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 : //! neon.signal
36 : //! zenith.signal - copy of neon.signal, for backward compatibility
37 : //! <other PostgreSQL files>
38 : //! ```
39 : //!
40 : use std::collections::BTreeMap;
41 : use std::fmt::Display;
42 : use std::net::{IpAddr, Ipv4Addr, SocketAddr, TcpStream};
43 : use std::path::PathBuf;
44 : use std::process::Command;
45 : use std::str::FromStr;
46 : use std::sync::Arc;
47 : use std::time::{Duration, Instant};
48 :
49 : use anyhow::{Context, Result, anyhow, bail};
50 : use base64::Engine;
51 : use base64::prelude::BASE64_URL_SAFE_NO_PAD;
52 : use compute_api::requests::{
53 : COMPUTE_AUDIENCE, ComputeClaims, ComputeClaimsScope, ConfigurationRequest,
54 : };
55 : use compute_api::responses::{
56 : ComputeConfig, ComputeCtlConfig, ComputeStatus, ComputeStatusResponse, TerminateResponse,
57 : TlsConfig,
58 : };
59 : use compute_api::spec::{
60 : Cluster, ComputeAudit, ComputeFeature, ComputeMode, ComputeSpec, Database, PageserverProtocol,
61 : PgIdent, RemoteExtSpec, Role,
62 : };
63 : use jsonwebtoken::jwk::{
64 : AlgorithmParameters, CommonParameters, EllipticCurve, Jwk, JwkSet, KeyAlgorithm, KeyOperations,
65 : OctetKeyPairParameters, OctetKeyPairType, PublicKeyUse,
66 : };
67 : use nix::sys::signal::{Signal, kill};
68 : use pageserver_api::shard::ShardStripeSize;
69 : use pem::Pem;
70 : use reqwest::header::CONTENT_TYPE;
71 : use safekeeper_api::PgMajorVersion;
72 : use safekeeper_api::membership::SafekeeperGeneration;
73 : use serde::{Deserialize, Serialize};
74 : use sha2::{Digest, Sha256};
75 : use spki::der::Decode;
76 : use spki::{SubjectPublicKeyInfo, SubjectPublicKeyInfoRef};
77 : use tracing::debug;
78 : use url::Host;
79 : use utils::id::{NodeId, TenantId, TimelineId};
80 :
81 : use crate::local_env::LocalEnv;
82 : use crate::postgresql_conf::PostgresConf;
83 :
84 : // contents of a endpoint.json file
85 0 : #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
86 : pub struct EndpointConf {
87 : endpoint_id: String,
88 : tenant_id: TenantId,
89 : timeline_id: TimelineId,
90 : mode: ComputeMode,
91 : pg_port: u16,
92 : external_http_port: u16,
93 : internal_http_port: u16,
94 : pg_version: PgMajorVersion,
95 : grpc: bool,
96 : skip_pg_catalog_updates: bool,
97 : reconfigure_concurrency: usize,
98 : drop_subscriptions_before_start: bool,
99 : features: Vec<ComputeFeature>,
100 : cluster: Option<Cluster>,
101 : compute_ctl_config: ComputeCtlConfig,
102 : }
103 :
104 : //
105 : // ComputeControlPlane
106 : //
107 : pub struct ComputeControlPlane {
108 : base_port: u16,
109 :
110 : // endpoint ID is the key
111 : pub endpoints: BTreeMap<String, Arc<Endpoint>>,
112 :
113 : env: LocalEnv,
114 : }
115 :
116 : impl ComputeControlPlane {
117 : // Load current endpoints from the endpoints/ subdirectories
118 0 : pub fn load(env: LocalEnv) -> Result<ComputeControlPlane> {
119 0 : let mut endpoints = BTreeMap::default();
120 0 : for endpoint_dir in std::fs::read_dir(env.endpoints_path())
121 0 : .with_context(|| format!("failed to list {}", env.endpoints_path().display()))?
122 : {
123 0 : let ep_res = Endpoint::from_dir_entry(endpoint_dir?, &env);
124 0 : let ep = match ep_res {
125 0 : Ok(ep) => ep,
126 0 : Err(e) => match e.downcast::<std::io::Error>() {
127 0 : Ok(e) => {
128 : // A parallel task could delete an endpoint while we have just scanned the directory
129 0 : if e.kind() == std::io::ErrorKind::NotFound {
130 0 : continue;
131 : } else {
132 0 : Err(e)?
133 : }
134 : }
135 0 : Err(e) => Err(e)?,
136 : },
137 : };
138 0 : endpoints.insert(ep.endpoint_id.clone(), Arc::new(ep));
139 : }
140 :
141 0 : Ok(ComputeControlPlane {
142 0 : base_port: 55431,
143 0 : endpoints,
144 0 : env,
145 0 : })
146 0 : }
147 :
148 0 : fn get_port(&mut self) -> u16 {
149 0 : 1 + self
150 0 : .endpoints
151 0 : .values()
152 0 : .map(|ep| std::cmp::max(ep.pg_address.port(), ep.external_http_address.port()))
153 0 : .max()
154 0 : .unwrap_or(self.base_port)
155 0 : }
156 :
157 : /// Create a JSON Web Key Set. This ideally matches the way we create a JWKS
158 : /// from the production control plane.
159 0 : fn create_jwks_from_pem(pem: &Pem) -> Result<JwkSet> {
160 0 : let spki: SubjectPublicKeyInfoRef = SubjectPublicKeyInfo::from_der(pem.contents())?;
161 0 : let public_key = spki.subject_public_key.raw_bytes();
162 :
163 0 : let mut hasher = Sha256::new();
164 0 : hasher.update(public_key);
165 0 : let key_hash = hasher.finalize();
166 :
167 0 : Ok(JwkSet {
168 0 : keys: vec![Jwk {
169 0 : common: CommonParameters {
170 0 : public_key_use: Some(PublicKeyUse::Signature),
171 0 : key_operations: Some(vec![KeyOperations::Verify]),
172 0 : key_algorithm: Some(KeyAlgorithm::EdDSA),
173 0 : key_id: Some(BASE64_URL_SAFE_NO_PAD.encode(key_hash)),
174 0 : x509_url: None::<String>,
175 0 : x509_chain: None::<Vec<String>>,
176 0 : x509_sha1_fingerprint: None::<String>,
177 0 : x509_sha256_fingerprint: None::<String>,
178 0 : },
179 0 : algorithm: AlgorithmParameters::OctetKeyPair(OctetKeyPairParameters {
180 0 : key_type: OctetKeyPairType::OctetKeyPair,
181 0 : curve: EllipticCurve::Ed25519,
182 0 : x: BASE64_URL_SAFE_NO_PAD.encode(public_key),
183 0 : }),
184 0 : }],
185 0 : })
186 0 : }
187 :
188 : #[allow(clippy::too_many_arguments)]
189 0 : pub fn new_endpoint(
190 0 : &mut self,
191 0 : endpoint_id: &str,
192 0 : tenant_id: TenantId,
193 0 : timeline_id: TimelineId,
194 0 : pg_port: Option<u16>,
195 0 : external_http_port: Option<u16>,
196 0 : internal_http_port: Option<u16>,
197 0 : pg_version: PgMajorVersion,
198 0 : mode: ComputeMode,
199 0 : grpc: bool,
200 0 : skip_pg_catalog_updates: bool,
201 0 : drop_subscriptions_before_start: bool,
202 0 : ) -> Result<Arc<Endpoint>> {
203 0 : let pg_port = pg_port.unwrap_or_else(|| self.get_port());
204 0 : let external_http_port = external_http_port.unwrap_or_else(|| self.get_port() + 1);
205 0 : let internal_http_port = internal_http_port.unwrap_or_else(|| external_http_port + 1);
206 0 : let compute_ctl_config = ComputeCtlConfig {
207 0 : jwks: Self::create_jwks_from_pem(&self.env.read_public_key()?)?,
208 0 : tls: None::<TlsConfig>,
209 : };
210 0 : let ep = Arc::new(Endpoint {
211 0 : endpoint_id: endpoint_id.to_owned(),
212 0 : pg_address: SocketAddr::new(IpAddr::from(Ipv4Addr::LOCALHOST), pg_port),
213 0 : external_http_address: SocketAddr::new(
214 0 : IpAddr::from(Ipv4Addr::UNSPECIFIED),
215 0 : external_http_port,
216 0 : ),
217 0 : internal_http_address: SocketAddr::new(
218 0 : IpAddr::from(Ipv4Addr::LOCALHOST),
219 0 : internal_http_port,
220 0 : ),
221 0 : env: self.env.clone(),
222 0 : timeline_id,
223 0 : mode,
224 0 : tenant_id,
225 0 : pg_version,
226 0 : // We don't setup roles and databases in the spec locally, so we don't need to
227 0 : // do catalog updates. Catalog updates also include check availability
228 0 : // data creation. Yet, we have tests that check that size and db dump
229 0 : // before and after start are the same. So, skip catalog updates,
230 0 : // with this we basically test a case of waking up an idle compute, where
231 0 : // we also skip catalog updates in the cloud.
232 0 : skip_pg_catalog_updates,
233 0 : drop_subscriptions_before_start,
234 0 : grpc,
235 0 : reconfigure_concurrency: 1,
236 0 : features: vec![],
237 0 : cluster: None,
238 0 : compute_ctl_config: compute_ctl_config.clone(),
239 0 : });
240 :
241 0 : ep.create_endpoint_dir()?;
242 0 : std::fs::write(
243 0 : ep.endpoint_path().join("endpoint.json"),
244 0 : serde_json::to_string_pretty(&EndpointConf {
245 0 : endpoint_id: endpoint_id.to_string(),
246 0 : tenant_id,
247 0 : timeline_id,
248 0 : mode,
249 0 : external_http_port,
250 0 : internal_http_port,
251 0 : pg_port,
252 0 : pg_version,
253 0 : grpc,
254 0 : skip_pg_catalog_updates,
255 0 : drop_subscriptions_before_start,
256 0 : reconfigure_concurrency: 1,
257 0 : features: vec![],
258 0 : cluster: None,
259 0 : compute_ctl_config,
260 0 : })?,
261 0 : )?;
262 0 : std::fs::write(
263 0 : ep.endpoint_path().join("postgresql.conf"),
264 0 : ep.setup_pg_conf()?.to_string(),
265 0 : )?;
266 :
267 0 : self.endpoints
268 0 : .insert(ep.endpoint_id.clone(), Arc::clone(&ep));
269 :
270 0 : Ok(ep)
271 0 : }
272 :
273 0 : pub fn check_conflicting_endpoints(
274 0 : &self,
275 0 : mode: ComputeMode,
276 0 : tenant_id: TenantId,
277 0 : timeline_id: TimelineId,
278 0 : ) -> Result<()> {
279 0 : if matches!(mode, ComputeMode::Primary) {
280 : // this check is not complete, as you could have a concurrent attempt at
281 : // creating another primary, both reading the state before checking it here,
282 : // but it's better than nothing.
283 0 : let mut duplicates = self.endpoints.iter().filter(|(_k, v)| {
284 0 : v.tenant_id == tenant_id
285 0 : && v.timeline_id == timeline_id
286 0 : && v.mode == mode
287 0 : && v.status() != EndpointStatus::Stopped
288 0 : });
289 :
290 0 : if let Some((key, _)) = duplicates.next() {
291 0 : bail!(
292 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."
293 : );
294 0 : }
295 0 : }
296 0 : Ok(())
297 0 : }
298 : }
299 :
300 : ///////////////////////////////////////////////////////////////////////////////
301 :
302 : pub struct Endpoint {
303 : /// used as the directory name
304 : endpoint_id: String,
305 : pub tenant_id: TenantId,
306 : pub timeline_id: TimelineId,
307 : pub mode: ComputeMode,
308 : /// If true, the endpoint should use gRPC to communicate with Pageservers.
309 : pub grpc: bool,
310 :
311 : // port and address of the Postgres server and `compute_ctl`'s HTTP APIs
312 : pub pg_address: SocketAddr,
313 : pub external_http_address: SocketAddr,
314 : pub internal_http_address: SocketAddr,
315 :
316 : // postgres major version in the format: 14, 15, etc.
317 : pg_version: PgMajorVersion,
318 :
319 : // These are not part of the endpoint as such, but the environment
320 : // the endpoint runs in.
321 : pub env: LocalEnv,
322 :
323 : // Optimizations
324 : skip_pg_catalog_updates: bool,
325 :
326 : drop_subscriptions_before_start: bool,
327 : reconfigure_concurrency: usize,
328 : // Feature flags
329 : features: Vec<ComputeFeature>,
330 : // Cluster settings
331 : cluster: Option<Cluster>,
332 :
333 : /// The compute_ctl config for the endpoint's compute.
334 : compute_ctl_config: ComputeCtlConfig,
335 : }
336 :
337 : #[derive(PartialEq, Eq)]
338 : pub enum EndpointStatus {
339 : Running,
340 : Stopped,
341 : Crashed,
342 : RunningNoPidfile,
343 : }
344 :
345 : impl Display for EndpointStatus {
346 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
347 0 : writer.write_str(match self {
348 0 : Self::Running => "running",
349 0 : Self::Stopped => "stopped",
350 0 : Self::Crashed => "crashed",
351 0 : Self::RunningNoPidfile => "running, no pidfile",
352 : })
353 0 : }
354 : }
355 :
356 : #[derive(Default, Clone, Copy, clap::ValueEnum)]
357 : pub enum EndpointTerminateMode {
358 : #[default]
359 : /// Use pg_ctl stop -m fast
360 : Fast,
361 : /// Use pg_ctl stop -m immediate
362 : Immediate,
363 : /// Use /terminate?mode=immediate
364 : ImmediateTerminate,
365 : }
366 :
367 : impl std::fmt::Display for EndpointTerminateMode {
368 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
369 0 : f.write_str(match &self {
370 0 : EndpointTerminateMode::Fast => "fast",
371 0 : EndpointTerminateMode::Immediate => "immediate",
372 0 : EndpointTerminateMode::ImmediateTerminate => "immediate-terminate",
373 : })
374 0 : }
375 : }
376 :
377 : pub struct EndpointStartArgs {
378 : pub auth_token: Option<String>,
379 : pub endpoint_storage_token: String,
380 : pub endpoint_storage_addr: String,
381 : pub safekeepers_generation: Option<SafekeeperGeneration>,
382 : pub safekeepers: Vec<NodeId>,
383 : pub pageservers: Vec<(PageserverProtocol, Host, u16)>,
384 : pub remote_ext_base_url: Option<String>,
385 : pub shard_stripe_size: usize,
386 : pub create_test_user: bool,
387 : pub start_timeout: Duration,
388 : pub autoprewarm: bool,
389 : pub offload_lfc_interval_seconds: Option<std::num::NonZeroU64>,
390 : pub dev: bool,
391 : }
392 :
393 : impl Endpoint {
394 0 : fn from_dir_entry(entry: std::fs::DirEntry, env: &LocalEnv) -> Result<Endpoint> {
395 0 : if !entry.file_type()?.is_dir() {
396 0 : anyhow::bail!(
397 0 : "Endpoint::from_dir_entry failed: '{}' is not a directory",
398 0 : entry.path().display()
399 : );
400 0 : }
401 :
402 : // parse data directory name
403 0 : let fname = entry.file_name();
404 0 : let endpoint_id = fname.to_str().unwrap().to_string();
405 :
406 : // Read the endpoint.json file
407 0 : let conf: EndpointConf =
408 0 : serde_json::from_slice(&std::fs::read(entry.path().join("endpoint.json"))?)?;
409 :
410 0 : debug!("serialized endpoint conf: {:?}", conf);
411 :
412 0 : Ok(Endpoint {
413 0 : pg_address: SocketAddr::new(IpAddr::from(Ipv4Addr::LOCALHOST), conf.pg_port),
414 0 : external_http_address: SocketAddr::new(
415 0 : IpAddr::from(Ipv4Addr::UNSPECIFIED),
416 0 : conf.external_http_port,
417 0 : ),
418 0 : internal_http_address: SocketAddr::new(
419 0 : IpAddr::from(Ipv4Addr::LOCALHOST),
420 0 : conf.internal_http_port,
421 0 : ),
422 0 : endpoint_id,
423 0 : env: env.clone(),
424 0 : timeline_id: conf.timeline_id,
425 0 : mode: conf.mode,
426 0 : tenant_id: conf.tenant_id,
427 0 : pg_version: conf.pg_version,
428 0 : grpc: conf.grpc,
429 0 : skip_pg_catalog_updates: conf.skip_pg_catalog_updates,
430 0 : reconfigure_concurrency: conf.reconfigure_concurrency,
431 0 : drop_subscriptions_before_start: conf.drop_subscriptions_before_start,
432 0 : features: conf.features,
433 0 : cluster: conf.cluster,
434 0 : compute_ctl_config: conf.compute_ctl_config,
435 0 : })
436 0 : }
437 :
438 0 : fn create_endpoint_dir(&self) -> Result<()> {
439 0 : std::fs::create_dir_all(self.endpoint_path()).with_context(|| {
440 0 : format!(
441 0 : "could not create endpoint directory {}",
442 0 : self.endpoint_path().display()
443 : )
444 0 : })
445 0 : }
446 :
447 : // Generate postgresql.conf with default configuration
448 0 : fn setup_pg_conf(&self) -> Result<PostgresConf> {
449 0 : let mut conf = PostgresConf::new();
450 0 : conf.append("max_wal_senders", "10");
451 0 : conf.append("wal_log_hints", "off");
452 0 : conf.append("max_replication_slots", "10");
453 0 : conf.append("hot_standby", "on");
454 : // Set to 1MB to both exercise getPage requests/LFC, and still have enough room for
455 : // Postgres to operate. Everything smaller might be not enough for Postgres under load,
456 : // and can cause errors like 'no unpinned buffers available', see
457 : // <https://github.com/neondatabase/neon/issues/9956>
458 0 : conf.append("shared_buffers", "1MB");
459 : // Postgres defaults to effective_io_concurrency=1, which does not exercise the pageserver's
460 : // batching logic. Set this to 2 so that we exercise the code a bit without letting
461 : // individual tests do a lot of concurrent work on underpowered test machines
462 0 : conf.append("effective_io_concurrency", "2");
463 0 : conf.append("fsync", "off");
464 0 : conf.append("max_connections", "100");
465 0 : conf.append("wal_level", "logical");
466 : // wal_sender_timeout is the maximum time to wait for WAL replication.
467 : // It also defines how often the walreciever will send a feedback message to the wal sender.
468 0 : conf.append("wal_sender_timeout", "5s");
469 0 : conf.append("listen_addresses", &self.pg_address.ip().to_string());
470 0 : conf.append("port", &self.pg_address.port().to_string());
471 0 : conf.append("wal_keep_size", "0");
472 : // walproposer panics when basebackup is invalid, it is pointless to restart in this case.
473 0 : conf.append("restart_after_crash", "off");
474 :
475 : // Load the 'neon' extension
476 0 : conf.append("shared_preload_libraries", "neon");
477 :
478 0 : conf.append_line("");
479 : // Replication-related configurations, such as WAL sending
480 0 : match &self.mode {
481 : ComputeMode::Primary => {
482 : // Configure backpressure
483 : // - Replication write lag depends on how fast the walreceiver can process incoming WAL.
484 : // This lag determines latency of get_page_at_lsn. Speed of applying WAL is about 10MB/sec,
485 : // so to avoid expiration of 1 minute timeout, this lag should not be larger than 600MB.
486 : // Actually latency should be much smaller (better if < 1sec). But we assume that recently
487 : // updates pages are not requested from pageserver.
488 : // - Replication flush lag depends on speed of persisting data by checkpointer (creation of
489 : // delta/image layers) and advancing disk_consistent_lsn. Safekeepers are able to
490 : // remove/archive WAL only beyond disk_consistent_lsn. Too large a lag can cause long
491 : // recovery time (in case of pageserver crash) and disk space overflow at safekeepers.
492 : // - Replication apply lag depends on speed of uploading changes to S3 by uploader thread.
493 : // To be able to restore database in case of pageserver node crash, safekeeper should not
494 : // remove WAL beyond this point. Too large lag can cause space exhaustion in safekeepers
495 : // (if they are not able to upload WAL to S3).
496 0 : conf.append("max_replication_write_lag", "15MB");
497 0 : conf.append("max_replication_flush_lag", "10GB");
498 :
499 0 : if !self.env.safekeepers.is_empty() {
500 : // Configure Postgres to connect to the safekeepers
501 0 : conf.append("synchronous_standby_names", "walproposer");
502 :
503 0 : let safekeepers = self
504 0 : .env
505 0 : .safekeepers
506 0 : .iter()
507 0 : .map(|sk| format!("localhost:{}", sk.get_compute_port()))
508 0 : .collect::<Vec<String>>()
509 0 : .join(",");
510 0 : conf.append("neon.safekeepers", &safekeepers);
511 0 : } else {
512 0 : // We only use setup without safekeepers for tests,
513 0 : // and don't care about data durability on pageserver,
514 0 : // so set more relaxed synchronous_commit.
515 0 : conf.append("synchronous_commit", "remote_write");
516 0 :
517 0 : // Configure the node to stream WAL directly to the pageserver
518 0 : // This isn't really a supported configuration, but can be useful for
519 0 : // testing.
520 0 : conf.append("synchronous_standby_names", "pageserver");
521 0 : }
522 : }
523 0 : ComputeMode::Static(lsn) => {
524 0 : conf.append("recovery_target_lsn", &lsn.to_string());
525 0 : }
526 : ComputeMode::Replica => {
527 0 : assert!(!self.env.safekeepers.is_empty());
528 :
529 : // TODO: use future host field from safekeeper spec
530 : // Pass the list of safekeepers to the replica so that it can connect to any of them,
531 : // whichever is available.
532 0 : let sk_ports = self
533 0 : .env
534 0 : .safekeepers
535 0 : .iter()
536 0 : .map(|x| x.get_compute_port().to_string())
537 0 : .collect::<Vec<_>>()
538 0 : .join(",");
539 0 : let sk_hosts = vec!["localhost"; self.env.safekeepers.len()].join(",");
540 :
541 0 : let connstr = format!(
542 0 : "host={} port={} options='-c timeline_id={} tenant_id={}' application_name=replica replication=true",
543 : sk_hosts,
544 : sk_ports,
545 0 : &self.timeline_id.to_string(),
546 0 : &self.tenant_id.to_string(),
547 : );
548 :
549 0 : let slot_name = format!("repl_{}_", self.timeline_id);
550 0 : conf.append("primary_conninfo", connstr.as_str());
551 0 : conf.append("primary_slot_name", slot_name.as_str());
552 0 : conf.append("hot_standby", "on");
553 : // prefetching of blocks referenced in WAL doesn't make sense for us
554 : // Neon hot standby ignores pages that are not in the shared_buffers
555 0 : if self.pg_version >= PgMajorVersion::PG15 {
556 0 : conf.append("recovery_prefetch", "off");
557 0 : }
558 : }
559 : }
560 :
561 0 : Ok(conf)
562 0 : }
563 :
564 0 : pub fn endpoint_path(&self) -> PathBuf {
565 0 : self.env.endpoints_path().join(&self.endpoint_id)
566 0 : }
567 :
568 0 : pub fn pgdata(&self) -> PathBuf {
569 0 : self.endpoint_path().join("pgdata")
570 0 : }
571 :
572 0 : pub fn status(&self) -> EndpointStatus {
573 0 : let timeout = Duration::from_millis(300);
574 0 : let has_pidfile = self.pgdata().join("postmaster.pid").exists();
575 0 : let can_connect = TcpStream::connect_timeout(&self.pg_address, timeout).is_ok();
576 :
577 0 : match (has_pidfile, can_connect) {
578 0 : (true, true) => EndpointStatus::Running,
579 0 : (false, false) => EndpointStatus::Stopped,
580 0 : (true, false) => EndpointStatus::Crashed,
581 0 : (false, true) => EndpointStatus::RunningNoPidfile,
582 : }
583 0 : }
584 :
585 0 : fn pg_ctl(&self, args: &[&str], auth_token: &Option<String>) -> Result<()> {
586 0 : let pg_ctl_path = self.env.pg_bin_dir(self.pg_version)?.join("pg_ctl");
587 0 : let mut cmd = Command::new(&pg_ctl_path);
588 0 : cmd.args(
589 0 : [
590 0 : &[
591 0 : "-D",
592 0 : self.pgdata().to_str().unwrap(),
593 0 : "-w", //wait till pg_ctl actually does what was asked
594 0 : ],
595 0 : args,
596 0 : ]
597 0 : .concat(),
598 0 : )
599 0 : .env_clear()
600 0 : .env(
601 : "LD_LIBRARY_PATH",
602 0 : self.env.pg_lib_dir(self.pg_version)?.to_str().unwrap(),
603 : )
604 0 : .env(
605 : "DYLD_LIBRARY_PATH",
606 0 : self.env.pg_lib_dir(self.pg_version)?.to_str().unwrap(),
607 : );
608 :
609 : // Pass authentication token used for the connections to pageserver and safekeepers
610 0 : if let Some(token) = auth_token {
611 0 : cmd.env("NEON_AUTH_TOKEN", token);
612 0 : }
613 :
614 0 : let pg_ctl = cmd
615 0 : .output()
616 0 : .context(format!("{} failed", pg_ctl_path.display()))?;
617 0 : if !pg_ctl.status.success() {
618 0 : anyhow::bail!(
619 0 : "pg_ctl failed, exit code: {}, stdout: {}, stderr: {}",
620 : pg_ctl.status,
621 0 : String::from_utf8_lossy(&pg_ctl.stdout),
622 0 : String::from_utf8_lossy(&pg_ctl.stderr),
623 : );
624 0 : }
625 :
626 0 : Ok(())
627 0 : }
628 :
629 0 : fn wait_for_compute_ctl_to_exit(&self, send_sigterm: bool) -> Result<()> {
630 : // TODO use background_process::stop_process instead: https://github.com/neondatabase/neon/pull/6482
631 0 : let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
632 0 : let pid: u32 = std::fs::read_to_string(pidfile_path)?.parse()?;
633 0 : let pid = nix::unistd::Pid::from_raw(pid as i32);
634 0 : if send_sigterm {
635 0 : kill(pid, Signal::SIGTERM).ok();
636 0 : }
637 0 : crate::background_process::wait_until_stopped("compute_ctl", pid)?;
638 0 : Ok(())
639 0 : }
640 :
641 0 : fn read_postgresql_conf(&self) -> Result<String> {
642 : // Slurp the endpoints/<endpoint id>/postgresql.conf file into
643 : // memory. We will include it in the spec file that we pass to
644 : // `compute_ctl`, and `compute_ctl` will write it to the postgresql.conf
645 : // in the data directory.
646 0 : let postgresql_conf_path = self.endpoint_path().join("postgresql.conf");
647 0 : match std::fs::read(&postgresql_conf_path) {
648 0 : Ok(content) => Ok(String::from_utf8(content)?),
649 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok("".to_string()),
650 0 : Err(e) => Err(anyhow::Error::new(e).context(format!(
651 0 : "failed to read config file in {}",
652 0 : postgresql_conf_path.to_str().unwrap()
653 0 : ))),
654 : }
655 0 : }
656 :
657 0 : fn build_pageserver_connstr(pageservers: &[(PageserverProtocol, Host, u16)]) -> String {
658 0 : pageservers
659 0 : .iter()
660 0 : .map(|(scheme, host, port)| format!("{scheme}://no_user@{host}:{port}"))
661 0 : .collect::<Vec<_>>()
662 0 : .join(",")
663 0 : }
664 :
665 : /// Map safekeepers ids to the actual connection strings.
666 0 : fn build_safekeepers_connstrs(&self, sk_ids: Vec<NodeId>) -> Result<Vec<String>> {
667 0 : let mut safekeeper_connstrings = Vec::new();
668 0 : if self.mode == ComputeMode::Primary {
669 0 : for sk_id in sk_ids {
670 0 : let sk = self
671 0 : .env
672 0 : .safekeepers
673 0 : .iter()
674 0 : .find(|node| node.id == sk_id)
675 0 : .ok_or_else(|| anyhow!("safekeeper {sk_id} does not exist"))?;
676 0 : safekeeper_connstrings.push(format!("127.0.0.1:{}", sk.get_compute_port()));
677 : }
678 0 : }
679 0 : Ok(safekeeper_connstrings)
680 0 : }
681 :
682 : /// Generate a JWT with the correct claims.
683 0 : pub fn generate_jwt(&self, scope: Option<ComputeClaimsScope>) -> Result<String> {
684 0 : self.env.generate_auth_token(&ComputeClaims {
685 0 : audience: match scope {
686 0 : Some(ComputeClaimsScope::Admin) => Some(vec![COMPUTE_AUDIENCE.to_owned()]),
687 0 : _ => None,
688 : },
689 0 : compute_id: match scope {
690 0 : Some(ComputeClaimsScope::Admin) => None,
691 0 : _ => Some(self.endpoint_id.clone()),
692 : },
693 0 : scope,
694 : })
695 0 : }
696 :
697 0 : pub async fn start(&self, args: EndpointStartArgs) -> Result<()> {
698 0 : if self.status() == EndpointStatus::Running {
699 0 : anyhow::bail!("The endpoint is already running");
700 0 : }
701 :
702 0 : let postgresql_conf = self.read_postgresql_conf()?;
703 :
704 : // We always start the compute node from scratch, so if the Postgres
705 : // data dir exists from a previous launch, remove it first.
706 0 : if self.pgdata().exists() {
707 0 : std::fs::remove_dir_all(self.pgdata())?;
708 0 : }
709 :
710 0 : let pageserver_connstring = Self::build_pageserver_connstr(&args.pageservers);
711 0 : assert!(!pageserver_connstring.is_empty());
712 :
713 0 : let safekeeper_connstrings = self.build_safekeepers_connstrs(args.safekeepers)?;
714 :
715 : // check for file remote_extensions_spec.json
716 : // if it is present, read it and pass to compute_ctl
717 0 : let remote_extensions_spec_path = self.endpoint_path().join("remote_extensions_spec.json");
718 0 : let remote_extensions_spec = std::fs::File::open(remote_extensions_spec_path);
719 : let remote_extensions: Option<RemoteExtSpec>;
720 :
721 0 : if let Ok(spec_file) = remote_extensions_spec {
722 0 : remote_extensions = serde_json::from_reader(spec_file).ok();
723 0 : } else {
724 0 : remote_extensions = None;
725 0 : };
726 :
727 : // Create config file
728 0 : let config = {
729 0 : let mut spec = ComputeSpec {
730 0 : skip_pg_catalog_updates: self.skip_pg_catalog_updates,
731 : format_version: 1.0,
732 0 : operation_uuid: None,
733 0 : features: self.features.clone(),
734 0 : swap_size_bytes: None,
735 0 : disk_quota_bytes: None,
736 0 : disable_lfc_resizing: None,
737 : cluster: Cluster {
738 0 : cluster_id: None, // project ID: not used
739 0 : name: None, // project name: not used
740 0 : state: None,
741 0 : roles: if args.create_test_user {
742 0 : vec![Role {
743 0 : name: PgIdent::from_str("test").unwrap(),
744 0 : encrypted_password: None,
745 0 : options: None,
746 0 : }]
747 : } else {
748 0 : Vec::new()
749 : },
750 0 : databases: if args.create_test_user {
751 0 : vec![Database {
752 0 : name: PgIdent::from_str("neondb").unwrap(),
753 0 : owner: PgIdent::from_str("test").unwrap(),
754 0 : options: None,
755 0 : restrict_conn: false,
756 0 : invalid: false,
757 0 : }]
758 : } else {
759 0 : Vec::new()
760 : },
761 0 : settings: None,
762 0 : postgresql_conf: Some(postgresql_conf.clone()),
763 : },
764 0 : delta_operations: None,
765 0 : tenant_id: Some(self.tenant_id),
766 0 : timeline_id: Some(self.timeline_id),
767 0 : project_id: None,
768 0 : branch_id: None,
769 0 : endpoint_id: Some(self.endpoint_id.clone()),
770 0 : mode: self.mode,
771 0 : pageserver_connstring: Some(pageserver_connstring),
772 0 : safekeepers_generation: args.safekeepers_generation.map(|g| g.into_inner()),
773 0 : safekeeper_connstrings,
774 0 : storage_auth_token: args.auth_token.clone(),
775 0 : remote_extensions,
776 0 : pgbouncer_settings: None,
777 0 : shard_stripe_size: Some(args.shard_stripe_size),
778 0 : local_proxy_config: None,
779 0 : reconfigure_concurrency: self.reconfigure_concurrency,
780 0 : drop_subscriptions_before_start: self.drop_subscriptions_before_start,
781 0 : audit_log_level: ComputeAudit::Disabled,
782 0 : logs_export_host: None::<String>,
783 0 : endpoint_storage_addr: Some(args.endpoint_storage_addr),
784 0 : endpoint_storage_token: Some(args.endpoint_storage_token),
785 0 : autoprewarm: args.autoprewarm,
786 0 : offload_lfc_interval_seconds: args.offload_lfc_interval_seconds,
787 : suspend_timeout_seconds: -1, // Only used in neon_local.
788 : };
789 :
790 : // this strange code is needed to support respec() in tests
791 0 : if self.cluster.is_some() {
792 0 : debug!("Cluster is already set in the endpoint spec, using it");
793 0 : spec.cluster = self.cluster.clone().unwrap();
794 :
795 0 : debug!("spec.cluster {:?}", spec.cluster);
796 :
797 : // fill missing fields again
798 0 : if args.create_test_user {
799 0 : spec.cluster.roles.push(Role {
800 0 : name: PgIdent::from_str("test").unwrap(),
801 0 : encrypted_password: None,
802 0 : options: None,
803 0 : });
804 0 : spec.cluster.databases.push(Database {
805 0 : name: PgIdent::from_str("neondb").unwrap(),
806 0 : owner: PgIdent::from_str("test").unwrap(),
807 0 : options: None,
808 0 : restrict_conn: false,
809 0 : invalid: false,
810 0 : });
811 0 : }
812 0 : spec.cluster.postgresql_conf = Some(postgresql_conf);
813 0 : }
814 :
815 0 : ComputeConfig {
816 0 : spec: Some(spec),
817 0 : compute_ctl_config: self.compute_ctl_config.clone(),
818 0 : }
819 : };
820 :
821 0 : let config_path = self.endpoint_path().join("config.json");
822 0 : std::fs::write(config_path, serde_json::to_string_pretty(&config)?)?;
823 :
824 : // Open log file. We'll redirect the stdout and stderr of `compute_ctl` to it.
825 0 : let logfile = std::fs::OpenOptions::new()
826 0 : .create(true)
827 0 : .append(true)
828 0 : .open(self.endpoint_path().join("compute.log"))?;
829 :
830 : // Launch compute_ctl
831 0 : let conn_str = self.connstr("cloud_admin", "postgres");
832 0 : println!("Starting postgres node at '{conn_str}'");
833 0 : if args.create_test_user {
834 0 : let conn_str = self.connstr("test", "neondb");
835 0 : println!("Also at '{conn_str}'");
836 0 : }
837 0 : let mut cmd = Command::new(self.env.neon_distrib_dir.join("compute_ctl"));
838 0 : cmd.args([
839 0 : "--external-http-port",
840 0 : &self.external_http_address.port().to_string(),
841 0 : ])
842 0 : .args([
843 0 : "--internal-http-port",
844 0 : &self.internal_http_address.port().to_string(),
845 0 : ])
846 0 : .args(["--pgdata", self.pgdata().to_str().unwrap()])
847 0 : .args(["--connstr", &conn_str])
848 0 : .arg("--config")
849 0 : .arg(self.endpoint_path().join("config.json").as_os_str())
850 0 : .args([
851 : "--pgbin",
852 0 : self.env
853 0 : .pg_bin_dir(self.pg_version)?
854 0 : .join("postgres")
855 0 : .to_str()
856 0 : .unwrap(),
857 : ])
858 : // TODO: It would be nice if we generated compute IDs with the same
859 : // algorithm as the real control plane.
860 0 : .args(["--compute-id", &self.endpoint_id])
861 0 : .stdin(std::process::Stdio::null())
862 0 : .stderr(logfile.try_clone()?)
863 0 : .stdout(logfile);
864 :
865 0 : if let Some(remote_ext_base_url) = args.remote_ext_base_url {
866 0 : cmd.args(["--remote-ext-base-url", &remote_ext_base_url]);
867 0 : }
868 :
869 0 : if args.dev {
870 0 : cmd.arg("--dev");
871 0 : }
872 :
873 0 : let child = cmd.spawn()?;
874 : // set up a scopeguard to kill & wait for the child in case we panic or bail below
875 0 : let child = scopeguard::guard(child, |mut child| {
876 0 : println!("SIGKILL & wait the started process");
877 0 : (|| {
878 : // TODO: use another signal that can be caught by the child so it can clean up any children it spawned
879 0 : child.kill().context("SIGKILL child")?;
880 0 : child.wait().context("wait() for child process")?;
881 0 : anyhow::Ok(())
882 : })()
883 0 : .with_context(|| format!("scopeguard kill&wait child {child:?}"))
884 0 : .unwrap();
885 0 : });
886 :
887 : // Write down the pid so we can wait for it when we want to stop
888 : // TODO use background_process::start_process instead: https://github.com/neondatabase/neon/pull/6482
889 0 : let pid = child.id();
890 0 : let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
891 0 : std::fs::write(pidfile_path, pid.to_string())?;
892 :
893 : // Wait for it to start
894 : const ATTEMPT_INTERVAL: Duration = Duration::from_millis(100);
895 0 : let start_at = Instant::now();
896 : loop {
897 0 : match self.get_status().await {
898 0 : Ok(state) => {
899 0 : match state.status {
900 : ComputeStatus::Init => {
901 0 : let timeout = args.start_timeout;
902 0 : if Instant::now().duration_since(start_at) > timeout {
903 0 : bail!(
904 0 : "compute startup timed out {:?}; still in Init state",
905 : timeout
906 : );
907 0 : }
908 : // keep retrying
909 : }
910 : ComputeStatus::Running => {
911 : // All good!
912 0 : break;
913 : }
914 : ComputeStatus::Failed => {
915 0 : bail!(
916 0 : "compute startup failed: {}",
917 0 : state
918 0 : .error
919 0 : .as_deref()
920 0 : .unwrap_or("<no error from compute_ctl>")
921 : );
922 : }
923 : ComputeStatus::Empty
924 : | ComputeStatus::ConfigurationPending
925 : | ComputeStatus::Configuration
926 : | ComputeStatus::TerminationPendingFast
927 : | ComputeStatus::TerminationPendingImmediate
928 : | ComputeStatus::Terminated => {
929 0 : bail!("unexpected compute status: {:?}", state.status)
930 : }
931 : }
932 : }
933 0 : Err(e) => {
934 0 : if Instant::now().duration_since(start_at) > args.start_timeout {
935 0 : return Err(e).context(format!(
936 0 : "timed out {:?} waiting to connect to compute_ctl HTTP",
937 : args.start_timeout
938 : ));
939 0 : }
940 : }
941 : }
942 0 : tokio::time::sleep(ATTEMPT_INTERVAL).await;
943 : }
944 :
945 : // disarm the scopeguard, let the child outlive this function (and neon_local invoction)
946 0 : drop(scopeguard::ScopeGuard::into_inner(child));
947 :
948 0 : Ok(())
949 0 : }
950 :
951 : // Call the /status HTTP API
952 0 : pub async fn get_status(&self) -> Result<ComputeStatusResponse> {
953 0 : let client = reqwest::Client::new();
954 :
955 0 : let response = client
956 0 : .request(
957 0 : reqwest::Method::GET,
958 0 : format!(
959 0 : "http://{}:{}/status",
960 0 : self.external_http_address.ip(),
961 0 : self.external_http_address.port()
962 : ),
963 : )
964 0 : .bearer_auth(self.generate_jwt(None::<ComputeClaimsScope>)?)
965 0 : .send()
966 0 : .await?;
967 :
968 : // Interpret the response
969 0 : let status = response.status();
970 0 : if !(status.is_client_error() || status.is_server_error()) {
971 0 : Ok(response.json().await?)
972 : } else {
973 : // reqwest does not export its error construction utility functions, so let's craft the message ourselves
974 0 : let url = response.url().to_owned();
975 0 : let msg = match response.text().await {
976 0 : Ok(err_body) => format!("Error: {err_body}"),
977 0 : Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
978 : };
979 0 : Err(anyhow::anyhow!(msg))
980 : }
981 0 : }
982 :
983 0 : pub async fn reconfigure(
984 0 : &self,
985 0 : pageservers: Option<Vec<(PageserverProtocol, Host, u16)>>,
986 0 : stripe_size: Option<ShardStripeSize>,
987 0 : safekeepers: Option<Vec<NodeId>>,
988 0 : safekeeper_generation: Option<SafekeeperGeneration>,
989 0 : ) -> Result<()> {
990 0 : let (mut spec, compute_ctl_config) = {
991 0 : let config_path = self.endpoint_path().join("config.json");
992 0 : let file = std::fs::File::open(config_path)?;
993 0 : let config: ComputeConfig = serde_json::from_reader(file)?;
994 :
995 0 : (config.spec.unwrap(), config.compute_ctl_config)
996 : };
997 :
998 0 : let postgresql_conf = self.read_postgresql_conf()?;
999 0 : spec.cluster.postgresql_conf = Some(postgresql_conf);
1000 :
1001 : // If pageservers are not specified, don't change them.
1002 0 : if let Some(pageservers) = pageservers {
1003 0 : anyhow::ensure!(!pageservers.is_empty(), "no pageservers provided");
1004 :
1005 0 : let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
1006 0 : spec.pageserver_connstring = Some(pageserver_connstr);
1007 0 : if stripe_size.is_some() {
1008 0 : spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
1009 0 : }
1010 0 : }
1011 :
1012 : // If safekeepers are not specified, don't change them.
1013 0 : if let Some(safekeepers) = safekeepers {
1014 0 : let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
1015 0 : spec.safekeeper_connstrings = safekeeper_connstrings;
1016 0 : if let Some(g) = safekeeper_generation {
1017 0 : spec.safekeepers_generation = Some(g.into_inner());
1018 0 : }
1019 0 : }
1020 :
1021 0 : let client = reqwest::Client::builder()
1022 0 : .timeout(Duration::from_secs(120))
1023 0 : .build()
1024 0 : .unwrap();
1025 0 : let response = client
1026 0 : .post(format!(
1027 0 : "http://{}:{}/configure",
1028 0 : self.external_http_address.ip(),
1029 0 : self.external_http_address.port()
1030 : ))
1031 0 : .header(CONTENT_TYPE.as_str(), "application/json")
1032 0 : .bearer_auth(self.generate_jwt(None::<ComputeClaimsScope>)?)
1033 0 : .body(
1034 0 : serde_json::to_string(&ConfigurationRequest {
1035 0 : spec,
1036 0 : compute_ctl_config,
1037 0 : })
1038 0 : .unwrap(),
1039 : )
1040 0 : .send()
1041 0 : .await?;
1042 :
1043 0 : let status = response.status();
1044 0 : if !(status.is_client_error() || status.is_server_error()) {
1045 0 : Ok(())
1046 : } else {
1047 0 : let url = response.url().to_owned();
1048 0 : let msg = match response.text().await {
1049 0 : Ok(err_body) => format!("Error: {err_body}"),
1050 0 : Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
1051 : };
1052 0 : Err(anyhow::anyhow!(msg))
1053 : }
1054 0 : }
1055 :
1056 0 : pub async fn reconfigure_pageservers(
1057 0 : &self,
1058 0 : pageservers: Vec<(PageserverProtocol, Host, u16)>,
1059 0 : stripe_size: Option<ShardStripeSize>,
1060 0 : ) -> Result<()> {
1061 0 : self.reconfigure(Some(pageservers), stripe_size, None, None)
1062 0 : .await
1063 0 : }
1064 :
1065 0 : pub async fn reconfigure_safekeepers(
1066 0 : &self,
1067 0 : safekeepers: Vec<NodeId>,
1068 0 : generation: SafekeeperGeneration,
1069 0 : ) -> Result<()> {
1070 0 : self.reconfigure(None, None, Some(safekeepers), Some(generation))
1071 0 : .await
1072 0 : }
1073 :
1074 0 : pub async fn stop(
1075 0 : &self,
1076 0 : mode: EndpointTerminateMode,
1077 0 : destroy: bool,
1078 0 : ) -> Result<TerminateResponse> {
1079 : // pg_ctl stop is fast but doesn't allow us to collect LSN. /terminate is
1080 : // slow, and test runs time out. Solution: special mode "immediate-terminate"
1081 : // which uses /terminate
1082 0 : let response = if let EndpointTerminateMode::ImmediateTerminate = mode {
1083 0 : let ip = self.external_http_address.ip();
1084 0 : let port = self.external_http_address.port();
1085 0 : let url = format!("http://{ip}:{port}/terminate?mode=immediate");
1086 0 : let token = self.generate_jwt(Some(ComputeClaimsScope::Admin))?;
1087 0 : let request = reqwest::Client::new().post(url).bearer_auth(token);
1088 0 : let response = request.send().await.context("/terminate")?;
1089 0 : let text = response.text().await.context("/terminate result")?;
1090 0 : serde_json::from_str(&text).with_context(|| format!("deserializing {text}"))?
1091 : } else {
1092 0 : self.pg_ctl(&["-m", &mode.to_string(), "stop"], &None)?;
1093 0 : TerminateResponse { lsn: None }
1094 : };
1095 :
1096 : // Also wait for the compute_ctl process to die. It might have some
1097 : // cleanup work to do after postgres stops, like syncing safekeepers,
1098 : // etc.
1099 : //
1100 : // If destroying or stop mode is immediate, send it SIGTERM before
1101 : // waiting. Sometimes we do *not* want this cleanup: tests intentionally
1102 : // do stop when majority of safekeepers is down, so sync-safekeepers
1103 : // would hang otherwise. This could be a separate flag though.
1104 0 : let send_sigterm = destroy || !matches!(mode, EndpointTerminateMode::Fast);
1105 0 : self.wait_for_compute_ctl_to_exit(send_sigterm)?;
1106 0 : if destroy {
1107 0 : println!(
1108 0 : "Destroying postgres data directory '{}'",
1109 0 : self.pgdata().to_str().unwrap()
1110 : );
1111 0 : std::fs::remove_dir_all(self.endpoint_path())?;
1112 0 : }
1113 0 : Ok(response)
1114 0 : }
1115 :
1116 0 : pub fn connstr(&self, user: &str, db_name: &str) -> String {
1117 0 : format!(
1118 0 : "postgresql://{}@{}:{}/{}",
1119 : user,
1120 0 : self.pg_address.ip(),
1121 0 : self.pg_address.port(),
1122 : db_name
1123 : )
1124 0 : }
1125 : }
|
