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, HashMap};
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 : PageserverShardInfo, PgIdent, RemoteExtSpec, Role,
62 : };
63 :
64 : // re-export these, because they're used in the reconfigure() function
65 : pub use compute_api::spec::{PageserverConnectionInfo, PageserverShardConnectionInfo};
66 :
67 : use jsonwebtoken::jwk::{
68 : AlgorithmParameters, CommonParameters, EllipticCurve, Jwk, JwkSet, KeyAlgorithm, KeyOperations,
69 : OctetKeyPairParameters, OctetKeyPairType, PublicKeyUse,
70 : };
71 : use nix::sys::signal::{Signal, kill};
72 : use pem::Pem;
73 : use reqwest::header::CONTENT_TYPE;
74 : use safekeeper_api::PgMajorVersion;
75 : use safekeeper_api::membership::SafekeeperGeneration;
76 : use serde::{Deserialize, Serialize};
77 : use sha2::{Digest, Sha256};
78 : use spki::der::Decode;
79 : use spki::{SubjectPublicKeyInfo, SubjectPublicKeyInfoRef};
80 : use tracing::debug;
81 : use utils::id::{NodeId, TenantId, TimelineId};
82 : use utils::shard::{ShardCount, ShardIndex, ShardNumber};
83 :
84 : use pageserver_api::config::DEFAULT_GRPC_LISTEN_PORT as DEFAULT_PAGESERVER_GRPC_PORT;
85 : use postgres_connection::parse_host_port;
86 :
87 : use crate::local_env::LocalEnv;
88 : use crate::postgresql_conf::PostgresConf;
89 :
90 : // contents of a endpoint.json file
91 0 : #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
92 : pub struct EndpointConf {
93 : endpoint_id: String,
94 : tenant_id: TenantId,
95 : timeline_id: TimelineId,
96 : mode: ComputeMode,
97 : pg_port: u16,
98 : external_http_port: u16,
99 : internal_http_port: u16,
100 : pg_version: PgMajorVersion,
101 : grpc: bool,
102 : skip_pg_catalog_updates: bool,
103 : reconfigure_concurrency: usize,
104 : drop_subscriptions_before_start: bool,
105 : features: Vec<ComputeFeature>,
106 : cluster: Option<Cluster>,
107 : compute_ctl_config: ComputeCtlConfig,
108 : privileged_role_name: Option<String>,
109 : }
110 :
111 : //
112 : // ComputeControlPlane
113 : //
114 : pub struct ComputeControlPlane {
115 : base_port: u16,
116 :
117 : // endpoint ID is the key
118 : pub endpoints: BTreeMap<String, Arc<Endpoint>>,
119 :
120 : env: LocalEnv,
121 : }
122 :
123 : impl ComputeControlPlane {
124 : // Load current endpoints from the endpoints/ subdirectories
125 0 : pub fn load(env: LocalEnv) -> Result<ComputeControlPlane> {
126 0 : let mut endpoints = BTreeMap::default();
127 0 : for endpoint_dir in std::fs::read_dir(env.endpoints_path())
128 0 : .with_context(|| format!("failed to list {}", env.endpoints_path().display()))?
129 : {
130 0 : let ep_res = Endpoint::from_dir_entry(endpoint_dir?, &env);
131 0 : let ep = match ep_res {
132 0 : Ok(ep) => ep,
133 0 : Err(e) => match e.downcast::<std::io::Error>() {
134 0 : Ok(e) => {
135 : // A parallel task could delete an endpoint while we have just scanned the directory
136 0 : if e.kind() == std::io::ErrorKind::NotFound {
137 0 : continue;
138 : } else {
139 0 : Err(e)?
140 : }
141 : }
142 0 : Err(e) => Err(e)?,
143 : },
144 : };
145 0 : endpoints.insert(ep.endpoint_id.clone(), Arc::new(ep));
146 : }
147 :
148 0 : Ok(ComputeControlPlane {
149 0 : base_port: 55431,
150 0 : endpoints,
151 0 : env,
152 0 : })
153 0 : }
154 :
155 0 : fn get_port(&mut self) -> u16 {
156 0 : 1 + self
157 0 : .endpoints
158 0 : .values()
159 0 : .map(|ep| std::cmp::max(ep.pg_address.port(), ep.external_http_address.port()))
160 0 : .max()
161 0 : .unwrap_or(self.base_port)
162 0 : }
163 :
164 : /// Create a JSON Web Key Set. This ideally matches the way we create a JWKS
165 : /// from the production control plane.
166 0 : fn create_jwks_from_pem(pem: &Pem) -> Result<JwkSet> {
167 0 : let spki: SubjectPublicKeyInfoRef = SubjectPublicKeyInfo::from_der(pem.contents())?;
168 0 : let public_key = spki.subject_public_key.raw_bytes();
169 :
170 0 : let mut hasher = Sha256::new();
171 0 : hasher.update(public_key);
172 0 : let key_hash = hasher.finalize();
173 :
174 0 : Ok(JwkSet {
175 0 : keys: vec![Jwk {
176 0 : common: CommonParameters {
177 0 : public_key_use: Some(PublicKeyUse::Signature),
178 0 : key_operations: Some(vec![KeyOperations::Verify]),
179 0 : key_algorithm: Some(KeyAlgorithm::EdDSA),
180 0 : key_id: Some(BASE64_URL_SAFE_NO_PAD.encode(key_hash)),
181 0 : x509_url: None::<String>,
182 0 : x509_chain: None::<Vec<String>>,
183 0 : x509_sha1_fingerprint: None::<String>,
184 0 : x509_sha256_fingerprint: None::<String>,
185 0 : },
186 0 : algorithm: AlgorithmParameters::OctetKeyPair(OctetKeyPairParameters {
187 0 : key_type: OctetKeyPairType::OctetKeyPair,
188 0 : curve: EllipticCurve::Ed25519,
189 0 : x: BASE64_URL_SAFE_NO_PAD.encode(public_key),
190 0 : }),
191 0 : }],
192 0 : })
193 0 : }
194 :
195 : #[allow(clippy::too_many_arguments)]
196 0 : pub fn new_endpoint(
197 0 : &mut self,
198 0 : endpoint_id: &str,
199 0 : tenant_id: TenantId,
200 0 : timeline_id: TimelineId,
201 0 : pg_port: Option<u16>,
202 0 : external_http_port: Option<u16>,
203 0 : internal_http_port: Option<u16>,
204 0 : pg_version: PgMajorVersion,
205 0 : mode: ComputeMode,
206 0 : grpc: bool,
207 0 : skip_pg_catalog_updates: bool,
208 0 : drop_subscriptions_before_start: bool,
209 0 : privileged_role_name: Option<String>,
210 0 : ) -> Result<Arc<Endpoint>> {
211 0 : let pg_port = pg_port.unwrap_or_else(|| self.get_port());
212 0 : let external_http_port = external_http_port.unwrap_or_else(|| self.get_port() + 1);
213 0 : let internal_http_port = internal_http_port.unwrap_or_else(|| external_http_port + 1);
214 0 : let compute_ctl_config = ComputeCtlConfig {
215 0 : jwks: Self::create_jwks_from_pem(&self.env.read_public_key()?)?,
216 0 : tls: None::<TlsConfig>,
217 : };
218 0 : let ep = Arc::new(Endpoint {
219 0 : endpoint_id: endpoint_id.to_owned(),
220 0 : pg_address: SocketAddr::new(IpAddr::from(Ipv4Addr::LOCALHOST), pg_port),
221 0 : external_http_address: SocketAddr::new(
222 0 : IpAddr::from(Ipv4Addr::UNSPECIFIED),
223 0 : external_http_port,
224 0 : ),
225 0 : internal_http_address: SocketAddr::new(
226 0 : IpAddr::from(Ipv4Addr::LOCALHOST),
227 0 : internal_http_port,
228 0 : ),
229 0 : env: self.env.clone(),
230 0 : timeline_id,
231 0 : mode,
232 0 : tenant_id,
233 0 : pg_version,
234 0 : // We don't setup roles and databases in the spec locally, so we don't need to
235 0 : // do catalog updates. Catalog updates also include check availability
236 0 : // data creation. Yet, we have tests that check that size and db dump
237 0 : // before and after start are the same. So, skip catalog updates,
238 0 : // with this we basically test a case of waking up an idle compute, where
239 0 : // we also skip catalog updates in the cloud.
240 0 : skip_pg_catalog_updates,
241 0 : drop_subscriptions_before_start,
242 0 : grpc,
243 0 : reconfigure_concurrency: 1,
244 0 : features: vec![],
245 0 : cluster: None,
246 0 : compute_ctl_config: compute_ctl_config.clone(),
247 0 : privileged_role_name: privileged_role_name.clone(),
248 0 : });
249 :
250 0 : ep.create_endpoint_dir()?;
251 0 : std::fs::write(
252 0 : ep.endpoint_path().join("endpoint.json"),
253 0 : serde_json::to_string_pretty(&EndpointConf {
254 0 : endpoint_id: endpoint_id.to_string(),
255 0 : tenant_id,
256 0 : timeline_id,
257 0 : mode,
258 0 : external_http_port,
259 0 : internal_http_port,
260 0 : pg_port,
261 0 : pg_version,
262 0 : grpc,
263 0 : skip_pg_catalog_updates,
264 0 : drop_subscriptions_before_start,
265 0 : reconfigure_concurrency: 1,
266 0 : features: vec![],
267 0 : cluster: None,
268 0 : compute_ctl_config,
269 0 : privileged_role_name,
270 0 : })?,
271 0 : )?;
272 0 : std::fs::write(
273 0 : ep.endpoint_path().join("postgresql.conf"),
274 0 : ep.setup_pg_conf()?.to_string(),
275 0 : )?;
276 :
277 0 : self.endpoints
278 0 : .insert(ep.endpoint_id.clone(), Arc::clone(&ep));
279 :
280 0 : Ok(ep)
281 0 : }
282 :
283 0 : pub fn check_conflicting_endpoints(
284 0 : &self,
285 0 : mode: ComputeMode,
286 0 : tenant_id: TenantId,
287 0 : timeline_id: TimelineId,
288 0 : ) -> Result<()> {
289 0 : if matches!(mode, ComputeMode::Primary) {
290 : // this check is not complete, as you could have a concurrent attempt at
291 : // creating another primary, both reading the state before checking it here,
292 : // but it's better than nothing.
293 0 : let mut duplicates = self.endpoints.iter().filter(|(_k, v)| {
294 0 : v.tenant_id == tenant_id
295 0 : && v.timeline_id == timeline_id
296 0 : && v.mode == mode
297 0 : && v.status() != EndpointStatus::Stopped
298 0 : });
299 :
300 0 : if let Some((key, _)) = duplicates.next() {
301 0 : bail!(
302 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."
303 : );
304 0 : }
305 0 : }
306 0 : Ok(())
307 0 : }
308 : }
309 :
310 : ///////////////////////////////////////////////////////////////////////////////
311 :
312 : pub struct Endpoint {
313 : /// used as the directory name
314 : endpoint_id: String,
315 : pub tenant_id: TenantId,
316 : pub timeline_id: TimelineId,
317 : pub mode: ComputeMode,
318 : /// If true, the endpoint should use gRPC to communicate with Pageservers.
319 : pub grpc: bool,
320 :
321 : // port and address of the Postgres server and `compute_ctl`'s HTTP APIs
322 : pub pg_address: SocketAddr,
323 : pub external_http_address: SocketAddr,
324 : pub internal_http_address: SocketAddr,
325 :
326 : // postgres major version in the format: 14, 15, etc.
327 : pg_version: PgMajorVersion,
328 :
329 : // These are not part of the endpoint as such, but the environment
330 : // the endpoint runs in.
331 : pub env: LocalEnv,
332 :
333 : // Optimizations
334 : skip_pg_catalog_updates: bool,
335 :
336 : drop_subscriptions_before_start: bool,
337 : reconfigure_concurrency: usize,
338 : // Feature flags
339 : features: Vec<ComputeFeature>,
340 : // Cluster settings
341 : cluster: Option<Cluster>,
342 :
343 : /// The compute_ctl config for the endpoint's compute.
344 : compute_ctl_config: ComputeCtlConfig,
345 :
346 : /// The name of the privileged role for the endpoint.
347 : privileged_role_name: Option<String>,
348 : }
349 :
350 : #[derive(PartialEq, Eq)]
351 : pub enum EndpointStatus {
352 : Running,
353 : Stopped,
354 : Crashed,
355 : RunningNoPidfile,
356 : }
357 :
358 : impl Display for EndpointStatus {
359 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
360 0 : writer.write_str(match self {
361 0 : Self::Running => "running",
362 0 : Self::Stopped => "stopped",
363 0 : Self::Crashed => "crashed",
364 0 : Self::RunningNoPidfile => "running, no pidfile",
365 : })
366 0 : }
367 : }
368 :
369 : #[derive(Default, Clone, Copy, clap::ValueEnum)]
370 : pub enum EndpointTerminateMode {
371 : #[default]
372 : /// Use pg_ctl stop -m fast
373 : Fast,
374 : /// Use pg_ctl stop -m immediate
375 : Immediate,
376 : /// Use /terminate?mode=immediate
377 : ImmediateTerminate,
378 : }
379 :
380 : impl std::fmt::Display for EndpointTerminateMode {
381 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
382 0 : f.write_str(match &self {
383 0 : EndpointTerminateMode::Fast => "fast",
384 0 : EndpointTerminateMode::Immediate => "immediate",
385 0 : EndpointTerminateMode::ImmediateTerminate => "immediate-terminate",
386 : })
387 0 : }
388 : }
389 :
390 : pub struct EndpointStartArgs {
391 : pub auth_token: Option<String>,
392 : pub endpoint_storage_token: String,
393 : pub endpoint_storage_addr: String,
394 : pub safekeepers_generation: Option<SafekeeperGeneration>,
395 : pub safekeepers: Vec<NodeId>,
396 : pub pageserver_conninfo: PageserverConnectionInfo,
397 : pub remote_ext_base_url: Option<String>,
398 : pub create_test_user: bool,
399 : pub start_timeout: Duration,
400 : pub autoprewarm: bool,
401 : pub offload_lfc_interval_seconds: Option<std::num::NonZeroU64>,
402 : pub dev: bool,
403 : }
404 :
405 : impl Endpoint {
406 0 : fn from_dir_entry(entry: std::fs::DirEntry, env: &LocalEnv) -> Result<Endpoint> {
407 0 : if !entry.file_type()?.is_dir() {
408 0 : anyhow::bail!(
409 0 : "Endpoint::from_dir_entry failed: '{}' is not a directory",
410 0 : entry.path().display()
411 : );
412 0 : }
413 :
414 : // parse data directory name
415 0 : let fname = entry.file_name();
416 0 : let endpoint_id = fname.to_str().unwrap().to_string();
417 :
418 : // Read the endpoint.json file
419 0 : let conf: EndpointConf =
420 0 : serde_json::from_slice(&std::fs::read(entry.path().join("endpoint.json"))?)?;
421 :
422 0 : debug!("serialized endpoint conf: {:?}", conf);
423 :
424 0 : Ok(Endpoint {
425 0 : pg_address: SocketAddr::new(IpAddr::from(Ipv4Addr::LOCALHOST), conf.pg_port),
426 0 : external_http_address: SocketAddr::new(
427 0 : IpAddr::from(Ipv4Addr::UNSPECIFIED),
428 0 : conf.external_http_port,
429 0 : ),
430 0 : internal_http_address: SocketAddr::new(
431 0 : IpAddr::from(Ipv4Addr::LOCALHOST),
432 0 : conf.internal_http_port,
433 0 : ),
434 0 : endpoint_id,
435 0 : env: env.clone(),
436 0 : timeline_id: conf.timeline_id,
437 0 : mode: conf.mode,
438 0 : tenant_id: conf.tenant_id,
439 0 : pg_version: conf.pg_version,
440 0 : grpc: conf.grpc,
441 0 : skip_pg_catalog_updates: conf.skip_pg_catalog_updates,
442 0 : reconfigure_concurrency: conf.reconfigure_concurrency,
443 0 : drop_subscriptions_before_start: conf.drop_subscriptions_before_start,
444 0 : features: conf.features,
445 0 : cluster: conf.cluster,
446 0 : compute_ctl_config: conf.compute_ctl_config,
447 0 : privileged_role_name: conf.privileged_role_name,
448 0 : })
449 0 : }
450 :
451 0 : fn create_endpoint_dir(&self) -> Result<()> {
452 0 : std::fs::create_dir_all(self.endpoint_path()).with_context(|| {
453 0 : format!(
454 0 : "could not create endpoint directory {}",
455 0 : self.endpoint_path().display()
456 : )
457 0 : })
458 0 : }
459 :
460 : // Generate postgresql.conf with default configuration
461 0 : fn setup_pg_conf(&self) -> Result<PostgresConf> {
462 0 : let mut conf = PostgresConf::new();
463 0 : conf.append("max_wal_senders", "10");
464 0 : conf.append("wal_log_hints", "off");
465 0 : conf.append("max_replication_slots", "10");
466 0 : conf.append("hot_standby", "on");
467 : // Set to 1MB to both exercise getPage requests/LFC, and still have enough room for
468 : // Postgres to operate. Everything smaller might be not enough for Postgres under load,
469 : // and can cause errors like 'no unpinned buffers available', see
470 : // <https://github.com/neondatabase/neon/issues/9956>
471 0 : conf.append("shared_buffers", "1MB");
472 : // Postgres defaults to effective_io_concurrency=1, which does not exercise the pageserver's
473 : // batching logic. Set this to 2 so that we exercise the code a bit without letting
474 : // individual tests do a lot of concurrent work on underpowered test machines
475 0 : conf.append("effective_io_concurrency", "2");
476 0 : conf.append("fsync", "off");
477 0 : conf.append("max_connections", "100");
478 0 : conf.append("wal_level", "logical");
479 : // wal_sender_timeout is the maximum time to wait for WAL replication.
480 : // It also defines how often the walreceiver will send a feedback message to the wal sender.
481 0 : conf.append("wal_sender_timeout", "5s");
482 0 : conf.append("listen_addresses", &self.pg_address.ip().to_string());
483 0 : conf.append("port", &self.pg_address.port().to_string());
484 0 : conf.append("wal_keep_size", "0");
485 : // walproposer panics when basebackup is invalid, it is pointless to restart in this case.
486 0 : conf.append("restart_after_crash", "off");
487 :
488 : // Load the 'neon' extension
489 0 : conf.append("shared_preload_libraries", "neon");
490 :
491 0 : conf.append_line("");
492 : // Replication-related configurations, such as WAL sending
493 0 : match &self.mode {
494 : ComputeMode::Primary => {
495 : // Configure backpressure
496 : // - Replication write lag depends on how fast the walreceiver can process incoming WAL.
497 : // This lag determines latency of get_page_at_lsn. Speed of applying WAL is about 10MB/sec,
498 : // so to avoid expiration of 1 minute timeout, this lag should not be larger than 600MB.
499 : // Actually latency should be much smaller (better if < 1sec). But we assume that recently
500 : // updates pages are not requested from pageserver.
501 : // - Replication flush lag depends on speed of persisting data by checkpointer (creation of
502 : // delta/image layers) and advancing disk_consistent_lsn. Safekeepers are able to
503 : // remove/archive WAL only beyond disk_consistent_lsn. Too large a lag can cause long
504 : // recovery time (in case of pageserver crash) and disk space overflow at safekeepers.
505 : // - Replication apply lag depends on speed of uploading changes to S3 by uploader thread.
506 : // To be able to restore database in case of pageserver node crash, safekeeper should not
507 : // remove WAL beyond this point. Too large lag can cause space exhaustion in safekeepers
508 : // (if they are not able to upload WAL to S3).
509 0 : conf.append("max_replication_write_lag", "15MB");
510 0 : conf.append("max_replication_flush_lag", "10GB");
511 :
512 0 : if !self.env.safekeepers.is_empty() {
513 : // Configure Postgres to connect to the safekeepers
514 0 : conf.append("synchronous_standby_names", "walproposer");
515 :
516 0 : let safekeepers = self
517 0 : .env
518 0 : .safekeepers
519 0 : .iter()
520 0 : .map(|sk| format!("localhost:{}", sk.get_compute_port()))
521 0 : .collect::<Vec<String>>()
522 0 : .join(",");
523 0 : conf.append("neon.safekeepers", &safekeepers);
524 0 : } else {
525 0 : // We only use setup without safekeepers for tests,
526 0 : // and don't care about data durability on pageserver,
527 0 : // so set more relaxed synchronous_commit.
528 0 : conf.append("synchronous_commit", "remote_write");
529 0 :
530 0 : // Configure the node to stream WAL directly to the pageserver
531 0 : // This isn't really a supported configuration, but can be useful for
532 0 : // testing.
533 0 : conf.append("synchronous_standby_names", "pageserver");
534 0 : }
535 : }
536 0 : ComputeMode::Static(lsn) => {
537 0 : conf.append("recovery_target_lsn", &lsn.to_string());
538 0 : }
539 : ComputeMode::Replica => {
540 0 : assert!(!self.env.safekeepers.is_empty());
541 :
542 : // TODO: use future host field from safekeeper spec
543 : // Pass the list of safekeepers to the replica so that it can connect to any of them,
544 : // whichever is available.
545 0 : let sk_ports = self
546 0 : .env
547 0 : .safekeepers
548 0 : .iter()
549 0 : .map(|x| x.get_compute_port().to_string())
550 0 : .collect::<Vec<_>>()
551 0 : .join(",");
552 0 : let sk_hosts = vec!["localhost"; self.env.safekeepers.len()].join(",");
553 :
554 0 : let connstr = format!(
555 0 : "host={} port={} options='-c timeline_id={} tenant_id={}' application_name=replica replication=true",
556 : sk_hosts,
557 : sk_ports,
558 0 : &self.timeline_id.to_string(),
559 0 : &self.tenant_id.to_string(),
560 : );
561 :
562 0 : let slot_name = format!("repl_{}_", self.timeline_id);
563 0 : conf.append("primary_conninfo", connstr.as_str());
564 0 : conf.append("primary_slot_name", slot_name.as_str());
565 0 : conf.append("hot_standby", "on");
566 : // prefetching of blocks referenced in WAL doesn't make sense for us
567 : // Neon hot standby ignores pages that are not in the shared_buffers
568 0 : if self.pg_version >= PgMajorVersion::PG15 {
569 0 : conf.append("recovery_prefetch", "off");
570 0 : }
571 : }
572 : }
573 :
574 0 : Ok(conf)
575 0 : }
576 :
577 0 : pub fn endpoint_path(&self) -> PathBuf {
578 0 : self.env.endpoints_path().join(&self.endpoint_id)
579 0 : }
580 :
581 0 : pub fn pgdata(&self) -> PathBuf {
582 0 : self.endpoint_path().join("pgdata")
583 0 : }
584 :
585 0 : pub fn status(&self) -> EndpointStatus {
586 0 : let timeout = Duration::from_millis(300);
587 0 : let has_pidfile = self.pgdata().join("postmaster.pid").exists();
588 0 : let can_connect = TcpStream::connect_timeout(&self.pg_address, timeout).is_ok();
589 :
590 0 : match (has_pidfile, can_connect) {
591 0 : (true, true) => EndpointStatus::Running,
592 0 : (false, false) => EndpointStatus::Stopped,
593 0 : (true, false) => EndpointStatus::Crashed,
594 0 : (false, true) => EndpointStatus::RunningNoPidfile,
595 : }
596 0 : }
597 :
598 0 : fn pg_ctl(&self, args: &[&str], auth_token: &Option<String>) -> Result<()> {
599 0 : let pg_ctl_path = self.env.pg_bin_dir(self.pg_version)?.join("pg_ctl");
600 0 : let mut cmd = Command::new(&pg_ctl_path);
601 0 : cmd.args(
602 0 : [
603 0 : &[
604 0 : "-D",
605 0 : self.pgdata().to_str().unwrap(),
606 0 : "-w", //wait till pg_ctl actually does what was asked
607 0 : ],
608 0 : args,
609 0 : ]
610 0 : .concat(),
611 0 : )
612 0 : .env_clear()
613 0 : .env(
614 : "LD_LIBRARY_PATH",
615 0 : self.env.pg_lib_dir(self.pg_version)?.to_str().unwrap(),
616 : )
617 0 : .env(
618 : "DYLD_LIBRARY_PATH",
619 0 : self.env.pg_lib_dir(self.pg_version)?.to_str().unwrap(),
620 : );
621 :
622 : // Pass authentication token used for the connections to pageserver and safekeepers
623 0 : if let Some(token) = auth_token {
624 0 : cmd.env("NEON_AUTH_TOKEN", token);
625 0 : }
626 :
627 0 : let pg_ctl = cmd
628 0 : .output()
629 0 : .context(format!("{} failed", pg_ctl_path.display()))?;
630 0 : if !pg_ctl.status.success() {
631 0 : anyhow::bail!(
632 0 : "pg_ctl failed, exit code: {}, stdout: {}, stderr: {}",
633 : pg_ctl.status,
634 0 : String::from_utf8_lossy(&pg_ctl.stdout),
635 0 : String::from_utf8_lossy(&pg_ctl.stderr),
636 : );
637 0 : }
638 :
639 0 : Ok(())
640 0 : }
641 :
642 0 : fn wait_for_compute_ctl_to_exit(&self, send_sigterm: bool) -> Result<()> {
643 : // TODO use background_process::stop_process instead: https://github.com/neondatabase/neon/pull/6482
644 0 : let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
645 0 : let pid: u32 = std::fs::read_to_string(pidfile_path)?.parse()?;
646 0 : let pid = nix::unistd::Pid::from_raw(pid as i32);
647 0 : if send_sigterm {
648 0 : kill(pid, Signal::SIGTERM).ok();
649 0 : }
650 0 : crate::background_process::wait_until_stopped("compute_ctl", pid)?;
651 0 : Ok(())
652 0 : }
653 :
654 0 : fn read_postgresql_conf(&self) -> Result<String> {
655 : // Slurp the endpoints/<endpoint id>/postgresql.conf file into
656 : // memory. We will include it in the spec file that we pass to
657 : // `compute_ctl`, and `compute_ctl` will write it to the postgresql.conf
658 : // in the data directory.
659 0 : let postgresql_conf_path = self.endpoint_path().join("postgresql.conf");
660 0 : match std::fs::read(&postgresql_conf_path) {
661 0 : Ok(content) => Ok(String::from_utf8(content)?),
662 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok("".to_string()),
663 0 : Err(e) => Err(anyhow::Error::new(e).context(format!(
664 0 : "failed to read config file in {}",
665 0 : postgresql_conf_path.to_str().unwrap()
666 0 : ))),
667 : }
668 0 : }
669 :
670 : /// Map safekeepers ids to the actual connection strings.
671 0 : fn build_safekeepers_connstrs(&self, sk_ids: Vec<NodeId>) -> Result<Vec<String>> {
672 0 : let mut safekeeper_connstrings = Vec::new();
673 0 : if self.mode == ComputeMode::Primary {
674 0 : for sk_id in sk_ids {
675 0 : let sk = self
676 0 : .env
677 0 : .safekeepers
678 0 : .iter()
679 0 : .find(|node| node.id == sk_id)
680 0 : .ok_or_else(|| anyhow!("safekeeper {sk_id} does not exist"))?;
681 0 : safekeeper_connstrings.push(format!("127.0.0.1:{}", sk.get_compute_port()));
682 : }
683 0 : }
684 0 : Ok(safekeeper_connstrings)
685 0 : }
686 :
687 : /// Generate a JWT with the correct claims.
688 0 : pub fn generate_jwt(&self, scope: Option<ComputeClaimsScope>) -> Result<String> {
689 0 : self.env.generate_auth_token(&ComputeClaims {
690 0 : audience: match scope {
691 0 : Some(ComputeClaimsScope::Admin) => Some(vec![COMPUTE_AUDIENCE.to_owned()]),
692 0 : _ => None,
693 : },
694 0 : compute_id: match scope {
695 0 : Some(ComputeClaimsScope::Admin) => None,
696 0 : _ => Some(self.endpoint_id.clone()),
697 : },
698 0 : scope,
699 : })
700 0 : }
701 :
702 0 : pub async fn start(&self, args: EndpointStartArgs) -> Result<()> {
703 0 : if self.status() == EndpointStatus::Running {
704 0 : anyhow::bail!("The endpoint is already running");
705 0 : }
706 :
707 0 : let postgresql_conf = self.read_postgresql_conf()?;
708 :
709 : // We always start the compute node from scratch, so if the Postgres
710 : // data dir exists from a previous launch, remove it first.
711 0 : if self.pgdata().exists() {
712 0 : std::fs::remove_dir_all(self.pgdata())?;
713 0 : }
714 :
715 0 : let safekeeper_connstrings = self.build_safekeepers_connstrs(args.safekeepers)?;
716 :
717 : // check for file remote_extensions_spec.json
718 : // if it is present, read it and pass to compute_ctl
719 0 : let remote_extensions_spec_path = self.endpoint_path().join("remote_extensions_spec.json");
720 0 : let remote_extensions_spec = std::fs::File::open(remote_extensions_spec_path);
721 : let remote_extensions: Option<RemoteExtSpec>;
722 :
723 0 : if let Ok(spec_file) = remote_extensions_spec {
724 0 : remote_extensions = serde_json::from_reader(spec_file).ok();
725 0 : } else {
726 0 : remote_extensions = None;
727 0 : };
728 :
729 : // For the sake of backwards-compatibility, also fill in 'pageserver_connstring'
730 : //
731 : // XXX: I believe this is not really needed, except to make
732 : // test_forward_compatibility happy.
733 : //
734 : // Use a closure so that we can conviniently return None in the middle of the
735 : // loop.
736 0 : let pageserver_connstring: Option<String> = (|| {
737 0 : let num_shards = args.pageserver_conninfo.shard_count.count();
738 0 : let mut connstrings = Vec::new();
739 0 : for shard_no in 0..num_shards {
740 0 : let shard_index = ShardIndex {
741 0 : shard_count: args.pageserver_conninfo.shard_count,
742 0 : shard_number: ShardNumber(shard_no),
743 0 : };
744 0 : let shard = args
745 0 : .pageserver_conninfo
746 0 : .shards
747 0 : .get(&shard_index)
748 0 : .ok_or_else(|| {
749 0 : anyhow!(
750 0 : "shard {} not found in pageserver_connection_info",
751 : shard_index
752 : )
753 0 : })?;
754 0 : let pageserver = shard
755 0 : .pageservers
756 0 : .first()
757 0 : .ok_or(anyhow!("must have at least one pageserver"))?;
758 0 : if let Some(libpq_url) = &pageserver.libpq_url {
759 0 : connstrings.push(libpq_url.clone());
760 0 : } else {
761 0 : return Ok::<_, anyhow::Error>(None);
762 : }
763 : }
764 0 : Ok(Some(connstrings.join(",")))
765 0 : })()?;
766 :
767 : // Create config file
768 0 : let config = {
769 0 : let mut spec = ComputeSpec {
770 0 : skip_pg_catalog_updates: self.skip_pg_catalog_updates,
771 : format_version: 1.0,
772 0 : operation_uuid: None,
773 0 : features: self.features.clone(),
774 0 : swap_size_bytes: None,
775 0 : disk_quota_bytes: None,
776 0 : disable_lfc_resizing: None,
777 : cluster: Cluster {
778 0 : cluster_id: None, // project ID: not used
779 0 : name: None, // project name: not used
780 0 : state: None,
781 0 : roles: if args.create_test_user {
782 0 : vec![Role {
783 0 : name: PgIdent::from_str("test").unwrap(),
784 0 : encrypted_password: None,
785 0 : options: None,
786 0 : }]
787 : } else {
788 0 : Vec::new()
789 : },
790 0 : databases: if args.create_test_user {
791 0 : vec![Database {
792 0 : name: PgIdent::from_str("neondb").unwrap(),
793 0 : owner: PgIdent::from_str("test").unwrap(),
794 0 : options: None,
795 0 : restrict_conn: false,
796 0 : invalid: false,
797 0 : }]
798 : } else {
799 0 : Vec::new()
800 : },
801 0 : settings: None,
802 0 : postgresql_conf: Some(postgresql_conf.clone()),
803 : },
804 0 : delta_operations: None,
805 0 : tenant_id: Some(self.tenant_id),
806 0 : timeline_id: Some(self.timeline_id),
807 0 : project_id: None,
808 0 : branch_id: None,
809 0 : endpoint_id: Some(self.endpoint_id.clone()),
810 0 : mode: self.mode,
811 0 : pageserver_connection_info: Some(args.pageserver_conninfo.clone()),
812 0 : pageserver_connstring,
813 0 : safekeepers_generation: args.safekeepers_generation.map(|g| g.into_inner()),
814 0 : safekeeper_connstrings,
815 0 : storage_auth_token: args.auth_token.clone(),
816 0 : remote_extensions,
817 0 : pgbouncer_settings: None,
818 0 : shard_stripe_size: args.pageserver_conninfo.stripe_size, // redundant with pageserver_connection_info.stripe_size
819 0 : local_proxy_config: None,
820 0 : reconfigure_concurrency: self.reconfigure_concurrency,
821 0 : drop_subscriptions_before_start: self.drop_subscriptions_before_start,
822 0 : audit_log_level: ComputeAudit::Disabled,
823 0 : logs_export_host: None::<String>,
824 0 : endpoint_storage_addr: Some(args.endpoint_storage_addr),
825 0 : endpoint_storage_token: Some(args.endpoint_storage_token),
826 0 : autoprewarm: args.autoprewarm,
827 0 : offload_lfc_interval_seconds: args.offload_lfc_interval_seconds,
828 : suspend_timeout_seconds: -1, // Only used in neon_local.
829 0 : databricks_settings: None,
830 : };
831 :
832 : // this strange code is needed to support respec() in tests
833 0 : if self.cluster.is_some() {
834 0 : debug!("Cluster is already set in the endpoint spec, using it");
835 0 : spec.cluster = self.cluster.clone().unwrap();
836 :
837 0 : debug!("spec.cluster {:?}", spec.cluster);
838 :
839 : // fill missing fields again
840 0 : if args.create_test_user {
841 0 : spec.cluster.roles.push(Role {
842 0 : name: PgIdent::from_str("test").unwrap(),
843 0 : encrypted_password: None,
844 0 : options: None,
845 0 : });
846 0 : spec.cluster.databases.push(Database {
847 0 : name: PgIdent::from_str("neondb").unwrap(),
848 0 : owner: PgIdent::from_str("test").unwrap(),
849 0 : options: None,
850 0 : restrict_conn: false,
851 0 : invalid: false,
852 0 : });
853 0 : }
854 0 : spec.cluster.postgresql_conf = Some(postgresql_conf);
855 0 : }
856 :
857 0 : ComputeConfig {
858 0 : spec: Some(spec),
859 0 : compute_ctl_config: self.compute_ctl_config.clone(),
860 0 : }
861 : };
862 :
863 0 : let config_path = self.endpoint_path().join("config.json");
864 0 : std::fs::write(config_path, serde_json::to_string_pretty(&config)?)?;
865 :
866 : // Open log file. We'll redirect the stdout and stderr of `compute_ctl` to it.
867 0 : let logfile = std::fs::OpenOptions::new()
868 0 : .create(true)
869 0 : .append(true)
870 0 : .open(self.endpoint_path().join("compute.log"))?;
871 :
872 : // Launch compute_ctl
873 0 : let conn_str = self.connstr("cloud_admin", "postgres");
874 0 : println!("Starting postgres node at '{conn_str}'");
875 0 : if args.create_test_user {
876 0 : let conn_str = self.connstr("test", "neondb");
877 0 : println!("Also at '{conn_str}'");
878 0 : }
879 0 : let mut cmd = Command::new(self.env.neon_distrib_dir.join("compute_ctl"));
880 0 : cmd.args([
881 0 : "--external-http-port",
882 0 : &self.external_http_address.port().to_string(),
883 0 : ])
884 0 : .args([
885 0 : "--internal-http-port",
886 0 : &self.internal_http_address.port().to_string(),
887 0 : ])
888 0 : .args(["--pgdata", self.pgdata().to_str().unwrap()])
889 0 : .args(["--connstr", &conn_str])
890 0 : .arg("--config")
891 0 : .arg(self.endpoint_path().join("config.json").as_os_str())
892 0 : .args([
893 : "--pgbin",
894 0 : self.env
895 0 : .pg_bin_dir(self.pg_version)?
896 0 : .join("postgres")
897 0 : .to_str()
898 0 : .unwrap(),
899 : ])
900 : // TODO: It would be nice if we generated compute IDs with the same
901 : // algorithm as the real control plane.
902 0 : .args(["--compute-id", &self.endpoint_id])
903 0 : .stdin(std::process::Stdio::null())
904 0 : .stderr(logfile.try_clone()?)
905 0 : .stdout(logfile);
906 :
907 0 : if let Some(remote_ext_base_url) = args.remote_ext_base_url {
908 0 : cmd.args(["--remote-ext-base-url", &remote_ext_base_url]);
909 0 : }
910 :
911 0 : if args.dev {
912 0 : cmd.arg("--dev");
913 0 : }
914 :
915 0 : if let Some(privileged_role_name) = self.privileged_role_name.clone() {
916 0 : cmd.args(["--privileged-role-name", &privileged_role_name]);
917 0 : }
918 :
919 0 : let child = cmd.spawn()?;
920 : // set up a scopeguard to kill & wait for the child in case we panic or bail below
921 0 : let child = scopeguard::guard(child, |mut child| {
922 0 : println!("SIGKILL & wait the started process");
923 0 : (|| {
924 : // TODO: use another signal that can be caught by the child so it can clean up any children it spawned
925 0 : child.kill().context("SIGKILL child")?;
926 0 : child.wait().context("wait() for child process")?;
927 0 : anyhow::Ok(())
928 : })()
929 0 : .with_context(|| format!("scopeguard kill&wait child {child:?}"))
930 0 : .unwrap();
931 0 : });
932 :
933 : // Write down the pid so we can wait for it when we want to stop
934 : // TODO use background_process::start_process instead: https://github.com/neondatabase/neon/pull/6482
935 0 : let pid = child.id();
936 0 : let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
937 0 : std::fs::write(pidfile_path, pid.to_string())?;
938 :
939 : // Wait for it to start
940 : const ATTEMPT_INTERVAL: Duration = Duration::from_millis(100);
941 0 : let start_at = Instant::now();
942 : loop {
943 0 : match self.get_status().await {
944 0 : Ok(state) => {
945 0 : match state.status {
946 : ComputeStatus::Init => {
947 0 : let timeout = args.start_timeout;
948 0 : if Instant::now().duration_since(start_at) > timeout {
949 0 : bail!(
950 0 : "compute startup timed out {:?}; still in Init state",
951 : timeout
952 : );
953 0 : }
954 : // keep retrying
955 : }
956 : ComputeStatus::Running => {
957 : // All good!
958 0 : break;
959 : }
960 : ComputeStatus::Failed => {
961 0 : bail!(
962 0 : "compute startup failed: {}",
963 0 : state
964 0 : .error
965 0 : .as_deref()
966 0 : .unwrap_or("<no error from compute_ctl>")
967 : );
968 : }
969 : ComputeStatus::Empty
970 : | ComputeStatus::ConfigurationPending
971 : | ComputeStatus::Configuration
972 : | ComputeStatus::TerminationPendingFast
973 : | ComputeStatus::TerminationPendingImmediate
974 : | ComputeStatus::Terminated
975 : | ComputeStatus::RefreshConfigurationPending
976 : | ComputeStatus::RefreshConfiguration => {
977 0 : bail!("unexpected compute status: {:?}", state.status)
978 : }
979 : }
980 : }
981 0 : Err(e) => {
982 0 : if Instant::now().duration_since(start_at) > args.start_timeout {
983 0 : return Err(e).context(format!(
984 0 : "timed out {:?} waiting to connect to compute_ctl HTTP",
985 : args.start_timeout
986 : ));
987 0 : }
988 : }
989 : }
990 0 : tokio::time::sleep(ATTEMPT_INTERVAL).await;
991 : }
992 :
993 : // disarm the scopeguard, let the child outlive this function (and neon_local invoction)
994 0 : drop(scopeguard::ScopeGuard::into_inner(child));
995 :
996 0 : Ok(())
997 0 : }
998 :
999 : // Update the pageservers in the spec file of the endpoint. This is useful to test the spec refresh scenario.
1000 0 : pub async fn update_pageservers_in_config(
1001 0 : &self,
1002 0 : pageserver_conninfo: &PageserverConnectionInfo,
1003 0 : ) -> Result<()> {
1004 0 : let config_path = self.endpoint_path().join("config.json");
1005 0 : let mut config: ComputeConfig = {
1006 0 : let file = std::fs::File::open(&config_path)?;
1007 0 : serde_json::from_reader(file)?
1008 : };
1009 :
1010 0 : let mut spec = config.spec.unwrap();
1011 0 : spec.pageserver_connection_info = Some(pageserver_conninfo.clone());
1012 0 : config.spec = Some(spec);
1013 :
1014 0 : let file = std::fs::File::create(&config_path)?;
1015 0 : serde_json::to_writer_pretty(file, &config)?;
1016 :
1017 0 : Ok(())
1018 0 : }
1019 :
1020 : // Call the /status HTTP API
1021 0 : pub async fn get_status(&self) -> Result<ComputeStatusResponse> {
1022 0 : let client = reqwest::Client::new();
1023 :
1024 0 : let response = client
1025 0 : .request(
1026 0 : reqwest::Method::GET,
1027 0 : format!(
1028 0 : "http://{}:{}/status",
1029 0 : self.external_http_address.ip(),
1030 0 : self.external_http_address.port()
1031 : ),
1032 : )
1033 0 : .bearer_auth(self.generate_jwt(None::<ComputeClaimsScope>)?)
1034 0 : .send()
1035 0 : .await?;
1036 :
1037 : // Interpret the response
1038 0 : let status = response.status();
1039 0 : if !(status.is_client_error() || status.is_server_error()) {
1040 0 : Ok(response.json().await?)
1041 : } else {
1042 : // reqwest does not export its error construction utility functions, so let's craft the message ourselves
1043 0 : let url = response.url().to_owned();
1044 0 : let msg = match response.text().await {
1045 0 : Ok(err_body) => format!("Error: {err_body}"),
1046 0 : Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
1047 : };
1048 0 : Err(anyhow::anyhow!(msg))
1049 : }
1050 0 : }
1051 :
1052 0 : pub async fn reconfigure(
1053 0 : &self,
1054 0 : pageserver_conninfo: Option<&PageserverConnectionInfo>,
1055 0 : safekeepers: Option<Vec<NodeId>>,
1056 0 : safekeeper_generation: Option<SafekeeperGeneration>,
1057 0 : ) -> Result<()> {
1058 0 : let (mut spec, compute_ctl_config) = {
1059 0 : let config_path = self.endpoint_path().join("config.json");
1060 0 : let file = std::fs::File::open(config_path)?;
1061 0 : let config: ComputeConfig = serde_json::from_reader(file)?;
1062 :
1063 0 : (config.spec.unwrap(), config.compute_ctl_config)
1064 : };
1065 :
1066 0 : let postgresql_conf = self.read_postgresql_conf()?;
1067 0 : spec.cluster.postgresql_conf = Some(postgresql_conf);
1068 :
1069 0 : if let Some(pageserver_conninfo) = pageserver_conninfo {
1070 : // If pageservers are provided, we need to ensure that they are not empty.
1071 : // This is a requirement for the compute_ctl configuration.
1072 0 : anyhow::ensure!(
1073 0 : !pageserver_conninfo.shards.is_empty(),
1074 0 : "no pageservers provided"
1075 : );
1076 0 : spec.pageserver_connection_info = Some(pageserver_conninfo.clone());
1077 0 : spec.shard_stripe_size = pageserver_conninfo.stripe_size;
1078 0 : }
1079 :
1080 : // If safekeepers are not specified, don't change them.
1081 0 : if let Some(safekeepers) = safekeepers {
1082 0 : let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
1083 0 : spec.safekeeper_connstrings = safekeeper_connstrings;
1084 0 : if let Some(g) = safekeeper_generation {
1085 0 : spec.safekeepers_generation = Some(g.into_inner());
1086 0 : }
1087 0 : }
1088 :
1089 0 : let client = reqwest::Client::builder()
1090 0 : .timeout(Duration::from_secs(120))
1091 0 : .build()
1092 0 : .unwrap();
1093 0 : let response = client
1094 0 : .post(format!(
1095 0 : "http://{}:{}/configure",
1096 0 : self.external_http_address.ip(),
1097 0 : self.external_http_address.port()
1098 : ))
1099 0 : .header(CONTENT_TYPE.as_str(), "application/json")
1100 0 : .bearer_auth(self.generate_jwt(None::<ComputeClaimsScope>)?)
1101 0 : .body(
1102 0 : serde_json::to_string(&ConfigurationRequest {
1103 0 : spec,
1104 0 : compute_ctl_config,
1105 0 : })
1106 0 : .unwrap(),
1107 : )
1108 0 : .send()
1109 0 : .await?;
1110 :
1111 0 : let status = response.status();
1112 0 : if !(status.is_client_error() || status.is_server_error()) {
1113 0 : Ok(())
1114 : } else {
1115 0 : let url = response.url().to_owned();
1116 0 : let msg = match response.text().await {
1117 0 : Ok(err_body) => format!("Error: {err_body}"),
1118 0 : Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
1119 : };
1120 0 : Err(anyhow::anyhow!(msg))
1121 : }
1122 0 : }
1123 :
1124 0 : pub async fn reconfigure_pageservers(
1125 0 : &self,
1126 0 : pageservers: &PageserverConnectionInfo,
1127 0 : ) -> Result<()> {
1128 0 : self.reconfigure(Some(pageservers), None, None).await
1129 0 : }
1130 :
1131 0 : pub async fn reconfigure_safekeepers(
1132 0 : &self,
1133 0 : safekeepers: Vec<NodeId>,
1134 0 : generation: SafekeeperGeneration,
1135 0 : ) -> Result<()> {
1136 0 : self.reconfigure(None, Some(safekeepers), Some(generation))
1137 0 : .await
1138 0 : }
1139 :
1140 0 : pub async fn stop(
1141 0 : &self,
1142 0 : mode: EndpointTerminateMode,
1143 0 : destroy: bool,
1144 0 : ) -> Result<TerminateResponse> {
1145 : // pg_ctl stop is fast but doesn't allow us to collect LSN. /terminate is
1146 : // slow, and test runs time out. Solution: special mode "immediate-terminate"
1147 : // which uses /terminate
1148 0 : let response = if let EndpointTerminateMode::ImmediateTerminate = mode {
1149 0 : let ip = self.external_http_address.ip();
1150 0 : let port = self.external_http_address.port();
1151 0 : let url = format!("http://{ip}:{port}/terminate?mode=immediate");
1152 0 : let token = self.generate_jwt(Some(ComputeClaimsScope::Admin))?;
1153 0 : let request = reqwest::Client::new().post(url).bearer_auth(token);
1154 0 : let response = request.send().await.context("/terminate")?;
1155 0 : let text = response.text().await.context("/terminate result")?;
1156 0 : serde_json::from_str(&text).with_context(|| format!("deserializing {text}"))?
1157 : } else {
1158 0 : self.pg_ctl(&["-m", &mode.to_string(), "stop"], &None)?;
1159 0 : TerminateResponse { lsn: None }
1160 : };
1161 :
1162 : // Also wait for the compute_ctl process to die. It might have some
1163 : // cleanup work to do after postgres stops, like syncing safekeepers,
1164 : // etc.
1165 : //
1166 : // If destroying or stop mode is immediate, send it SIGTERM before
1167 : // waiting. Sometimes we do *not* want this cleanup: tests intentionally
1168 : // do stop when majority of safekeepers is down, so sync-safekeepers
1169 : // would hang otherwise. This could be a separate flag though.
1170 0 : let send_sigterm = destroy || !matches!(mode, EndpointTerminateMode::Fast);
1171 0 : self.wait_for_compute_ctl_to_exit(send_sigterm)?;
1172 0 : if destroy {
1173 0 : println!(
1174 0 : "Destroying postgres data directory '{}'",
1175 0 : self.pgdata().to_str().unwrap()
1176 : );
1177 0 : std::fs::remove_dir_all(self.endpoint_path())?;
1178 0 : }
1179 0 : Ok(response)
1180 0 : }
1181 :
1182 0 : pub async fn refresh_configuration(&self) -> Result<()> {
1183 0 : let client = reqwest::Client::builder()
1184 0 : .timeout(Duration::from_secs(30))
1185 0 : .build()
1186 0 : .unwrap();
1187 0 : let response = client
1188 0 : .post(format!(
1189 0 : "http://{}:{}/refresh_configuration",
1190 0 : self.internal_http_address.ip(),
1191 0 : self.internal_http_address.port()
1192 0 : ))
1193 0 : .send()
1194 0 : .await?;
1195 :
1196 0 : let status = response.status();
1197 0 : if !(status.is_client_error() || status.is_server_error()) {
1198 0 : Ok(())
1199 : } else {
1200 0 : let url = response.url().to_owned();
1201 0 : let msg = match response.text().await {
1202 0 : Ok(err_body) => format!("Error: {err_body}"),
1203 0 : Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
1204 : };
1205 0 : Err(anyhow::anyhow!(msg))
1206 : }
1207 0 : }
1208 :
1209 0 : pub fn connstr(&self, user: &str, db_name: &str) -> String {
1210 0 : format!(
1211 0 : "postgresql://{}@{}:{}/{}",
1212 : user,
1213 0 : self.pg_address.ip(),
1214 0 : self.pg_address.port(),
1215 : db_name
1216 : )
1217 0 : }
1218 : }
1219 :
1220 : /// If caller is telling us what pageserver to use, this is not a tenant which is
1221 : /// fully managed by storage controller, therefore not sharded.
1222 0 : pub fn local_pageserver_conf_to_conn_info(
1223 0 : conf: &crate::local_env::PageServerConf,
1224 0 : ) -> Result<PageserverConnectionInfo> {
1225 0 : let libpq_url = {
1226 0 : let (host, port) = parse_host_port(&conf.listen_pg_addr)?;
1227 0 : let port = port.unwrap_or(5432);
1228 0 : Some(format!("postgres://no_user@{host}:{port}"))
1229 : };
1230 0 : let grpc_url = if let Some(grpc_addr) = &conf.listen_grpc_addr {
1231 0 : let (host, port) = parse_host_port(grpc_addr)?;
1232 0 : let port = port.unwrap_or(DEFAULT_PAGESERVER_GRPC_PORT);
1233 0 : Some(format!("grpc://no_user@{host}:{port}"))
1234 : } else {
1235 0 : None
1236 : };
1237 0 : let ps_conninfo = PageserverShardConnectionInfo {
1238 0 : id: Some(conf.id),
1239 0 : libpq_url,
1240 0 : grpc_url,
1241 0 : };
1242 :
1243 0 : let shard_info = PageserverShardInfo {
1244 0 : pageservers: vec![ps_conninfo],
1245 0 : };
1246 :
1247 0 : let shards: HashMap<_, _> = vec![(ShardIndex::unsharded(), shard_info)]
1248 0 : .into_iter()
1249 0 : .collect();
1250 0 : Ok(PageserverConnectionInfo {
1251 0 : shard_count: ShardCount::unsharded(),
1252 0 : stripe_size: None,
1253 0 : shards,
1254 0 : prefer_protocol: PageserverProtocol::default(),
1255 0 : })
1256 0 : }
1257 :
1258 0 : pub fn tenant_locate_response_to_conn_info(
1259 0 : response: &pageserver_api::controller_api::TenantLocateResponse,
1260 0 : ) -> Result<PageserverConnectionInfo> {
1261 0 : let mut shards = HashMap::new();
1262 0 : for shard in response.shards.iter() {
1263 0 : tracing::info!("parsing {}", shard.listen_pg_addr);
1264 0 : let libpq_url = {
1265 0 : let host = &shard.listen_pg_addr;
1266 0 : let port = shard.listen_pg_port;
1267 0 : Some(format!("postgres://no_user@{host}:{port}"))
1268 : };
1269 0 : let grpc_url = if let Some(grpc_addr) = &shard.listen_grpc_addr {
1270 0 : let host = grpc_addr;
1271 0 : let port = shard.listen_grpc_port.expect("no gRPC port");
1272 0 : Some(format!("grpc://no_user@{host}:{port}"))
1273 : } else {
1274 0 : None
1275 : };
1276 :
1277 0 : let shard_info = PageserverShardInfo {
1278 0 : pageservers: vec![PageserverShardConnectionInfo {
1279 0 : id: Some(shard.node_id),
1280 0 : libpq_url,
1281 0 : grpc_url,
1282 0 : }],
1283 0 : };
1284 :
1285 0 : shards.insert(shard.shard_id.to_index(), shard_info);
1286 : }
1287 :
1288 0 : let stripe_size = if response.shard_params.count.is_unsharded() {
1289 0 : None
1290 : } else {
1291 0 : Some(response.shard_params.stripe_size)
1292 : };
1293 0 : Ok(PageserverConnectionInfo {
1294 0 : shard_count: response.shard_params.count,
1295 0 : stripe_size,
1296 0 : shards,
1297 0 : prefer_protocol: PageserverProtocol::default(),
1298 0 : })
1299 0 : }
|
