conf);
364 :
365 0 : Ok(Endpoint {
366 0 : pg_address: SocketAddr::new(IpAddr::from(Ipv4Addr::LOCALHOST), conf.pg_port),
367 0 : external_http_address: SocketAddr::new(
368 0 : IpAddr::from(Ipv4Addr::UNSPECIFIED),
369 0 : conf.external_http_port,
370 0 : ),
371 0 : internal_http_address: SocketAddr::new(
372 0 : IpAddr::from(Ipv4Addr::LOCALHOST),
373 0 : conf.internal_http_port,
374 0 : ),
375 0 : endpoint_id,
376 0 : env: env.clone(),
377 0 : timeline_id: conf.timeline_id,
378 0 : mode: conf.mode,
379 0 : tenant_id: conf.tenant_id,
380 0 : pg_version: conf.pg_version,
381 0 : skip_pg_catalog_updates: conf.skip_pg_catalog_updates,
382 0 : reconfigure_concurrency: conf.reconfigure_concurrency,
383 0 : drop_subscriptions_before_start: conf.drop_subscriptions_before_start,
384 0 : features: conf.features,
385 0 : cluster: conf.cluster,
386 0 : compute_ctl_config: conf.compute_ctl_config,
387 0 : })
388 0 : }
389 :
390 0 : fn create_endpoint_dir(&self) -> Result<()> {
391 0 : std::fs::create_dir_all(self.endpoint_path()).with_context(|| {
392 0 : format!(
393 0 : "could not create endpoint directory {}",
394 0 : self.endpoint_path().display()
395 0 : )
396 0 : })
397 0 : }
398 :
399 : // Generate postgresql.conf with default configuration
400 0 : fn setup_pg_conf(&self) -> Result<PostgresConf> {
401 0 : let mut conf = PostgresConf::new();
402 0 : conf.append("max_wal_senders", "10");
403 0 : conf.append("wal_log_hints", "off");
404 0 : conf.append("max_replication_slots", "10");
405 0 : conf.append("hot_standby", "on");
406 0 : // Set to 1MB to both exercise getPage requests/LFC, and still have enough room for
407 0 : // Postgres to operate. Everything smaller might be not enough for Postgres under load,
408 0 : // and can cause errors like 'no unpinned buffers available', see
409 0 : // <https://github.com/neondatabase/neon/issues/9956>
410 0 : conf.append("shared_buffers", "1MB");
411 0 : // Postgres defaults to effective_io_concurrency=1, which does not exercise the pageserver's
412 0 : // batching logic. Set this to 2 so that we exercise the code a bit without letting
413 0 : // individual tests do a lot of concurrent work on underpowered test machines
414 0 : conf.append("effective_io_concurrency", "2");
415 0 : conf.append("fsync", "off");
416 0 : conf.append("max_connections", "100");
417 0 : conf.append("wal_level", "logical");
418 0 : // wal_sender_timeout is the maximum time to wait for WAL replication.
419 0 : // It also defines how often the walreciever will send a feedback message to the wal sender.
420 0 : conf.append("wal_sender_timeout", "5s");
421 0 : conf.append("listen_addresses", &self.pg_address.ip().to_string());
422 0 : conf.append("port", &self.pg_address.port().to_string());
423 0 : conf.append("wal_keep_size", "0");
424 0 : // walproposer panics when basebackup is invalid, it is pointless to restart in this case.
425 0 : conf.append("restart_after_crash", "off");
426 0 :
427 0 : // Load the 'neon' extension
428 0 : conf.append("shared_preload_libraries", "neon");
429 0 :
430 0 : conf.append_line("");
431 0 : // Replication-related configurations, such as WAL sending
432 0 : match &self.mode {
433 : ComputeMode::Primary => {
434 : // Configure backpressure
435 : // - Replication write lag depends on how fast the walreceiver can process incoming WAL.
436 : // This lag determines latency of get_page_at_lsn. Speed of applying WAL is about 10MB/sec,
437 : // so to avoid expiration of 1 minute timeout, this lag should not be larger than 600MB.
438 : // Actually latency should be much smaller (better if < 1sec). But we assume that recently
439 : // updates pages are not requested from pageserver.
440 : // - Replication flush lag depends on speed of persisting data by checkpointer (creation of
441 : // delta/image layers) and advancing disk_consistent_lsn. Safekeepers are able to
442 : // remove/archive WAL only beyond disk_consistent_lsn. Too large a lag can cause long
443 : // recovery time (in case of pageserver crash) and disk space overflow at safekeepers.
444 : // - Replication apply lag depends on speed of uploading changes to S3 by uploader thread.
445 : // To be able to restore database in case of pageserver node crash, safekeeper should not
446 : // remove WAL beyond this point. Too large lag can cause space exhaustion in safekeepers
447 : // (if they are not able to upload WAL to S3).
448 0 : conf.append("max_replication_write_lag", "15MB");
449 0 : conf.append("max_replication_flush_lag", "10GB");
450 0 :
451 0 : if !self.env.safekeepers.is_empty() {
452 0 : // Configure Postgres to connect to the safekeepers
453 0 : conf.append("synchronous_standby_names", "walproposer");
454 0 :
455 0 : let safekeepers = self
456 0 : .env
457 0 : .safekeepers
458 0 : .iter()
459 0 : .map(|sk| format!("localhost:{}", sk.get_compute_port()))
460 0 : .collect::<Vec<String>>()
461 0 : .join(",");
462 0 : conf.append("neon.safekeepers", &safekeepers);
463 0 : } else {
464 0 : // We only use setup without safekeepers for tests,
465 0 : // and don't care about data durability on pageserver,
466 0 : // so set more relaxed synchronous_commit.
467 0 : conf.append("synchronous_commit", "remote_write");
468 0 :
469 0 : // Configure the node to stream WAL directly to the pageserver
470 0 : // This isn't really a supported configuration, but can be useful for
471 0 : // testing.
472 0 : conf.append("synchronous_standby_names", "pageserver");
473 0 : }
474 : }
475 0 : ComputeMode::Static(lsn) => {
476 0 : conf.append("recovery_target_lsn", &lsn.to_string());
477 0 : }
478 : ComputeMode::Replica => {
479 0 : assert!(!self.env.safekeepers.is_empty());
480 :
481 : // TODO: use future host field from safekeeper spec
482 : // Pass the list of safekeepers to the replica so that it can connect to any of them,
483 : // whichever is available.
484 0 : let sk_ports = self
485 0 : .env
486 0 : .safekeepers
487 0 : .iter()
488 0 : .map(|x| x.get_compute_port().to_string())
489 0 : .collect::<Vec<_>>()
490 0 : .join(",");
491 0 : let sk_hosts = vec!["localhost"; self.env.safekeepers.len()].join(",");
492 0 :
493 0 : let connstr = format!(
494 0 : "host={} port={} options='-c timeline_id={} tenant_id={}' application_name=replica replication=true",
495 0 : sk_hosts,
496 0 : sk_ports,
497 0 : &self.timeline_id.to_string(),
498 0 : &self.tenant_id.to_string(),
499 0 : );
500 0 :
501 0 : let slot_name = format!("repl_{}_", self.timeline_id);
502 0 : conf.append("primary_conninfo", connstr.as_str());
503 0 : conf.append("primary_slot_name", slot_name.as_str());
504 0 : conf.append("hot_standby", "on");
505 0 : // prefetching of blocks referenced in WAL doesn't make sense for us
506 0 : // Neon hot standby ignores pages that are not in the shared_buffers
507 0 : if self.pg_version >= 15 {
508 0 : conf.append("recovery_prefetch", "off");
509 0 : }
510 : }
511 : }
512 :
513 0 : Ok(conf)
514 0 : }
515 :
516 0 : pub fn endpoint_path(&self) -> PathBuf {
517 0 : self.env.endpoints_path().join(&self.endpoint_id)
518 0 : }
519 :
520 0 : pub fn pgdata(&self) -> PathBuf {
521 0 : self.endpoint_path().join("pgdata")
522 0 : }
523 :
524 0 : pub fn status(&self) -> EndpointStatus {
525 0 : let timeout = Duration::from_millis(300);
526 0 : let has_pidfile = self.pgdata().join("postmaster.pid").exists();
527 0 : let can_connect = TcpStream::connect_timeout(&self.pg_address, timeout).is_ok();
528 0 :
529 0 : match (has_pidfile, can_connect) {
530 0 : (true, true) => EndpointStatus::Running,
531 0 : (false, false) => EndpointStatus::Stopped,
532 0 : (true, false) => EndpointStatus::Crashed,
533 0 : (false, true) => EndpointStatus::RunningNoPidfile,
534 : }
535 0 : }
536 :
537 0 : fn pg_ctl(&self, args: &[&str], auth_token: &Option<String>) -> Result<()> {
538 0 : let pg_ctl_path = self.env.pg_bin_dir(self.pg_version)?.join("pg_ctl");
539 0 : let mut cmd = Command::new(&pg_ctl_path);
540 0 : cmd.args(
541 0 : [
542 0 : &[
543 0 : "-D",
544 0 : self.pgdata().to_str().unwrap(),
545 0 : "-w", //wait till pg_ctl actually does what was asked
546 0 : ],
547 0 : args,
548 0 : ]
549 0 : .concat(),
550 0 : )
551 0 : .env_clear()
552 0 : .env(
553 0 : "LD_LIBRARY_PATH",
554 0 : self.env.pg_lib_dir(self.pg_version)?.to_str().unwrap(),
555 0 : )
556 0 : .env(
557 0 : "DYLD_LIBRARY_PATH",
558 0 : self.env.pg_lib_dir(self.pg_version)?.to_str().unwrap(),
559 : );
560 :
561 : // Pass authentication token used for the connections to pageserver and safekeepers
562 0 : if let Some(token) = auth_token {
563 0 : cmd.env("NEON_AUTH_TOKEN", token);
564 0 : }
565 :
566 0 : let pg_ctl = cmd
567 0 : .output()
568 0 : .context(format!("{} failed", pg_ctl_path.display()))?;
569 0 : if !pg_ctl.status.success() {
570 0 : anyhow::bail!(
571 0 : "pg_ctl failed, exit code: {}, stdout: {}, stderr: {}",
572 0 : pg_ctl.status,
573 0 : String::from_utf8_lossy(&pg_ctl.stdout),
574 0 : String::from_utf8_lossy(&pg_ctl.stderr),
575 0 : );
576 0 : }
577 0 :
578 0 : Ok(())
579 0 : }
580 :
581 0 : fn wait_for_compute_ctl_to_exit(&self, send_sigterm: bool) -> Result<()> {
582 0 : // TODO use background_process::stop_process instead: https://github.com/neondatabase/neon/pull/6482
583 0 : let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
584 0 : let pid: u32 = std::fs::read_to_string(pidfile_path)?.parse()?;
585 0 : let pid = nix::unistd::Pid::from_raw(pid as i32);
586 0 : if send_sigterm {
587 0 : kill(pid, Signal::SIGTERM).ok();
588 0 : }
589 0 : crate::background_process::wait_until_stopped("compute_ctl", pid)?;
590 0 : Ok(())
591 0 : }
592 :
593 0 : fn read_postgresql_conf(&self) -> Result<String> {
594 0 : // Slurp the endpoints/<endpoint id>/postgresql.conf file into
595 0 : // memory. We will include it in the spec file that we pass to
596 0 : // `compute_ctl`, and `compute_ctl` will write it to the postgresql.conf
597 0 : // in the data directory.
598 0 : let postgresql_conf_path = self.endpoint_path().join("postgresql.conf");
599 0 : match std::fs::read(&postgresql_conf_path) {
600 0 : Ok(content) => Ok(String::from_utf8(content)?),
601 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok("".to_string()),
602 0 : Err(e) => Err(anyhow::Error::new(e).context(format!(
603 0 : "failed to read config file in {}",
604 0 : postgresql_conf_path.to_str().unwrap()
605 0 : ))),
606 : }
607 0 : }
608 :
609 0 : fn build_pageserver_connstr(pageservers: &[(Host, u16)]) -> String {
610 0 : pageservers
611 0 : .iter()
612 0 : .map(|(host, port)| format!("postgresql://no_user@{host}:{port}"))
613 0 : .collect::<Vec<_>>()
614 0 : .join(",")
615 0 : }
616 :
617 : /// Map safekeepers ids to the actual connection strings.
618 0 : fn build_safekeepers_connstrs(&self, sk_ids: Vec<NodeId>) -> Result<Vec<String>> {
619 0 : let mut safekeeper_connstrings = Vec::new();
620 0 : if self.mode == ComputeMode::Primary {
621 0 : for sk_id in sk_ids {
622 0 : let sk = self
623 0 : .env
624 0 : .safekeepers
625 0 : .iter()
626 0 : .find(|node| node.id == sk_id)
627 0 : .ok_or_else(|| anyhow!("safekeeper {sk_id} does not exist"))?;
628 0 : safekeeper_connstrings.push(format!("127.0.0.1:{}", sk.get_compute_port()));
629 : }
630 0 : }
631 0 : Ok(safekeeper_connstrings)
632 0 : }
633 :
634 : /// Generate a JWT with the correct claims.
635 0 : pub fn generate_jwt(&self, scope: Option<ComputeClaimsScope>) -> Result<String> {
636 0 : self.env.generate_auth_token(&ComputeClaims {
637 0 : audience: match scope {
638 0 : Some(ComputeClaimsScope::Admin) => Some(vec![COMPUTE_AUDIENCE.to_owned()]),
639 0 : _ => None,
640 : },
641 0 : compute_id: match scope {
642 0 : Some(ComputeClaimsScope::Admin) => None,
643 0 : _ => Some(self.endpoint_id.clone()),
644 : },
645 0 : scope,
646 0 : })
647 0 : }
648 :
649 : #[allow(clippy::too_many_arguments)]
650 0 : pub async fn start(
651 0 : &self,
652 0 : auth_token: &Option<String>,
653 0 : endpoint_storage_token: String,
654 0 : endpoint_storage_addr: String,
655 0 : safekeepers_generation: Option<SafekeeperGeneration>,
656 0 : safekeepers: Vec<NodeId>,
657 0 : pageservers: Vec<(Host, u16)>,
658 0 : remote_ext_base_url: Option<&String>,
659 0 : shard_stripe_size: usize,
660 0 : create_test_user: bool,
661 0 : start_timeout: Duration,
662 0 : ) -> Result<()> {
663 0 : if self.status() == EndpointStatus::Running {
664 0 : anyhow::bail!("The endpoint is already running");
665 0 : }
666 :
667 0 : let postgresql_conf = self.read_postgresql_conf()?;
668 :
669 : // We always start the compute node from scratch, so if the Postgres
670 : // data dir exists from a previous launch, remove it first.
671 0 : if self.pgdata().exists() {
672 0 : std::fs::remove_dir_all(self.pgdata())?;
673 0 : }
674 :
675 0 : let pageserver_connstring = Self::build_pageserver_connstr(&pageservers);
676 0 : assert!(!pageserver_connstring.is_empty());
677 :
678 0 : let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
679 :
680 : // check for file remote_extensions_spec.json
681 : // if it is present, read it and pass to compute_ctl
682 0 : let remote_extensions_spec_path = self.endpoint_path().join("remote_extensions_spec.json");
683 0 : let remote_extensions_spec = std::fs::File::open(remote_extensions_spec_path);
684 : let remote_extensions: Option<RemoteExtSpec>;
685 :
686 0 : if let Ok(spec_file) = remote_extensions_spec {
687 0 : remote_extensions = serde_json::from_reader(spec_file).ok();
688 0 : } else {
689 0 : remote_extensions = None;
690 0 : };
691 :
692 : // Create config file
693 0 : let config = {
694 0 : let mut spec = ComputeSpec {
695 0 : skip_pg_catalog_updates: self.skip_pg_catalog_updates,
696 0 : format_version: 1.0,
697 0 : operation_uuid: None,
698 0 : features: self.features.clone(),
699 0 : swap_size_bytes: None,
700 0 : disk_quota_bytes: None,
701 0 : disable_lfc_resizing: None,
702 0 : cluster: Cluster {
703 0 : cluster_id: None, // project ID: not used
704 0 : name: None, // project name: not used
705 0 : state: None,
706 0 : roles: if create_test_user {
707 0 : vec![Role {
708 0 : name: PgIdent::from_str("test").unwrap(),
709 0 : encrypted_password: None,
710 0 : options: None,
711 0 : }]
712 : } else {
713 0 : Vec::new()
714 : },
715 0 : databases: if create_test_user {
716 0 : vec![Database {
717 0 : name: PgIdent::from_str("neondb").unwrap(),
718 0 : owner: PgIdent::from_str("test").unwrap(),
719 0 : options: None,
720 0 : restrict_conn: false,
721 0 : invalid: false,
722 0 : }]
723 : } else {
724 0 : Vec::new()
725 : },
726 0 : settings: None,
727 0 : postgresql_conf: Some(postgresql_conf.clone()),
728 0 : },
729 0 : delta_operations: None,
730 0 : tenant_id: Some(self.tenant_id),
731 0 : timeline_id: Some(self.timeline_id),
732 0 : project_id: None,
733 0 : branch_id: None,
734 0 : endpoint_id: Some(self.endpoint_id.clone()),
735 0 : mode: self.mode,
736 0 : pageserver_connstring: Some(pageserver_connstring),
737 0 : safekeepers_generation: safekeepers_generation.map(|g| g.into_inner()),
738 0 : safekeeper_connstrings,
739 0 : storage_auth_token: auth_token.clone(),
740 0 : remote_extensions,
741 0 : pgbouncer_settings: None,
742 0 : shard_stripe_size: Some(shard_stripe_size),
743 0 : local_proxy_config: None,
744 0 : reconfigure_concurrency: self.reconfigure_concurrency,
745 0 : drop_subscriptions_before_start: self.drop_subscriptions_before_start,
746 0 : audit_log_level: ComputeAudit::Disabled,
747 0 : logs_export_host: None::<String>,
748 0 : endpoint_storage_addr: Some(endpoint_storage_addr),
749 0 : endpoint_storage_token: Some(endpoint_storage_token),
750 0 : prewarm_lfc_on_startup: false,
751 0 : };
752 0 :
753 0 : // this strange code is needed to support respec() in tests
754 0 : if self.cluster.is_some() {
755 0 : debug!("Cluster is already set in the endpoint spec, using it");
756 0 : spec.cluster = self.cluster.clone().unwrap();
757 0 :
758 0 : debug!("spec.cluster {:?}", spec.cluster);
759 :
760 : // fill missing fields again
761 0 : if create_test_user {
762 0 : spec.cluster.roles.push(Role {
763 0 : name: PgIdent::from_str("test").unwrap(),
764 0 : encrypted_password: None,
765 0 : options: None,
766 0 : });
767 0 : spec.cluster.databases.push(Database {
768 0 : name: PgIdent::from_str("neondb").unwrap(),
769 0 : owner: PgIdent::from_str("test").unwrap(),
770 0 : options: None,
771 0 : restrict_conn: false,
772 0 : invalid: false,
773 0 : });
774 0 : }
775 0 : spec.cluster.postgresql_conf = Some(postgresql_conf);
776 0 : }
777 :
778 0 : ComputeConfig {
779 0 : spec: Some(spec),
780 0 : compute_ctl_config: self.compute_ctl_config.clone(),
781 0 : }
782 0 : };
783 0 :
784 0 : let config_path = self.endpoint_path().join("config.json");
785 0 : std::fs::write(config_path, serde_json::to_string_pretty(&config)?)?;
786 :
787 : // Open log file. We'll redirect the stdout and stderr of `compute_ctl` to it.
788 0 : let logfile = std::fs::OpenOptions::new()
789 0 : .create(true)
790 0 : .append(true)
791 0 : .open(self.endpoint_path().join("compute.log"))?;
792 :
793 : // Launch compute_ctl
794 0 : let conn_str = self.connstr("cloud_admin", "postgres");
795 0 : println!("Starting postgres node at '{}'", conn_str);
796 0 : if create_test_user {
797 0 : let conn_str = self.connstr("test", "neondb");
798 0 : println!("Also at '{}'", conn_str);
799 0 : }
800 0 : let mut cmd = Command::new(self.env.neon_distrib_dir.join("compute_ctl"));
801 0 : cmd.args([
802 0 : "--external-http-port",
803 0 : &self.external_http_address.port().to_string(),
804 0 : ])
805 0 : .args([
806 0 : "--internal-http-port",
807 0 : &self.internal_http_address.port().to_string(),
808 0 : ])
809 0 : .args(["--pgdata", self.pgdata().to_str().unwrap()])
810 0 : .args(["--connstr", &conn_str])
811 0 : .arg("--config")
812 0 : .arg(self.endpoint_path().join("config.json").as_os_str())
813 0 : .args([
814 0 : "--pgbin",
815 0 : self.env
816 0 : .pg_bin_dir(self.pg_version)?
817 0 : .join("postgres")
818 0 : .to_str()
819 0 : .unwrap(),
820 0 : ])
821 0 : // TODO: It would be nice if we generated compute IDs with the same
822 0 : // algorithm as the real control plane.
823 0 : .args(["--compute-id", &self.endpoint_id])
824 0 : .stdin(std::process::Stdio::null())
825 0 : .stderr(logfile.try_clone()?)
826 0 : .stdout(logfile);
827 :
828 0 : if let Some(remote_ext_base_url) = remote_ext_base_url {
829 0 : cmd.args(["--remote-ext-base-url", remote_ext_base_url]);
830 0 : }
831 :
832 0 : let child = cmd.spawn()?;
833 : // set up a scopeguard to kill & wait for the child in case we panic or bail below
834 0 : let child = scopeguard::guard(child, |mut child| {
835 0 : println!("SIGKILL & wait the started process");
836 0 : (|| {
837 0 : // TODO: use another signal that can be caught by the child so it can clean up any children it spawned
838 0 : child.kill().context("SIGKILL child")?;
839 0 : child.wait().context("wait() for child process")?;
840 0 : anyhow::Ok(())
841 0 : })()
842 0 : .with_context(|| format!("scopeguard kill&wait child {child:?}"))
843 0 : .unwrap();
844 0 : });
845 0 :
846 0 : // Write down the pid so we can wait for it when we want to stop
847 0 : // TODO use background_process::start_process instead: https://github.com/neondatabase/neon/pull/6482
848 0 : let pid = child.id();
849 0 : let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
850 0 : std::fs::write(pidfile_path, pid.to_string())?;
851 :
852 : // Wait for it to start
853 : const ATTEMPT_INTERVAL: Duration = Duration::from_millis(100);
854 0 : let start_at = Instant::now();
855 : loop {
856 0 : match self.get_status().await {
857 0 : Ok(state) => {
858 0 : match state.status {
859 : ComputeStatus::Init => {
860 0 : if Instant::now().duration_since(start_at) > start_timeout {
861 0 : bail!(
862 0 : "compute startup timed out {:?}; still in Init state",
863 0 : start_timeout
864 0 : );
865 0 : }
866 : // keep retrying
867 : }
868 : ComputeStatus::Running => {
869 : // All good!
870 0 : break;
871 : }
872 : ComputeStatus::Failed => {
873 0 : bail!(
874 0 : "compute startup failed: {}",
875 0 : state
876 0 : .error
877 0 : .as_deref()
878 0 : .unwrap_or("<no error from compute_ctl>")
879 0 : );
880 : }
881 : ComputeStatus::Empty
882 : | ComputeStatus::ConfigurationPending
883 : | ComputeStatus::Configuration
884 : | ComputeStatus::TerminationPending
885 : | ComputeStatus::Terminated => {
886 0 : bail!("unexpected compute status: {:?}", state.status)
887 : }
888 : }
889 : }
890 0 : Err(e) => {
891 0 : if Instant::now().duration_since(start_at) > start_timeout {
892 0 : return Err(e).context(format!(
893 0 : "timed out {:?} waiting to connect to compute_ctl HTTP",
894 0 : start_timeout,
895 0 : ));
896 0 : }
897 : }
898 : }
899 0 : tokio::time::sleep(ATTEMPT_INTERVAL).await;
900 : }
901 :
902 : // disarm the scopeguard, let the child outlive this function (and neon_local invoction)
903 0 : drop(scopeguard::ScopeGuard::into_inner(child));
904 0 :
905 0 : Ok(())
906 0 : }
907 :
908 : // Call the /status HTTP API
909 0 : pub async fn get_status(&self) -> Result<ComputeStatusResponse> {
910 0 : let client = reqwest::Client::new();
911 :
912 0 : let response = client
913 0 : .request(
914 0 : reqwest::Method::GET,
915 0 : format!(
916 0 : "http://{}:{}/status",
917 0 : self.external_http_address.ip(),
918 0 : self.external_http_address.port()
919 0 : ),
920 0 : )
921 0 : .bearer_auth(self.generate_jwt(None::<ComputeClaimsScope>)?)
922 0 : .send()
923 0 : .await?;
924 :
925 : // Interpret the response
926 0 : let status = response.status();
927 0 : if !(status.is_client_error() || status.is_server_error()) {
928 0 : Ok(response.json().await?)
929 : } else {
930 : // reqwest does not export its error construction utility functions, so let's craft the message ourselves
931 0 : let url = response.url().to_owned();
932 0 : let msg = match response.text().await {
933 0 : Ok(err_body) => format!("Error: {}", err_body),
934 0 : Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
935 : };
936 0 : Err(anyhow::anyhow!(msg))
937 : }
938 0 : }
939 :
940 0 : pub async fn reconfigure(
941 0 : &self,
942 0 : mut pageservers: Vec<(Host, u16)>,
943 0 : stripe_size: Option<ShardStripeSize>,
944 0 : safekeepers: Option<Vec<NodeId>>,
945 0 : ) -> Result<()> {
946 0 : let (mut spec, compute_ctl_config) = {
947 0 : let config_path = self.endpoint_path().join("config.json");
948 0 : let file = std::fs::File::open(config_path)?;
949 0 : let config: ComputeConfig = serde_json::from_reader(file)?;
950 :
951 0 : (config.spec.unwrap(), config.compute_ctl_config)
952 : };
953 :
954 0 : let postgresql_conf = self.read_postgresql_conf()?;
955 0 : spec.cluster.postgresql_conf = Some(postgresql_conf);
956 0 :
957 0 : // If we weren't given explicit pageservers, query the storage controller
958 0 : if pageservers.is_empty() {
959 0 : let storage_controller = StorageController::from_env(&self.env);
960 0 : let locate_result = storage_controller.tenant_locate(self.tenant_id).await?;
961 0 : pageservers = locate_result
962 0 : .shards
963 0 : .into_iter()
964 0 : .map(|shard| {
965 0 : (
966 0 : Host::parse(&shard.listen_pg_addr)
967 0 : .expect("Storage controller reported bad hostname"),
968 0 : shard.listen_pg_port,
969 0 : )
970 0 : })
971 0 : .collect::<Vec<_>>();
972 0 : }
973 :
974 0 : let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
975 0 : assert!(!pageserver_connstr.is_empty());
976 0 : spec.pageserver_connstring = Some(pageserver_connstr);
977 0 : if stripe_size.is_some() {
978 0 : spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
979 0 : }
980 :
981 : // If safekeepers are not specified, don't change them.
982 0 : if let Some(safekeepers) = safekeepers {
983 0 : let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
984 0 : spec.safekeeper_connstrings = safekeeper_connstrings;
985 0 : }
986 :
987 0 : let client = reqwest::Client::builder()
988 0 : .timeout(Duration::from_secs(120))
989 0 : .build()
990 0 : .unwrap();
991 0 : let response = client
992 0 : .post(format!(
993 0 : "http://{}:{}/configure",
994 0 : self.external_http_address.ip(),
995 0 : self.external_http_address.port()
996 0 : ))
997 0 : .header(CONTENT_TYPE.as_str(), "application/json")
998 0 : .bearer_auth(self.generate_jwt(None::<ComputeClaimsScope>)?)
999 0 : .body(
1000 0 : serde_json::to_string(&ConfigurationRequest {
1001 0 : spec,
1002 0 : compute_ctl_config,
1003 0 : })
1004 0 : .unwrap(),
1005 0 : )
1006 0 : .send()
1007 0 : .await?;
1008 :
1009 0 : let status = response.status();
1010 0 : if !(status.is_client_error() || status.is_server_error()) {
1011 0 : Ok(())
1012 : } else {
1013 0 : let url = response.url().to_owned();
1014 0 : let msg = match response.text().await {
1015 0 : Ok(err_body) => format!("Error: {}", err_body),
1016 0 : Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
1017 : };
1018 0 : Err(anyhow::anyhow!(msg))
1019 : }
1020 0 : }
1021 :
1022 0 : pub fn stop(&self, mode: &str, destroy: bool) -> Result<()> {
1023 0 : self.pg_ctl(&["-m", mode, "stop"], &None)?;
1024 :
1025 : // Also wait for the compute_ctl process to die. It might have some
1026 : // cleanup work to do after postgres stops, like syncing safekeepers,
1027 : // etc.
1028 : //
1029 : // If destroying or stop mode is immediate, send it SIGTERM before
1030 : // waiting. Sometimes we do *not* want this cleanup: tests intentionally
1031 : // do stop when majority of safekeepers is down, so sync-safekeepers
1032 : // would hang otherwise. This could be a separate flag though.
1033 0 : let send_sigterm = destroy || mode == "immediate";
1034 0 : self.wait_for_compute_ctl_to_exit(send_sigterm)?;
1035 0 : if destroy {
1036 0 : println!(
1037 0 : "Destroying postgres data directory '{}'",
1038 0 : self.pgdata().to_str().unwrap()
1039 0 : );
1040 0 : std::fs::remove_dir_all(self.endpoint_path())?;
1041 0 : }
1042 0 : Ok(())
1043 0 : }
1044 :
1045 0 : pub fn connstr(&self, user: &str, db_name: &str) -> String {
1046 0 : format!(
1047 0 : "postgresql://{}@{}:{}/{}",
1048 0 : user,
1049 0 : self.pg_address.ip(),
1050 0 : self.pg_address.port(),
1051 0 : db_name
1052 0 : )
1053 0 : }
1054 : }
|
