Line data Source code
1 : //! The Page Service listens for client connections and serves their GetPage@LSN
2 : //! requests.
3 :
4 : use anyhow::{bail, Context};
5 : use async_compression::tokio::write::GzipEncoder;
6 : use bytes::Buf;
7 : use futures::FutureExt;
8 : use itertools::Itertools;
9 : use once_cell::sync::OnceCell;
10 : use pageserver_api::config::{
11 : PageServicePipeliningConfig, PageServicePipeliningConfigPipelined,
12 : PageServiceProtocolPipelinedExecutionStrategy,
13 : };
14 : use pageserver_api::models::{self, TenantState};
15 : use pageserver_api::models::{
16 : PagestreamBeMessage, PagestreamDbSizeRequest, PagestreamDbSizeResponse,
17 : PagestreamErrorResponse, PagestreamExistsRequest, PagestreamExistsResponse,
18 : PagestreamFeMessage, PagestreamGetPageRequest, PagestreamGetSlruSegmentRequest,
19 : PagestreamGetSlruSegmentResponse, PagestreamNblocksRequest, PagestreamNblocksResponse,
20 : PagestreamProtocolVersion,
21 : };
22 : use pageserver_api::shard::TenantShardId;
23 : use postgres_backend::{
24 : is_expected_io_error, AuthType, PostgresBackend, PostgresBackendReader, QueryError,
25 : };
26 : use pq_proto::framed::ConnectionError;
27 : use pq_proto::FeStartupPacket;
28 : use pq_proto::{BeMessage, FeMessage, RowDescriptor};
29 : use std::borrow::Cow;
30 : use std::io;
31 : use std::num::NonZeroUsize;
32 : use std::str;
33 : use std::str::FromStr;
34 : use std::sync::Arc;
35 : use std::time::SystemTime;
36 : use std::time::{Duration, Instant};
37 : use tokio::io::{AsyncRead, AsyncWrite};
38 : use tokio::io::{AsyncWriteExt, BufWriter};
39 : use tokio::task::JoinHandle;
40 : use tokio_util::sync::CancellationToken;
41 : use tracing::*;
42 : use utils::sync::spsc_fold;
43 : use utils::{
44 : auth::{Claims, Scope, SwappableJwtAuth},
45 : id::{TenantId, TimelineId},
46 : lsn::Lsn,
47 : simple_rcu::RcuReadGuard,
48 : };
49 :
50 : use crate::auth::check_permission;
51 : use crate::basebackup::BasebackupError;
52 : use crate::config::PageServerConf;
53 : use crate::context::{DownloadBehavior, RequestContext};
54 : use crate::metrics::{self, SmgrOpTimer};
55 : use crate::metrics::{ComputeCommandKind, COMPUTE_COMMANDS_COUNTERS, LIVE_CONNECTIONS};
56 : use crate::pgdatadir_mapping::Version;
57 : use crate::span::debug_assert_current_span_has_tenant_and_timeline_id;
58 : use crate::span::debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id;
59 : use crate::task_mgr::TaskKind;
60 : use crate::task_mgr::{self, COMPUTE_REQUEST_RUNTIME};
61 : use crate::tenant::mgr::ShardSelector;
62 : use crate::tenant::mgr::TenantManager;
63 : use crate::tenant::mgr::{GetActiveTenantError, GetTenantError, ShardResolveResult};
64 : use crate::tenant::timeline::{self, WaitLsnError};
65 : use crate::tenant::GetTimelineError;
66 : use crate::tenant::PageReconstructError;
67 : use crate::tenant::Timeline;
68 : use crate::{basebackup, timed_after_cancellation};
69 : use pageserver_api::key::rel_block_to_key;
70 : use pageserver_api::reltag::{BlockNumber, RelTag, SlruKind};
71 : use postgres_ffi::pg_constants::DEFAULTTABLESPACE_OID;
72 : use postgres_ffi::BLCKSZ;
73 :
74 : /// How long we may wait for a [`crate::tenant::mgr::TenantSlot::InProgress`]` and/or a [`crate::tenant::Tenant`] which
75 : /// is not yet in state [`TenantState::Active`].
76 : ///
77 : /// NB: this is a different value than [`crate::http::routes::ACTIVE_TENANT_TIMEOUT`].
78 : const ACTIVE_TENANT_TIMEOUT: Duration = Duration::from_millis(30000);
79 :
80 : ///////////////////////////////////////////////////////////////////////////////
81 :
82 : pub struct Listener {
83 : cancel: CancellationToken,
84 : /// Cancel the listener task through `listen_cancel` to shut down the listener
85 : /// and get a handle on the existing connections.
86 : task: JoinHandle<Connections>,
87 : }
88 :
89 : pub struct Connections {
90 : cancel: CancellationToken,
91 : tasks: tokio::task::JoinSet<ConnectionHandlerResult>,
92 : }
93 :
94 0 : pub fn spawn(
95 0 : conf: &'static PageServerConf,
96 0 : tenant_manager: Arc<TenantManager>,
97 0 : pg_auth: Option<Arc<SwappableJwtAuth>>,
98 0 : tcp_listener: tokio::net::TcpListener,
99 0 : ) -> Listener {
100 0 : let cancel = CancellationToken::new();
101 0 : let libpq_ctx = RequestContext::todo_child(
102 0 : TaskKind::LibpqEndpointListener,
103 0 : // listener task shouldn't need to download anything. (We will
104 0 : // create a separate sub-contexts for each connection, with their
105 0 : // own download behavior. This context is used only to listen and
106 0 : // accept connections.)
107 0 : DownloadBehavior::Error,
108 0 : );
109 0 : let task = COMPUTE_REQUEST_RUNTIME.spawn(task_mgr::exit_on_panic_or_error(
110 0 : "libpq listener",
111 0 : libpq_listener_main(
112 0 : tenant_manager,
113 0 : pg_auth,
114 0 : tcp_listener,
115 0 : conf.pg_auth_type,
116 0 : conf.page_service_pipelining.clone(),
117 0 : libpq_ctx,
118 0 : cancel.clone(),
119 0 : )
120 0 : .map(anyhow::Ok),
121 0 : ));
122 0 :
123 0 : Listener { cancel, task }
124 0 : }
125 :
126 : impl Listener {
127 0 : pub async fn stop_accepting(self) -> Connections {
128 0 : self.cancel.cancel();
129 0 : self.task
130 0 : .await
131 0 : .expect("unreachable: we wrap the listener task in task_mgr::exit_on_panic_or_error")
132 0 : }
133 : }
134 : impl Connections {
135 0 : pub(crate) async fn shutdown(self) {
136 0 : let Self { cancel, mut tasks } = self;
137 0 : cancel.cancel();
138 0 : while let Some(res) = tasks.join_next().await {
139 0 : Self::handle_connection_completion(res);
140 0 : }
141 0 : }
142 :
143 0 : fn handle_connection_completion(res: Result<anyhow::Result<()>, tokio::task::JoinError>) {
144 0 : match res {
145 0 : Ok(Ok(())) => {}
146 0 : Ok(Err(e)) => error!("error in page_service connection task: {:?}", e),
147 0 : Err(e) => error!("page_service connection task panicked: {:?}", e),
148 : }
149 0 : }
150 : }
151 :
152 : ///
153 : /// Main loop of the page service.
154 : ///
155 : /// Listens for connections, and launches a new handler task for each.
156 : ///
157 : /// Returns Ok(()) upon cancellation via `cancel`, returning the set of
158 : /// open connections.
159 : ///
160 0 : pub async fn libpq_listener_main(
161 0 : tenant_manager: Arc<TenantManager>,
162 0 : auth: Option<Arc<SwappableJwtAuth>>,
163 0 : listener: tokio::net::TcpListener,
164 0 : auth_type: AuthType,
165 0 : pipelining_config: PageServicePipeliningConfig,
166 0 : listener_ctx: RequestContext,
167 0 : listener_cancel: CancellationToken,
168 0 : ) -> Connections {
169 0 : let connections_cancel = CancellationToken::new();
170 0 : let mut connection_handler_tasks = tokio::task::JoinSet::default();
171 :
172 : loop {
173 0 : let accepted = tokio::select! {
174 : biased;
175 0 : _ = listener_cancel.cancelled() => break,
176 0 : next = connection_handler_tasks.join_next(), if !connection_handler_tasks.is_empty() => {
177 0 : let res = next.expect("we dont poll while empty");
178 0 : Connections::handle_connection_completion(res);
179 0 : continue;
180 : }
181 0 : accepted = listener.accept() => accepted,
182 0 : };
183 0 :
184 0 : match accepted {
185 0 : Ok((socket, peer_addr)) => {
186 0 : // Connection established. Spawn a new task to handle it.
187 0 : debug!("accepted connection from {}", peer_addr);
188 0 : let local_auth = auth.clone();
189 0 : let connection_ctx = listener_ctx
190 0 : .detached_child(TaskKind::PageRequestHandler, DownloadBehavior::Download);
191 0 : connection_handler_tasks.spawn(page_service_conn_main(
192 0 : tenant_manager.clone(),
193 0 : local_auth,
194 0 : socket,
195 0 : auth_type,
196 0 : pipelining_config.clone(),
197 0 : connection_ctx,
198 0 : connections_cancel.child_token(),
199 0 : ));
200 : }
201 0 : Err(err) => {
202 0 : // accept() failed. Log the error, and loop back to retry on next connection.
203 0 : error!("accept() failed: {:?}", err);
204 : }
205 : }
206 : }
207 :
208 0 : debug!("page_service listener loop terminated");
209 :
210 0 : Connections {
211 0 : cancel: connections_cancel,
212 0 : tasks: connection_handler_tasks,
213 0 : }
214 0 : }
215 :
216 : type ConnectionHandlerResult = anyhow::Result<()>;
217 :
218 0 : #[instrument(skip_all, fields(peer_addr))]
219 : async fn page_service_conn_main(
220 : tenant_manager: Arc<TenantManager>,
221 : auth: Option<Arc<SwappableJwtAuth>>,
222 : socket: tokio::net::TcpStream,
223 : auth_type: AuthType,
224 : pipelining_config: PageServicePipeliningConfig,
225 : connection_ctx: RequestContext,
226 : cancel: CancellationToken,
227 : ) -> ConnectionHandlerResult {
228 : let _guard = LIVE_CONNECTIONS
229 : .with_label_values(&["page_service"])
230 : .guard();
231 :
232 : socket
233 : .set_nodelay(true)
234 : .context("could not set TCP_NODELAY")?;
235 :
236 : let peer_addr = socket.peer_addr().context("get peer address")?;
237 : tracing::Span::current().record("peer_addr", field::display(peer_addr));
238 :
239 : // setup read timeout of 10 minutes. the timeout is rather arbitrary for requirements:
240 : // - long enough for most valid compute connections
241 : // - less than infinite to stop us from "leaking" connections to long-gone computes
242 : //
243 : // no write timeout is used, because the kernel is assumed to error writes after some time.
244 : let mut socket = tokio_io_timeout::TimeoutReader::new(socket);
245 :
246 : let default_timeout_ms = 10 * 60 * 1000; // 10 minutes by default
247 0 : let socket_timeout_ms = (|| {
248 0 : fail::fail_point!("simulated-bad-compute-connection", |avg_timeout_ms| {
249 : // Exponential distribution for simulating
250 : // poor network conditions, expect about avg_timeout_ms to be around 15
251 : // in tests
252 0 : if let Some(avg_timeout_ms) = avg_timeout_ms {
253 0 : let avg = avg_timeout_ms.parse::<i64>().unwrap() as f32;
254 0 : let u = rand::random::<f32>();
255 0 : ((1.0 - u).ln() / (-avg)) as u64
256 : } else {
257 0 : default_timeout_ms
258 : }
259 0 : });
260 0 : default_timeout_ms
261 : })();
262 :
263 : // A timeout here does not mean the client died, it can happen if it's just idle for
264 : // a while: we will tear down this PageServerHandler and instantiate a new one if/when
265 : // they reconnect.
266 : socket.set_timeout(Some(std::time::Duration::from_millis(socket_timeout_ms)));
267 : let socket = Box::pin(socket);
268 :
269 : fail::fail_point!("ps::connection-start::pre-login");
270 :
271 : // XXX: pgbackend.run() should take the connection_ctx,
272 : // and create a child per-query context when it invokes process_query.
273 : // But it's in a shared crate, so, we store connection_ctx inside PageServerHandler
274 : // and create the per-query context in process_query ourselves.
275 : let mut conn_handler = PageServerHandler::new(
276 : tenant_manager,
277 : auth,
278 : pipelining_config,
279 : connection_ctx,
280 : cancel.clone(),
281 : );
282 : let pgbackend = PostgresBackend::new_from_io(socket, peer_addr, auth_type, None)?;
283 :
284 : match pgbackend.run(&mut conn_handler, &cancel).await {
285 : Ok(()) => {
286 : // we've been requested to shut down
287 : Ok(())
288 : }
289 : Err(QueryError::Disconnected(ConnectionError::Io(io_error))) => {
290 : if is_expected_io_error(&io_error) {
291 : info!("Postgres client disconnected ({io_error})");
292 : Ok(())
293 : } else {
294 : let tenant_id = conn_handler.timeline_handles.as_ref().unwrap().tenant_id();
295 : Err(io_error).context(format!(
296 : "Postgres connection error for tenant_id={:?} client at peer_addr={}",
297 : tenant_id, peer_addr
298 : ))
299 : }
300 : }
301 : other => {
302 : let tenant_id = conn_handler.timeline_handles.as_ref().unwrap().tenant_id();
303 : other.context(format!(
304 : "Postgres query error for tenant_id={:?} client peer_addr={}",
305 : tenant_id, peer_addr
306 : ))
307 : }
308 : }
309 : }
310 :
311 : struct PageServerHandler {
312 : auth: Option<Arc<SwappableJwtAuth>>,
313 : claims: Option<Claims>,
314 :
315 : /// The context created for the lifetime of the connection
316 : /// services by this PageServerHandler.
317 : /// For each query received over the connection,
318 : /// `process_query` creates a child context from this one.
319 : connection_ctx: RequestContext,
320 :
321 : cancel: CancellationToken,
322 :
323 : /// None only while pagestream protocol is being processed.
324 : timeline_handles: Option<TimelineHandles>,
325 :
326 : pipelining_config: PageServicePipeliningConfig,
327 : }
328 :
329 : struct TimelineHandles {
330 : wrapper: TenantManagerWrapper,
331 : /// Note on size: the typical size of this map is 1. The largest size we expect
332 : /// to see is the number of shards divided by the number of pageservers (typically < 2),
333 : /// or the ratio used when splitting shards (i.e. how many children created from one)
334 : /// parent shard, where a "large" number might be ~8.
335 : handles: timeline::handle::Cache<TenantManagerTypes>,
336 : }
337 :
338 : impl TimelineHandles {
339 0 : fn new(tenant_manager: Arc<TenantManager>) -> Self {
340 0 : Self {
341 0 : wrapper: TenantManagerWrapper {
342 0 : tenant_manager,
343 0 : tenant_id: OnceCell::new(),
344 0 : },
345 0 : handles: Default::default(),
346 0 : }
347 0 : }
348 0 : async fn get(
349 0 : &mut self,
350 0 : tenant_id: TenantId,
351 0 : timeline_id: TimelineId,
352 0 : shard_selector: ShardSelector,
353 0 : ) -> Result<timeline::handle::Handle<TenantManagerTypes>, GetActiveTimelineError> {
354 0 : if *self.wrapper.tenant_id.get_or_init(|| tenant_id) != tenant_id {
355 0 : return Err(GetActiveTimelineError::Tenant(
356 0 : GetActiveTenantError::SwitchedTenant,
357 0 : ));
358 0 : }
359 0 : self.handles
360 0 : .get(timeline_id, shard_selector, &self.wrapper)
361 0 : .await
362 0 : .map_err(|e| match e {
363 0 : timeline::handle::GetError::TenantManager(e) => e,
364 : timeline::handle::GetError::TimelineGateClosed => {
365 0 : trace!("timeline gate closed");
366 0 : GetActiveTimelineError::Timeline(GetTimelineError::ShuttingDown)
367 : }
368 : timeline::handle::GetError::PerTimelineStateShutDown => {
369 0 : trace!("per-timeline state shut down");
370 0 : GetActiveTimelineError::Timeline(GetTimelineError::ShuttingDown)
371 : }
372 0 : })
373 0 : }
374 :
375 0 : fn tenant_id(&self) -> Option<TenantId> {
376 0 : self.wrapper.tenant_id.get().copied()
377 0 : }
378 : }
379 :
380 : pub(crate) struct TenantManagerWrapper {
381 : tenant_manager: Arc<TenantManager>,
382 : // We do not support switching tenant_id on a connection at this point.
383 : // We can can add support for this later if needed without changing
384 : // the protocol.
385 : tenant_id: once_cell::sync::OnceCell<TenantId>,
386 : }
387 :
388 : #[derive(Debug)]
389 : pub(crate) struct TenantManagerTypes;
390 :
391 : impl timeline::handle::Types for TenantManagerTypes {
392 : type TenantManagerError = GetActiveTimelineError;
393 : type TenantManager = TenantManagerWrapper;
394 : type Timeline = Arc<Timeline>;
395 : }
396 :
397 : impl timeline::handle::ArcTimeline<TenantManagerTypes> for Arc<Timeline> {
398 0 : fn gate(&self) -> &utils::sync::gate::Gate {
399 0 : &self.gate
400 0 : }
401 :
402 0 : fn shard_timeline_id(&self) -> timeline::handle::ShardTimelineId {
403 0 : Timeline::shard_timeline_id(self)
404 0 : }
405 :
406 0 : fn per_timeline_state(&self) -> &timeline::handle::PerTimelineState<TenantManagerTypes> {
407 0 : &self.handles
408 0 : }
409 :
410 0 : fn get_shard_identity(&self) -> &pageserver_api::shard::ShardIdentity {
411 0 : Timeline::get_shard_identity(self)
412 0 : }
413 : }
414 :
415 : impl timeline::handle::TenantManager<TenantManagerTypes> for TenantManagerWrapper {
416 0 : async fn resolve(
417 0 : &self,
418 0 : timeline_id: TimelineId,
419 0 : shard_selector: ShardSelector,
420 0 : ) -> Result<Arc<Timeline>, GetActiveTimelineError> {
421 0 : let tenant_id = self.tenant_id.get().expect("we set this in get()");
422 0 : let timeout = ACTIVE_TENANT_TIMEOUT;
423 0 : let wait_start = Instant::now();
424 0 : let deadline = wait_start + timeout;
425 0 : let tenant_shard = loop {
426 0 : let resolved = self
427 0 : .tenant_manager
428 0 : .resolve_attached_shard(tenant_id, shard_selector);
429 0 : match resolved {
430 0 : ShardResolveResult::Found(tenant_shard) => break tenant_shard,
431 : ShardResolveResult::NotFound => {
432 0 : return Err(GetActiveTimelineError::Tenant(
433 0 : GetActiveTenantError::NotFound(GetTenantError::NotFound(*tenant_id)),
434 0 : ));
435 : }
436 0 : ShardResolveResult::InProgress(barrier) => {
437 0 : // We can't authoritatively answer right now: wait for InProgress state
438 0 : // to end, then try again
439 0 : tokio::select! {
440 0 : _ = barrier.wait() => {
441 0 : // The barrier completed: proceed around the loop to try looking up again
442 0 : },
443 0 : _ = tokio::time::sleep(deadline.duration_since(Instant::now())) => {
444 0 : return Err(GetActiveTimelineError::Tenant(GetActiveTenantError::WaitForActiveTimeout {
445 0 : latest_state: None,
446 0 : wait_time: timeout,
447 0 : }));
448 : }
449 : }
450 : }
451 : };
452 : };
453 :
454 0 : tracing::debug!("Waiting for tenant to enter active state...");
455 0 : tenant_shard
456 0 : .wait_to_become_active(deadline.duration_since(Instant::now()))
457 0 : .await
458 0 : .map_err(GetActiveTimelineError::Tenant)?;
459 :
460 0 : let timeline = tenant_shard
461 0 : .get_timeline(timeline_id, true)
462 0 : .map_err(GetActiveTimelineError::Timeline)?;
463 0 : set_tracing_field_shard_id(&timeline);
464 0 : Ok(timeline)
465 0 : }
466 : }
467 :
468 : #[derive(thiserror::Error, Debug)]
469 : enum PageStreamError {
470 : /// We encountered an error that should prompt the client to reconnect:
471 : /// in practice this means we drop the connection without sending a response.
472 : #[error("Reconnect required: {0}")]
473 : Reconnect(Cow<'static, str>),
474 :
475 : /// We were instructed to shutdown while processing the query
476 : #[error("Shutting down")]
477 : Shutdown,
478 :
479 : /// Something went wrong reading a page: this likely indicates a pageserver bug
480 : #[error("Read error")]
481 : Read(#[source] PageReconstructError),
482 :
483 : /// Ran out of time waiting for an LSN
484 : #[error("LSN timeout: {0}")]
485 : LsnTimeout(WaitLsnError),
486 :
487 : /// The entity required to serve the request (tenant or timeline) is not found,
488 : /// or is not found in a suitable state to serve a request.
489 : #[error("Not found: {0}")]
490 : NotFound(Cow<'static, str>),
491 :
492 : /// Request asked for something that doesn't make sense, like an invalid LSN
493 : #[error("Bad request: {0}")]
494 : BadRequest(Cow<'static, str>),
495 : }
496 :
497 : impl From<PageReconstructError> for PageStreamError {
498 0 : fn from(value: PageReconstructError) -> Self {
499 0 : match value {
500 0 : PageReconstructError::Cancelled => Self::Shutdown,
501 0 : e => Self::Read(e),
502 : }
503 0 : }
504 : }
505 :
506 : impl From<GetActiveTimelineError> for PageStreamError {
507 0 : fn from(value: GetActiveTimelineError) -> Self {
508 0 : match value {
509 : GetActiveTimelineError::Tenant(GetActiveTenantError::Cancelled)
510 : | GetActiveTimelineError::Tenant(GetActiveTenantError::WillNotBecomeActive(
511 : TenantState::Stopping { .. },
512 : ))
513 0 : | GetActiveTimelineError::Timeline(GetTimelineError::ShuttingDown) => Self::Shutdown,
514 0 : GetActiveTimelineError::Tenant(e) => Self::NotFound(format!("{e}").into()),
515 0 : GetActiveTimelineError::Timeline(e) => Self::NotFound(format!("{e}").into()),
516 : }
517 0 : }
518 : }
519 :
520 : impl From<WaitLsnError> for PageStreamError {
521 0 : fn from(value: WaitLsnError) -> Self {
522 0 : match value {
523 0 : e @ WaitLsnError::Timeout(_) => Self::LsnTimeout(e),
524 0 : WaitLsnError::Shutdown => Self::Shutdown,
525 0 : e @ WaitLsnError::BadState { .. } => Self::Reconnect(format!("{e}").into()),
526 : }
527 0 : }
528 : }
529 :
530 : impl From<WaitLsnError> for QueryError {
531 0 : fn from(value: WaitLsnError) -> Self {
532 0 : match value {
533 0 : e @ WaitLsnError::Timeout(_) => Self::Other(anyhow::Error::new(e)),
534 0 : WaitLsnError::Shutdown => Self::Shutdown,
535 0 : WaitLsnError::BadState { .. } => Self::Reconnect,
536 : }
537 0 : }
538 : }
539 :
540 : enum BatchedFeMessage {
541 : Exists {
542 : span: Span,
543 : timer: SmgrOpTimer,
544 : shard: timeline::handle::Handle<TenantManagerTypes>,
545 : req: models::PagestreamExistsRequest,
546 : },
547 : Nblocks {
548 : span: Span,
549 : timer: SmgrOpTimer,
550 : shard: timeline::handle::Handle<TenantManagerTypes>,
551 : req: models::PagestreamNblocksRequest,
552 : },
553 : GetPage {
554 : span: Span,
555 : shard: timeline::handle::Handle<TenantManagerTypes>,
556 : effective_request_lsn: Lsn,
557 : pages: smallvec::SmallVec<[(RelTag, BlockNumber, SmgrOpTimer); 1]>,
558 : },
559 : DbSize {
560 : span: Span,
561 : timer: SmgrOpTimer,
562 : shard: timeline::handle::Handle<TenantManagerTypes>,
563 : req: models::PagestreamDbSizeRequest,
564 : },
565 : GetSlruSegment {
566 : span: Span,
567 : timer: SmgrOpTimer,
568 : shard: timeline::handle::Handle<TenantManagerTypes>,
569 : req: models::PagestreamGetSlruSegmentRequest,
570 : },
571 : RespondError {
572 : span: Span,
573 : error: PageStreamError,
574 : },
575 : }
576 :
577 : impl BatchedFeMessage {
578 0 : async fn throttle_and_record_start_processing(
579 0 : &mut self,
580 0 : cancel: &CancellationToken,
581 0 : ) -> Result<(), QueryError> {
582 0 : let (shard, tokens, timers) = match self {
583 0 : BatchedFeMessage::Exists { shard, timer, .. }
584 0 : | BatchedFeMessage::Nblocks { shard, timer, .. }
585 0 : | BatchedFeMessage::DbSize { shard, timer, .. }
586 0 : | BatchedFeMessage::GetSlruSegment { shard, timer, .. } => {
587 0 : (
588 0 : shard,
589 0 : // 1 token is probably under-estimating because these
590 0 : // request handlers typically do several Timeline::get calls.
591 0 : 1,
592 0 : itertools::Either::Left(std::iter::once(timer)),
593 0 : )
594 : }
595 0 : BatchedFeMessage::GetPage { shard, pages, .. } => (
596 0 : shard,
597 0 : pages.len(),
598 0 : itertools::Either::Right(pages.iter_mut().map(|(_, _, timer)| timer)),
599 0 : ),
600 0 : BatchedFeMessage::RespondError { .. } => return Ok(()),
601 : };
602 0 : let throttled = tokio::select! {
603 0 : throttled = shard.pagestream_throttle.throttle(tokens) => { throttled }
604 0 : _ = cancel.cancelled() => {
605 0 : return Err(QueryError::Shutdown);
606 : }
607 : };
608 0 : for timer in timers {
609 0 : timer.observe_throttle_done_execution_starting(&throttled);
610 0 : }
611 0 : Ok(())
612 0 : }
613 : }
614 :
615 : impl PageServerHandler {
616 0 : pub fn new(
617 0 : tenant_manager: Arc<TenantManager>,
618 0 : auth: Option<Arc<SwappableJwtAuth>>,
619 0 : pipelining_config: PageServicePipeliningConfig,
620 0 : connection_ctx: RequestContext,
621 0 : cancel: CancellationToken,
622 0 : ) -> Self {
623 0 : PageServerHandler {
624 0 : auth,
625 0 : claims: None,
626 0 : connection_ctx,
627 0 : timeline_handles: Some(TimelineHandles::new(tenant_manager)),
628 0 : cancel,
629 0 : pipelining_config,
630 0 : }
631 0 : }
632 :
633 : /// This function always respects cancellation of any timeline in `[Self::shard_timelines]`. Pass in
634 : /// a cancellation token at the next scope up (such as a tenant cancellation token) to ensure we respect
635 : /// cancellation if there aren't any timelines in the cache.
636 : ///
637 : /// If calling from a function that doesn't use the `[Self::shard_timelines]` cache, then pass in the
638 : /// timeline cancellation token.
639 0 : async fn flush_cancellable<IO>(
640 0 : &self,
641 0 : pgb: &mut PostgresBackend<IO>,
642 0 : cancel: &CancellationToken,
643 0 : ) -> Result<(), QueryError>
644 0 : where
645 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
646 0 : {
647 0 : tokio::select!(
648 0 : flush_r = pgb.flush() => {
649 0 : Ok(flush_r?)
650 : },
651 0 : _ = cancel.cancelled() => {
652 0 : Err(QueryError::Shutdown)
653 : }
654 : )
655 0 : }
656 :
657 0 : async fn pagestream_read_message<IO>(
658 0 : pgb: &mut PostgresBackendReader<IO>,
659 0 : tenant_id: TenantId,
660 0 : timeline_id: TimelineId,
661 0 : timeline_handles: &mut TimelineHandles,
662 0 : cancel: &CancellationToken,
663 0 : ctx: &RequestContext,
664 0 : parent_span: Span,
665 0 : ) -> Result<Option<BatchedFeMessage>, QueryError>
666 0 : where
667 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
668 0 : {
669 0 : let msg = tokio::select! {
670 : biased;
671 0 : _ = cancel.cancelled() => {
672 0 : return Err(QueryError::Shutdown)
673 : }
674 0 : msg = pgb.read_message() => { msg }
675 0 : };
676 0 :
677 0 : let received_at = Instant::now();
678 :
679 0 : let copy_data_bytes = match msg? {
680 0 : Some(FeMessage::CopyData(bytes)) => bytes,
681 : Some(FeMessage::Terminate) => {
682 0 : return Ok(None);
683 : }
684 0 : Some(m) => {
685 0 : return Err(QueryError::Other(anyhow::anyhow!(
686 0 : "unexpected message: {m:?} during COPY"
687 0 : )));
688 : }
689 : None => {
690 0 : return Ok(None);
691 : } // client disconnected
692 : };
693 0 : trace!("query: {copy_data_bytes:?}");
694 :
695 0 : fail::fail_point!("ps::handle-pagerequest-message");
696 :
697 : // parse request
698 0 : let neon_fe_msg = PagestreamFeMessage::parse(&mut copy_data_bytes.reader())?;
699 :
700 0 : let batched_msg = match neon_fe_msg {
701 0 : PagestreamFeMessage::Exists(req) => {
702 0 : let span = tracing::info_span!(parent: parent_span, "handle_get_rel_exists_request", rel = %req.rel, req_lsn = %req.request_lsn);
703 0 : let shard = timeline_handles
704 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
705 0 : .instrument(span.clone()) // sets `shard_id` field
706 0 : .await?;
707 0 : let timer = shard
708 0 : .query_metrics
709 0 : .start_smgr_op(metrics::SmgrQueryType::GetRelExists, received_at);
710 0 : BatchedFeMessage::Exists {
711 0 : span,
712 0 : timer,
713 0 : shard,
714 0 : req,
715 0 : }
716 : }
717 0 : PagestreamFeMessage::Nblocks(req) => {
718 0 : let span = tracing::info_span!(parent: parent_span, "handle_get_nblocks_request", rel = %req.rel, req_lsn = %req.request_lsn);
719 0 : let shard = timeline_handles
720 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
721 0 : .instrument(span.clone()) // sets `shard_id` field
722 0 : .await?;
723 0 : let timer = shard
724 0 : .query_metrics
725 0 : .start_smgr_op(metrics::SmgrQueryType::GetRelSize, received_at);
726 0 : BatchedFeMessage::Nblocks {
727 0 : span,
728 0 : timer,
729 0 : shard,
730 0 : req,
731 0 : }
732 : }
733 0 : PagestreamFeMessage::DbSize(req) => {
734 0 : let span = tracing::info_span!(parent: parent_span, "handle_db_size_request", dbnode = %req.dbnode, req_lsn = %req.request_lsn);
735 0 : let shard = timeline_handles
736 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
737 0 : .instrument(span.clone()) // sets `shard_id` field
738 0 : .await?;
739 0 : let timer = shard
740 0 : .query_metrics
741 0 : .start_smgr_op(metrics::SmgrQueryType::GetDbSize, received_at);
742 0 : BatchedFeMessage::DbSize {
743 0 : span,
744 0 : timer,
745 0 : shard,
746 0 : req,
747 0 : }
748 : }
749 0 : PagestreamFeMessage::GetSlruSegment(req) => {
750 0 : let span = tracing::info_span!(parent: parent_span, "handle_get_slru_segment_request", kind = %req.kind, segno = %req.segno, req_lsn = %req.request_lsn);
751 0 : let shard = timeline_handles
752 0 : .get(tenant_id, timeline_id, ShardSelector::Zero)
753 0 : .instrument(span.clone()) // sets `shard_id` field
754 0 : .await?;
755 0 : let timer = shard
756 0 : .query_metrics
757 0 : .start_smgr_op(metrics::SmgrQueryType::GetSlruSegment, received_at);
758 0 : BatchedFeMessage::GetSlruSegment {
759 0 : span,
760 0 : timer,
761 0 : shard,
762 0 : req,
763 0 : }
764 : }
765 : PagestreamFeMessage::GetPage(PagestreamGetPageRequest {
766 0 : request_lsn,
767 0 : not_modified_since,
768 0 : rel,
769 0 : blkno,
770 : }) => {
771 0 : let span = tracing::info_span!(parent: parent_span, "handle_get_page_at_lsn_request_batched", req_lsn = %request_lsn);
772 :
773 : macro_rules! respond_error {
774 : ($error:expr) => {{
775 : let error = BatchedFeMessage::RespondError {
776 : span,
777 : error: $error,
778 : };
779 : Ok(Some(error))
780 : }};
781 : }
782 :
783 0 : let key = rel_block_to_key(rel, blkno);
784 0 : let shard = match timeline_handles
785 0 : .get(tenant_id, timeline_id, ShardSelector::Page(key))
786 0 : .instrument(span.clone()) // sets `shard_id` field
787 0 : .await
788 : {
789 0 : Ok(tl) => tl,
790 : Err(GetActiveTimelineError::Tenant(GetActiveTenantError::NotFound(_))) => {
791 : // We already know this tenant exists in general, because we resolved it at
792 : // start of connection. Getting a NotFound here indicates that the shard containing
793 : // the requested page is not present on this node: the client's knowledge of shard->pageserver
794 : // mapping is out of date.
795 : //
796 : // Closing the connection by returning ``::Reconnect` has the side effect of rate-limiting above message, via
797 : // client's reconnect backoff, as well as hopefully prompting the client to load its updated configuration
798 : // and talk to a different pageserver.
799 0 : return respond_error!(PageStreamError::Reconnect(
800 0 : "getpage@lsn request routed to wrong shard".into()
801 0 : ));
802 : }
803 0 : Err(e) => {
804 0 : return respond_error!(e.into());
805 : }
806 : };
807 :
808 : // It's important to start the timer before waiting for the LSN
809 : // so that the _started counters are incremented before we do
810 : // any serious waiting, e.g., for LSNs.
811 0 : let timer = shard
812 0 : .query_metrics
813 0 : .start_smgr_op(metrics::SmgrQueryType::GetPageAtLsn, received_at);
814 :
815 0 : let effective_request_lsn = match Self::wait_or_get_last_lsn(
816 0 : &shard,
817 0 : request_lsn,
818 0 : not_modified_since,
819 0 : &shard.get_latest_gc_cutoff_lsn(),
820 0 : ctx,
821 0 : )
822 0 : // TODO: if we actually need to wait for lsn here, it delays the entire batch which doesn't need to wait
823 0 : .await
824 : {
825 0 : Ok(lsn) => lsn,
826 0 : Err(e) => {
827 0 : return respond_error!(e);
828 : }
829 : };
830 : BatchedFeMessage::GetPage {
831 0 : span,
832 0 : shard,
833 0 : effective_request_lsn,
834 0 : pages: smallvec::smallvec![(rel, blkno, timer)],
835 : }
836 : }
837 : };
838 0 : Ok(Some(batched_msg))
839 0 : }
840 :
841 : /// Post-condition: `batch` is Some()
842 0 : #[instrument(skip_all, level = tracing::Level::TRACE)]
843 : #[allow(clippy::boxed_local)]
844 : fn pagestream_do_batch(
845 : max_batch_size: NonZeroUsize,
846 : batch: &mut Result<BatchedFeMessage, QueryError>,
847 : this_msg: Result<BatchedFeMessage, QueryError>,
848 : ) -> Result<(), Result<BatchedFeMessage, QueryError>> {
849 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
850 :
851 : let this_msg = match this_msg {
852 : Ok(this_msg) => this_msg,
853 : Err(e) => return Err(Err(e)),
854 : };
855 :
856 : match (&mut *batch, this_msg) {
857 : // something batched already, let's see if we can add this message to the batch
858 : (
859 : Ok(BatchedFeMessage::GetPage {
860 : span: _,
861 : shard: accum_shard,
862 : pages: ref mut accum_pages,
863 : effective_request_lsn: accum_lsn,
864 : }),
865 : BatchedFeMessage::GetPage {
866 : span: _,
867 : shard: this_shard,
868 : pages: this_pages,
869 : effective_request_lsn: this_lsn,
870 : },
871 0 : ) if (|| {
872 0 : assert_eq!(this_pages.len(), 1);
873 0 : if accum_pages.len() >= max_batch_size.get() {
874 0 : trace!(%accum_lsn, %this_lsn, %max_batch_size, "stopping batching because of batch size");
875 0 : assert_eq!(accum_pages.len(), max_batch_size.get());
876 0 : return false;
877 0 : }
878 0 : if (accum_shard.tenant_shard_id, accum_shard.timeline_id)
879 0 : != (this_shard.tenant_shard_id, this_shard.timeline_id)
880 : {
881 0 : trace!(%accum_lsn, %this_lsn, "stopping batching because timeline object mismatch");
882 : // TODO: we _could_ batch & execute each shard seperately (and in parallel).
883 : // But the current logic for keeping responses in order does not support that.
884 0 : return false;
885 0 : }
886 0 : // the vectored get currently only supports a single LSN, so, bounce as soon
887 0 : // as the effective request_lsn changes
888 0 : if *accum_lsn != this_lsn {
889 0 : trace!(%accum_lsn, %this_lsn, "stopping batching because LSN changed");
890 0 : return false;
891 0 : }
892 0 : true
893 : })() =>
894 : {
895 : // ok to batch
896 : accum_pages.extend(this_pages);
897 : Ok(())
898 : }
899 : // something batched already but this message is unbatchable
900 : (_, this_msg) => {
901 : // by default, don't continue batching
902 : Err(Ok(this_msg))
903 : }
904 : }
905 : }
906 :
907 0 : #[instrument(level = tracing::Level::DEBUG, skip_all)]
908 : async fn pagesteam_handle_batched_message<IO>(
909 : &mut self,
910 : pgb_writer: &mut PostgresBackend<IO>,
911 : batch: BatchedFeMessage,
912 : cancel: &CancellationToken,
913 : ctx: &RequestContext,
914 : ) -> Result<(), QueryError>
915 : where
916 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
917 : {
918 : // invoke handler function
919 : let (handler_results, span): (
920 : Vec<Result<(PagestreamBeMessage, SmgrOpTimer), PageStreamError>>,
921 : _,
922 : ) = match batch {
923 : BatchedFeMessage::Exists {
924 : span,
925 : timer,
926 : shard,
927 : req,
928 : } => {
929 : fail::fail_point!("ps::handle-pagerequest-message::exists");
930 : (
931 : vec![self
932 : .handle_get_rel_exists_request(&shard, &req, ctx)
933 : .instrument(span.clone())
934 : .await
935 0 : .map(|msg| (msg, timer))],
936 : span,
937 : )
938 : }
939 : BatchedFeMessage::Nblocks {
940 : span,
941 : timer,
942 : shard,
943 : req,
944 : } => {
945 : fail::fail_point!("ps::handle-pagerequest-message::nblocks");
946 : (
947 : vec![self
948 : .handle_get_nblocks_request(&shard, &req, ctx)
949 : .instrument(span.clone())
950 : .await
951 0 : .map(|msg| (msg, timer))],
952 : span,
953 : )
954 : }
955 : BatchedFeMessage::GetPage {
956 : span,
957 : shard,
958 : effective_request_lsn,
959 : pages,
960 : } => {
961 : fail::fail_point!("ps::handle-pagerequest-message::getpage");
962 : (
963 : {
964 : let npages = pages.len();
965 : trace!(npages, "handling getpage request");
966 : let res = self
967 : .handle_get_page_at_lsn_request_batched(
968 : &shard,
969 : effective_request_lsn,
970 : pages,
971 : ctx,
972 : )
973 : .instrument(span.clone())
974 : .await;
975 : assert_eq!(res.len(), npages);
976 : res
977 : },
978 : span,
979 : )
980 : }
981 : BatchedFeMessage::DbSize {
982 : span,
983 : timer,
984 : shard,
985 : req,
986 : } => {
987 : fail::fail_point!("ps::handle-pagerequest-message::dbsize");
988 : (
989 : vec![self
990 : .handle_db_size_request(&shard, &req, ctx)
991 : .instrument(span.clone())
992 : .await
993 0 : .map(|msg| (msg, timer))],
994 : span,
995 : )
996 : }
997 : BatchedFeMessage::GetSlruSegment {
998 : span,
999 : timer,
1000 : shard,
1001 : req,
1002 : } => {
1003 : fail::fail_point!("ps::handle-pagerequest-message::slrusegment");
1004 : (
1005 : vec![self
1006 : .handle_get_slru_segment_request(&shard, &req, ctx)
1007 : .instrument(span.clone())
1008 : .await
1009 0 : .map(|msg| (msg, timer))],
1010 : span,
1011 : )
1012 : }
1013 : BatchedFeMessage::RespondError { span, error } => {
1014 : // We've already decided to respond with an error, so we don't need to
1015 : // call the handler.
1016 : (vec![Err(error)], span)
1017 : }
1018 : };
1019 :
1020 : // Map handler result to protocol behavior.
1021 : // Some handler errors cause exit from pagestream protocol.
1022 : // Other handler errors are sent back as an error message and we stay in pagestream protocol.
1023 : for handler_result in handler_results {
1024 : let (response_msg, timer) = match handler_result {
1025 : Err(e) => match &e {
1026 : PageStreamError::Shutdown => {
1027 : // If we fail to fulfil a request during shutdown, which may be _because_ of
1028 : // shutdown, then do not send the error to the client. Instead just drop the
1029 : // connection.
1030 0 : span.in_scope(|| info!("dropping connection due to shutdown"));
1031 : return Err(QueryError::Shutdown);
1032 : }
1033 : PageStreamError::Reconnect(reason) => {
1034 0 : span.in_scope(|| info!("handler requested reconnect: {reason}"));
1035 : return Err(QueryError::Reconnect);
1036 : }
1037 : PageStreamError::Read(_)
1038 : | PageStreamError::LsnTimeout(_)
1039 : | PageStreamError::NotFound(_)
1040 : | PageStreamError::BadRequest(_) => {
1041 : // print the all details to the log with {:#}, but for the client the
1042 : // error message is enough. Do not log if shutting down, as the anyhow::Error
1043 : // here includes cancellation which is not an error.
1044 : let full = utils::error::report_compact_sources(&e);
1045 0 : span.in_scope(|| {
1046 0 : error!("error reading relation or page version: {full:#}")
1047 0 : });
1048 : (
1049 : PagestreamBeMessage::Error(PagestreamErrorResponse {
1050 : message: e.to_string(),
1051 : }),
1052 : None, // TODO: measure errors
1053 : )
1054 : }
1055 : },
1056 : Ok((response_msg, timer)) => (response_msg, Some(timer)),
1057 : };
1058 :
1059 : //
1060 : // marshal & transmit response message
1061 : //
1062 :
1063 : pgb_writer.write_message_noflush(&BeMessage::CopyData(&response_msg.serialize()))?;
1064 :
1065 : // We purposefully don't count flush time into the timer.
1066 : //
1067 : // The reason is that current compute client will not perform protocol processing
1068 : // if the postgres backend process is doing things other than `->smgr_read()`.
1069 : // This is especially the case for prefetch.
1070 : //
1071 : // If the compute doesn't read from the connection, eventually TCP will backpressure
1072 : // all the way into our flush call below.
1073 : //
1074 : // The timer's underlying metric is used for a storage-internal latency SLO and
1075 : // we don't want to include latency in it that we can't control.
1076 : // And as pointed out above, in this case, we don't control the time that flush will take.
1077 : let flushing_timer =
1078 0 : timer.map(|timer| timer.observe_smgr_op_completion_and_start_flushing());
1079 :
1080 : // what we want to do
1081 : let flush_fut = pgb_writer.flush();
1082 : // metric for how long flushing takes
1083 : let flush_fut = match flushing_timer {
1084 : Some(flushing_timer) => {
1085 : futures::future::Either::Left(flushing_timer.measure(flush_fut))
1086 : }
1087 : None => futures::future::Either::Right(flush_fut),
1088 : };
1089 : // do it while respecting cancellation
1090 0 : let _: () = async move {
1091 0 : tokio::select! {
1092 : biased;
1093 0 : _ = cancel.cancelled() => {
1094 : // We were requested to shut down.
1095 0 : info!("shutdown request received in page handler");
1096 0 : return Err(QueryError::Shutdown)
1097 : }
1098 0 : res = flush_fut => {
1099 0 : res?;
1100 : }
1101 : }
1102 0 : Ok(())
1103 0 : }
1104 : // and log the info! line inside the request span
1105 : .instrument(span.clone())
1106 : .await?;
1107 : }
1108 : Ok(())
1109 : }
1110 :
1111 : /// Pagestream sub-protocol handler.
1112 : ///
1113 : /// It is a simple request-response protocol inside a COPYBOTH session.
1114 : ///
1115 : /// # Coding Discipline
1116 : ///
1117 : /// Coding discipline within this function: all interaction with the `pgb` connection
1118 : /// needs to be sensitive to connection shutdown, currently signalled via [`Self::cancel`].
1119 : /// This is so that we can shutdown page_service quickly.
1120 0 : #[instrument(skip_all)]
1121 : async fn handle_pagerequests<IO>(
1122 : &mut self,
1123 : pgb: &mut PostgresBackend<IO>,
1124 : tenant_id: TenantId,
1125 : timeline_id: TimelineId,
1126 : _protocol_version: PagestreamProtocolVersion,
1127 : ctx: RequestContext,
1128 : ) -> Result<(), QueryError>
1129 : where
1130 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
1131 : {
1132 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
1133 :
1134 : // switch client to COPYBOTH
1135 : pgb.write_message_noflush(&BeMessage::CopyBothResponse)?;
1136 : tokio::select! {
1137 : biased;
1138 : _ = self.cancel.cancelled() => {
1139 : return Err(QueryError::Shutdown)
1140 : }
1141 : res = pgb.flush() => {
1142 : res?;
1143 : }
1144 : }
1145 :
1146 : let pgb_reader = pgb
1147 : .split()
1148 : .context("implementation error: split pgb into reader and writer")?;
1149 :
1150 : let timeline_handles = self
1151 : .timeline_handles
1152 : .take()
1153 : .expect("implementation error: timeline_handles should not be locked");
1154 :
1155 : let request_span = info_span!("request", shard_id = tracing::field::Empty);
1156 : let ((pgb_reader, timeline_handles), result) = match self.pipelining_config.clone() {
1157 : PageServicePipeliningConfig::Pipelined(pipelining_config) => {
1158 : self.handle_pagerequests_pipelined(
1159 : pgb,
1160 : pgb_reader,
1161 : tenant_id,
1162 : timeline_id,
1163 : timeline_handles,
1164 : request_span,
1165 : pipelining_config,
1166 : &ctx,
1167 : )
1168 : .await
1169 : }
1170 : PageServicePipeliningConfig::Serial => {
1171 : self.handle_pagerequests_serial(
1172 : pgb,
1173 : pgb_reader,
1174 : tenant_id,
1175 : timeline_id,
1176 : timeline_handles,
1177 : request_span,
1178 : &ctx,
1179 : )
1180 : .await
1181 : }
1182 : };
1183 :
1184 : debug!("pagestream subprotocol shut down cleanly");
1185 :
1186 : pgb.unsplit(pgb_reader)
1187 : .context("implementation error: unsplit pgb")?;
1188 :
1189 : let replaced = self.timeline_handles.replace(timeline_handles);
1190 : assert!(replaced.is_none());
1191 :
1192 : result
1193 : }
1194 :
1195 : #[allow(clippy::too_many_arguments)]
1196 0 : async fn handle_pagerequests_serial<IO>(
1197 0 : &mut self,
1198 0 : pgb_writer: &mut PostgresBackend<IO>,
1199 0 : mut pgb_reader: PostgresBackendReader<IO>,
1200 0 : tenant_id: TenantId,
1201 0 : timeline_id: TimelineId,
1202 0 : mut timeline_handles: TimelineHandles,
1203 0 : request_span: Span,
1204 0 : ctx: &RequestContext,
1205 0 : ) -> (
1206 0 : (PostgresBackendReader<IO>, TimelineHandles),
1207 0 : Result<(), QueryError>,
1208 0 : )
1209 0 : where
1210 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
1211 0 : {
1212 0 : let cancel = self.cancel.clone();
1213 0 : let err = loop {
1214 0 : let msg = Self::pagestream_read_message(
1215 0 : &mut pgb_reader,
1216 0 : tenant_id,
1217 0 : timeline_id,
1218 0 : &mut timeline_handles,
1219 0 : &cancel,
1220 0 : ctx,
1221 0 : request_span.clone(),
1222 0 : )
1223 0 : .await;
1224 0 : let msg = match msg {
1225 0 : Ok(msg) => msg,
1226 0 : Err(e) => break e,
1227 : };
1228 0 : let mut msg = match msg {
1229 0 : Some(msg) => msg,
1230 : None => {
1231 0 : debug!("pagestream subprotocol end observed");
1232 0 : return ((pgb_reader, timeline_handles), Ok(()));
1233 : }
1234 : };
1235 :
1236 0 : if let Err(cancelled) = msg.throttle_and_record_start_processing(&self.cancel).await {
1237 0 : break cancelled;
1238 0 : }
1239 :
1240 0 : let err = self
1241 0 : .pagesteam_handle_batched_message(pgb_writer, msg, &cancel, ctx)
1242 0 : .await;
1243 0 : match err {
1244 0 : Ok(()) => {}
1245 0 : Err(e) => break e,
1246 : }
1247 : };
1248 0 : ((pgb_reader, timeline_handles), Err(err))
1249 0 : }
1250 :
1251 : /// # Cancel-Safety
1252 : ///
1253 : /// May leak tokio tasks if not polled to completion.
1254 : #[allow(clippy::too_many_arguments)]
1255 0 : async fn handle_pagerequests_pipelined<IO>(
1256 0 : &mut self,
1257 0 : pgb_writer: &mut PostgresBackend<IO>,
1258 0 : pgb_reader: PostgresBackendReader<IO>,
1259 0 : tenant_id: TenantId,
1260 0 : timeline_id: TimelineId,
1261 0 : mut timeline_handles: TimelineHandles,
1262 0 : request_span: Span,
1263 0 : pipelining_config: PageServicePipeliningConfigPipelined,
1264 0 : ctx: &RequestContext,
1265 0 : ) -> (
1266 0 : (PostgresBackendReader<IO>, TimelineHandles),
1267 0 : Result<(), QueryError>,
1268 0 : )
1269 0 : where
1270 0 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
1271 0 : {
1272 0 : //
1273 0 : // Pipelined pagestream handling consists of
1274 0 : // - a Batcher that reads requests off the wire and
1275 0 : // and batches them if possible,
1276 0 : // - an Executor that processes the batched requests.
1277 0 : //
1278 0 : // The batch is built up inside an `spsc_fold` channel,
1279 0 : // shared betwen Batcher (Sender) and Executor (Receiver).
1280 0 : //
1281 0 : // The Batcher continously folds client requests into the batch,
1282 0 : // while the Executor can at any time take out what's in the batch
1283 0 : // in order to process it.
1284 0 : // This means the next batch builds up while the Executor
1285 0 : // executes the last batch.
1286 0 : //
1287 0 : // CANCELLATION
1288 0 : //
1289 0 : // We run both Batcher and Executor futures to completion before
1290 0 : // returning from this function.
1291 0 : //
1292 0 : // If Executor exits first, it signals cancellation to the Batcher
1293 0 : // via a CancellationToken that is child of `self.cancel`.
1294 0 : // If Batcher exits first, it signals cancellation to the Executor
1295 0 : // by dropping the spsc_fold channel Sender.
1296 0 : //
1297 0 : // CLEAN SHUTDOWN
1298 0 : //
1299 0 : // Clean shutdown means that the client ends the COPYBOTH session.
1300 0 : // In response to such a client message, the Batcher exits.
1301 0 : // The Executor continues to run, draining the spsc_fold channel.
1302 0 : // Once drained, the spsc_fold recv will fail with a distinct error
1303 0 : // indicating that the sender disconnected.
1304 0 : // The Executor exits with Ok(()) in response to that error.
1305 0 : //
1306 0 : // Server initiated shutdown is not clean shutdown, but instead
1307 0 : // is an error Err(QueryError::Shutdown) that is propagated through
1308 0 : // error propagation.
1309 0 : //
1310 0 : // ERROR PROPAGATION
1311 0 : //
1312 0 : // When the Batcher encounter an error, it sends it as a value
1313 0 : // through the spsc_fold channel and exits afterwards.
1314 0 : // When the Executor observes such an error in the channel,
1315 0 : // it exits returning that error value.
1316 0 : //
1317 0 : // This design ensures that the Executor stage will still process
1318 0 : // the batch that was in flight when the Batcher encountered an error,
1319 0 : // thereby beahving identical to a serial implementation.
1320 0 :
1321 0 : let PageServicePipeliningConfigPipelined {
1322 0 : max_batch_size,
1323 0 : execution,
1324 0 : } = pipelining_config;
1325 :
1326 : // Macro to _define_ a pipeline stage.
1327 : macro_rules! pipeline_stage {
1328 : ($name:literal, $cancel:expr, $make_fut:expr) => {{
1329 : let cancel: CancellationToken = $cancel;
1330 : let stage_fut = $make_fut(cancel.clone());
1331 0 : async move {
1332 0 : scopeguard::defer! {
1333 0 : debug!("exiting");
1334 0 : }
1335 0 : timed_after_cancellation(stage_fut, $name, Duration::from_millis(100), &cancel)
1336 0 : .await
1337 0 : }
1338 : .instrument(tracing::info_span!($name))
1339 : }};
1340 : }
1341 :
1342 : //
1343 : // Batcher
1344 : //
1345 :
1346 0 : let cancel_batcher = self.cancel.child_token();
1347 0 : let (mut batch_tx, mut batch_rx) = spsc_fold::channel();
1348 0 : let batcher = pipeline_stage!("batcher", cancel_batcher.clone(), move |cancel_batcher| {
1349 0 : let ctx = ctx.attached_child();
1350 0 : async move {
1351 0 : let mut pgb_reader = pgb_reader;
1352 0 : let mut exit = false;
1353 0 : while !exit {
1354 0 : let read_res = Self::pagestream_read_message(
1355 0 : &mut pgb_reader,
1356 0 : tenant_id,
1357 0 : timeline_id,
1358 0 : &mut timeline_handles,
1359 0 : &cancel_batcher,
1360 0 : &ctx,
1361 0 : request_span.clone(),
1362 0 : )
1363 0 : .await;
1364 0 : let Some(read_res) = read_res.transpose() else {
1365 0 : debug!("client-initiated shutdown");
1366 0 : break;
1367 : };
1368 0 : exit |= read_res.is_err();
1369 0 : let could_send = batch_tx
1370 0 : .send(read_res, |batch, res| {
1371 0 : Self::pagestream_do_batch(max_batch_size, batch, res)
1372 0 : })
1373 0 : .await;
1374 0 : exit |= could_send.is_err();
1375 : }
1376 0 : (pgb_reader, timeline_handles)
1377 0 : }
1378 0 : });
1379 :
1380 : //
1381 : // Executor
1382 : //
1383 :
1384 0 : let executor = pipeline_stage!("executor", self.cancel.clone(), move |cancel| {
1385 0 : let ctx = ctx.attached_child();
1386 0 : async move {
1387 0 : let _cancel_batcher = cancel_batcher.drop_guard();
1388 : loop {
1389 0 : let maybe_batch = batch_rx.recv().await;
1390 0 : let batch = match maybe_batch {
1391 0 : Ok(batch) => batch,
1392 : Err(spsc_fold::RecvError::SenderGone) => {
1393 0 : debug!("upstream gone");
1394 0 : return Ok(());
1395 : }
1396 : };
1397 0 : let mut batch = match batch {
1398 0 : Ok(batch) => batch,
1399 0 : Err(e) => {
1400 0 : return Err(e);
1401 : }
1402 : };
1403 0 : batch
1404 0 : .throttle_and_record_start_processing(&self.cancel)
1405 0 : .await?;
1406 0 : self.pagesteam_handle_batched_message(pgb_writer, batch, &cancel, &ctx)
1407 0 : .await?;
1408 : }
1409 0 : }
1410 0 : });
1411 :
1412 : //
1413 : // Execute the stages.
1414 : //
1415 :
1416 0 : match execution {
1417 : PageServiceProtocolPipelinedExecutionStrategy::ConcurrentFutures => {
1418 0 : tokio::join!(batcher, executor)
1419 : }
1420 : PageServiceProtocolPipelinedExecutionStrategy::Tasks => {
1421 : // These tasks are not tracked anywhere.
1422 0 : let read_messages_task = tokio::spawn(batcher);
1423 0 : let (read_messages_task_res, executor_res_) =
1424 0 : tokio::join!(read_messages_task, executor,);
1425 0 : (
1426 0 : read_messages_task_res.expect("propagated panic from read_messages"),
1427 0 : executor_res_,
1428 0 : )
1429 : }
1430 : }
1431 0 : }
1432 :
1433 : /// Helper function to handle the LSN from client request.
1434 : ///
1435 : /// Each GetPage (and Exists and Nblocks) request includes information about
1436 : /// which version of the page is being requested. The primary compute node
1437 : /// will always request the latest page version, by setting 'request_lsn' to
1438 : /// the last inserted or flushed WAL position, while a standby will request
1439 : /// a version at the LSN that it's currently caught up to.
1440 : ///
1441 : /// In either case, if the page server hasn't received the WAL up to the
1442 : /// requested LSN yet, we will wait for it to arrive. The return value is
1443 : /// the LSN that should be used to look up the page versions.
1444 : ///
1445 : /// In addition to the request LSN, each request carries another LSN,
1446 : /// 'not_modified_since', which is a hint to the pageserver that the client
1447 : /// knows that the page has not been modified between 'not_modified_since'
1448 : /// and the request LSN. This allows skipping the wait, as long as the WAL
1449 : /// up to 'not_modified_since' has arrived. If the client doesn't have any
1450 : /// information about when the page was modified, it will use
1451 : /// not_modified_since == lsn. If the client lies and sends a too low
1452 : /// not_modified_hint such that there are in fact later page versions, the
1453 : /// behavior is undefined: the pageserver may return any of the page versions
1454 : /// or an error.
1455 0 : async fn wait_or_get_last_lsn(
1456 0 : timeline: &Timeline,
1457 0 : request_lsn: Lsn,
1458 0 : not_modified_since: Lsn,
1459 0 : latest_gc_cutoff_lsn: &RcuReadGuard<Lsn>,
1460 0 : ctx: &RequestContext,
1461 0 : ) -> Result<Lsn, PageStreamError> {
1462 0 : let last_record_lsn = timeline.get_last_record_lsn();
1463 0 :
1464 0 : // Sanity check the request
1465 0 : if request_lsn < not_modified_since {
1466 0 : return Err(PageStreamError::BadRequest(
1467 0 : format!(
1468 0 : "invalid request with request LSN {} and not_modified_since {}",
1469 0 : request_lsn, not_modified_since,
1470 0 : )
1471 0 : .into(),
1472 0 : ));
1473 0 : }
1474 0 :
1475 0 : // Check explicitly for INVALID just to get a less scary error message if the request is obviously bogus
1476 0 : if request_lsn == Lsn::INVALID {
1477 0 : return Err(PageStreamError::BadRequest(
1478 0 : "invalid LSN(0) in request".into(),
1479 0 : ));
1480 0 : }
1481 0 :
1482 0 : // Clients should only read from recent LSNs on their timeline, or from locations holding an LSN lease.
1483 0 : //
1484 0 : // We may have older data available, but we make a best effort to detect this case and return an error,
1485 0 : // to distinguish a misbehaving client (asking for old LSN) from a storage issue (data missing at a legitimate LSN).
1486 0 : if request_lsn < **latest_gc_cutoff_lsn && !timeline.is_gc_blocked_by_lsn_lease_deadline() {
1487 0 : let gc_info = &timeline.gc_info.read().unwrap();
1488 0 : if !gc_info.leases.contains_key(&request_lsn) {
1489 0 : return Err(
1490 0 : PageStreamError::BadRequest(format!(
1491 0 : "tried to request a page version that was garbage collected. requested at {} gc cutoff {}",
1492 0 : request_lsn, **latest_gc_cutoff_lsn
1493 0 : ).into())
1494 0 : );
1495 0 : }
1496 0 : }
1497 :
1498 : // Wait for WAL up to 'not_modified_since' to arrive, if necessary
1499 0 : if not_modified_since > last_record_lsn {
1500 0 : timeline
1501 0 : .wait_lsn(
1502 0 : not_modified_since,
1503 0 : crate::tenant::timeline::WaitLsnWaiter::PageService,
1504 0 : ctx,
1505 0 : )
1506 0 : .await?;
1507 : // Since we waited for 'not_modified_since' to arrive, that is now the last
1508 : // record LSN. (Or close enough for our purposes; the last-record LSN can
1509 : // advance immediately after we return anyway)
1510 0 : Ok(not_modified_since)
1511 : } else {
1512 : // It might be better to use max(not_modified_since, latest_gc_cutoff_lsn)
1513 : // here instead. That would give the same result, since we know that there
1514 : // haven't been any modifications since 'not_modified_since'. Using an older
1515 : // LSN might be faster, because that could allow skipping recent layers when
1516 : // finding the page. However, we have historically used 'last_record_lsn', so
1517 : // stick to that for now.
1518 0 : Ok(std::cmp::min(last_record_lsn, request_lsn))
1519 : }
1520 0 : }
1521 :
1522 : /// Handles the lsn lease request.
1523 : /// If a lease cannot be obtained, the client will receive NULL.
1524 0 : #[instrument(skip_all, fields(shard_id, %lsn))]
1525 : async fn handle_make_lsn_lease<IO>(
1526 : &mut self,
1527 : pgb: &mut PostgresBackend<IO>,
1528 : tenant_shard_id: TenantShardId,
1529 : timeline_id: TimelineId,
1530 : lsn: Lsn,
1531 : ctx: &RequestContext,
1532 : ) -> Result<(), QueryError>
1533 : where
1534 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
1535 : {
1536 : let timeline = self
1537 : .timeline_handles
1538 : .as_mut()
1539 : .unwrap()
1540 : .get(
1541 : tenant_shard_id.tenant_id,
1542 : timeline_id,
1543 : ShardSelector::Known(tenant_shard_id.to_index()),
1544 : )
1545 : .await?;
1546 : set_tracing_field_shard_id(&timeline);
1547 :
1548 : let lease = timeline
1549 : .renew_lsn_lease(lsn, timeline.get_lsn_lease_length(), ctx)
1550 0 : .inspect_err(|e| {
1551 0 : warn!("{e}");
1552 0 : })
1553 : .ok();
1554 0 : let valid_until_str = lease.map(|l| {
1555 0 : l.valid_until
1556 0 : .duration_since(SystemTime::UNIX_EPOCH)
1557 0 : .expect("valid_until is earlier than UNIX_EPOCH")
1558 0 : .as_millis()
1559 0 : .to_string()
1560 0 : });
1561 0 : let bytes = valid_until_str.as_ref().map(|x| x.as_bytes());
1562 :
1563 : pgb.write_message_noflush(&BeMessage::RowDescription(&[RowDescriptor::text_col(
1564 : b"valid_until",
1565 : )]))?
1566 : .write_message_noflush(&BeMessage::DataRow(&[bytes]))?;
1567 :
1568 : Ok(())
1569 : }
1570 :
1571 0 : #[instrument(skip_all, fields(shard_id))]
1572 : async fn handle_get_rel_exists_request(
1573 : &mut self,
1574 : timeline: &Timeline,
1575 : req: &PagestreamExistsRequest,
1576 : ctx: &RequestContext,
1577 : ) -> Result<PagestreamBeMessage, PageStreamError> {
1578 : let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
1579 : let lsn = Self::wait_or_get_last_lsn(
1580 : timeline,
1581 : req.request_lsn,
1582 : req.not_modified_since,
1583 : &latest_gc_cutoff_lsn,
1584 : ctx,
1585 : )
1586 : .await?;
1587 :
1588 : let exists = timeline
1589 : .get_rel_exists(req.rel, Version::Lsn(lsn), ctx)
1590 : .await?;
1591 :
1592 : Ok(PagestreamBeMessage::Exists(PagestreamExistsResponse {
1593 : exists,
1594 : }))
1595 : }
1596 :
1597 0 : #[instrument(skip_all, fields(shard_id))]
1598 : async fn handle_get_nblocks_request(
1599 : &mut self,
1600 : timeline: &Timeline,
1601 : req: &PagestreamNblocksRequest,
1602 : ctx: &RequestContext,
1603 : ) -> Result<PagestreamBeMessage, PageStreamError> {
1604 : let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
1605 : let lsn = Self::wait_or_get_last_lsn(
1606 : timeline,
1607 : req.request_lsn,
1608 : req.not_modified_since,
1609 : &latest_gc_cutoff_lsn,
1610 : ctx,
1611 : )
1612 : .await?;
1613 :
1614 : let n_blocks = timeline
1615 : .get_rel_size(req.rel, Version::Lsn(lsn), ctx)
1616 : .await?;
1617 :
1618 : Ok(PagestreamBeMessage::Nblocks(PagestreamNblocksResponse {
1619 : n_blocks,
1620 : }))
1621 : }
1622 :
1623 0 : #[instrument(skip_all, fields(shard_id))]
1624 : async fn handle_db_size_request(
1625 : &mut self,
1626 : timeline: &Timeline,
1627 : req: &PagestreamDbSizeRequest,
1628 : ctx: &RequestContext,
1629 : ) -> Result<PagestreamBeMessage, PageStreamError> {
1630 : let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
1631 : let lsn = Self::wait_or_get_last_lsn(
1632 : timeline,
1633 : req.request_lsn,
1634 : req.not_modified_since,
1635 : &latest_gc_cutoff_lsn,
1636 : ctx,
1637 : )
1638 : .await?;
1639 :
1640 : let total_blocks = timeline
1641 : .get_db_size(DEFAULTTABLESPACE_OID, req.dbnode, Version::Lsn(lsn), ctx)
1642 : .await?;
1643 : let db_size = total_blocks as i64 * BLCKSZ as i64;
1644 :
1645 : Ok(PagestreamBeMessage::DbSize(PagestreamDbSizeResponse {
1646 : db_size,
1647 : }))
1648 : }
1649 :
1650 0 : #[instrument(skip_all)]
1651 : async fn handle_get_page_at_lsn_request_batched(
1652 : &mut self,
1653 : timeline: &Timeline,
1654 : effective_lsn: Lsn,
1655 : requests: smallvec::SmallVec<[(RelTag, BlockNumber, SmgrOpTimer); 1]>,
1656 : ctx: &RequestContext,
1657 : ) -> Vec<Result<(PagestreamBeMessage, SmgrOpTimer), PageStreamError>> {
1658 : debug_assert_current_span_has_tenant_and_timeline_id();
1659 :
1660 : timeline
1661 : .query_metrics
1662 : .observe_getpage_batch_start(requests.len());
1663 :
1664 : let results = timeline
1665 : .get_rel_page_at_lsn_batched(
1666 0 : requests.iter().map(|(reltag, blkno, _)| (reltag, blkno)),
1667 : effective_lsn,
1668 : ctx,
1669 : )
1670 : .await;
1671 : assert_eq!(results.len(), requests.len());
1672 :
1673 : // TODO: avoid creating the new Vec here
1674 : Vec::from_iter(
1675 : requests
1676 : .into_iter()
1677 : .zip(results.into_iter())
1678 0 : .map(|((_, _, timer), res)| {
1679 0 : res.map(|page| {
1680 0 : (
1681 0 : PagestreamBeMessage::GetPage(models::PagestreamGetPageResponse {
1682 0 : page,
1683 0 : }),
1684 0 : timer,
1685 0 : )
1686 0 : })
1687 0 : .map_err(PageStreamError::from)
1688 0 : }),
1689 : )
1690 : }
1691 :
1692 0 : #[instrument(skip_all, fields(shard_id))]
1693 : async fn handle_get_slru_segment_request(
1694 : &mut self,
1695 : timeline: &Timeline,
1696 : req: &PagestreamGetSlruSegmentRequest,
1697 : ctx: &RequestContext,
1698 : ) -> Result<PagestreamBeMessage, PageStreamError> {
1699 : let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
1700 : let lsn = Self::wait_or_get_last_lsn(
1701 : timeline,
1702 : req.request_lsn,
1703 : req.not_modified_since,
1704 : &latest_gc_cutoff_lsn,
1705 : ctx,
1706 : )
1707 : .await?;
1708 :
1709 : let kind = SlruKind::from_repr(req.kind)
1710 : .ok_or(PageStreamError::BadRequest("invalid SLRU kind".into()))?;
1711 : let segment = timeline.get_slru_segment(kind, req.segno, lsn, ctx).await?;
1712 :
1713 : Ok(PagestreamBeMessage::GetSlruSegment(
1714 : PagestreamGetSlruSegmentResponse { segment },
1715 : ))
1716 : }
1717 :
1718 : /// Note on "fullbackup":
1719 : /// Full basebackups should only be used for debugging purposes.
1720 : /// Originally, it was introduced to enable breaking storage format changes,
1721 : /// but that is not applicable anymore.
1722 : ///
1723 : /// # Coding Discipline
1724 : ///
1725 : /// Coding discipline within this function: all interaction with the `pgb` connection
1726 : /// needs to be sensitive to connection shutdown, currently signalled via [`Self::cancel`].
1727 : /// This is so that we can shutdown page_service quickly.
1728 : ///
1729 : /// TODO: wrap the pgb that we pass to the basebackup handler so that it's sensitive
1730 : /// to connection cancellation.
1731 : #[allow(clippy::too_many_arguments)]
1732 0 : #[instrument(skip_all, fields(shard_id, ?lsn, ?prev_lsn, %full_backup))]
1733 : async fn handle_basebackup_request<IO>(
1734 : &mut self,
1735 : pgb: &mut PostgresBackend<IO>,
1736 : tenant_id: TenantId,
1737 : timeline_id: TimelineId,
1738 : lsn: Option<Lsn>,
1739 : prev_lsn: Option<Lsn>,
1740 : full_backup: bool,
1741 : gzip: bool,
1742 : replica: bool,
1743 : ctx: &RequestContext,
1744 : ) -> Result<(), QueryError>
1745 : where
1746 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin,
1747 : {
1748 0 : fn map_basebackup_error(err: BasebackupError) -> QueryError {
1749 0 : match err {
1750 0 : BasebackupError::Client(e) => QueryError::Disconnected(ConnectionError::Io(e)),
1751 0 : BasebackupError::Server(e) => QueryError::Other(e),
1752 : }
1753 0 : }
1754 :
1755 : let started = std::time::Instant::now();
1756 :
1757 : let timeline = self
1758 : .timeline_handles
1759 : .as_mut()
1760 : .unwrap()
1761 : .get(tenant_id, timeline_id, ShardSelector::Zero)
1762 : .await?;
1763 :
1764 : let latest_gc_cutoff_lsn = timeline.get_latest_gc_cutoff_lsn();
1765 : if let Some(lsn) = lsn {
1766 : // Backup was requested at a particular LSN. Wait for it to arrive.
1767 : info!("waiting for {}", lsn);
1768 : timeline
1769 : .wait_lsn(
1770 : lsn,
1771 : crate::tenant::timeline::WaitLsnWaiter::PageService,
1772 : ctx,
1773 : )
1774 : .await?;
1775 : timeline
1776 : .check_lsn_is_in_scope(lsn, &latest_gc_cutoff_lsn)
1777 : .context("invalid basebackup lsn")?;
1778 : }
1779 :
1780 : let lsn_awaited_after = started.elapsed();
1781 :
1782 : // switch client to COPYOUT
1783 : pgb.write_message_noflush(&BeMessage::CopyOutResponse)
1784 : .map_err(QueryError::Disconnected)?;
1785 : self.flush_cancellable(pgb, &self.cancel).await?;
1786 :
1787 : // Send a tarball of the latest layer on the timeline. Compress if not
1788 : // fullbackup. TODO Compress in that case too (tests need to be updated)
1789 : if full_backup {
1790 : let mut writer = pgb.copyout_writer();
1791 : basebackup::send_basebackup_tarball(
1792 : &mut writer,
1793 : &timeline,
1794 : lsn,
1795 : prev_lsn,
1796 : full_backup,
1797 : replica,
1798 : ctx,
1799 : )
1800 : .await
1801 : .map_err(map_basebackup_error)?;
1802 : } else {
1803 : let mut writer = BufWriter::new(pgb.copyout_writer());
1804 : if gzip {
1805 : let mut encoder = GzipEncoder::with_quality(
1806 : &mut writer,
1807 : // NOTE using fast compression because it's on the critical path
1808 : // for compute startup. For an empty database, we get
1809 : // <100KB with this method. The Level::Best compression method
1810 : // gives us <20KB, but maybe we should add basebackup caching
1811 : // on compute shutdown first.
1812 : async_compression::Level::Fastest,
1813 : );
1814 : basebackup::send_basebackup_tarball(
1815 : &mut encoder,
1816 : &timeline,
1817 : lsn,
1818 : prev_lsn,
1819 : full_backup,
1820 : replica,
1821 : ctx,
1822 : )
1823 : .await
1824 : .map_err(map_basebackup_error)?;
1825 : // shutdown the encoder to ensure the gzip footer is written
1826 : encoder
1827 : .shutdown()
1828 : .await
1829 0 : .map_err(|e| QueryError::Disconnected(ConnectionError::Io(e)))?;
1830 : } else {
1831 : basebackup::send_basebackup_tarball(
1832 : &mut writer,
1833 : &timeline,
1834 : lsn,
1835 : prev_lsn,
1836 : full_backup,
1837 : replica,
1838 : ctx,
1839 : )
1840 : .await
1841 : .map_err(map_basebackup_error)?;
1842 : }
1843 : writer
1844 : .flush()
1845 : .await
1846 0 : .map_err(|e| map_basebackup_error(BasebackupError::Client(e)))?;
1847 : }
1848 :
1849 : pgb.write_message_noflush(&BeMessage::CopyDone)
1850 : .map_err(QueryError::Disconnected)?;
1851 : self.flush_cancellable(pgb, &timeline.cancel).await?;
1852 :
1853 : let basebackup_after = started
1854 : .elapsed()
1855 : .checked_sub(lsn_awaited_after)
1856 : .unwrap_or(Duration::ZERO);
1857 :
1858 : info!(
1859 : lsn_await_millis = lsn_awaited_after.as_millis(),
1860 : basebackup_millis = basebackup_after.as_millis(),
1861 : "basebackup complete"
1862 : );
1863 :
1864 : Ok(())
1865 : }
1866 :
1867 : // when accessing management api supply None as an argument
1868 : // when using to authorize tenant pass corresponding tenant id
1869 0 : fn check_permission(&self, tenant_id: Option<TenantId>) -> Result<(), QueryError> {
1870 0 : if self.auth.is_none() {
1871 : // auth is set to Trust, nothing to check so just return ok
1872 0 : return Ok(());
1873 0 : }
1874 0 : // auth is some, just checked above, when auth is some
1875 0 : // then claims are always present because of checks during connection init
1876 0 : // so this expect won't trigger
1877 0 : let claims = self
1878 0 : .claims
1879 0 : .as_ref()
1880 0 : .expect("claims presence already checked");
1881 0 : check_permission(claims, tenant_id).map_err(|e| QueryError::Unauthorized(e.0))
1882 0 : }
1883 : }
1884 :
1885 : /// `basebackup tenant timeline [lsn] [--gzip] [--replica]`
1886 : #[derive(Debug, Clone, Eq, PartialEq)]
1887 : struct BaseBackupCmd {
1888 : tenant_id: TenantId,
1889 : timeline_id: TimelineId,
1890 : lsn: Option<Lsn>,
1891 : gzip: bool,
1892 : replica: bool,
1893 : }
1894 :
1895 : /// `fullbackup tenant timeline [lsn] [prev_lsn]`
1896 : #[derive(Debug, Clone, Eq, PartialEq)]
1897 : struct FullBackupCmd {
1898 : tenant_id: TenantId,
1899 : timeline_id: TimelineId,
1900 : lsn: Option<Lsn>,
1901 : prev_lsn: Option<Lsn>,
1902 : }
1903 :
1904 : /// `pagestream_v2 tenant timeline`
1905 : #[derive(Debug, Clone, Eq, PartialEq)]
1906 : struct PageStreamCmd {
1907 : tenant_id: TenantId,
1908 : timeline_id: TimelineId,
1909 : }
1910 :
1911 : /// `lease lsn tenant timeline lsn`
1912 : #[derive(Debug, Clone, Eq, PartialEq)]
1913 : struct LeaseLsnCmd {
1914 : tenant_shard_id: TenantShardId,
1915 : timeline_id: TimelineId,
1916 : lsn: Lsn,
1917 : }
1918 :
1919 : #[derive(Debug, Clone, Eq, PartialEq)]
1920 : enum PageServiceCmd {
1921 : Set,
1922 : PageStream(PageStreamCmd),
1923 : BaseBackup(BaseBackupCmd),
1924 : FullBackup(FullBackupCmd),
1925 : LeaseLsn(LeaseLsnCmd),
1926 : }
1927 :
1928 : impl PageStreamCmd {
1929 6 : fn parse(query: &str) -> anyhow::Result<Self> {
1930 6 : let parameters = query.split_whitespace().collect_vec();
1931 6 : if parameters.len() != 2 {
1932 2 : bail!(
1933 2 : "invalid number of parameters for pagestream command: {}",
1934 2 : query
1935 2 : );
1936 4 : }
1937 4 : let tenant_id = TenantId::from_str(parameters[0])
1938 4 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
1939 2 : let timeline_id = TimelineId::from_str(parameters[1])
1940 2 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
1941 2 : Ok(Self {
1942 2 : tenant_id,
1943 2 : timeline_id,
1944 2 : })
1945 6 : }
1946 : }
1947 :
1948 : impl FullBackupCmd {
1949 4 : fn parse(query: &str) -> anyhow::Result<Self> {
1950 4 : let parameters = query.split_whitespace().collect_vec();
1951 4 : if parameters.len() < 2 || parameters.len() > 4 {
1952 0 : bail!(
1953 0 : "invalid number of parameters for basebackup command: {}",
1954 0 : query
1955 0 : );
1956 4 : }
1957 4 : let tenant_id = TenantId::from_str(parameters[0])
1958 4 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
1959 4 : let timeline_id = TimelineId::from_str(parameters[1])
1960 4 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
1961 : // The caller is responsible for providing correct lsn and prev_lsn.
1962 4 : let lsn = if let Some(lsn_str) = parameters.get(2) {
1963 : Some(
1964 2 : Lsn::from_str(lsn_str)
1965 2 : .with_context(|| format!("Failed to parse Lsn from {lsn_str}"))?,
1966 : )
1967 : } else {
1968 2 : None
1969 : };
1970 4 : let prev_lsn = if let Some(prev_lsn_str) = parameters.get(3) {
1971 : Some(
1972 2 : Lsn::from_str(prev_lsn_str)
1973 2 : .with_context(|| format!("Failed to parse Lsn from {prev_lsn_str}"))?,
1974 : )
1975 : } else {
1976 2 : None
1977 : };
1978 4 : Ok(Self {
1979 4 : tenant_id,
1980 4 : timeline_id,
1981 4 : lsn,
1982 4 : prev_lsn,
1983 4 : })
1984 4 : }
1985 : }
1986 :
1987 : impl BaseBackupCmd {
1988 18 : fn parse(query: &str) -> anyhow::Result<Self> {
1989 18 : let parameters = query.split_whitespace().collect_vec();
1990 18 : if parameters.len() < 2 {
1991 0 : bail!(
1992 0 : "invalid number of parameters for basebackup command: {}",
1993 0 : query
1994 0 : );
1995 18 : }
1996 18 : let tenant_id = TenantId::from_str(parameters[0])
1997 18 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
1998 18 : let timeline_id = TimelineId::from_str(parameters[1])
1999 18 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
2000 : let lsn;
2001 : let flags_parse_from;
2002 18 : if let Some(maybe_lsn) = parameters.get(2) {
2003 16 : if *maybe_lsn == "latest" {
2004 2 : lsn = None;
2005 2 : flags_parse_from = 3;
2006 14 : } else if maybe_lsn.starts_with("--") {
2007 10 : lsn = None;
2008 10 : flags_parse_from = 2;
2009 10 : } else {
2010 : lsn = Some(
2011 4 : Lsn::from_str(maybe_lsn)
2012 4 : .with_context(|| format!("Failed to parse lsn from {maybe_lsn}"))?,
2013 : );
2014 4 : flags_parse_from = 3;
2015 : }
2016 2 : } else {
2017 2 : lsn = None;
2018 2 : flags_parse_from = 2;
2019 2 : }
2020 :
2021 18 : let mut gzip = false;
2022 18 : let mut replica = false;
2023 :
2024 22 : for ¶m in ¶meters[flags_parse_from..] {
2025 22 : match param {
2026 22 : "--gzip" => {
2027 14 : if gzip {
2028 2 : bail!("duplicate parameter for basebackup command: {param}")
2029 12 : }
2030 12 : gzip = true
2031 : }
2032 8 : "--replica" => {
2033 4 : if replica {
2034 0 : bail!("duplicate parameter for basebackup command: {param}")
2035 4 : }
2036 4 : replica = true
2037 : }
2038 4 : _ => bail!("invalid parameter for basebackup command: {param}"),
2039 : }
2040 : }
2041 12 : Ok(Self {
2042 12 : tenant_id,
2043 12 : timeline_id,
2044 12 : lsn,
2045 12 : gzip,
2046 12 : replica,
2047 12 : })
2048 18 : }
2049 : }
2050 :
2051 : impl LeaseLsnCmd {
2052 4 : fn parse(query: &str) -> anyhow::Result<Self> {
2053 4 : let parameters = query.split_whitespace().collect_vec();
2054 4 : if parameters.len() != 3 {
2055 0 : bail!(
2056 0 : "invalid number of parameters for lease lsn command: {}",
2057 0 : query
2058 0 : );
2059 4 : }
2060 4 : let tenant_shard_id = TenantShardId::from_str(parameters[0])
2061 4 : .with_context(|| format!("Failed to parse tenant id from {}", parameters[0]))?;
2062 4 : let timeline_id = TimelineId::from_str(parameters[1])
2063 4 : .with_context(|| format!("Failed to parse timeline id from {}", parameters[1]))?;
2064 4 : let lsn = Lsn::from_str(parameters[2])
2065 4 : .with_context(|| format!("Failed to parse lsn from {}", parameters[2]))?;
2066 4 : Ok(Self {
2067 4 : tenant_shard_id,
2068 4 : timeline_id,
2069 4 : lsn,
2070 4 : })
2071 4 : }
2072 : }
2073 :
2074 : impl PageServiceCmd {
2075 42 : fn parse(query: &str) -> anyhow::Result<Self> {
2076 42 : let query = query.trim();
2077 42 : let Some((cmd, other)) = query.split_once(' ') else {
2078 4 : bail!("cannot parse query: {query}")
2079 : };
2080 38 : match cmd.to_ascii_lowercase().as_str() {
2081 38 : "pagestream_v2" => Ok(Self::PageStream(PageStreamCmd::parse(other)?)),
2082 32 : "basebackup" => Ok(Self::BaseBackup(BaseBackupCmd::parse(other)?)),
2083 14 : "fullbackup" => Ok(Self::FullBackup(FullBackupCmd::parse(other)?)),
2084 10 : "lease" => {
2085 6 : let Some((cmd2, other)) = other.split_once(' ') else {
2086 0 : bail!("invalid lease command: {cmd}");
2087 : };
2088 6 : let cmd2 = cmd2.to_ascii_lowercase();
2089 6 : if cmd2 == "lsn" {
2090 4 : Ok(Self::LeaseLsn(LeaseLsnCmd::parse(other)?))
2091 : } else {
2092 2 : bail!("invalid lease command: {cmd}");
2093 : }
2094 : }
2095 4 : "set" => Ok(Self::Set),
2096 0 : _ => Err(anyhow::anyhow!("unsupported command {cmd} in {query}")),
2097 : }
2098 42 : }
2099 : }
2100 :
2101 : impl<IO> postgres_backend::Handler<IO> for PageServerHandler
2102 : where
2103 : IO: AsyncRead + AsyncWrite + Send + Sync + Unpin + 'static,
2104 : {
2105 0 : fn check_auth_jwt(
2106 0 : &mut self,
2107 0 : _pgb: &mut PostgresBackend<IO>,
2108 0 : jwt_response: &[u8],
2109 0 : ) -> Result<(), QueryError> {
2110 : // this unwrap is never triggered, because check_auth_jwt only called when auth_type is NeonJWT
2111 : // which requires auth to be present
2112 0 : let data = self
2113 0 : .auth
2114 0 : .as_ref()
2115 0 : .unwrap()
2116 0 : .decode(str::from_utf8(jwt_response).context("jwt response is not UTF-8")?)
2117 0 : .map_err(|e| QueryError::Unauthorized(e.0))?;
2118 :
2119 0 : if matches!(data.claims.scope, Scope::Tenant) && data.claims.tenant_id.is_none() {
2120 0 : return Err(QueryError::Unauthorized(
2121 0 : "jwt token scope is Tenant, but tenant id is missing".into(),
2122 0 : ));
2123 0 : }
2124 0 :
2125 0 : debug!(
2126 0 : "jwt scope check succeeded for scope: {:#?} by tenant id: {:?}",
2127 : data.claims.scope, data.claims.tenant_id,
2128 : );
2129 :
2130 0 : self.claims = Some(data.claims);
2131 0 : Ok(())
2132 0 : }
2133 :
2134 0 : fn startup(
2135 0 : &mut self,
2136 0 : _pgb: &mut PostgresBackend<IO>,
2137 0 : _sm: &FeStartupPacket,
2138 0 : ) -> Result<(), QueryError> {
2139 0 : fail::fail_point!("ps::connection-start::startup-packet");
2140 0 : Ok(())
2141 0 : }
2142 :
2143 0 : #[instrument(skip_all, fields(tenant_id, timeline_id))]
2144 : async fn process_query(
2145 : &mut self,
2146 : pgb: &mut PostgresBackend<IO>,
2147 : query_string: &str,
2148 : ) -> Result<(), QueryError> {
2149 0 : fail::fail_point!("simulated-bad-compute-connection", |_| {
2150 0 : info!("Hit failpoint for bad connection");
2151 0 : Err(QueryError::SimulatedConnectionError)
2152 0 : });
2153 :
2154 : fail::fail_point!("ps::connection-start::process-query");
2155 :
2156 : let ctx = self.connection_ctx.attached_child();
2157 : debug!("process query {query_string}");
2158 : let query = PageServiceCmd::parse(query_string)?;
2159 : match query {
2160 : PageServiceCmd::PageStream(PageStreamCmd {
2161 : tenant_id,
2162 : timeline_id,
2163 : }) => {
2164 : tracing::Span::current()
2165 : .record("tenant_id", field::display(tenant_id))
2166 : .record("timeline_id", field::display(timeline_id));
2167 :
2168 : self.check_permission(Some(tenant_id))?;
2169 :
2170 : COMPUTE_COMMANDS_COUNTERS
2171 : .for_command(ComputeCommandKind::PageStreamV2)
2172 : .inc();
2173 :
2174 : self.handle_pagerequests(
2175 : pgb,
2176 : tenant_id,
2177 : timeline_id,
2178 : PagestreamProtocolVersion::V2,
2179 : ctx,
2180 : )
2181 : .await?;
2182 : }
2183 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2184 : tenant_id,
2185 : timeline_id,
2186 : lsn,
2187 : gzip,
2188 : replica,
2189 : }) => {
2190 : tracing::Span::current()
2191 : .record("tenant_id", field::display(tenant_id))
2192 : .record("timeline_id", field::display(timeline_id));
2193 :
2194 : self.check_permission(Some(tenant_id))?;
2195 :
2196 : COMPUTE_COMMANDS_COUNTERS
2197 : .for_command(ComputeCommandKind::Basebackup)
2198 : .inc();
2199 : let metric_recording = metrics::BASEBACKUP_QUERY_TIME.start_recording();
2200 0 : let res = async {
2201 0 : self.handle_basebackup_request(
2202 0 : pgb,
2203 0 : tenant_id,
2204 0 : timeline_id,
2205 0 : lsn,
2206 0 : None,
2207 0 : false,
2208 0 : gzip,
2209 0 : replica,
2210 0 : &ctx,
2211 0 : )
2212 0 : .await?;
2213 0 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
2214 0 : Result::<(), QueryError>::Ok(())
2215 0 : }
2216 : .await;
2217 : metric_recording.observe(&res);
2218 : res?;
2219 : }
2220 : // same as basebackup, but result includes relational data as well
2221 : PageServiceCmd::FullBackup(FullBackupCmd {
2222 : tenant_id,
2223 : timeline_id,
2224 : lsn,
2225 : prev_lsn,
2226 : }) => {
2227 : tracing::Span::current()
2228 : .record("tenant_id", field::display(tenant_id))
2229 : .record("timeline_id", field::display(timeline_id));
2230 :
2231 : self.check_permission(Some(tenant_id))?;
2232 :
2233 : COMPUTE_COMMANDS_COUNTERS
2234 : .for_command(ComputeCommandKind::Fullbackup)
2235 : .inc();
2236 :
2237 : // Check that the timeline exists
2238 : self.handle_basebackup_request(
2239 : pgb,
2240 : tenant_id,
2241 : timeline_id,
2242 : lsn,
2243 : prev_lsn,
2244 : true,
2245 : false,
2246 : false,
2247 : &ctx,
2248 : )
2249 : .await?;
2250 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
2251 : }
2252 : PageServiceCmd::Set => {
2253 : // important because psycopg2 executes "SET datestyle TO 'ISO'"
2254 : // on connect
2255 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?;
2256 : }
2257 : PageServiceCmd::LeaseLsn(LeaseLsnCmd {
2258 : tenant_shard_id,
2259 : timeline_id,
2260 : lsn,
2261 : }) => {
2262 : tracing::Span::current()
2263 : .record("tenant_id", field::display(tenant_shard_id))
2264 : .record("timeline_id", field::display(timeline_id));
2265 :
2266 : self.check_permission(Some(tenant_shard_id.tenant_id))?;
2267 :
2268 : COMPUTE_COMMANDS_COUNTERS
2269 : .for_command(ComputeCommandKind::LeaseLsn)
2270 : .inc();
2271 :
2272 : match self
2273 : .handle_make_lsn_lease(pgb, tenant_shard_id, timeline_id, lsn, &ctx)
2274 : .await
2275 : {
2276 : Ok(()) => {
2277 : pgb.write_message_noflush(&BeMessage::CommandComplete(b"SELECT 1"))?
2278 : }
2279 : Err(e) => {
2280 : error!("error obtaining lsn lease for {lsn}: {e:?}");
2281 : pgb.write_message_noflush(&BeMessage::ErrorResponse(
2282 : &e.to_string(),
2283 : Some(e.pg_error_code()),
2284 : ))?
2285 : }
2286 : };
2287 : }
2288 : }
2289 :
2290 : Ok(())
2291 : }
2292 : }
2293 :
2294 : impl From<GetActiveTenantError> for QueryError {
2295 0 : fn from(e: GetActiveTenantError) -> Self {
2296 0 : match e {
2297 0 : GetActiveTenantError::WaitForActiveTimeout { .. } => QueryError::Disconnected(
2298 0 : ConnectionError::Io(io::Error::new(io::ErrorKind::TimedOut, e.to_string())),
2299 0 : ),
2300 : GetActiveTenantError::Cancelled
2301 : | GetActiveTenantError::WillNotBecomeActive(TenantState::Stopping { .. }) => {
2302 0 : QueryError::Shutdown
2303 : }
2304 0 : e @ GetActiveTenantError::NotFound(_) => QueryError::NotFound(format!("{e}").into()),
2305 0 : e => QueryError::Other(anyhow::anyhow!(e)),
2306 : }
2307 0 : }
2308 : }
2309 :
2310 : #[derive(Debug, thiserror::Error)]
2311 : pub(crate) enum GetActiveTimelineError {
2312 : #[error(transparent)]
2313 : Tenant(GetActiveTenantError),
2314 : #[error(transparent)]
2315 : Timeline(#[from] GetTimelineError),
2316 : }
2317 :
2318 : impl From<GetActiveTimelineError> for QueryError {
2319 0 : fn from(e: GetActiveTimelineError) -> Self {
2320 0 : match e {
2321 0 : GetActiveTimelineError::Tenant(GetActiveTenantError::Cancelled) => QueryError::Shutdown,
2322 0 : GetActiveTimelineError::Tenant(e) => e.into(),
2323 0 : GetActiveTimelineError::Timeline(e) => QueryError::NotFound(format!("{e}").into()),
2324 : }
2325 0 : }
2326 : }
2327 :
2328 0 : fn set_tracing_field_shard_id(timeline: &Timeline) {
2329 0 : debug_assert_current_span_has_tenant_and_timeline_id_no_shard_id();
2330 0 : tracing::Span::current().record(
2331 0 : "shard_id",
2332 0 : tracing::field::display(timeline.tenant_shard_id.shard_slug()),
2333 0 : );
2334 0 : debug_assert_current_span_has_tenant_and_timeline_id();
2335 0 : }
2336 :
2337 : struct WaitedForLsn(Lsn);
2338 : impl From<WaitedForLsn> for Lsn {
2339 0 : fn from(WaitedForLsn(lsn): WaitedForLsn) -> Self {
2340 0 : lsn
2341 0 : }
2342 : }
2343 :
2344 : #[cfg(test)]
2345 : mod tests {
2346 : use utils::shard::ShardCount;
2347 :
2348 : use super::*;
2349 :
2350 : #[test]
2351 2 : fn pageservice_cmd_parse() {
2352 2 : let tenant_id = TenantId::generate();
2353 2 : let timeline_id = TimelineId::generate();
2354 2 : let cmd =
2355 2 : PageServiceCmd::parse(&format!("pagestream_v2 {tenant_id} {timeline_id}")).unwrap();
2356 2 : assert_eq!(
2357 2 : cmd,
2358 2 : PageServiceCmd::PageStream(PageStreamCmd {
2359 2 : tenant_id,
2360 2 : timeline_id
2361 2 : })
2362 2 : );
2363 2 : let cmd = PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id}")).unwrap();
2364 2 : assert_eq!(
2365 2 : cmd,
2366 2 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2367 2 : tenant_id,
2368 2 : timeline_id,
2369 2 : lsn: None,
2370 2 : gzip: false,
2371 2 : replica: false
2372 2 : })
2373 2 : );
2374 2 : let cmd =
2375 2 : PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id} --gzip")).unwrap();
2376 2 : assert_eq!(
2377 2 : cmd,
2378 2 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2379 2 : tenant_id,
2380 2 : timeline_id,
2381 2 : lsn: None,
2382 2 : gzip: true,
2383 2 : replica: false
2384 2 : })
2385 2 : );
2386 2 : let cmd =
2387 2 : PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id} latest")).unwrap();
2388 2 : assert_eq!(
2389 2 : cmd,
2390 2 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2391 2 : tenant_id,
2392 2 : timeline_id,
2393 2 : lsn: None,
2394 2 : gzip: false,
2395 2 : replica: false
2396 2 : })
2397 2 : );
2398 2 : let cmd = PageServiceCmd::parse(&format!("basebackup {tenant_id} {timeline_id} 0/16ABCDE"))
2399 2 : .unwrap();
2400 2 : assert_eq!(
2401 2 : cmd,
2402 2 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2403 2 : tenant_id,
2404 2 : timeline_id,
2405 2 : lsn: Some(Lsn::from_str("0/16ABCDE").unwrap()),
2406 2 : gzip: false,
2407 2 : replica: false
2408 2 : })
2409 2 : );
2410 2 : let cmd = PageServiceCmd::parse(&format!(
2411 2 : "basebackup {tenant_id} {timeline_id} --replica --gzip"
2412 2 : ))
2413 2 : .unwrap();
2414 2 : assert_eq!(
2415 2 : cmd,
2416 2 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2417 2 : tenant_id,
2418 2 : timeline_id,
2419 2 : lsn: None,
2420 2 : gzip: true,
2421 2 : replica: true
2422 2 : })
2423 2 : );
2424 2 : let cmd = PageServiceCmd::parse(&format!(
2425 2 : "basebackup {tenant_id} {timeline_id} 0/16ABCDE --replica --gzip"
2426 2 : ))
2427 2 : .unwrap();
2428 2 : assert_eq!(
2429 2 : cmd,
2430 2 : PageServiceCmd::BaseBackup(BaseBackupCmd {
2431 2 : tenant_id,
2432 2 : timeline_id,
2433 2 : lsn: Some(Lsn::from_str("0/16ABCDE").unwrap()),
2434 2 : gzip: true,
2435 2 : replica: true
2436 2 : })
2437 2 : );
2438 2 : let cmd = PageServiceCmd::parse(&format!("fullbackup {tenant_id} {timeline_id}")).unwrap();
2439 2 : assert_eq!(
2440 2 : cmd,
2441 2 : PageServiceCmd::FullBackup(FullBackupCmd {
2442 2 : tenant_id,
2443 2 : timeline_id,
2444 2 : lsn: None,
2445 2 : prev_lsn: None
2446 2 : })
2447 2 : );
2448 2 : let cmd = PageServiceCmd::parse(&format!(
2449 2 : "fullbackup {tenant_id} {timeline_id} 0/16ABCDE 0/16ABCDF"
2450 2 : ))
2451 2 : .unwrap();
2452 2 : assert_eq!(
2453 2 : cmd,
2454 2 : PageServiceCmd::FullBackup(FullBackupCmd {
2455 2 : tenant_id,
2456 2 : timeline_id,
2457 2 : lsn: Some(Lsn::from_str("0/16ABCDE").unwrap()),
2458 2 : prev_lsn: Some(Lsn::from_str("0/16ABCDF").unwrap()),
2459 2 : })
2460 2 : );
2461 2 : let tenant_shard_id = TenantShardId::unsharded(tenant_id);
2462 2 : let cmd = PageServiceCmd::parse(&format!(
2463 2 : "lease lsn {tenant_shard_id} {timeline_id} 0/16ABCDE"
2464 2 : ))
2465 2 : .unwrap();
2466 2 : assert_eq!(
2467 2 : cmd,
2468 2 : PageServiceCmd::LeaseLsn(LeaseLsnCmd {
2469 2 : tenant_shard_id,
2470 2 : timeline_id,
2471 2 : lsn: Lsn::from_str("0/16ABCDE").unwrap(),
2472 2 : })
2473 2 : );
2474 2 : let tenant_shard_id = TenantShardId::split(&tenant_shard_id, ShardCount(8))[1];
2475 2 : let cmd = PageServiceCmd::parse(&format!(
2476 2 : "lease lsn {tenant_shard_id} {timeline_id} 0/16ABCDE"
2477 2 : ))
2478 2 : .unwrap();
2479 2 : assert_eq!(
2480 2 : cmd,
2481 2 : PageServiceCmd::LeaseLsn(LeaseLsnCmd {
2482 2 : tenant_shard_id,
2483 2 : timeline_id,
2484 2 : lsn: Lsn::from_str("0/16ABCDE").unwrap(),
2485 2 : })
2486 2 : );
2487 2 : let cmd = PageServiceCmd::parse("set a = b").unwrap();
2488 2 : assert_eq!(cmd, PageServiceCmd::Set);
2489 2 : let cmd = PageServiceCmd::parse("SET foo").unwrap();
2490 2 : assert_eq!(cmd, PageServiceCmd::Set);
2491 2 : }
2492 :
2493 : #[test]
2494 2 : fn pageservice_cmd_err_handling() {
2495 2 : let tenant_id = TenantId::generate();
2496 2 : let timeline_id = TimelineId::generate();
2497 2 : let cmd = PageServiceCmd::parse("unknown_command");
2498 2 : assert!(cmd.is_err());
2499 2 : let cmd = PageServiceCmd::parse("pagestream_v2");
2500 2 : assert!(cmd.is_err());
2501 2 : let cmd = PageServiceCmd::parse(&format!("pagestream_v2 {tenant_id}xxx"));
2502 2 : assert!(cmd.is_err());
2503 2 : let cmd = PageServiceCmd::parse(&format!("pagestream_v2 {tenant_id}xxx {timeline_id}xxx"));
2504 2 : assert!(cmd.is_err());
2505 2 : let cmd = PageServiceCmd::parse(&format!(
2506 2 : "basebackup {tenant_id} {timeline_id} --gzip --gzip"
2507 2 : ));
2508 2 : assert!(cmd.is_err());
2509 2 : let cmd = PageServiceCmd::parse(&format!(
2510 2 : "basebackup {tenant_id} {timeline_id} --gzip --unknown"
2511 2 : ));
2512 2 : assert!(cmd.is_err());
2513 2 : let cmd = PageServiceCmd::parse(&format!(
2514 2 : "basebackup {tenant_id} {timeline_id} --gzip 0/16ABCDE"
2515 2 : ));
2516 2 : assert!(cmd.is_err());
2517 2 : let cmd = PageServiceCmd::parse(&format!("lease {tenant_id} {timeline_id} gzip 0/16ABCDE"));
2518 2 : assert!(cmd.is_err());
2519 2 : }
2520 : }
|
