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