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