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 : //! spec.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;
41 : use std::net::Ipv4Addr;
42 : use std::net::SocketAddr;
43 : use std::net::TcpStream;
44 : use std::path::PathBuf;
45 : use std::process::Command;
46 : use std::str::FromStr;
47 : use std::sync::Arc;
48 : use std::time::Duration;
49 :
50 : use anyhow::{anyhow, bail, Context, Result};
51 : use compute_api::spec::Database;
52 : use compute_api::spec::PgIdent;
53 : use compute_api::spec::RemoteExtSpec;
54 : use compute_api::spec::Role;
55 : use nix::sys::signal::kill;
56 : use nix::sys::signal::Signal;
57 : use pageserver_api::shard::ShardStripeSize;
58 : use reqwest::header::CONTENT_TYPE;
59 : use serde::{Deserialize, Serialize};
60 : use url::Host;
61 : use utils::id::{NodeId, TenantId, TimelineId};
62 :
63 : use crate::local_env::LocalEnv;
64 : use crate::postgresql_conf::PostgresConf;
65 : use crate::storage_controller::StorageController;
66 :
67 : use compute_api::responses::{ComputeStatus, ComputeStatusResponse};
68 : use compute_api::spec::{Cluster, ComputeFeature, ComputeMode, ComputeSpec};
69 :
70 : // contents of a endpoint.json file
71 0 : #[derive(Serialize, Deserialize, PartialEq, Eq, Clone, Debug)]
72 : pub struct EndpointConf {
73 : endpoint_id: String,
74 : tenant_id: TenantId,
75 : timeline_id: TimelineId,
76 : mode: ComputeMode,
77 : pg_port: u16,
78 : external_http_port: u16,
79 : internal_http_port: u16,
80 : pg_version: u32,
81 : skip_pg_catalog_updates: bool,
82 : drop_subscriptions_before_start: bool,
83 : features: Vec<ComputeFeature>,
84 : }
85 :
86 : //
87 : // ComputeControlPlane
88 : //
89 : pub struct ComputeControlPlane {
90 : base_port: u16,
91 :
92 : // endpoint ID is the key
93 : pub endpoints: BTreeMap<String, Arc<Endpoint>>,
94 :
95 : env: LocalEnv,
96 : }
97 :
98 : impl ComputeControlPlane {
99 : // Load current endpoints from the endpoints/ subdirectories
100 0 : pub fn load(env: LocalEnv) -> Result<ComputeControlPlane> {
101 0 : let mut endpoints = BTreeMap::default();
102 0 : for endpoint_dir in std::fs::read_dir(env.endpoints_path())
103 0 : .with_context(|| format!("failed to list {}", env.endpoints_path().display()))?
104 : {
105 0 : let ep_res = Endpoint::from_dir_entry(endpoint_dir?, &env);
106 0 : let ep = match ep_res {
107 0 : Ok(ep) => ep,
108 0 : Err(e) => match e.downcast::<std::io::Error>() {
109 0 : Ok(e) => {
110 0 : // A parallel task could delete an endpoint while we have just scanned the directory
111 0 : if e.kind() == std::io::ErrorKind::NotFound {
112 0 : continue;
113 : } else {
114 0 : Err(e)?
115 : }
116 : }
117 0 : Err(e) => Err(e)?,
118 : },
119 : };
120 0 : endpoints.insert(ep.endpoint_id.clone(), Arc::new(ep));
121 : }
122 :
123 0 : Ok(ComputeControlPlane {
124 0 : base_port: 55431,
125 0 : endpoints,
126 0 : env,
127 0 : })
128 0 : }
129 :
130 0 : fn get_port(&mut self) -> u16 {
131 0 : 1 + self
132 0 : .endpoints
133 0 : .values()
134 0 : .map(|ep| std::cmp::max(ep.pg_address.port(), ep.external_http_address.port()))
135 0 : .max()
136 0 : .unwrap_or(self.base_port)
137 0 : }
138 :
139 : #[allow(clippy::too_many_arguments)]
140 0 : pub fn new_endpoint(
141 0 : &mut self,
142 0 : endpoint_id: &str,
143 0 : tenant_id: TenantId,
144 0 : timeline_id: TimelineId,
145 0 : pg_port: Option<u16>,
146 0 : external_http_port: Option<u16>,
147 0 : internal_http_port: Option<u16>,
148 0 : pg_version: u32,
149 0 : mode: ComputeMode,
150 0 : skip_pg_catalog_updates: bool,
151 0 : drop_subscriptions_before_start: bool,
152 0 : ) -> Result<Arc<Endpoint>> {
153 0 : let pg_port = pg_port.unwrap_or_else(|| self.get_port());
154 0 : let external_http_port = external_http_port.unwrap_or_else(|| self.get_port() + 1);
155 0 : let internal_http_port = internal_http_port.unwrap_or_else(|| external_http_port + 1);
156 0 : let ep = Arc::new(Endpoint {
157 0 : endpoint_id: endpoint_id.to_owned(),
158 0 : pg_address: SocketAddr::new(IpAddr::from(Ipv4Addr::LOCALHOST), pg_port),
159 0 : external_http_address: SocketAddr::new(
160 0 : IpAddr::from(Ipv4Addr::UNSPECIFIED),
161 0 : external_http_port,
162 0 : ),
163 0 : internal_http_address: SocketAddr::new(
164 0 : IpAddr::from(Ipv4Addr::LOCALHOST),
165 0 : internal_http_port,
166 0 : ),
167 0 : env: self.env.clone(),
168 0 : timeline_id,
169 0 : mode,
170 0 : tenant_id,
171 0 : pg_version,
172 0 : // We don't setup roles and databases in the spec locally, so we don't need to
173 0 : // do catalog updates. Catalog updates also include check availability
174 0 : // data creation. Yet, we have tests that check that size and db dump
175 0 : // before and after start are the same. So, skip catalog updates,
176 0 : // with this we basically test a case of waking up an idle compute, where
177 0 : // we also skip catalog updates in the cloud.
178 0 : skip_pg_catalog_updates,
179 0 : drop_subscriptions_before_start,
180 0 : features: vec![],
181 0 : });
182 0 :
183 0 : ep.create_endpoint_dir()?;
184 : std::fs::write(
185 0 : ep.endpoint_path().join("endpoint.json"),
186 0 : serde_json::to_string_pretty(&EndpointConf {
187 0 : endpoint_id: endpoint_id.to_string(),
188 0 : tenant_id,
189 0 : timeline_id,
190 0 : mode,
191 0 : external_http_port,
192 0 : internal_http_port,
193 0 : pg_port,
194 0 : pg_version,
195 0 : skip_pg_catalog_updates,
196 0 : drop_subscriptions_before_start,
197 0 : features: vec![],
198 0 : })?,
199 0 : )?;
200 : std::fs::write(
201 0 : ep.endpoint_path().join("postgresql.conf"),
202 0 : ep.setup_pg_conf()?.to_string(),
203 0 : )?;
204 :
205 0 : self.endpoints
206 0 : .insert(ep.endpoint_id.clone(), Arc::clone(&ep));
207 0 :
208 0 : Ok(ep)
209 0 : }
210 :
211 0 : pub fn check_conflicting_endpoints(
212 0 : &self,
213 0 : mode: ComputeMode,
214 0 : tenant_id: TenantId,
215 0 : timeline_id: TimelineId,
216 0 : ) -> Result<()> {
217 0 : if matches!(mode, ComputeMode::Primary) {
218 : // this check is not complete, as you could have a concurrent attempt at
219 : // creating another primary, both reading the state before checking it here,
220 : // but it's better than nothing.
221 0 : let mut duplicates = self.endpoints.iter().filter(|(_k, v)| {
222 0 : v.tenant_id == tenant_id
223 0 : && v.timeline_id == timeline_id
224 0 : && v.mode == mode
225 0 : && v.status() != EndpointStatus::Stopped
226 0 : });
227 :
228 0 : if let Some((key, _)) = duplicates.next() {
229 0 : bail!("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.");
230 0 : }
231 0 : }
232 0 : Ok(())
233 0 : }
234 : }
235 :
236 : ///////////////////////////////////////////////////////////////////////////////
237 :
238 : #[derive(Debug)]
239 : pub struct Endpoint {
240 : /// used as the directory name
241 : endpoint_id: String,
242 : pub tenant_id: TenantId,
243 : pub timeline_id: TimelineId,
244 : pub mode: ComputeMode,
245 :
246 : // port and address of the Postgres server and `compute_ctl`'s HTTP APIs
247 : pub pg_address: SocketAddr,
248 : pub external_http_address: SocketAddr,
249 : pub internal_http_address: SocketAddr,
250 :
251 : // postgres major version in the format: 14, 15, etc.
252 : pg_version: u32,
253 :
254 : // These are not part of the endpoint as such, but the environment
255 : // the endpoint runs in.
256 : pub env: LocalEnv,
257 :
258 : // Optimizations
259 : skip_pg_catalog_updates: bool,
260 :
261 : drop_subscriptions_before_start: bool,
262 : // Feature flags
263 : features: Vec<ComputeFeature>,
264 : }
265 :
266 : #[derive(PartialEq, Eq)]
267 : pub enum EndpointStatus {
268 : Running,
269 : Stopped,
270 : Crashed,
271 : RunningNoPidfile,
272 : }
273 :
274 : impl std::fmt::Display for EndpointStatus {
275 0 : fn fmt(&self, writer: &mut std::fmt::Formatter) -> std::fmt::Result {
276 0 : let s = match self {
277 0 : Self::Running => "running",
278 0 : Self::Stopped => "stopped",
279 0 : Self::Crashed => "crashed",
280 0 : Self::RunningNoPidfile => "running, no pidfile",
281 : };
282 0 : write!(writer, "{}", s)
283 0 : }
284 : }
285 :
286 : impl Endpoint {
287 0 : fn from_dir_entry(entry: std::fs::DirEntry, env: &LocalEnv) -> Result<Endpoint> {
288 0 : if !entry.file_type()?.is_dir() {
289 0 : anyhow::bail!(
290 0 : "Endpoint::from_dir_entry failed: '{}' is not a directory",
291 0 : entry.path().display()
292 0 : );
293 0 : }
294 0 :
295 0 : // parse data directory name
296 0 : let fname = entry.file_name();
297 0 : let endpoint_id = fname.to_str().unwrap().to_string();
298 :
299 : // Read the endpoint.json file
300 0 : let conf: EndpointConf =
301 0 : serde_json::from_slice(&std::fs::read(entry.path().join("endpoint.json"))?)?;
302 :
303 0 : Ok(Endpoint {
304 0 : pg_address: SocketAddr::new(IpAddr::from(Ipv4Addr::LOCALHOST), conf.pg_port),
305 0 : external_http_address: SocketAddr::new(
306 0 : IpAddr::from(Ipv4Addr::UNSPECIFIED),
307 0 : conf.external_http_port,
308 0 : ),
309 0 : internal_http_address: SocketAddr::new(
310 0 : IpAddr::from(Ipv4Addr::LOCALHOST),
311 0 : conf.internal_http_port,
312 0 : ),
313 0 : endpoint_id,
314 0 : env: env.clone(),
315 0 : timeline_id: conf.timeline_id,
316 0 : mode: conf.mode,
317 0 : tenant_id: conf.tenant_id,
318 0 : pg_version: conf.pg_version,
319 0 : skip_pg_catalog_updates: conf.skip_pg_catalog_updates,
320 0 : drop_subscriptions_before_start: conf.drop_subscriptions_before_start,
321 0 : features: conf.features,
322 0 : })
323 0 : }
324 :
325 0 : fn create_endpoint_dir(&self) -> Result<()> {
326 0 : std::fs::create_dir_all(self.endpoint_path()).with_context(|| {
327 0 : format!(
328 0 : "could not create endpoint directory {}",
329 0 : self.endpoint_path().display()
330 0 : )
331 0 : })
332 0 : }
333 :
334 : // Generate postgresql.conf with default configuration
335 0 : fn setup_pg_conf(&self) -> Result<PostgresConf> {
336 0 : let mut conf = PostgresConf::new();
337 0 : conf.append("max_wal_senders", "10");
338 0 : conf.append("wal_log_hints", "off");
339 0 : conf.append("max_replication_slots", "10");
340 0 : conf.append("hot_standby", "on");
341 0 : // Set to 1MB to both exercise getPage requests/LFC, and still have enough room for
342 0 : // Postgres to operate. Everything smaller might be not enough for Postgres under load,
343 0 : // and can cause errors like 'no unpinned buffers available', see
344 0 : // <https://github.com/neondatabase/neon/issues/9956>
345 0 : conf.append("shared_buffers", "1MB");
346 0 : // Postgres defaults to effective_io_concurrency=1, which does not exercise the pageserver's
347 0 : // batching logic. Set this to 2 so that we exercise the code a bit without letting
348 0 : // individual tests do a lot of concurrent work on underpowered test machines
349 0 : conf.append("effective_io_concurrency", "2");
350 0 : conf.append("fsync", "off");
351 0 : conf.append("max_connections", "100");
352 0 : conf.append("wal_level", "logical");
353 0 : // wal_sender_timeout is the maximum time to wait for WAL replication.
354 0 : // It also defines how often the walreciever will send a feedback message to the wal sender.
355 0 : conf.append("wal_sender_timeout", "5s");
356 0 : conf.append("listen_addresses", &self.pg_address.ip().to_string());
357 0 : conf.append("port", &self.pg_address.port().to_string());
358 0 : conf.append("wal_keep_size", "0");
359 0 : // walproposer panics when basebackup is invalid, it is pointless to restart in this case.
360 0 : conf.append("restart_after_crash", "off");
361 0 :
362 0 : // Load the 'neon' extension
363 0 : conf.append("shared_preload_libraries", "neon");
364 0 :
365 0 : conf.append_line("");
366 0 : // Replication-related configurations, such as WAL sending
367 0 : match &self.mode {
368 : ComputeMode::Primary => {
369 : // Configure backpressure
370 : // - Replication write lag depends on how fast the walreceiver can process incoming WAL.
371 : // This lag determines latency of get_page_at_lsn. Speed of applying WAL is about 10MB/sec,
372 : // so to avoid expiration of 1 minute timeout, this lag should not be larger than 600MB.
373 : // Actually latency should be much smaller (better if < 1sec). But we assume that recently
374 : // updates pages are not requested from pageserver.
375 : // - Replication flush lag depends on speed of persisting data by checkpointer (creation of
376 : // delta/image layers) and advancing disk_consistent_lsn. Safekeepers are able to
377 : // remove/archive WAL only beyond disk_consistent_lsn. Too large a lag can cause long
378 : // recovery time (in case of pageserver crash) and disk space overflow at safekeepers.
379 : // - Replication apply lag depends on speed of uploading changes to S3 by uploader thread.
380 : // To be able to restore database in case of pageserver node crash, safekeeper should not
381 : // remove WAL beyond this point. Too large lag can cause space exhaustion in safekeepers
382 : // (if they are not able to upload WAL to S3).
383 0 : conf.append("max_replication_write_lag", "15MB");
384 0 : conf.append("max_replication_flush_lag", "10GB");
385 0 :
386 0 : if !self.env.safekeepers.is_empty() {
387 0 : // Configure Postgres to connect to the safekeepers
388 0 : conf.append("synchronous_standby_names", "walproposer");
389 0 :
390 0 : let safekeepers = self
391 0 : .env
392 0 : .safekeepers
393 0 : .iter()
394 0 : .map(|sk| format!("localhost:{}", sk.get_compute_port()))
395 0 : .collect::<Vec<String>>()
396 0 : .join(",");
397 0 : conf.append("neon.safekeepers", &safekeepers);
398 0 : } else {
399 0 : // We only use setup without safekeepers for tests,
400 0 : // and don't care about data durability on pageserver,
401 0 : // so set more relaxed synchronous_commit.
402 0 : conf.append("synchronous_commit", "remote_write");
403 0 :
404 0 : // Configure the node to stream WAL directly to the pageserver
405 0 : // This isn't really a supported configuration, but can be useful for
406 0 : // testing.
407 0 : conf.append("synchronous_standby_names", "pageserver");
408 0 : }
409 : }
410 0 : ComputeMode::Static(lsn) => {
411 0 : conf.append("recovery_target_lsn", &lsn.to_string());
412 0 : }
413 : ComputeMode::Replica => {
414 0 : assert!(!self.env.safekeepers.is_empty());
415 :
416 : // TODO: use future host field from safekeeper spec
417 : // Pass the list of safekeepers to the replica so that it can connect to any of them,
418 : // whichever is available.
419 0 : let sk_ports = self
420 0 : .env
421 0 : .safekeepers
422 0 : .iter()
423 0 : .map(|x| x.get_compute_port().to_string())
424 0 : .collect::<Vec<_>>()
425 0 : .join(",");
426 0 : let sk_hosts = vec!["localhost"; self.env.safekeepers.len()].join(",");
427 0 :
428 0 : let connstr = format!(
429 0 : "host={} port={} options='-c timeline_id={} tenant_id={}' application_name=replica replication=true",
430 0 : sk_hosts,
431 0 : sk_ports,
432 0 : &self.timeline_id.to_string(),
433 0 : &self.tenant_id.to_string(),
434 0 : );
435 0 :
436 0 : let slot_name = format!("repl_{}_", self.timeline_id);
437 0 : conf.append("primary_conninfo", connstr.as_str());
438 0 : conf.append("primary_slot_name", slot_name.as_str());
439 0 : conf.append("hot_standby", "on");
440 0 : // prefetching of blocks referenced in WAL doesn't make sense for us
441 0 : // Neon hot standby ignores pages that are not in the shared_buffers
442 0 : if self.pg_version >= 15 {
443 0 : conf.append("recovery_prefetch", "off");
444 0 : }
445 : }
446 : }
447 :
448 0 : Ok(conf)
449 0 : }
450 :
451 0 : pub fn endpoint_path(&self) -> PathBuf {
452 0 : self.env.endpoints_path().join(&self.endpoint_id)
453 0 : }
454 :
455 0 : pub fn pgdata(&self) -> PathBuf {
456 0 : self.endpoint_path().join("pgdata")
457 0 : }
458 :
459 0 : pub fn status(&self) -> EndpointStatus {
460 0 : let timeout = Duration::from_millis(300);
461 0 : let has_pidfile = self.pgdata().join("postmaster.pid").exists();
462 0 : let can_connect = TcpStream::connect_timeout(&self.pg_address, timeout).is_ok();
463 0 :
464 0 : match (has_pidfile, can_connect) {
465 0 : (true, true) => EndpointStatus::Running,
466 0 : (false, false) => EndpointStatus::Stopped,
467 0 : (true, false) => EndpointStatus::Crashed,
468 0 : (false, true) => EndpointStatus::RunningNoPidfile,
469 : }
470 0 : }
471 :
472 0 : fn pg_ctl(&self, args: &[&str], auth_token: &Option<String>) -> Result<()> {
473 0 : let pg_ctl_path = self.env.pg_bin_dir(self.pg_version)?.join("pg_ctl");
474 0 : let mut cmd = Command::new(&pg_ctl_path);
475 0 : cmd.args(
476 0 : [
477 0 : &[
478 0 : "-D",
479 0 : self.pgdata().to_str().unwrap(),
480 0 : "-w", //wait till pg_ctl actually does what was asked
481 0 : ],
482 0 : args,
483 0 : ]
484 0 : .concat(),
485 0 : )
486 0 : .env_clear()
487 0 : .env(
488 0 : "LD_LIBRARY_PATH",
489 0 : self.env.pg_lib_dir(self.pg_version)?.to_str().unwrap(),
490 0 : )
491 0 : .env(
492 0 : "DYLD_LIBRARY_PATH",
493 0 : self.env.pg_lib_dir(self.pg_version)?.to_str().unwrap(),
494 : );
495 :
496 : // Pass authentication token used for the connections to pageserver and safekeepers
497 0 : if let Some(token) = auth_token {
498 0 : cmd.env("NEON_AUTH_TOKEN", token);
499 0 : }
500 :
501 0 : let pg_ctl = cmd
502 0 : .output()
503 0 : .context(format!("{} failed", pg_ctl_path.display()))?;
504 0 : if !pg_ctl.status.success() {
505 0 : anyhow::bail!(
506 0 : "pg_ctl failed, exit code: {}, stdout: {}, stderr: {}",
507 0 : pg_ctl.status,
508 0 : String::from_utf8_lossy(&pg_ctl.stdout),
509 0 : String::from_utf8_lossy(&pg_ctl.stderr),
510 0 : );
511 0 : }
512 0 :
513 0 : Ok(())
514 0 : }
515 :
516 0 : fn wait_for_compute_ctl_to_exit(&self, send_sigterm: bool) -> Result<()> {
517 0 : // TODO use background_process::stop_process instead: https://github.com/neondatabase/neon/pull/6482
518 0 : let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
519 0 : let pid: u32 = std::fs::read_to_string(pidfile_path)?.parse()?;
520 0 : let pid = nix::unistd::Pid::from_raw(pid as i32);
521 0 : if send_sigterm {
522 0 : kill(pid, Signal::SIGTERM).ok();
523 0 : }
524 0 : crate::background_process::wait_until_stopped("compute_ctl", pid)?;
525 0 : Ok(())
526 0 : }
527 :
528 0 : fn read_postgresql_conf(&self) -> Result<String> {
529 0 : // Slurp the endpoints/<endpoint id>/postgresql.conf file into
530 0 : // memory. We will include it in the spec file that we pass to
531 0 : // `compute_ctl`, and `compute_ctl` will write it to the postgresql.conf
532 0 : // in the data directory.
533 0 : let postgresql_conf_path = self.endpoint_path().join("postgresql.conf");
534 0 : match std::fs::read(&postgresql_conf_path) {
535 0 : Ok(content) => Ok(String::from_utf8(content)?),
536 0 : Err(e) if e.kind() == std::io::ErrorKind::NotFound => Ok("".to_string()),
537 0 : Err(e) => Err(anyhow::Error::new(e).context(format!(
538 0 : "failed to read config file in {}",
539 0 : postgresql_conf_path.to_str().unwrap()
540 0 : ))),
541 : }
542 0 : }
543 :
544 0 : fn build_pageserver_connstr(pageservers: &[(Host, u16)]) -> String {
545 0 : pageservers
546 0 : .iter()
547 0 : .map(|(host, port)| format!("postgresql://no_user@{host}:{port}"))
548 0 : .collect::<Vec<_>>()
549 0 : .join(",")
550 0 : }
551 :
552 : /// Map safekeepers ids to the actual connection strings.
553 0 : fn build_safekeepers_connstrs(&self, sk_ids: Vec<NodeId>) -> Result<Vec<String>> {
554 0 : let mut safekeeper_connstrings = Vec::new();
555 0 : if self.mode == ComputeMode::Primary {
556 0 : for sk_id in sk_ids {
557 0 : let sk = self
558 0 : .env
559 0 : .safekeepers
560 0 : .iter()
561 0 : .find(|node| node.id == sk_id)
562 0 : .ok_or_else(|| anyhow!("safekeeper {sk_id} does not exist"))?;
563 0 : safekeeper_connstrings.push(format!("127.0.0.1:{}", sk.get_compute_port()));
564 : }
565 0 : }
566 0 : Ok(safekeeper_connstrings)
567 0 : }
568 :
569 0 : pub async fn start(
570 0 : &self,
571 0 : auth_token: &Option<String>,
572 0 : safekeepers: Vec<NodeId>,
573 0 : pageservers: Vec<(Host, u16)>,
574 0 : remote_ext_config: Option<&String>,
575 0 : shard_stripe_size: usize,
576 0 : create_test_user: bool,
577 0 : ) -> Result<()> {
578 0 : if self.status() == EndpointStatus::Running {
579 0 : anyhow::bail!("The endpoint is already running");
580 0 : }
581 :
582 0 : let postgresql_conf = self.read_postgresql_conf()?;
583 :
584 : // We always start the compute node from scratch, so if the Postgres
585 : // data dir exists from a previous launch, remove it first.
586 0 : if self.pgdata().exists() {
587 0 : std::fs::remove_dir_all(self.pgdata())?;
588 0 : }
589 :
590 0 : let pageserver_connstring = Self::build_pageserver_connstr(&pageservers);
591 0 : assert!(!pageserver_connstring.is_empty());
592 :
593 0 : let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
594 :
595 : // check for file remote_extensions_spec.json
596 : // if it is present, read it and pass to compute_ctl
597 0 : let remote_extensions_spec_path = self.endpoint_path().join("remote_extensions_spec.json");
598 0 : let remote_extensions_spec = std::fs::File::open(remote_extensions_spec_path);
599 : let remote_extensions: Option<RemoteExtSpec>;
600 :
601 0 : if let Ok(spec_file) = remote_extensions_spec {
602 0 : remote_extensions = serde_json::from_reader(spec_file).ok();
603 0 : } else {
604 0 : remote_extensions = None;
605 0 : };
606 :
607 : // Create spec file
608 0 : let spec = ComputeSpec {
609 0 : skip_pg_catalog_updates: self.skip_pg_catalog_updates,
610 0 : format_version: 1.0,
611 0 : operation_uuid: None,
612 0 : features: self.features.clone(),
613 0 : swap_size_bytes: None,
614 0 : disk_quota_bytes: None,
615 0 : disable_lfc_resizing: None,
616 0 : cluster: Cluster {
617 0 : cluster_id: None, // project ID: not used
618 0 : name: None, // project name: not used
619 0 : state: None,
620 0 : roles: if create_test_user {
621 0 : vec![Role {
622 0 : name: PgIdent::from_str("test").unwrap(),
623 0 : encrypted_password: None,
624 0 : options: None,
625 0 : }]
626 : } else {
627 0 : Vec::new()
628 : },
629 0 : databases: if create_test_user {
630 0 : vec![Database {
631 0 : name: PgIdent::from_str("neondb").unwrap(),
632 0 : owner: PgIdent::from_str("test").unwrap(),
633 0 : options: None,
634 0 : restrict_conn: false,
635 0 : invalid: false,
636 0 : }]
637 : } else {
638 0 : Vec::new()
639 : },
640 0 : settings: None,
641 0 : postgresql_conf: Some(postgresql_conf),
642 0 : },
643 0 : delta_operations: None,
644 0 : tenant_id: Some(self.tenant_id),
645 0 : timeline_id: Some(self.timeline_id),
646 0 : mode: self.mode,
647 0 : pageserver_connstring: Some(pageserver_connstring),
648 0 : safekeeper_connstrings,
649 0 : storage_auth_token: auth_token.clone(),
650 0 : remote_extensions,
651 0 : pgbouncer_settings: None,
652 0 : shard_stripe_size: Some(shard_stripe_size),
653 0 : local_proxy_config: None,
654 0 : reconfigure_concurrency: 1,
655 0 : drop_subscriptions_before_start: self.drop_subscriptions_before_start,
656 0 : };
657 0 : let spec_path = self.endpoint_path().join("spec.json");
658 0 : std::fs::write(spec_path, serde_json::to_string_pretty(&spec)?)?;
659 :
660 : // Open log file. We'll redirect the stdout and stderr of `compute_ctl` to it.
661 0 : let logfile = std::fs::OpenOptions::new()
662 0 : .create(true)
663 0 : .append(true)
664 0 : .open(self.endpoint_path().join("compute.log"))?;
665 :
666 : // Launch compute_ctl
667 0 : let conn_str = self.connstr("cloud_admin", "postgres");
668 0 : println!("Starting postgres node at '{}'", conn_str);
669 0 : if create_test_user {
670 0 : let conn_str = self.connstr("test", "neondb");
671 0 : println!("Also at '{}'", conn_str);
672 0 : }
673 0 : let mut cmd = Command::new(self.env.neon_distrib_dir.join("compute_ctl"));
674 0 : //cmd.args([
675 0 : // "--external-http-port",
676 0 : // &self.external_http_address.port().to_string(),
677 0 : //])
678 0 : //.args([
679 0 : // "--internal-http-port",
680 0 : // &self.internal_http_address.port().to_string(),
681 0 : //])
682 0 : cmd.args([
683 0 : "--http-port",
684 0 : &self.external_http_address.port().to_string(),
685 0 : ])
686 0 : .args(["--pgdata", self.pgdata().to_str().unwrap()])
687 0 : .args(["--connstr", &conn_str])
688 0 : .args([
689 0 : "--spec-path",
690 0 : self.endpoint_path().join("spec.json").to_str().unwrap(),
691 0 : ])
692 0 : .args([
693 0 : "--pgbin",
694 0 : self.env
695 0 : .pg_bin_dir(self.pg_version)?
696 0 : .join("postgres")
697 0 : .to_str()
698 0 : .unwrap(),
699 0 : ])
700 0 : // TODO: It would be nice if we generated compute IDs with the same
701 0 : // algorithm as the real control plane.
702 0 : //
703 0 : // TODO: Add this back when
704 0 : // https://github.com/neondatabase/neon/pull/10747 is merged.
705 0 : //
706 0 : //.args([
707 0 : // "--compute-id",
708 0 : // &format!(
709 0 : // "compute-{}",
710 0 : // SystemTime::now()
711 0 : // .duration_since(UNIX_EPOCH)
712 0 : // .unwrap()
713 0 : // .as_secs()
714 0 : // ),
715 0 : //])
716 0 : .stdin(std::process::Stdio::null())
717 0 : .stderr(logfile.try_clone()?)
718 0 : .stdout(logfile);
719 :
720 0 : if let Some(remote_ext_config) = remote_ext_config {
721 0 : cmd.args(["--remote-ext-config", remote_ext_config]);
722 0 : }
723 :
724 0 : let child = cmd.spawn()?;
725 : // set up a scopeguard to kill & wait for the child in case we panic or bail below
726 0 : let child = scopeguard::guard(child, |mut child| {
727 0 : println!("SIGKILL & wait the started process");
728 0 : (|| {
729 0 : // TODO: use another signal that can be caught by the child so it can clean up any children it spawned
730 0 : child.kill().context("SIGKILL child")?;
731 0 : child.wait().context("wait() for child process")?;
732 0 : anyhow::Ok(())
733 0 : })()
734 0 : .with_context(|| format!("scopeguard kill&wait child {child:?}"))
735 0 : .unwrap();
736 0 : });
737 0 :
738 0 : // Write down the pid so we can wait for it when we want to stop
739 0 : // TODO use background_process::start_process instead: https://github.com/neondatabase/neon/pull/6482
740 0 : let pid = child.id();
741 0 : let pidfile_path = self.endpoint_path().join("compute_ctl.pid");
742 0 : std::fs::write(pidfile_path, pid.to_string())?;
743 :
744 : // Wait for it to start
745 0 : let mut attempt = 0;
746 : const ATTEMPT_INTERVAL: Duration = Duration::from_millis(100);
747 : const MAX_ATTEMPTS: u32 = 10 * 90; // Wait up to 1.5 min
748 : loop {
749 0 : attempt += 1;
750 0 : match self.get_status().await {
751 0 : Ok(state) => {
752 0 : match state.status {
753 : ComputeStatus::Init => {
754 0 : if attempt == MAX_ATTEMPTS {
755 0 : bail!("compute startup timed out; still in Init state");
756 0 : }
757 : // keep retrying
758 : }
759 : ComputeStatus::Running => {
760 : // All good!
761 0 : break;
762 : }
763 : ComputeStatus::Failed => {
764 0 : bail!(
765 0 : "compute startup failed: {}",
766 0 : state
767 0 : .error
768 0 : .as_deref()
769 0 : .unwrap_or("<no error from compute_ctl>")
770 0 : );
771 : }
772 : ComputeStatus::Empty
773 : | ComputeStatus::ConfigurationPending
774 : | ComputeStatus::Configuration
775 : | ComputeStatus::TerminationPending
776 : | ComputeStatus::Terminated => {
777 0 : bail!("unexpected compute status: {:?}", state.status)
778 : }
779 : }
780 : }
781 0 : Err(e) => {
782 0 : if attempt == MAX_ATTEMPTS {
783 0 : return Err(e).context("timed out waiting to connect to compute_ctl HTTP");
784 0 : }
785 : }
786 : }
787 0 : tokio::time::sleep(ATTEMPT_INTERVAL).await;
788 : }
789 :
790 : // disarm the scopeguard, let the child outlive this function (and neon_local invoction)
791 0 : drop(scopeguard::ScopeGuard::into_inner(child));
792 0 :
793 0 : Ok(())
794 0 : }
795 :
796 : // Call the /status HTTP API
797 0 : pub async fn get_status(&self) -> Result<ComputeStatusResponse> {
798 0 : let client = reqwest::Client::new();
799 :
800 0 : let response = client
801 0 : .request(
802 0 : reqwest::Method::GET,
803 0 : format!(
804 0 : "http://{}:{}/status",
805 0 : self.external_http_address.ip(),
806 0 : self.external_http_address.port()
807 0 : ),
808 0 : )
809 0 : .send()
810 0 : .await?;
811 :
812 : // Interpret the response
813 0 : let status = response.status();
814 0 : if !(status.is_client_error() || status.is_server_error()) {
815 0 : Ok(response.json().await?)
816 : } else {
817 : // reqwest does not export its error construction utility functions, so let's craft the message ourselves
818 0 : let url = response.url().to_owned();
819 0 : let msg = match response.text().await {
820 0 : Ok(err_body) => format!("Error: {}", err_body),
821 0 : Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
822 : };
823 0 : Err(anyhow::anyhow!(msg))
824 : }
825 0 : }
826 :
827 0 : pub async fn reconfigure(
828 0 : &self,
829 0 : mut pageservers: Vec<(Host, u16)>,
830 0 : stripe_size: Option<ShardStripeSize>,
831 0 : safekeepers: Option<Vec<NodeId>>,
832 0 : ) -> Result<()> {
833 0 : let mut spec: ComputeSpec = {
834 0 : let spec_path = self.endpoint_path().join("spec.json");
835 0 : let file = std::fs::File::open(spec_path)?;
836 0 : serde_json::from_reader(file)?
837 : };
838 :
839 0 : let postgresql_conf = self.read_postgresql_conf()?;
840 0 : spec.cluster.postgresql_conf = Some(postgresql_conf);
841 0 :
842 0 : // If we weren't given explicit pageservers, query the storage controller
843 0 : if pageservers.is_empty() {
844 0 : let storage_controller = StorageController::from_env(&self.env);
845 0 : let locate_result = storage_controller.tenant_locate(self.tenant_id).await?;
846 0 : pageservers = locate_result
847 0 : .shards
848 0 : .into_iter()
849 0 : .map(|shard| {
850 0 : (
851 0 : Host::parse(&shard.listen_pg_addr)
852 0 : .expect("Storage controller reported bad hostname"),
853 0 : shard.listen_pg_port,
854 0 : )
855 0 : })
856 0 : .collect::<Vec<_>>();
857 0 : }
858 :
859 0 : let pageserver_connstr = Self::build_pageserver_connstr(&pageservers);
860 0 : assert!(!pageserver_connstr.is_empty());
861 0 : spec.pageserver_connstring = Some(pageserver_connstr);
862 0 : if stripe_size.is_some() {
863 0 : spec.shard_stripe_size = stripe_size.map(|s| s.0 as usize);
864 0 : }
865 :
866 : // If safekeepers are not specified, don't change them.
867 0 : if let Some(safekeepers) = safekeepers {
868 0 : let safekeeper_connstrings = self.build_safekeepers_connstrs(safekeepers)?;
869 0 : spec.safekeeper_connstrings = safekeeper_connstrings;
870 0 : }
871 :
872 0 : let client = reqwest::Client::builder()
873 0 : .timeout(Duration::from_secs(120))
874 0 : .build()
875 0 : .unwrap();
876 0 : let response = client
877 0 : .post(format!(
878 0 : "http://{}:{}/configure",
879 0 : self.external_http_address.ip(),
880 0 : self.external_http_address.port()
881 0 : ))
882 0 : .header(CONTENT_TYPE.as_str(), "application/json")
883 0 : .body(format!(
884 0 : "{{\"spec\":{}}}",
885 0 : serde_json::to_string_pretty(&spec)?
886 : ))
887 0 : .send()
888 0 : .await?;
889 :
890 0 : let status = response.status();
891 0 : if !(status.is_client_error() || status.is_server_error()) {
892 0 : Ok(())
893 : } else {
894 0 : let url = response.url().to_owned();
895 0 : let msg = match response.text().await {
896 0 : Ok(err_body) => format!("Error: {}", err_body),
897 0 : Err(_) => format!("Http error ({}) at {}.", status.as_u16(), url),
898 : };
899 0 : Err(anyhow::anyhow!(msg))
900 : }
901 0 : }
902 :
903 0 : pub fn stop(&self, mode: &str, destroy: bool) -> Result<()> {
904 0 : self.pg_ctl(&["-m", mode, "stop"], &None)?;
905 :
906 : // Also wait for the compute_ctl process to die. It might have some
907 : // cleanup work to do after postgres stops, like syncing safekeepers,
908 : // etc.
909 : //
910 : // If destroying or stop mode is immediate, send it SIGTERM before
911 : // waiting. Sometimes we do *not* want this cleanup: tests intentionally
912 : // do stop when majority of safekeepers is down, so sync-safekeepers
913 : // would hang otherwise. This could be a separate flag though.
914 0 : let send_sigterm = destroy || mode == "immediate";
915 0 : self.wait_for_compute_ctl_to_exit(send_sigterm)?;
916 0 : if destroy {
917 0 : println!(
918 0 : "Destroying postgres data directory '{}'",
919 0 : self.pgdata().to_str().unwrap()
920 0 : );
921 0 : std::fs::remove_dir_all(self.endpoint_path())?;
922 0 : }
923 0 : Ok(())
924 0 : }
925 :
926 0 : pub fn connstr(&self, user: &str, db_name: &str) -> String {
927 0 : format!(
928 0 : "postgresql://{}@{}:{}/{}",
929 0 : user,
930 0 : self.pg_address.ip(),
931 0 : self.pg_address.port(),
932 0 : db_name
933 0 : )
934 0 : }
935 : }
|
