Line data Source code
1 : use std::collections::{HashMap, HashSet, VecDeque};
2 : use std::fmt::Debug;
3 : use std::sync::atomic::AtomicU32;
4 : use std::sync::Arc;
5 :
6 : use super::remote_timeline_client::is_same_remote_layer_path;
7 : use super::storage_layer::AsLayerDesc as _;
8 : use super::storage_layer::LayerName;
9 : use super::storage_layer::ResidentLayer;
10 : use crate::tenant::metadata::TimelineMetadata;
11 : use crate::tenant::remote_timeline_client::index::IndexPart;
12 : use crate::tenant::remote_timeline_client::index::LayerFileMetadata;
13 : use utils::generation::Generation;
14 : use utils::lsn::{AtomicLsn, Lsn};
15 :
16 : use chrono::NaiveDateTime;
17 : use once_cell::sync::Lazy;
18 : use tracing::info;
19 :
20 : /// Kill switch for upload queue reordering in case it causes problems.
21 : /// TODO: remove this once we have confidence in it.
22 : static DISABLE_UPLOAD_QUEUE_REORDERING: Lazy<bool> =
23 388 : Lazy::new(|| std::env::var("DISABLE_UPLOAD_QUEUE_REORDERING").as_deref() == Ok("true"));
24 :
25 : /// Kill switch for index upload coalescing in case it causes problems.
26 : /// TODO: remove this once we have confidence in it.
27 : static DISABLE_UPLOAD_QUEUE_INDEX_COALESCING: Lazy<bool> =
28 15 : Lazy::new(|| std::env::var("DISABLE_UPLOAD_QUEUE_INDEX_COALESCING").as_deref() == Ok("true"));
29 :
30 : // clippy warns that Uninitialized is much smaller than Initialized, which wastes
31 : // memory for Uninitialized variants. Doesn't matter in practice, there are not
32 : // that many upload queues in a running pageserver, and most of them are initialized
33 : // anyway.
34 : #[allow(clippy::large_enum_variant)]
35 : pub enum UploadQueue {
36 : Uninitialized,
37 : Initialized(UploadQueueInitialized),
38 : Stopped(UploadQueueStopped),
39 : }
40 :
41 : impl UploadQueue {
42 0 : pub fn as_str(&self) -> &'static str {
43 0 : match self {
44 0 : UploadQueue::Uninitialized => "Uninitialized",
45 0 : UploadQueue::Initialized(_) => "Initialized",
46 0 : UploadQueue::Stopped(_) => "Stopped",
47 : }
48 0 : }
49 : }
50 :
51 : #[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
52 : pub enum OpType {
53 : MayReorder,
54 : FlushDeletion,
55 : }
56 :
57 : /// This keeps track of queued and in-progress tasks.
58 : pub struct UploadQueueInitialized {
59 : /// Maximum number of inprogress tasks to schedule. 0 is no limit.
60 : pub(crate) inprogress_limit: usize,
61 :
62 : /// Counter to assign task IDs
63 : pub(crate) task_counter: u64,
64 :
65 : /// The next uploaded index_part.json; assumed to be dirty.
66 : ///
67 : /// Should not be read, directly except for layer file updates. Instead you should add a
68 : /// projected field.
69 : pub(crate) dirty: IndexPart,
70 :
71 : /// The latest remote persisted IndexPart.
72 : ///
73 : /// Each completed metadata upload will update this. The second item is the task_id which last
74 : /// updated the value, used to ensure we never store an older value over a newer one.
75 : pub(crate) clean: (IndexPart, Option<u64>),
76 :
77 : /// How many file uploads or deletions been scheduled, since the
78 : /// last (scheduling of) metadata index upload?
79 : pub(crate) latest_files_changes_since_metadata_upload_scheduled: u64,
80 :
81 : /// The Lsn is only updated after our generation has been validated with
82 : /// the control plane (unlesss a timeline's generation is None, in which case
83 : /// we skip validation)
84 : pub(crate) visible_remote_consistent_lsn: Arc<AtomicLsn>,
85 :
86 : /// Tasks that are currently in-progress. In-progress means that a tokio Task
87 : /// has been launched for it. An in-progress task can be busy uploading, but it can
88 : /// also be waiting on the `concurrency_limiter` Semaphore in S3Bucket, or it can
89 : /// be waiting for retry in `exponential_backoff`.
90 : pub inprogress_tasks: HashMap<u64, Arc<UploadTask>>,
91 :
92 : /// Queued operations that have not been launched yet. They might depend on previous
93 : /// tasks to finish. For example, metadata upload cannot be performed before all
94 : /// preceding layer file uploads have completed.
95 : pub queued_operations: VecDeque<UploadOp>,
96 :
97 : /// Files which have been unlinked but not yet had scheduled a deletion for. Only kept around
98 : /// for error logging.
99 : ///
100 : /// Putting this behind a testing feature to catch problems in tests, but assuming we could have a
101 : /// bug causing leaks, then it's better to not leave this enabled for production builds.
102 : #[cfg(feature = "testing")]
103 : pub(crate) dangling_files: HashMap<LayerName, Generation>,
104 :
105 : /// Ensure we order file operations correctly.
106 : pub(crate) recently_deleted: HashSet<(LayerName, Generation)>,
107 :
108 : /// Deletions that are blocked by the tenant configuration
109 : pub(crate) blocked_deletions: Vec<Delete>,
110 :
111 : /// Set to true when we have inserted the `UploadOp::Shutdown` into the `inprogress_tasks`.
112 : pub(crate) shutting_down: bool,
113 :
114 : /// Permitless semaphore on which any number of `RemoteTimelineClient::shutdown` futures can
115 : /// wait on until one of them stops the queue. The semaphore is closed when
116 : /// `RemoteTimelineClient::launch_queued_tasks` encounters `UploadOp::Shutdown`.
117 : pub(crate) shutdown_ready: Arc<tokio::sync::Semaphore>,
118 : }
119 :
120 : impl UploadQueueInitialized {
121 16 : pub(super) fn no_pending_work(&self) -> bool {
122 16 : self.inprogress_tasks.is_empty() && self.queued_operations.is_empty()
123 16 : }
124 :
125 0 : pub(super) fn get_last_remote_consistent_lsn_visible(&self) -> Lsn {
126 0 : self.visible_remote_consistent_lsn.load()
127 0 : }
128 :
129 0 : pub(super) fn get_last_remote_consistent_lsn_projected(&self) -> Option<Lsn> {
130 0 : let lsn = self.clean.0.metadata.disk_consistent_lsn();
131 0 : self.clean.1.map(|_| lsn)
132 0 : }
133 :
134 : /// Returns and removes the next ready operation from the queue, if any. This isn't necessarily
135 : /// the first operation in the queue, to avoid head-of-line blocking -- an operation can jump
136 : /// the queue if it doesn't conflict with operations ahead of it.
137 : ///
138 : /// Also returns any operations that were coalesced into this one, e.g. multiple index uploads.
139 : ///
140 : /// None may be returned even if the queue isn't empty, if no operations are ready yet.
141 : ///
142 : /// NB: this is quadratic, but queues are expected to be small, and bounded by inprogress_limit.
143 26827 : pub fn next_ready(&mut self) -> Option<(UploadOp, Vec<UploadOp>)> {
144 26827 : // If inprogress_tasks is already at limit, don't schedule anything more.
145 26827 : if self.inprogress_limit > 0 && self.inprogress_tasks.len() >= self.inprogress_limit {
146 12 : return None;
147 26815 : }
148 :
149 51127 : for (i, candidate) in self.queued_operations.iter().enumerate() {
150 : // If this candidate is ready, go for it. Otherwise, try the next one.
151 51127 : if self.is_ready(i) {
152 : // Shutdown operations are left at the head of the queue, to prevent further
153 : // operations from starting. Signal that we're ready to shut down.
154 10162 : if matches!(candidate, UploadOp::Shutdown) {
155 20 : assert!(self.inprogress_tasks.is_empty(), "shutdown with tasks");
156 20 : assert_eq!(i, 0, "shutdown not at head of queue");
157 20 : self.shutdown_ready.close();
158 20 : return None;
159 10142 : }
160 10142 :
161 10142 : let mut op = self.queued_operations.remove(i).expect("i can't disappear");
162 10142 :
163 10142 : // Coalesce any back-to-back index uploads by only uploading the newest one that's
164 10142 : // ready. This typically happens with layer/index/layer/index/... sequences, where
165 10142 : // the layers bypass the indexes, leaving the indexes queued.
166 10142 : //
167 10142 : // If other operations are interleaved between index uploads we don't try to
168 10142 : // coalesce them, since we may as well update the index concurrently with them.
169 10142 : // This keeps the index fresh and avoids starvation.
170 10142 : //
171 10142 : // NB: we assume that all uploaded indexes have the same remote path. This
172 10142 : // is true at the time of writing: the path only depends on the tenant,
173 10142 : // timeline and generation, all of which are static for a timeline instance.
174 10142 : // Otherwise, we must be careful not to coalesce different paths.
175 10142 : let mut coalesced_ops = Vec::new();
176 10142 : if matches!(op, UploadOp::UploadMetadata { .. }) {
177 3071 : while let Some(UploadOp::UploadMetadata { .. }) = self.queued_operations.get(i)
178 : {
179 31 : if *DISABLE_UPLOAD_QUEUE_INDEX_COALESCING {
180 0 : break;
181 31 : }
182 31 : if !self.is_ready(i) {
183 11 : break;
184 20 : }
185 20 : coalesced_ops.push(op);
186 20 : op = self.queued_operations.remove(i).expect("i can't disappear");
187 : }
188 7091 : }
189 :
190 10142 : return Some((op, coalesced_ops));
191 40965 : }
192 :
193 : // Nothing can bypass a barrier or shutdown. If it wasn't scheduled above, give up.
194 40965 : if matches!(candidate, UploadOp::Barrier(_) | UploadOp::Shutdown) {
195 4779 : return None;
196 36186 : }
197 36186 :
198 36186 : // If upload queue reordering is disabled, bail out after the first operation.
199 36186 : if *DISABLE_UPLOAD_QUEUE_REORDERING {
200 0 : return None;
201 36186 : }
202 : }
203 11874 : None
204 26827 : }
205 :
206 : /// Returns true if the queued operation at the given position is ready to be uploaded, i.e. if
207 : /// it doesn't conflict with any in-progress or queued operations ahead of it. Operations are
208 : /// allowed to skip the queue when it's safe to do so, to increase parallelism.
209 : ///
210 : /// The position must be valid for the queue size.
211 51158 : fn is_ready(&self, pos: usize) -> bool {
212 51158 : let candidate = self.queued_operations.get(pos).expect("invalid position");
213 51158 : self
214 51158 : // Look at in-progress operations, in random order.
215 51158 : .inprogress_tasks
216 51158 : .values()
217 2334420 : .map(|task| &task.op)
218 51158 : // Then queued operations ahead of the candidate, front-to-back.
219 51158 : .chain(self.queued_operations.iter().take(pos))
220 51158 : // Keep track of the active index ahead of each operation. This is used to ensure that
221 51158 : // an upload doesn't skip the queue too far, such that it modifies a layer that's
222 51158 : // referenced by an active index.
223 51158 : //
224 51158 : // It's okay that in-progress operations are emitted in random order above, since at
225 51158 : // most one of them can be an index upload (enforced by can_bypass).
226 2362423 : .scan(&self.clean.0, |next_active_index, op| {
227 2362423 : let active_index = *next_active_index;
228 2362423 : if let UploadOp::UploadMetadata { ref uploaded } = op {
229 32105 : *next_active_index = uploaded; // stash index for next operation after this
230 2330318 : }
231 2362423 : Some((op, active_index))
232 2362423 : })
233 51158 : // Check if the candidate can bypass all of them.
234 2362423 : .all(|(op, active_index)| candidate.can_bypass(op, active_index))
235 51158 : }
236 :
237 : /// Returns the number of in-progress deletion operations.
238 : #[cfg(test)]
239 4 : pub(crate) fn num_inprogress_deletions(&self) -> usize {
240 4 : self.inprogress_tasks
241 4 : .iter()
242 4 : .filter(|(_, t)| matches!(t.op, UploadOp::Delete(_)))
243 4 : .count()
244 4 : }
245 :
246 : /// Returns the number of in-progress layer uploads.
247 : #[cfg(test)]
248 8 : pub(crate) fn num_inprogress_layer_uploads(&self) -> usize {
249 8 : self.inprogress_tasks
250 8 : .iter()
251 12 : .filter(|(_, t)| matches!(t.op, UploadOp::UploadLayer(_, _, _)))
252 8 : .count()
253 8 : }
254 :
255 : /// Test helper that schedules all ready operations into inprogress_tasks, and returns
256 : /// references to them.
257 : ///
258 : /// TODO: the corresponding production logic should be moved from RemoteTimelineClient into
259 : /// UploadQueue, so we can use the same code path.
260 : #[cfg(test)]
261 156 : fn schedule_ready(&mut self) -> Vec<Arc<UploadTask>> {
262 156 : let mut tasks = Vec::new();
263 : // NB: schedule operations one by one, to handle conflicts with inprogress_tasks.
264 344 : while let Some((op, coalesced_ops)) = self.next_ready() {
265 188 : self.task_counter += 1;
266 188 : let task = Arc::new(UploadTask {
267 188 : task_id: self.task_counter,
268 188 : op,
269 188 : coalesced_ops,
270 188 : retries: 0.into(),
271 188 : });
272 188 : self.inprogress_tasks.insert(task.task_id, task.clone());
273 188 : tasks.push(task);
274 188 : }
275 156 : tasks
276 156 : }
277 :
278 : /// Test helper that marks an operation as completed, removing it from inprogress_tasks.
279 : ///
280 : /// TODO: the corresponding production logic should be moved from RemoteTimelineClient into
281 : /// UploadQueue, so we can use the same code path.
282 : #[cfg(test)]
283 116 : fn complete(&mut self, task_id: u64) {
284 116 : let Some(task) = self.inprogress_tasks.remove(&task_id) else {
285 0 : return;
286 : };
287 : // Update the clean index on uploads.
288 116 : if let UploadOp::UploadMetadata { ref uploaded } = task.op {
289 32 : if task.task_id > self.clean.1.unwrap_or_default() {
290 32 : self.clean = (*uploaded.clone(), Some(task.task_id));
291 32 : }
292 84 : }
293 116 : }
294 : }
295 :
296 : #[derive(Clone, Copy)]
297 : pub(super) enum SetDeletedFlagProgress {
298 : NotRunning,
299 : InProgress(NaiveDateTime),
300 : Successful(NaiveDateTime),
301 : }
302 :
303 : pub struct UploadQueueStoppedDeletable {
304 : pub(super) upload_queue_for_deletion: UploadQueueInitialized,
305 : pub(super) deleted_at: SetDeletedFlagProgress,
306 : }
307 :
308 : pub enum UploadQueueStopped {
309 : Deletable(UploadQueueStoppedDeletable),
310 : Uninitialized,
311 : }
312 :
313 : #[derive(thiserror::Error, Debug)]
314 : pub enum NotInitialized {
315 : #[error("queue is in state Uninitialized")]
316 : Uninitialized,
317 : #[error("queue is in state Stopped")]
318 : Stopped,
319 : #[error("queue is shutting down")]
320 : ShuttingDown,
321 : }
322 :
323 : impl NotInitialized {
324 0 : pub(crate) fn is_stopping(&self) -> bool {
325 : use NotInitialized::*;
326 0 : match self {
327 0 : Uninitialized => false,
328 0 : Stopped => true,
329 0 : ShuttingDown => true,
330 : }
331 0 : }
332 : }
333 :
334 : impl UploadQueue {
335 900 : pub fn initialize_empty_remote(
336 900 : &mut self,
337 900 : metadata: &TimelineMetadata,
338 900 : inprogress_limit: usize,
339 900 : ) -> anyhow::Result<&mut UploadQueueInitialized> {
340 900 : match self {
341 900 : UploadQueue::Uninitialized => (),
342 : UploadQueue::Initialized(_) | UploadQueue::Stopped(_) => {
343 0 : anyhow::bail!("already initialized, state {}", self.as_str())
344 : }
345 : }
346 :
347 900 : info!("initializing upload queue for empty remote");
348 :
349 900 : let index_part = IndexPart::empty(metadata.clone());
350 900 :
351 900 : let state = UploadQueueInitialized {
352 900 : inprogress_limit,
353 900 : dirty: index_part.clone(),
354 900 : clean: (index_part, None),
355 900 : latest_files_changes_since_metadata_upload_scheduled: 0,
356 900 : visible_remote_consistent_lsn: Arc::new(AtomicLsn::new(0)),
357 900 : // what follows are boring default initializations
358 900 : task_counter: 0,
359 900 : inprogress_tasks: HashMap::new(),
360 900 : queued_operations: VecDeque::new(),
361 900 : #[cfg(feature = "testing")]
362 900 : dangling_files: HashMap::new(),
363 900 : recently_deleted: HashSet::new(),
364 900 : blocked_deletions: Vec::new(),
365 900 : shutting_down: false,
366 900 : shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
367 900 : };
368 900 :
369 900 : *self = UploadQueue::Initialized(state);
370 900 : Ok(self.initialized_mut().expect("we just set it"))
371 900 : }
372 :
373 44 : pub fn initialize_with_current_remote_index_part(
374 44 : &mut self,
375 44 : index_part: &IndexPart,
376 44 : inprogress_limit: usize,
377 44 : ) -> anyhow::Result<&mut UploadQueueInitialized> {
378 44 : match self {
379 44 : UploadQueue::Uninitialized => (),
380 : UploadQueue::Initialized(_) | UploadQueue::Stopped(_) => {
381 0 : anyhow::bail!("already initialized, state {}", self.as_str())
382 : }
383 : }
384 :
385 44 : info!(
386 0 : "initializing upload queue with remote index_part.disk_consistent_lsn: {}",
387 0 : index_part.metadata.disk_consistent_lsn()
388 : );
389 :
390 44 : let state = UploadQueueInitialized {
391 44 : inprogress_limit,
392 44 : dirty: index_part.clone(),
393 44 : clean: (index_part.clone(), None),
394 44 : latest_files_changes_since_metadata_upload_scheduled: 0,
395 44 : visible_remote_consistent_lsn: Arc::new(
396 44 : index_part.metadata.disk_consistent_lsn().into(),
397 44 : ),
398 44 : // what follows are boring default initializations
399 44 : task_counter: 0,
400 44 : inprogress_tasks: HashMap::new(),
401 44 : queued_operations: VecDeque::new(),
402 44 : #[cfg(feature = "testing")]
403 44 : dangling_files: HashMap::new(),
404 44 : recently_deleted: HashSet::new(),
405 44 : blocked_deletions: Vec::new(),
406 44 : shutting_down: false,
407 44 : shutdown_ready: Arc::new(tokio::sync::Semaphore::new(0)),
408 44 : };
409 44 :
410 44 : *self = UploadQueue::Initialized(state);
411 44 : Ok(self.initialized_mut().expect("we just set it"))
412 44 : }
413 :
414 18320 : pub fn initialized_mut(&mut self) -> Result<&mut UploadQueueInitialized, NotInitialized> {
415 : use UploadQueue::*;
416 18320 : match self {
417 0 : Uninitialized => Err(NotInitialized::Uninitialized),
418 18320 : Initialized(x) => {
419 18320 : if x.shutting_down {
420 0 : Err(NotInitialized::ShuttingDown)
421 : } else {
422 18320 : Ok(x)
423 : }
424 : }
425 0 : Stopped(_) => Err(NotInitialized::Stopped),
426 : }
427 18320 : }
428 :
429 4 : pub(crate) fn stopped_mut(&mut self) -> anyhow::Result<&mut UploadQueueStoppedDeletable> {
430 4 : match self {
431 : UploadQueue::Initialized(_) | UploadQueue::Uninitialized => {
432 0 : anyhow::bail!("queue is in state {}", self.as_str())
433 : }
434 : UploadQueue::Stopped(UploadQueueStopped::Uninitialized) => {
435 0 : anyhow::bail!("queue is in state Stopped(Uninitialized)")
436 : }
437 4 : UploadQueue::Stopped(UploadQueueStopped::Deletable(deletable)) => Ok(deletable),
438 : }
439 4 : }
440 : }
441 :
442 : /// An in-progress upload or delete task.
443 : #[derive(Debug)]
444 : pub struct UploadTask {
445 : /// Unique ID of this task. Used as the key in `inprogress_tasks` above.
446 : pub task_id: u64,
447 : /// Number of task retries.
448 : pub retries: AtomicU32,
449 : /// The upload operation.
450 : pub op: UploadOp,
451 : /// Any upload operations that were coalesced into this operation. This typically happens with
452 : /// back-to-back index uploads, see `UploadQueueInitialized::next_ready()`.
453 : pub coalesced_ops: Vec<UploadOp>,
454 : }
455 :
456 : /// A deletion of some layers within the lifetime of a timeline. This is not used
457 : /// for timeline deletion, which skips this queue and goes directly to DeletionQueue.
458 : #[derive(Debug, Clone)]
459 : pub struct Delete {
460 : pub layers: Vec<(LayerName, LayerFileMetadata)>,
461 : }
462 :
463 : #[derive(Clone, Debug)]
464 : pub enum UploadOp {
465 : /// Upload a layer file. The last field indicates the last operation for thie file.
466 : UploadLayer(ResidentLayer, LayerFileMetadata, Option<OpType>),
467 :
468 : /// Upload a index_part.json file
469 : UploadMetadata {
470 : /// The next [`UploadQueueInitialized::clean`] after this upload succeeds.
471 : uploaded: Box<IndexPart>,
472 : },
473 :
474 : /// Delete layer files
475 : Delete(Delete),
476 :
477 : /// Barrier. When the barrier operation is reached, the channel is closed.
478 : Barrier(tokio::sync::watch::Sender<()>),
479 :
480 : /// Shutdown; upon encountering this operation no new operations will be spawned, otherwise
481 : /// this is the same as a Barrier.
482 : Shutdown,
483 : }
484 :
485 : impl std::fmt::Display for UploadOp {
486 0 : fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
487 0 : match self {
488 0 : UploadOp::UploadLayer(layer, metadata, mode) => {
489 0 : write!(
490 0 : f,
491 0 : "UploadLayer({}, size={:?}, gen={:?}, mode={:?})",
492 0 : layer, metadata.file_size, metadata.generation, mode
493 0 : )
494 : }
495 0 : UploadOp::UploadMetadata { uploaded, .. } => {
496 0 : write!(
497 0 : f,
498 0 : "UploadMetadata(lsn: {})",
499 0 : uploaded.metadata.disk_consistent_lsn()
500 0 : )
501 : }
502 0 : UploadOp::Delete(delete) => {
503 0 : write!(f, "Delete({} layers)", delete.layers.len())
504 : }
505 0 : UploadOp::Barrier(_) => write!(f, "Barrier"),
506 0 : UploadOp::Shutdown => write!(f, "Shutdown"),
507 : }
508 0 : }
509 : }
510 :
511 : impl UploadOp {
512 : /// Returns true if self can bypass other, i.e. if the operations don't conflict. index is the
513 : /// active index when other would be uploaded -- if we allow self to bypass other, this would
514 : /// be the active index when self is uploaded.
515 2362519 : pub fn can_bypass(&self, other: &UploadOp, index: &IndexPart) -> bool {
516 2362519 : match (self, other) {
517 : // Nothing can bypass a barrier or shutdown, and it can't bypass anything.
518 4795 : (UploadOp::Barrier(_), _) | (_, UploadOp::Barrier(_)) => false,
519 16 : (UploadOp::Shutdown, _) | (_, UploadOp::Shutdown) => false,
520 :
521 : // Uploads and deletes can bypass each other unless they're for the same file.
522 40047 : (UploadOp::UploadLayer(a, ameta, _), UploadOp::UploadLayer(b, bmeta, _)) => {
523 40047 : let aname = &a.layer_desc().layer_name();
524 40047 : let bname = &b.layer_desc().layer_name();
525 40047 : !is_same_remote_layer_path(aname, ameta, bname, bmeta)
526 : }
527 84 : (UploadOp::UploadLayer(u, umeta, _), UploadOp::Delete(d))
528 2275000 : | (UploadOp::Delete(d), UploadOp::UploadLayer(u, umeta, _)) => {
529 2275114 : d.layers.iter().all(|(dname, dmeta)| {
530 2275114 : !is_same_remote_layer_path(&u.layer_desc().layer_name(), umeta, dname, dmeta)
531 2275114 : })
532 : }
533 :
534 : // Deletes are idempotent and can always bypass each other.
535 5914 : (UploadOp::Delete(_), UploadOp::Delete(_)) => true,
536 :
537 : // Uploads and deletes can bypass an index upload as long as neither the uploaded index
538 : // nor the active index below it references the file. A layer can't be modified or
539 : // deleted while referenced by an index.
540 : //
541 : // Similarly, index uploads can bypass uploads and deletes as long as neither the
542 : // uploaded index nor the active index references the file (the latter would be
543 : // incorrect use by the caller).
544 181 : (UploadOp::UploadLayer(u, umeta, _), UploadOp::UploadMetadata { uploaded: i })
545 6870 : | (UploadOp::UploadMetadata { uploaded: i }, UploadOp::UploadLayer(u, umeta, _)) => {
546 7051 : let uname = u.layer_desc().layer_name();
547 7051 : !i.references(&uname, umeta) && !index.references(&uname, umeta)
548 : }
549 29286 : (UploadOp::Delete(d), UploadOp::UploadMetadata { uploaded: i })
550 38 : | (UploadOp::UploadMetadata { uploaded: i }, UploadOp::Delete(d)) => {
551 29324 : d.layers.iter().all(|(dname, dmeta)| {
552 29324 : !i.references(dname, dmeta) && !index.references(dname, dmeta)
553 29324 : })
554 : }
555 :
556 : // Indexes can never bypass each other. They can coalesce though, and
557 : // `UploadQueue::next_ready()` currently does this when possible.
558 288 : (UploadOp::UploadMetadata { .. }, UploadOp::UploadMetadata { .. }) => false,
559 : }
560 2362519 : }
561 : }
562 :
563 : #[cfg(test)]
564 : mod tests {
565 : use super::*;
566 : use crate::tenant::harness::{TenantHarness, TIMELINE_ID};
567 : use crate::tenant::storage_layer::layer::local_layer_path;
568 : use crate::tenant::storage_layer::Layer;
569 : use crate::tenant::Timeline;
570 : use crate::DEFAULT_PG_VERSION;
571 : use itertools::Itertools as _;
572 : use std::str::FromStr as _;
573 : use utils::shard::{ShardCount, ShardIndex, ShardNumber};
574 :
575 : /// Test helper which asserts that two operations are the same, in lieu of UploadOp PartialEq.
576 : #[track_caller]
577 196 : fn assert_same_op(a: &UploadOp, b: &UploadOp) {
578 : use UploadOp::*;
579 196 : match (a, b) {
580 88 : (UploadLayer(a, ameta, atype), UploadLayer(b, bmeta, btype)) => {
581 88 : assert_eq!(a.layer_desc().layer_name(), b.layer_desc().layer_name());
582 88 : assert_eq!(ameta, bmeta);
583 88 : assert_eq!(atype, btype);
584 : }
585 44 : (Delete(a), Delete(b)) => assert_eq!(a.layers, b.layers),
586 56 : (UploadMetadata { uploaded: a }, UploadMetadata { uploaded: b }) => assert_eq!(a, b),
587 8 : (Barrier(_), Barrier(_)) => {}
588 0 : (Shutdown, Shutdown) => {}
589 0 : (a, b) => panic!("{a:?} != {b:?}"),
590 : }
591 196 : }
592 :
593 : /// Test helper which asserts that two sets of operations are the same.
594 : #[track_caller]
595 44 : fn assert_same_ops<'a>(
596 44 : a: impl IntoIterator<Item = &'a UploadOp>,
597 44 : b: impl IntoIterator<Item = &'a UploadOp>,
598 44 : ) {
599 44 : a.into_iter()
600 44 : .zip_eq(b)
601 116 : .for_each(|(a, b)| assert_same_op(a, b))
602 44 : }
603 :
604 : /// Test helper to construct a test timeline.
605 : ///
606 : /// TODO: it really shouldn't be necessary to construct an entire tenant and timeline just to
607 : /// test the upload queue -- decouple ResidentLayer from Timeline.
608 : ///
609 : /// TODO: the upload queue uses TimelineMetadata::example() instead, because there's no way to
610 : /// obtain a TimelineMetadata from a Timeline.
611 48 : fn make_timeline() -> Arc<Timeline> {
612 48 : // Grab the current test name from the current thread name.
613 48 : // TODO: TenantHarness shouldn't take a &'static str, but just leak the test name for now.
614 48 : let test_name = std::thread::current().name().unwrap().to_string();
615 48 : let test_name = Box::leak(test_name.into_boxed_str());
616 48 :
617 48 : let runtime = tokio::runtime::Builder::new_current_thread()
618 48 : .enable_all()
619 48 : .build()
620 48 : .expect("failed to create runtime");
621 48 :
622 48 : runtime
623 48 : .block_on(async {
624 48 : let harness = TenantHarness::create(test_name).await?;
625 48 : let (tenant, ctx) = harness.load().await;
626 48 : tenant
627 48 : .create_test_timeline(TIMELINE_ID, Lsn(8), DEFAULT_PG_VERSION, &ctx)
628 48 : .await
629 48 : })
630 48 : .expect("failed to create timeline")
631 48 : }
632 :
633 : /// Test helper to construct an (empty) resident layer.
634 120 : fn make_layer(timeline: &Arc<Timeline>, name: &str) -> ResidentLayer {
635 120 : make_layer_with_size(timeline, name, 0)
636 120 : }
637 :
638 : /// Test helper to construct a resident layer with the given size.
639 132 : fn make_layer_with_size(timeline: &Arc<Timeline>, name: &str, size: usize) -> ResidentLayer {
640 132 : let metadata = LayerFileMetadata {
641 132 : generation: timeline.generation,
642 132 : shard: timeline.get_shard_index(),
643 132 : file_size: size as u64,
644 132 : };
645 132 : make_layer_with_metadata(timeline, name, metadata)
646 132 : }
647 :
648 : /// Test helper to construct a layer with the given metadata.
649 196 : fn make_layer_with_metadata(
650 196 : timeline: &Arc<Timeline>,
651 196 : name: &str,
652 196 : metadata: LayerFileMetadata,
653 196 : ) -> ResidentLayer {
654 196 : let name = LayerName::from_str(name).expect("invalid name");
655 196 : let local_path = local_layer_path(
656 196 : timeline.conf,
657 196 : &timeline.tenant_shard_id,
658 196 : &timeline.timeline_id,
659 196 : &name,
660 196 : &metadata.generation,
661 196 : );
662 196 : std::fs::write(&local_path, vec![0; metadata.file_size as usize])
663 196 : .expect("failed to write file");
664 196 : Layer::for_resident(timeline.conf, timeline, local_path, name, metadata)
665 196 : }
666 :
667 : /// Test helper to add a layer to an index and return a new index.
668 24 : fn index_with(index: &IndexPart, layer: &ResidentLayer) -> Box<IndexPart> {
669 24 : let mut index = index.clone();
670 24 : index
671 24 : .layer_metadata
672 24 : .insert(layer.layer_desc().layer_name(), layer.metadata());
673 24 : Box::new(index)
674 24 : }
675 :
676 : /// Test helper to remove a layer from an index and return a new index.
677 8 : fn index_without(index: &IndexPart, layer: &ResidentLayer) -> Box<IndexPart> {
678 8 : let mut index = index.clone();
679 8 : index
680 8 : .layer_metadata
681 8 : .remove(&layer.layer_desc().layer_name());
682 8 : Box::new(index)
683 8 : }
684 :
685 : /// Nothing can bypass a barrier, and it can't bypass inprogress tasks.
686 : #[test]
687 4 : fn schedule_barrier() -> anyhow::Result<()> {
688 4 : let mut queue = UploadQueue::Uninitialized;
689 4 : let queue = queue.initialize_empty_remote(&TimelineMetadata::example(), 0)?;
690 4 : let tli = make_timeline();
691 4 :
692 4 : let index = Box::new(queue.clean.0.clone()); // empty, doesn't matter
693 4 : let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
694 4 : let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
695 4 : let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
696 4 : let layer3 = make_layer(&tli, "300000000000000000000000000000000000-400000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
697 4 : let (barrier, _) = tokio::sync::watch::channel(());
698 4 :
699 4 : // Enqueue non-conflicting upload, delete, and index before and after a barrier.
700 4 : let ops = [
701 4 : UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None),
702 4 : UploadOp::Delete(Delete {
703 4 : layers: vec![(layer1.layer_desc().layer_name(), layer1.metadata())],
704 4 : }),
705 4 : UploadOp::UploadMetadata {
706 4 : uploaded: index.clone(),
707 4 : },
708 4 : UploadOp::Barrier(barrier),
709 4 : UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None),
710 4 : UploadOp::Delete(Delete {
711 4 : layers: vec![(layer3.layer_desc().layer_name(), layer3.metadata())],
712 4 : }),
713 4 : UploadOp::UploadMetadata {
714 4 : uploaded: index.clone(),
715 4 : },
716 4 : ];
717 4 :
718 4 : queue.queued_operations.extend(ops.clone());
719 4 :
720 4 : // Schedule the initial operations ahead of the barrier.
721 4 : let tasks = queue.schedule_ready();
722 4 :
723 12 : assert_same_ops(tasks.iter().map(|t| &t.op), &ops[0..3]);
724 4 : assert!(matches!(
725 4 : queue.queued_operations.front(),
726 : Some(&UploadOp::Barrier(_))
727 : ));
728 :
729 : // Complete the initial operations. The barrier isn't scheduled while they're pending.
730 16 : for task in tasks {
731 12 : assert!(queue.schedule_ready().is_empty());
732 12 : queue.complete(task.task_id);
733 : }
734 :
735 : // Schedule the barrier. The later tasks won't schedule until it completes.
736 4 : let tasks = queue.schedule_ready();
737 4 :
738 4 : assert_eq!(tasks.len(), 1);
739 4 : assert!(matches!(tasks[0].op, UploadOp::Barrier(_)));
740 4 : assert_eq!(queue.queued_operations.len(), 3);
741 :
742 : // Complete the barrier. The rest of the tasks schedule immediately.
743 4 : queue.complete(tasks[0].task_id);
744 4 :
745 4 : let tasks = queue.schedule_ready();
746 12 : assert_same_ops(tasks.iter().map(|t| &t.op), &ops[4..]);
747 4 : assert!(queue.queued_operations.is_empty());
748 :
749 4 : Ok(())
750 4 : }
751 :
752 : /// Deletes can be scheduled in parallel, even if they're for the same file.
753 : #[test]
754 4 : fn schedule_delete_parallel() -> anyhow::Result<()> {
755 4 : let mut queue = UploadQueue::Uninitialized;
756 4 : let queue = queue.initialize_empty_remote(&TimelineMetadata::example(), 0)?;
757 4 : let tli = make_timeline();
758 4 :
759 4 : // Enqueue a bunch of deletes, some with conflicting names.
760 4 : let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
761 4 : let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
762 4 : let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
763 4 : let layer3 = make_layer(&tli, "300000000000000000000000000000000000-400000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
764 4 :
765 4 : let ops = [
766 4 : UploadOp::Delete(Delete {
767 4 : layers: vec![(layer0.layer_desc().layer_name(), layer0.metadata())],
768 4 : }),
769 4 : UploadOp::Delete(Delete {
770 4 : layers: vec![(layer1.layer_desc().layer_name(), layer1.metadata())],
771 4 : }),
772 4 : UploadOp::Delete(Delete {
773 4 : layers: vec![
774 4 : (layer1.layer_desc().layer_name(), layer1.metadata()),
775 4 : (layer2.layer_desc().layer_name(), layer2.metadata()),
776 4 : ],
777 4 : }),
778 4 : UploadOp::Delete(Delete {
779 4 : layers: vec![(layer2.layer_desc().layer_name(), layer2.metadata())],
780 4 : }),
781 4 : UploadOp::Delete(Delete {
782 4 : layers: vec![(layer3.layer_desc().layer_name(), layer3.metadata())],
783 4 : }),
784 4 : ];
785 4 :
786 4 : queue.queued_operations.extend(ops.clone());
787 4 :
788 4 : // Schedule all ready operations. Since deletes don't conflict, they're all scheduled.
789 4 : let tasks = queue.schedule_ready();
790 4 :
791 20 : assert_same_ops(tasks.iter().map(|t| &t.op), &ops);
792 4 : assert!(queue.queued_operations.is_empty());
793 :
794 4 : Ok(())
795 4 : }
796 :
797 : /// Conflicting uploads are serialized.
798 : #[test]
799 4 : fn schedule_upload_conflicts() -> anyhow::Result<()> {
800 4 : let mut queue = UploadQueue::Uninitialized;
801 4 : let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?;
802 4 : let tli = make_timeline();
803 4 :
804 4 : // Enqueue three versions of the same layer, with different file sizes.
805 4 : let layer0a = make_layer_with_size(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51", 1);
806 4 : let layer0b = make_layer_with_size(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51", 2);
807 4 : let layer0c = make_layer_with_size(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51", 3);
808 4 :
809 4 : let ops = [
810 4 : UploadOp::UploadLayer(layer0a.clone(), layer0a.metadata(), None),
811 4 : UploadOp::UploadLayer(layer0b.clone(), layer0b.metadata(), None),
812 4 : UploadOp::UploadLayer(layer0c.clone(), layer0c.metadata(), None),
813 4 : ];
814 4 :
815 4 : queue.queued_operations.extend(ops.clone());
816 :
817 : // Only one version should be scheduled and uploaded at a time.
818 16 : for op in ops {
819 12 : let tasks = queue.schedule_ready();
820 12 : assert_eq!(tasks.len(), 1);
821 12 : assert_same_op(&tasks[0].op, &op);
822 12 : queue.complete(tasks[0].task_id);
823 : }
824 4 : assert!(queue.schedule_ready().is_empty());
825 4 : assert!(queue.queued_operations.is_empty());
826 :
827 4 : Ok(())
828 4 : }
829 :
830 : /// Conflicting uploads and deletes are serialized.
831 : #[test]
832 4 : fn schedule_upload_delete_conflicts() -> anyhow::Result<()> {
833 4 : let mut queue = UploadQueue::Uninitialized;
834 4 : let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?;
835 4 : let tli = make_timeline();
836 4 :
837 4 : // Enqueue two layer uploads, with a delete of both layers in between them. These should be
838 4 : // scheduled one at a time, since deletes can't bypass uploads and vice versa.
839 4 : let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
840 4 : let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
841 4 :
842 4 : let ops = [
843 4 : UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None),
844 4 : UploadOp::Delete(Delete {
845 4 : layers: vec![
846 4 : (layer0.layer_desc().layer_name(), layer0.metadata()),
847 4 : (layer1.layer_desc().layer_name(), layer1.metadata()),
848 4 : ],
849 4 : }),
850 4 : UploadOp::UploadLayer(layer1.clone(), layer1.metadata(), None),
851 4 : ];
852 4 :
853 4 : queue.queued_operations.extend(ops.clone());
854 :
855 : // Only one version should be scheduled and uploaded at a time.
856 16 : for op in ops {
857 12 : let tasks = queue.schedule_ready();
858 12 : assert_eq!(tasks.len(), 1);
859 12 : assert_same_op(&tasks[0].op, &op);
860 12 : queue.complete(tasks[0].task_id);
861 : }
862 4 : assert!(queue.schedule_ready().is_empty());
863 4 : assert!(queue.queued_operations.is_empty());
864 :
865 4 : Ok(())
866 4 : }
867 :
868 : /// Non-conflicting uploads and deletes can bypass the queue, avoiding the conflicting
869 : /// delete/upload operations at the head of the queue.
870 : #[test]
871 4 : fn schedule_upload_delete_conflicts_bypass() -> anyhow::Result<()> {
872 4 : let mut queue = UploadQueue::Uninitialized;
873 4 : let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?;
874 4 : let tli = make_timeline();
875 4 :
876 4 : // Enqueue two layer uploads, with a delete of both layers in between them. These should be
877 4 : // scheduled one at a time, since deletes can't bypass uploads and vice versa.
878 4 : //
879 4 : // Also enqueue non-conflicting uploads and deletes at the end. These can bypass the queue
880 4 : // and run immediately.
881 4 : let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
882 4 : let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
883 4 : let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
884 4 : let layer3 = make_layer(&tli, "300000000000000000000000000000000000-400000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
885 4 :
886 4 : let ops = [
887 4 : UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None),
888 4 : UploadOp::Delete(Delete {
889 4 : layers: vec![
890 4 : (layer0.layer_desc().layer_name(), layer0.metadata()),
891 4 : (layer1.layer_desc().layer_name(), layer1.metadata()),
892 4 : ],
893 4 : }),
894 4 : UploadOp::UploadLayer(layer1.clone(), layer1.metadata(), None),
895 4 : UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None),
896 4 : UploadOp::Delete(Delete {
897 4 : layers: vec![(layer3.layer_desc().layer_name(), layer3.metadata())],
898 4 : }),
899 4 : ];
900 4 :
901 4 : queue.queued_operations.extend(ops.clone());
902 4 :
903 4 : // Operations 0, 3, and 4 are scheduled immediately.
904 4 : let tasks = queue.schedule_ready();
905 12 : assert_same_ops(tasks.iter().map(|t| &t.op), [&ops[0], &ops[3], &ops[4]]);
906 4 : assert_eq!(queue.queued_operations.len(), 2);
907 :
908 4 : Ok(())
909 4 : }
910 :
911 : /// Non-conflicting uploads are parallelized.
912 : #[test]
913 4 : fn schedule_upload_parallel() -> anyhow::Result<()> {
914 4 : let mut queue = UploadQueue::Uninitialized;
915 4 : let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?;
916 4 : let tli = make_timeline();
917 4 :
918 4 : // Enqueue three different layer uploads.
919 4 : let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
920 4 : let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
921 4 : let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
922 4 :
923 4 : let ops = [
924 4 : UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None),
925 4 : UploadOp::UploadLayer(layer1.clone(), layer1.metadata(), None),
926 4 : UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None),
927 4 : ];
928 4 :
929 4 : queue.queued_operations.extend(ops.clone());
930 4 :
931 4 : // All uploads should be scheduled concurrently.
932 4 : let tasks = queue.schedule_ready();
933 4 :
934 12 : assert_same_ops(tasks.iter().map(|t| &t.op), &ops);
935 4 : assert!(queue.queued_operations.is_empty());
936 :
937 4 : Ok(())
938 4 : }
939 :
940 : /// Index uploads are coalesced.
941 : #[test]
942 4 : fn schedule_index_coalesce() -> anyhow::Result<()> {
943 4 : let mut queue = UploadQueue::Uninitialized;
944 4 : let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?;
945 :
946 : // Enqueue three uploads of the current empty index.
947 4 : let index = Box::new(queue.clean.0.clone());
948 4 :
949 4 : let ops = [
950 4 : UploadOp::UploadMetadata {
951 4 : uploaded: index.clone(),
952 4 : },
953 4 : UploadOp::UploadMetadata {
954 4 : uploaded: index.clone(),
955 4 : },
956 4 : UploadOp::UploadMetadata {
957 4 : uploaded: index.clone(),
958 4 : },
959 4 : ];
960 4 :
961 4 : queue.queued_operations.extend(ops.clone());
962 4 :
963 4 : // The index uploads are coalesced into a single operation.
964 4 : let tasks = queue.schedule_ready();
965 4 : assert_eq!(tasks.len(), 1);
966 4 : assert_same_op(&tasks[0].op, &ops[2]);
967 4 : assert_same_ops(&tasks[0].coalesced_ops, &ops[0..2]);
968 4 :
969 4 : assert!(queue.queued_operations.is_empty());
970 :
971 4 : Ok(())
972 4 : }
973 :
974 : /// Chains of upload/index operations lead to parallel layer uploads and serial index uploads.
975 : /// This is the common case with layer flushes.
976 : #[test]
977 4 : fn schedule_index_upload_chain() -> anyhow::Result<()> {
978 4 : let mut queue = UploadQueue::Uninitialized;
979 4 : let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?;
980 4 : let tli = make_timeline();
981 4 :
982 4 : // Enqueue three uploads of the current empty index.
983 4 : let index = Box::new(queue.clean.0.clone());
984 4 : let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
985 4 : let index0 = index_with(&index, &layer0);
986 4 : let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
987 4 : let index1 = index_with(&index0, &layer1);
988 4 : let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
989 4 : let index2 = index_with(&index1, &layer2);
990 4 :
991 4 : let ops = [
992 4 : UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None),
993 4 : UploadOp::UploadMetadata {
994 4 : uploaded: index0.clone(),
995 4 : },
996 4 : UploadOp::UploadLayer(layer1.clone(), layer1.metadata(), None),
997 4 : UploadOp::UploadMetadata {
998 4 : uploaded: index1.clone(),
999 4 : },
1000 4 : UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None),
1001 4 : UploadOp::UploadMetadata {
1002 4 : uploaded: index2.clone(),
1003 4 : },
1004 4 : ];
1005 4 :
1006 4 : queue.queued_operations.extend(ops.clone());
1007 4 :
1008 4 : // The layer uploads should be scheduled immediately. The indexes must wait.
1009 4 : let upload_tasks = queue.schedule_ready();
1010 4 : assert_same_ops(
1011 12 : upload_tasks.iter().map(|t| &t.op),
1012 4 : [&ops[0], &ops[2], &ops[4]],
1013 4 : );
1014 4 :
1015 4 : // layer2 completes first. None of the indexes can upload yet.
1016 4 : queue.complete(upload_tasks[2].task_id);
1017 4 : assert!(queue.schedule_ready().is_empty());
1018 :
1019 : // layer0 completes. index0 can upload. It completes.
1020 4 : queue.complete(upload_tasks[0].task_id);
1021 4 : let index_tasks = queue.schedule_ready();
1022 4 : assert_eq!(index_tasks.len(), 1);
1023 4 : assert_same_op(&index_tasks[0].op, &ops[1]);
1024 4 : queue.complete(index_tasks[0].task_id);
1025 4 :
1026 4 : // layer 1 completes. This unblocks index 1 and 2, which coalesce into
1027 4 : // a single upload for index 2.
1028 4 : queue.complete(upload_tasks[1].task_id);
1029 4 :
1030 4 : let index_tasks = queue.schedule_ready();
1031 4 : assert_eq!(index_tasks.len(), 1);
1032 4 : assert_same_op(&index_tasks[0].op, &ops[5]);
1033 4 : assert_same_ops(&index_tasks[0].coalesced_ops, &ops[3..4]);
1034 4 :
1035 4 : assert!(queue.queued_operations.is_empty());
1036 :
1037 4 : Ok(())
1038 4 : }
1039 :
1040 : /// A delete can't bypass an index upload if an index ahead of it still references it.
1041 : #[test]
1042 4 : fn schedule_index_delete_dereferenced() -> anyhow::Result<()> {
1043 4 : let mut queue = UploadQueue::Uninitialized;
1044 4 : let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?;
1045 4 : let tli = make_timeline();
1046 4 :
1047 4 : // Create a layer to upload.
1048 4 : let layer = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
1049 4 : let index_upload = index_with(&queue.clean.0, &layer);
1050 4 :
1051 4 : // Remove the layer reference in a new index, then delete the layer.
1052 4 : let index_deref = index_without(&index_upload, &layer);
1053 4 :
1054 4 : let ops = [
1055 4 : // Initial upload, with a barrier to prevent index coalescing.
1056 4 : UploadOp::UploadLayer(layer.clone(), layer.metadata(), None),
1057 4 : UploadOp::UploadMetadata {
1058 4 : uploaded: index_upload.clone(),
1059 4 : },
1060 4 : UploadOp::Barrier(tokio::sync::watch::channel(()).0),
1061 4 : // Dereference the layer and delete it.
1062 4 : UploadOp::UploadMetadata {
1063 4 : uploaded: index_deref.clone(),
1064 4 : },
1065 4 : UploadOp::Delete(Delete {
1066 4 : layers: vec![(layer.layer_desc().layer_name(), layer.metadata())],
1067 4 : }),
1068 4 : ];
1069 4 :
1070 4 : queue.queued_operations.extend(ops.clone());
1071 :
1072 : // Operations are serialized.
1073 24 : for op in ops {
1074 20 : let tasks = queue.schedule_ready();
1075 20 : assert_eq!(tasks.len(), 1);
1076 20 : assert_same_op(&tasks[0].op, &op);
1077 20 : queue.complete(tasks[0].task_id);
1078 : }
1079 4 : assert!(queue.queued_operations.is_empty());
1080 :
1081 4 : Ok(())
1082 4 : }
1083 :
1084 : /// An upload with a reused layer name doesn't clobber the previous layer. Specifically, a
1085 : /// dereference/upload/reference cycle can't allow the upload to bypass the reference.
1086 : #[test]
1087 4 : fn schedule_index_upload_dereferenced() -> anyhow::Result<()> {
1088 4 : let mut queue = UploadQueue::Uninitialized;
1089 4 : let queue = queue.initialize_with_current_remote_index_part(&IndexPart::example(), 0)?;
1090 4 : let tli = make_timeline();
1091 4 :
1092 4 : // Create a layer to upload.
1093 4 : let layer = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
1094 4 :
1095 4 : // Upload the layer. Then dereference the layer, and upload/reference it again.
1096 4 : let index_upload = index_with(&queue.clean.0, &layer);
1097 4 : let index_deref = index_without(&index_upload, &layer);
1098 4 : let index_ref = index_with(&index_deref, &layer);
1099 4 :
1100 4 : let ops = [
1101 4 : // Initial upload, with a barrier to prevent index coalescing.
1102 4 : UploadOp::UploadLayer(layer.clone(), layer.metadata(), None),
1103 4 : UploadOp::UploadMetadata {
1104 4 : uploaded: index_upload.clone(),
1105 4 : },
1106 4 : UploadOp::Barrier(tokio::sync::watch::channel(()).0),
1107 4 : // Dereference the layer.
1108 4 : UploadOp::UploadMetadata {
1109 4 : uploaded: index_deref.clone(),
1110 4 : },
1111 4 : // Replace and reference the layer.
1112 4 : UploadOp::UploadLayer(layer.clone(), layer.metadata(), None),
1113 4 : UploadOp::UploadMetadata {
1114 4 : uploaded: index_ref.clone(),
1115 4 : },
1116 4 : ];
1117 4 :
1118 4 : queue.queued_operations.extend(ops.clone());
1119 :
1120 : // Operations are serialized.
1121 28 : for op in ops {
1122 24 : let tasks = queue.schedule_ready();
1123 24 : assert_eq!(tasks.len(), 1);
1124 24 : assert_same_op(&tasks[0].op, &op);
1125 24 : queue.complete(tasks[0].task_id);
1126 : }
1127 4 : assert!(queue.queued_operations.is_empty());
1128 :
1129 4 : Ok(())
1130 4 : }
1131 :
1132 : /// Nothing can bypass a shutdown, and it waits for inprogress tasks. It's never returned from
1133 : /// next_ready(), but is left at the head of the queue.
1134 : #[test]
1135 4 : fn schedule_shutdown() -> anyhow::Result<()> {
1136 4 : let mut queue = UploadQueue::Uninitialized;
1137 4 : let queue = queue.initialize_empty_remote(&TimelineMetadata::example(), 0)?;
1138 4 : let tli = make_timeline();
1139 4 :
1140 4 : let index = Box::new(queue.clean.0.clone()); // empty, doesn't matter
1141 4 : let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
1142 4 : let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
1143 4 : let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
1144 4 : let layer3 = make_layer(&tli, "300000000000000000000000000000000000-400000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
1145 4 :
1146 4 : // Enqueue non-conflicting upload, delete, and index before and after a shutdown.
1147 4 : let ops = [
1148 4 : UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None),
1149 4 : UploadOp::Delete(Delete {
1150 4 : layers: vec![(layer1.layer_desc().layer_name(), layer1.metadata())],
1151 4 : }),
1152 4 : UploadOp::UploadMetadata {
1153 4 : uploaded: index.clone(),
1154 4 : },
1155 4 : UploadOp::Shutdown,
1156 4 : UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None),
1157 4 : UploadOp::Delete(Delete {
1158 4 : layers: vec![(layer3.layer_desc().layer_name(), layer3.metadata())],
1159 4 : }),
1160 4 : UploadOp::UploadMetadata {
1161 4 : uploaded: index.clone(),
1162 4 : },
1163 4 : ];
1164 4 :
1165 4 : queue.queued_operations.extend(ops.clone());
1166 4 :
1167 4 : // Schedule the initial operations ahead of the shutdown.
1168 4 : let tasks = queue.schedule_ready();
1169 4 :
1170 12 : assert_same_ops(tasks.iter().map(|t| &t.op), &ops[0..3]);
1171 4 : assert!(matches!(
1172 4 : queue.queued_operations.front(),
1173 : Some(&UploadOp::Shutdown)
1174 : ));
1175 :
1176 : // Complete the initial operations. The shutdown isn't triggered while they're pending.
1177 16 : for task in tasks {
1178 12 : assert!(queue.schedule_ready().is_empty());
1179 12 : queue.complete(task.task_id);
1180 : }
1181 :
1182 : // The shutdown is triggered the next time we try to pull an operation. It isn't returned,
1183 : // but is left in the queue.
1184 4 : assert!(!queue.shutdown_ready.is_closed());
1185 4 : assert!(queue.next_ready().is_none());
1186 4 : assert!(queue.shutdown_ready.is_closed());
1187 :
1188 4 : Ok(())
1189 4 : }
1190 :
1191 : /// Scheduling respects inprogress_limit.
1192 : #[test]
1193 4 : fn schedule_inprogress_limit() -> anyhow::Result<()> {
1194 4 : // Create a queue with inprogress_limit=2.
1195 4 : let mut queue = UploadQueue::Uninitialized;
1196 4 : let queue = queue.initialize_empty_remote(&TimelineMetadata::example(), 2)?;
1197 4 : let tli = make_timeline();
1198 4 :
1199 4 : // Enqueue a bunch of uploads.
1200 4 : let layer0 = make_layer(&tli, "000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
1201 4 : let layer1 = make_layer(&tli, "100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
1202 4 : let layer2 = make_layer(&tli, "200000000000000000000000000000000000-300000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
1203 4 : let layer3 = make_layer(&tli, "300000000000000000000000000000000000-400000000000000000000000000000000000__00000000016B59D8-00000000016B5A51");
1204 4 :
1205 4 : let ops = [
1206 4 : UploadOp::UploadLayer(layer0.clone(), layer0.metadata(), None),
1207 4 : UploadOp::UploadLayer(layer1.clone(), layer1.metadata(), None),
1208 4 : UploadOp::UploadLayer(layer2.clone(), layer2.metadata(), None),
1209 4 : UploadOp::UploadLayer(layer3.clone(), layer3.metadata(), None),
1210 4 : ];
1211 4 :
1212 4 : queue.queued_operations.extend(ops.clone());
1213 4 :
1214 4 : // Schedule all ready operations. Only 2 are scheduled.
1215 4 : let tasks = queue.schedule_ready();
1216 8 : assert_same_ops(tasks.iter().map(|t| &t.op), &ops[0..2]);
1217 4 : assert!(queue.next_ready().is_none());
1218 :
1219 : // When one completes, another is scheduled.
1220 4 : queue.complete(tasks[0].task_id);
1221 4 : let tasks = queue.schedule_ready();
1222 4 : assert_same_ops(tasks.iter().map(|t| &t.op), &ops[2..3]);
1223 4 :
1224 4 : Ok(())
1225 4 : }
1226 :
1227 : /// Tests that can_bypass takes name, generation and shard index into account for all operations.
1228 : #[test]
1229 4 : fn can_bypass_path() -> anyhow::Result<()> {
1230 4 : let tli = make_timeline();
1231 4 :
1232 4 : let name0 = &"000000000000000000000000000000000000-100000000000000000000000000000000000__00000000016B59D8-00000000016B5A51";
1233 4 : let name1 = &"100000000000000000000000000000000000-200000000000000000000000000000000000__00000000016B59D8-00000000016B5A51";
1234 :
1235 : // Asserts that layers a and b either can or can't bypass each other, for all combinations
1236 : // of operations (except Delete and UploadMetadata which are special-cased).
1237 : #[track_caller]
1238 32 : fn assert_can_bypass(a: ResidentLayer, b: ResidentLayer, can_bypass: bool) {
1239 32 : let index = IndexPart::empty(TimelineMetadata::example());
1240 96 : for (a, b) in make_ops(a).into_iter().zip(make_ops(b)) {
1241 96 : match (&a, &b) {
1242 : // Deletes can always bypass each other.
1243 32 : (UploadOp::Delete(_), UploadOp::Delete(_)) => assert!(a.can_bypass(&b, &index)),
1244 : // Indexes can never bypass each other.
1245 : (UploadOp::UploadMetadata { .. }, UploadOp::UploadMetadata { .. }) => {
1246 32 : assert!(!a.can_bypass(&b, &index))
1247 : }
1248 : // For other operations, assert as requested.
1249 32 : (a, b) => assert_eq!(a.can_bypass(b, &index), can_bypass),
1250 : }
1251 : }
1252 32 : }
1253 :
1254 64 : fn make_ops(layer: ResidentLayer) -> Vec<UploadOp> {
1255 64 : let mut index = IndexPart::empty(TimelineMetadata::example());
1256 64 : index
1257 64 : .layer_metadata
1258 64 : .insert(layer.layer_desc().layer_name(), layer.metadata());
1259 64 : vec![
1260 64 : UploadOp::UploadLayer(layer.clone(), layer.metadata(), None),
1261 64 : UploadOp::Delete(Delete {
1262 64 : layers: vec![(layer.layer_desc().layer_name(), layer.metadata())],
1263 64 : }),
1264 64 : UploadOp::UploadMetadata {
1265 64 : uploaded: Box::new(index),
1266 64 : },
1267 64 : ]
1268 64 : }
1269 :
1270 : // Makes a ResidentLayer.
1271 64 : let layer = |name: &'static str, shard: Option<u8>, generation: u32| -> ResidentLayer {
1272 64 : let shard = shard
1273 64 : .map(|n| ShardIndex::new(ShardNumber(n), ShardCount(8)))
1274 64 : .unwrap_or(ShardIndex::unsharded());
1275 64 : let metadata = LayerFileMetadata {
1276 64 : shard,
1277 64 : generation: Generation::Valid(generation),
1278 64 : file_size: 0,
1279 64 : };
1280 64 : make_layer_with_metadata(&tli, name, metadata)
1281 64 : };
1282 :
1283 : // Same name and metadata can't bypass. This goes both for unsharded and sharded, as well as
1284 : // 0 or >0 generation.
1285 4 : assert_can_bypass(layer(name0, None, 0), layer(name0, None, 0), false);
1286 4 : assert_can_bypass(layer(name0, Some(0), 0), layer(name0, Some(0), 0), false);
1287 4 : assert_can_bypass(layer(name0, None, 1), layer(name0, None, 1), false);
1288 4 :
1289 4 : // Different names can bypass.
1290 4 : assert_can_bypass(layer(name0, None, 0), layer(name1, None, 0), true);
1291 4 :
1292 4 : // Different shards can bypass. Shard 0 is different from unsharded.
1293 4 : assert_can_bypass(layer(name0, Some(0), 0), layer(name0, Some(1), 0), true);
1294 4 : assert_can_bypass(layer(name0, Some(0), 0), layer(name0, None, 0), true);
1295 4 :
1296 4 : // Different generations can bypass, both sharded and unsharded.
1297 4 : assert_can_bypass(layer(name0, None, 0), layer(name0, None, 1), true);
1298 4 : assert_can_bypass(layer(name0, Some(1), 0), layer(name0, Some(1), 1), true);
1299 4 :
1300 4 : Ok(())
1301 4 : }
1302 : }
|
