Line data Source code
1 : //! An ImageLayer represents an image or a snapshot of a key-range at
2 : //! one particular LSN. It contains an image of all key-value pairs
3 : //! in its key-range. Any key that falls into the image layer's range
4 : //! but does not exist in the layer, does not exist.
5 : //!
6 : //! An image layer is stored in a file on disk. The file is stored in
7 : //! timelines/<timeline_id> directory. Currently, there are no
8 : //! subdirectories, and each image layer file is named like this:
9 : //!
10 : //! ```text
11 : //! <key start>-<key end>__<LSN>
12 : //! ```
13 : //!
14 : //! For example:
15 : //!
16 : //! ```text
17 : //! 000000067F000032BE0000400000000070B6-000000067F000032BE0000400000000080B6__00000000346BC568
18 : //! ```
19 : //!
20 : //! Every image layer file consists of three parts: "summary",
21 : //! "index", and "values". The summary is a fixed size header at the
22 : //! beginning of the file, and it contains basic information about the
23 : //! layer, and offsets to the other parts. The "index" is a B-tree,
24 : //! mapping from Key to an offset in the "values" part. The
25 : //! actual page images are stored in the "values" part.
26 : use crate::config::PageServerConf;
27 : use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
28 : use crate::page_cache::{self, FileId, PAGE_SZ};
29 : use crate::repository::{Key, Value, KEY_SIZE};
30 : use crate::tenant::blob_io::BlobWriter;
31 : use crate::tenant::block_io::{BlockBuf, BlockReader, FileBlockReader};
32 : use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
33 : use crate::tenant::storage_layer::{
34 : LayerAccessStats, ValueReconstructResult, ValueReconstructState,
35 : };
36 : use crate::tenant::timeline::GetVectoredError;
37 : use crate::tenant::vectored_blob_io::{
38 : BlobFlag, MaxVectoredReadBytes, VectoredBlobReader, VectoredRead, VectoredReadPlanner,
39 : };
40 : use crate::tenant::{PageReconstructError, Timeline};
41 : use crate::virtual_file::{self, VirtualFile};
42 : use crate::{IMAGE_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
43 : use anyhow::{anyhow, bail, ensure, Context, Result};
44 : use bytes::{Bytes, BytesMut};
45 : use camino::{Utf8Path, Utf8PathBuf};
46 : use hex;
47 : use itertools::Itertools;
48 : use pageserver_api::keyspace::KeySpace;
49 : use pageserver_api::models::LayerAccessKind;
50 : use pageserver_api::shard::{ShardIdentity, TenantShardId};
51 : use rand::{distributions::Alphanumeric, Rng};
52 : use serde::{Deserialize, Serialize};
53 : use std::fs::File;
54 : use std::io::SeekFrom;
55 : use std::ops::Range;
56 : use std::os::unix::prelude::FileExt;
57 : use std::str::FromStr;
58 : use std::sync::Arc;
59 : use tokio::sync::OnceCell;
60 : use tokio_stream::StreamExt;
61 : use tracing::*;
62 :
63 : use utils::{
64 : bin_ser::BeSer,
65 : id::{TenantId, TimelineId},
66 : lsn::Lsn,
67 : };
68 :
69 : use super::layer_name::ImageLayerName;
70 : use super::{
71 : AsLayerDesc, Layer, LayerName, PersistentLayerDesc, ResidentLayer, ValuesReconstructState,
72 : };
73 :
74 : ///
75 : /// Header stored in the beginning of the file
76 : ///
77 : /// After this comes the 'values' part, starting on block 1. After that,
78 : /// the 'index' starts at the block indicated by 'index_start_blk'
79 : ///
80 90 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
81 : pub struct Summary {
82 : /// Magic value to identify this as a neon image file. Always IMAGE_FILE_MAGIC.
83 : pub magic: u16,
84 : pub format_version: u16,
85 :
86 : pub tenant_id: TenantId,
87 : pub timeline_id: TimelineId,
88 : pub key_range: Range<Key>,
89 : pub lsn: Lsn,
90 :
91 : /// Block number where the 'index' part of the file begins.
92 : pub index_start_blk: u32,
93 : /// Block within the 'index', where the B-tree root page is stored
94 : pub index_root_blk: u32,
95 : // the 'values' part starts after the summary header, on block 1.
96 : }
97 :
98 : impl From<&ImageLayer> for Summary {
99 0 : fn from(layer: &ImageLayer) -> Self {
100 0 : Self::expected(
101 0 : layer.desc.tenant_shard_id.tenant_id,
102 0 : layer.desc.timeline_id,
103 0 : layer.desc.key_range.clone(),
104 0 : layer.lsn,
105 0 : )
106 0 : }
107 : }
108 :
109 : impl Summary {
110 90 : pub(super) fn expected(
111 90 : tenant_id: TenantId,
112 90 : timeline_id: TimelineId,
113 90 : key_range: Range<Key>,
114 90 : lsn: Lsn,
115 90 : ) -> Self {
116 90 : Self {
117 90 : magic: IMAGE_FILE_MAGIC,
118 90 : format_version: STORAGE_FORMAT_VERSION,
119 90 : tenant_id,
120 90 : timeline_id,
121 90 : key_range,
122 90 : lsn,
123 90 :
124 90 : index_start_blk: 0,
125 90 : index_root_blk: 0,
126 90 : }
127 90 : }
128 : }
129 :
130 : /// This is used only from `pagectl`. Within pageserver, all layers are
131 : /// [`crate::tenant::storage_layer::Layer`], which can hold an [`ImageLayerInner`].
132 : pub struct ImageLayer {
133 : path: Utf8PathBuf,
134 : pub desc: PersistentLayerDesc,
135 : // This entry contains an image of all pages as of this LSN, should be the same as desc.lsn
136 : pub lsn: Lsn,
137 : access_stats: LayerAccessStats,
138 : inner: OnceCell<ImageLayerInner>,
139 : }
140 :
141 : impl std::fmt::Debug for ImageLayer {
142 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
143 0 : use super::RangeDisplayDebug;
144 0 :
145 0 : f.debug_struct("ImageLayer")
146 0 : .field("key_range", &RangeDisplayDebug(&self.desc.key_range))
147 0 : .field("file_size", &self.desc.file_size)
148 0 : .field("lsn", &self.lsn)
149 0 : .field("inner", &self.inner)
150 0 : .finish()
151 0 : }
152 : }
153 :
154 : /// ImageLayer is the in-memory data structure associated with an on-disk image
155 : /// file.
156 : pub struct ImageLayerInner {
157 : // values copied from summary
158 : index_start_blk: u32,
159 : index_root_blk: u32,
160 :
161 : key_range: Range<Key>,
162 : lsn: Lsn,
163 :
164 : file: VirtualFile,
165 : file_id: FileId,
166 :
167 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
168 : }
169 :
170 : impl std::fmt::Debug for ImageLayerInner {
171 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
172 0 : f.debug_struct("ImageLayerInner")
173 0 : .field("index_start_blk", &self.index_start_blk)
174 0 : .field("index_root_blk", &self.index_root_blk)
175 0 : .finish()
176 0 : }
177 : }
178 :
179 : impl ImageLayerInner {
180 0 : pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
181 0 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
182 0 : let tree_reader = DiskBtreeReader::<_, KEY_SIZE>::new(
183 0 : self.index_start_blk,
184 0 : self.index_root_blk,
185 0 : block_reader,
186 0 : );
187 0 :
188 0 : tree_reader.dump().await?;
189 :
190 0 : tree_reader
191 0 : .visit(
192 0 : &[0u8; KEY_SIZE],
193 0 : VisitDirection::Forwards,
194 0 : |key, value| {
195 0 : println!("key: {} offset {}", hex::encode(key), value);
196 0 : true
197 0 : },
198 0 : ctx,
199 0 : )
200 0 : .await?;
201 :
202 0 : Ok(())
203 0 : }
204 : }
205 :
206 : /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
207 : impl std::fmt::Display for ImageLayer {
208 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
209 0 : write!(f, "{}", self.layer_desc().short_id())
210 0 : }
211 : }
212 :
213 : impl AsLayerDesc for ImageLayer {
214 0 : fn layer_desc(&self) -> &PersistentLayerDesc {
215 0 : &self.desc
216 0 : }
217 : }
218 :
219 : impl ImageLayer {
220 0 : pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
221 0 : self.desc.dump();
222 0 :
223 0 : if !verbose {
224 0 : return Ok(());
225 0 : }
226 :
227 0 : let inner = self.load(LayerAccessKind::Dump, ctx).await?;
228 :
229 0 : inner.dump(ctx).await?;
230 :
231 0 : Ok(())
232 0 : }
233 :
234 244 : fn temp_path_for(
235 244 : conf: &PageServerConf,
236 244 : timeline_id: TimelineId,
237 244 : tenant_shard_id: TenantShardId,
238 244 : fname: &ImageLayerName,
239 244 : ) -> Utf8PathBuf {
240 244 : let rand_string: String = rand::thread_rng()
241 244 : .sample_iter(&Alphanumeric)
242 244 : .take(8)
243 244 : .map(char::from)
244 244 : .collect();
245 244 :
246 244 : conf.timeline_path(&tenant_shard_id, &timeline_id)
247 244 : .join(format!("{fname}.{rand_string}.{TEMP_FILE_SUFFIX}"))
248 244 : }
249 :
250 : ///
251 : /// Open the underlying file and read the metadata into memory, if it's
252 : /// not loaded already.
253 : ///
254 0 : async fn load(
255 0 : &self,
256 0 : access_kind: LayerAccessKind,
257 0 : ctx: &RequestContext,
258 0 : ) -> Result<&ImageLayerInner> {
259 0 : self.access_stats.record_access(access_kind, ctx);
260 0 : self.inner
261 0 : .get_or_try_init(|| self.load_inner(ctx))
262 0 : .await
263 0 : .with_context(|| format!("Failed to load image layer {}", self.path()))
264 0 : }
265 :
266 0 : async fn load_inner(&self, ctx: &RequestContext) -> Result<ImageLayerInner> {
267 0 : let path = self.path();
268 :
269 0 : let loaded = ImageLayerInner::load(&path, self.desc.image_layer_lsn(), None, None, ctx)
270 0 : .await
271 0 : .and_then(|res| res)?;
272 :
273 : // not production code
274 0 : let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
275 0 : let expected_layer_name = self.layer_desc().layer_name();
276 0 :
277 0 : if actual_layer_name != expected_layer_name {
278 0 : println!("warning: filename does not match what is expected from in-file summary");
279 0 : println!("actual: {:?}", actual_layer_name.to_string());
280 0 : println!("expected: {:?}", expected_layer_name.to_string());
281 0 : }
282 :
283 0 : Ok(loaded)
284 0 : }
285 :
286 : /// Create an ImageLayer struct representing an existing file on disk.
287 : ///
288 : /// This variant is only used for debugging purposes, by the 'pagectl' binary.
289 0 : pub fn new_for_path(path: &Utf8Path, file: File) -> Result<ImageLayer> {
290 0 : let mut summary_buf = vec![0; PAGE_SZ];
291 0 : file.read_exact_at(&mut summary_buf, 0)?;
292 0 : let summary = Summary::des_prefix(&summary_buf)?;
293 0 : let metadata = file
294 0 : .metadata()
295 0 : .context("get file metadata to determine size")?;
296 :
297 : // This function is never used for constructing layers in a running pageserver,
298 : // so it does not need an accurate TenantShardId.
299 0 : let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
300 0 :
301 0 : Ok(ImageLayer {
302 0 : path: path.to_path_buf(),
303 0 : desc: PersistentLayerDesc::new_img(
304 0 : tenant_shard_id,
305 0 : summary.timeline_id,
306 0 : summary.key_range,
307 0 : summary.lsn,
308 0 : metadata.len(),
309 0 : ), // Now we assume image layer ALWAYS covers the full range. This may change in the future.
310 0 : lsn: summary.lsn,
311 0 : access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
312 0 : inner: OnceCell::new(),
313 0 : })
314 0 : }
315 :
316 0 : fn path(&self) -> Utf8PathBuf {
317 0 : self.path.clone()
318 0 : }
319 : }
320 :
321 0 : #[derive(thiserror::Error, Debug)]
322 : pub enum RewriteSummaryError {
323 : #[error("magic mismatch")]
324 : MagicMismatch,
325 : #[error(transparent)]
326 : Other(#[from] anyhow::Error),
327 : }
328 :
329 : impl From<std::io::Error> for RewriteSummaryError {
330 0 : fn from(e: std::io::Error) -> Self {
331 0 : Self::Other(anyhow::anyhow!(e))
332 0 : }
333 : }
334 :
335 : impl ImageLayer {
336 0 : pub async fn rewrite_summary<F>(
337 0 : path: &Utf8Path,
338 0 : rewrite: F,
339 0 : ctx: &RequestContext,
340 0 : ) -> Result<(), RewriteSummaryError>
341 0 : where
342 0 : F: Fn(Summary) -> Summary,
343 0 : {
344 0 : let mut file = VirtualFile::open_with_options(
345 0 : path,
346 0 : virtual_file::OpenOptions::new().read(true).write(true),
347 0 : ctx,
348 0 : )
349 0 : .await
350 0 : .with_context(|| format!("Failed to open file '{}'", path))?;
351 0 : let file_id = page_cache::next_file_id();
352 0 : let block_reader = FileBlockReader::new(&file, file_id);
353 0 : let summary_blk = block_reader.read_blk(0, ctx).await?;
354 0 : let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
355 0 : if actual_summary.magic != IMAGE_FILE_MAGIC {
356 0 : return Err(RewriteSummaryError::MagicMismatch);
357 0 : }
358 0 :
359 0 : let new_summary = rewrite(actual_summary);
360 0 :
361 0 : let mut buf = Vec::with_capacity(PAGE_SZ);
362 0 : // TODO: could use smallvec here but it's a pain with Slice<T>
363 0 : Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
364 0 : file.seek(SeekFrom::Start(0)).await?;
365 0 : let (_buf, res) = file.write_all(buf, ctx).await;
366 0 : res?;
367 0 : Ok(())
368 0 : }
369 : }
370 :
371 : impl ImageLayerInner {
372 : /// Returns nested result following Result<Result<_, OpErr>, Critical>:
373 : /// - inner has the success or transient failure
374 : /// - outer has the permanent failure
375 90 : pub(super) async fn load(
376 90 : path: &Utf8Path,
377 90 : lsn: Lsn,
378 90 : summary: Option<Summary>,
379 90 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
380 90 : ctx: &RequestContext,
381 90 : ) -> Result<Result<Self, anyhow::Error>, anyhow::Error> {
382 90 : let file = match VirtualFile::open(path, ctx).await {
383 90 : Ok(file) => file,
384 0 : Err(e) => return Ok(Err(anyhow::Error::new(e).context("open layer file"))),
385 : };
386 90 : let file_id = page_cache::next_file_id();
387 90 : let block_reader = FileBlockReader::new(&file, file_id);
388 90 : let summary_blk = match block_reader.read_blk(0, ctx).await {
389 90 : Ok(blk) => blk,
390 0 : Err(e) => return Ok(Err(anyhow::Error::new(e).context("read first block"))),
391 : };
392 :
393 : // length is the only way how this could fail, so it's not actually likely at all unless
394 : // read_blk returns wrong sized block.
395 : //
396 : // TODO: confirm and make this into assertion
397 90 : let actual_summary =
398 90 : Summary::des_prefix(summary_blk.as_ref()).context("deserialize first block")?;
399 :
400 90 : if let Some(mut expected_summary) = summary {
401 : // production code path
402 90 : expected_summary.index_start_blk = actual_summary.index_start_blk;
403 90 : expected_summary.index_root_blk = actual_summary.index_root_blk;
404 90 : // mask out the timeline_id, but still require the layers to be from the same tenant
405 90 : expected_summary.timeline_id = actual_summary.timeline_id;
406 90 :
407 90 : if actual_summary != expected_summary {
408 0 : bail!(
409 0 : "in-file summary does not match expected summary. actual = {:?} expected = {:?}",
410 0 : actual_summary,
411 0 : expected_summary
412 0 : );
413 90 : }
414 0 : }
415 :
416 90 : Ok(Ok(ImageLayerInner {
417 90 : index_start_blk: actual_summary.index_start_blk,
418 90 : index_root_blk: actual_summary.index_root_blk,
419 90 : lsn,
420 90 : file,
421 90 : file_id,
422 90 : max_vectored_read_bytes,
423 90 : key_range: actual_summary.key_range,
424 90 : }))
425 90 : }
426 :
427 7147 : pub(super) async fn get_value_reconstruct_data(
428 7147 : &self,
429 7147 : key: Key,
430 7147 : reconstruct_state: &mut ValueReconstructState,
431 7147 : ctx: &RequestContext,
432 7147 : ) -> anyhow::Result<ValueReconstructResult> {
433 7147 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
434 7147 : let tree_reader =
435 7147 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, &block_reader);
436 7147 :
437 7147 : let mut keybuf: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
438 7147 : key.write_to_byte_slice(&mut keybuf);
439 7147 : if let Some(offset) = tree_reader
440 7147 : .get(
441 7147 : &keybuf,
442 7147 : &RequestContextBuilder::extend(ctx)
443 7147 : .page_content_kind(PageContentKind::ImageLayerBtreeNode)
444 7147 : .build(),
445 7147 : )
446 413 : .await?
447 : {
448 7143 : let blob = block_reader
449 7143 : .block_cursor()
450 7143 : .read_blob(
451 7143 : offset,
452 7143 : &RequestContextBuilder::extend(ctx)
453 7143 : .page_content_kind(PageContentKind::ImageLayerValue)
454 7143 : .build(),
455 7143 : )
456 308 : .await
457 7143 : .with_context(|| format!("failed to read value from offset {}", offset))?;
458 7143 : let value = Bytes::from(blob);
459 7143 :
460 7143 : reconstruct_state.img = Some((self.lsn, value));
461 7143 : Ok(ValueReconstructResult::Complete)
462 : } else {
463 4 : Ok(ValueReconstructResult::Missing)
464 : }
465 7147 : }
466 :
467 : // Look up the keys in the provided keyspace and update
468 : // the reconstruct state with whatever is found.
469 74 : pub(super) async fn get_values_reconstruct_data(
470 74 : &self,
471 74 : keyspace: KeySpace,
472 74 : reconstruct_state: &mut ValuesReconstructState,
473 74 : ctx: &RequestContext,
474 74 : ) -> Result<(), GetVectoredError> {
475 74 : let reads = self
476 74 : .plan_reads(keyspace, None, ctx)
477 349 : .await
478 74 : .map_err(GetVectoredError::Other)?;
479 :
480 74 : self.do_reads_and_update_state(reads, reconstruct_state, ctx)
481 895 : .await;
482 :
483 74 : reconstruct_state.on_image_layer_visited(&self.key_range);
484 74 :
485 74 : Ok(())
486 74 : }
487 :
488 : /// Load all key-values in the delta layer, should be replaced by an iterator-based interface in the future.
489 8 : pub(super) async fn load_key_values(
490 8 : &self,
491 8 : ctx: &RequestContext,
492 8 : ) -> anyhow::Result<Vec<(Key, Lsn, Value)>> {
493 8 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
494 8 : let tree_reader =
495 8 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, &block_reader);
496 8 : let mut result = Vec::new();
497 8 : let mut stream = Box::pin(tree_reader.into_stream(&[0; KEY_SIZE], ctx));
498 8 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
499 8 : let cursor = block_reader.block_cursor();
500 80 : while let Some(item) = stream.next().await {
501 : // TODO: dedup code with get_reconstruct_value
502 72 : let (raw_key, offset) = item?;
503 72 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
504 : // TODO: ctx handling and sharding
505 72 : let blob = cursor
506 72 : .read_blob(offset, ctx)
507 2 : .await
508 72 : .with_context(|| format!("failed to read value from offset {}", offset))?;
509 72 : let value = Bytes::from(blob);
510 72 : result.push((key, self.lsn, Value::Image(value)));
511 : }
512 8 : Ok(result)
513 8 : }
514 :
515 : /// Traverse the layer's index to build read operations on the overlap of the input keyspace
516 : /// and the keys in this layer.
517 : ///
518 : /// If shard_identity is provided, it will be used to filter keys down to those stored on
519 : /// this shard.
520 82 : async fn plan_reads(
521 82 : &self,
522 82 : keyspace: KeySpace,
523 82 : shard_identity: Option<&ShardIdentity>,
524 82 : ctx: &RequestContext,
525 82 : ) -> anyhow::Result<Vec<VectoredRead>> {
526 82 : let mut planner = VectoredReadPlanner::new(
527 82 : self.max_vectored_read_bytes
528 82 : .expect("Layer is loaded with max vectored bytes config")
529 82 : .0
530 82 : .into(),
531 82 : );
532 82 :
533 82 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
534 82 : let tree_reader =
535 82 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
536 82 :
537 82 : let ctx = RequestContextBuilder::extend(ctx)
538 82 : .page_content_kind(PageContentKind::ImageLayerBtreeNode)
539 82 : .build();
540 :
541 21697 : for range in keyspace.ranges.iter() {
542 21697 : let mut range_end_handled = false;
543 21697 : let mut search_key: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
544 21697 : range.start.write_to_byte_slice(&mut search_key);
545 21697 :
546 21697 : let index_stream = tree_reader.clone().into_stream(&search_key, &ctx);
547 21697 : let mut index_stream = std::pin::pin!(index_stream);
548 :
549 1080725 : while let Some(index_entry) = index_stream.next().await {
550 1080664 : let (raw_key, offset) = index_entry?;
551 :
552 1080664 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
553 1080664 : assert!(key >= range.start);
554 :
555 1080664 : let flag = if let Some(shard_identity) = shard_identity {
556 1048576 : if shard_identity.is_key_disposable(&key) {
557 786432 : BlobFlag::Ignore
558 : } else {
559 262144 : BlobFlag::None
560 : }
561 : } else {
562 32088 : BlobFlag::None
563 : };
564 :
565 1080664 : if key >= range.end {
566 21636 : planner.handle_range_end(offset);
567 21636 : range_end_handled = true;
568 21636 : break;
569 1059028 : } else {
570 1059028 : planner.handle(key, self.lsn, offset, flag);
571 1059028 : }
572 : }
573 :
574 21697 : if !range_end_handled {
575 61 : let payload_end = self.index_start_blk as u64 * PAGE_SZ as u64;
576 61 : planner.handle_range_end(payload_end);
577 21636 : }
578 : }
579 :
580 82 : Ok(planner.finish())
581 82 : }
582 :
583 : /// Given a key range, select the parts of that range that should be retained by the ShardIdentity,
584 : /// then execute vectored GET operations, passing the results of all read keys into the writer.
585 8 : pub(super) async fn filter(
586 8 : &self,
587 8 : shard_identity: &ShardIdentity,
588 8 : writer: &mut ImageLayerWriter,
589 8 : ctx: &RequestContext,
590 8 : ) -> anyhow::Result<usize> {
591 : // Fragment the range into the regions owned by this ShardIdentity
592 8 : let plan = self
593 8 : .plan_reads(
594 8 : KeySpace {
595 8 : // If asked for the total key space, plan_reads will give us all the keys in the layer
596 8 : ranges: vec![Key::MIN..Key::MAX],
597 8 : },
598 8 : Some(shard_identity),
599 8 : ctx,
600 8 : )
601 469 : .await?;
602 :
603 8 : let vectored_blob_reader = VectoredBlobReader::new(&self.file);
604 8 : let mut key_count = 0;
605 16 : for read in plan.into_iter() {
606 16 : let buf_size = read.size();
607 16 :
608 16 : let buf = BytesMut::with_capacity(buf_size);
609 16 : let blobs_buf = vectored_blob_reader.read_blobs(&read, buf, ctx).await?;
610 :
611 16 : let frozen_buf = blobs_buf.buf.freeze();
612 :
613 262144 : for meta in blobs_buf.blobs.iter() {
614 262144 : let img_buf = frozen_buf.slice(meta.start..meta.end);
615 262144 :
616 262144 : key_count += 1;
617 262144 : writer
618 262144 : .put_image(meta.meta.key, img_buf, ctx)
619 266240 : .await
620 262144 : .context(format!("Storing key {}", meta.meta.key))?;
621 : }
622 : }
623 :
624 8 : Ok(key_count)
625 8 : }
626 :
627 74 : async fn do_reads_and_update_state(
628 74 : &self,
629 74 : reads: Vec<VectoredRead>,
630 74 : reconstruct_state: &mut ValuesReconstructState,
631 74 : ctx: &RequestContext,
632 74 : ) {
633 74 : let max_vectored_read_bytes = self
634 74 : .max_vectored_read_bytes
635 74 : .expect("Layer is loaded with max vectored bytes config")
636 74 : .0
637 74 : .into();
638 74 :
639 74 : let vectored_blob_reader = VectoredBlobReader::new(&self.file);
640 1756 : for read in reads.into_iter() {
641 1756 : let buf_size = read.size();
642 1756 :
643 1756 : if buf_size > max_vectored_read_bytes {
644 : // If the read is oversized, it should only contain one key.
645 0 : let offenders = read
646 0 : .blobs_at
647 0 : .as_slice()
648 0 : .iter()
649 0 : .map(|(_, blob_meta)| format!("{}@{}", blob_meta.key, blob_meta.lsn))
650 0 : .join(", ");
651 0 : tracing::warn!(
652 0 : "Oversized vectored read ({} > {}) for keys {}",
653 : buf_size,
654 : max_vectored_read_bytes,
655 : offenders
656 : );
657 1756 : }
658 :
659 1756 : let buf = BytesMut::with_capacity(buf_size);
660 1756 : let res = vectored_blob_reader.read_blobs(&read, buf, ctx).await;
661 :
662 1756 : match res {
663 1756 : Ok(blobs_buf) => {
664 1756 : let frozen_buf = blobs_buf.buf.freeze();
665 :
666 10452 : for meta in blobs_buf.blobs.iter() {
667 10452 : let img_buf = frozen_buf.slice(meta.start..meta.end);
668 10452 : reconstruct_state.update_key(
669 10452 : &meta.meta.key,
670 10452 : self.lsn,
671 10452 : Value::Image(img_buf),
672 10452 : );
673 10452 : }
674 : }
675 0 : Err(err) => {
676 0 : let kind = err.kind();
677 0 : for (_, blob_meta) in read.blobs_at.as_slice() {
678 0 : reconstruct_state.on_key_error(
679 0 : blob_meta.key,
680 0 : PageReconstructError::from(anyhow!(
681 0 : "Failed to read blobs from virtual file {}: {}",
682 0 : self.file.path,
683 0 : kind
684 0 : )),
685 0 : );
686 0 : }
687 : }
688 : };
689 : }
690 74 : }
691 :
692 : #[cfg(test)]
693 56 : pub(crate) fn iter<'a>(&'a self, ctx: &'a RequestContext) -> ImageLayerIterator<'a> {
694 56 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
695 56 : let tree_reader =
696 56 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
697 56 : ImageLayerIterator {
698 56 : image_layer: self,
699 56 : ctx,
700 56 : index_iter: tree_reader.iter(&[0; KEY_SIZE], ctx),
701 56 : key_values_batch: std::collections::VecDeque::new(),
702 56 : is_end: false,
703 56 : planner: crate::tenant::vectored_blob_io::StreamingVectoredReadPlanner::new(
704 56 : 1024 * 8192, // The default value. Unit tests might use a different value. 1024 * 8K = 8MB buffer.
705 56 : 1024, // The default value. Unit tests might use a different value
706 56 : ),
707 56 : }
708 56 : }
709 : }
710 :
711 : /// A builder object for constructing a new image layer.
712 : ///
713 : /// Usage:
714 : ///
715 : /// 1. Create the ImageLayerWriter by calling ImageLayerWriter::new(...)
716 : ///
717 : /// 2. Write the contents by calling `put_page_image` for every key-value
718 : /// pair in the key range.
719 : ///
720 : /// 3. Call `finish`.
721 : ///
722 : struct ImageLayerWriterInner {
723 : conf: &'static PageServerConf,
724 : path: Utf8PathBuf,
725 : timeline_id: TimelineId,
726 : tenant_shard_id: TenantShardId,
727 : key_range: Range<Key>,
728 : lsn: Lsn,
729 :
730 : blob_writer: BlobWriter<false>,
731 : tree: DiskBtreeBuilder<BlockBuf, KEY_SIZE>,
732 : }
733 :
734 : impl ImageLayerWriterInner {
735 : ///
736 : /// Start building a new image layer.
737 : ///
738 244 : async fn new(
739 244 : conf: &'static PageServerConf,
740 244 : timeline_id: TimelineId,
741 244 : tenant_shard_id: TenantShardId,
742 244 : key_range: &Range<Key>,
743 244 : lsn: Lsn,
744 244 : ctx: &RequestContext,
745 244 : ) -> anyhow::Result<Self> {
746 244 : // Create the file initially with a temporary filename.
747 244 : // We'll atomically rename it to the final name when we're done.
748 244 : let path = ImageLayer::temp_path_for(
749 244 : conf,
750 244 : timeline_id,
751 244 : tenant_shard_id,
752 244 : &ImageLayerName {
753 244 : key_range: key_range.clone(),
754 244 : lsn,
755 244 : },
756 244 : );
757 244 : trace!("creating image layer {}", path);
758 244 : let mut file = {
759 244 : VirtualFile::open_with_options(
760 244 : &path,
761 244 : virtual_file::OpenOptions::new()
762 244 : .write(true)
763 244 : .create_new(true),
764 244 : ctx,
765 244 : )
766 191 : .await?
767 : };
768 : // make room for the header block
769 244 : file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
770 244 : let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
771 244 :
772 244 : // Initialize the b-tree index builder
773 244 : let block_buf = BlockBuf::new();
774 244 : let tree_builder = DiskBtreeBuilder::new(block_buf);
775 244 :
776 244 : let writer = Self {
777 244 : conf,
778 244 : path,
779 244 : timeline_id,
780 244 : tenant_shard_id,
781 244 : key_range: key_range.clone(),
782 244 : lsn,
783 244 : tree: tree_builder,
784 244 : blob_writer,
785 244 : };
786 244 :
787 244 : Ok(writer)
788 244 : }
789 :
790 : ///
791 : /// Write next value to the file.
792 : ///
793 : /// The page versions must be appended in blknum order.
794 : ///
795 537711 : async fn put_image(
796 537711 : &mut self,
797 537711 : key: Key,
798 537711 : img: Bytes,
799 537711 : ctx: &RequestContext,
800 537711 : ) -> anyhow::Result<()> {
801 537711 : ensure!(self.key_range.contains(&key));
802 546204 : let (_img, res) = self.blob_writer.write_blob(img, ctx).await;
803 : // TODO: re-use the buffer for `img` further upstack
804 537711 : let off = res?;
805 :
806 537711 : let mut keybuf: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
807 537711 : key.write_to_byte_slice(&mut keybuf);
808 537711 : self.tree.append(&keybuf, off)?;
809 :
810 537711 : Ok(())
811 537711 : }
812 :
813 : ///
814 : /// Finish writing the image layer.
815 : ///
816 238 : async fn finish(
817 238 : self,
818 238 : timeline: &Arc<Timeline>,
819 238 : ctx: &RequestContext,
820 238 : ) -> anyhow::Result<ResidentLayer> {
821 238 : let index_start_blk =
822 238 : ((self.blob_writer.size() + PAGE_SZ as u64 - 1) / PAGE_SZ as u64) as u32;
823 238 :
824 238 : let mut file = self.blob_writer.into_inner();
825 238 :
826 238 : // Write out the index
827 238 : file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
828 0 : .await?;
829 238 : let (index_root_blk, block_buf) = self.tree.finish()?;
830 932 : for buf in block_buf.blocks {
831 694 : let (_buf, res) = file.write_all(buf, ctx).await;
832 694 : res?;
833 : }
834 :
835 : // Fill in the summary on blk 0
836 238 : let summary = Summary {
837 238 : magic: IMAGE_FILE_MAGIC,
838 238 : format_version: STORAGE_FORMAT_VERSION,
839 238 : tenant_id: self.tenant_shard_id.tenant_id,
840 238 : timeline_id: self.timeline_id,
841 238 : key_range: self.key_range.clone(),
842 238 : lsn: self.lsn,
843 238 : index_start_blk,
844 238 : index_root_blk,
845 238 : };
846 238 :
847 238 : let mut buf = Vec::with_capacity(PAGE_SZ);
848 238 : // TODO: could use smallvec here but it's a pain with Slice<T>
849 238 : Summary::ser_into(&summary, &mut buf)?;
850 238 : file.seek(SeekFrom::Start(0)).await?;
851 238 : let (_buf, res) = file.write_all(buf, ctx).await;
852 238 : res?;
853 :
854 238 : let metadata = file
855 238 : .metadata()
856 121 : .await
857 238 : .context("get metadata to determine file size")?;
858 :
859 238 : let desc = PersistentLayerDesc::new_img(
860 238 : self.tenant_shard_id,
861 238 : self.timeline_id,
862 238 : self.key_range.clone(),
863 238 : self.lsn,
864 238 : metadata.len(),
865 238 : );
866 238 :
867 238 : // Note: Because we open the file in write-only mode, we cannot
868 238 : // reuse the same VirtualFile for reading later. That's why we don't
869 238 : // set inner.file here. The first read will have to re-open it.
870 238 :
871 238 : // fsync the file
872 238 : file.sync_all().await?;
873 :
874 : // FIXME: why not carry the virtualfile here, it supports renaming?
875 238 : let layer = Layer::finish_creating(self.conf, timeline, desc, &self.path)?;
876 :
877 237 : info!("created image layer {}", layer.local_path());
878 :
879 237 : Ok(layer)
880 238 : }
881 : }
882 :
883 : /// A builder object for constructing a new image layer.
884 : ///
885 : /// Usage:
886 : ///
887 : /// 1. Create the ImageLayerWriter by calling ImageLayerWriter::new(...)
888 : ///
889 : /// 2. Write the contents by calling `put_page_image` for every key-value
890 : /// pair in the key range.
891 : ///
892 : /// 3. Call `finish`.
893 : ///
894 : /// # Note
895 : ///
896 : /// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
897 : /// possible for the writer to drop before `finish` is actually called. So this
898 : /// could lead to odd temporary files in the directory, exhausting file system.
899 : /// This structure wraps `ImageLayerWriterInner` and also contains `Drop`
900 : /// implementation that cleans up the temporary file in failure. It's not
901 : /// possible to do this directly in `ImageLayerWriterInner` since `finish` moves
902 : /// out some fields, making it impossible to implement `Drop`.
903 : ///
904 : #[must_use]
905 : pub struct ImageLayerWriter {
906 : inner: Option<ImageLayerWriterInner>,
907 : }
908 :
909 : impl ImageLayerWriter {
910 : ///
911 : /// Start building a new image layer.
912 : ///
913 244 : pub async fn new(
914 244 : conf: &'static PageServerConf,
915 244 : timeline_id: TimelineId,
916 244 : tenant_shard_id: TenantShardId,
917 244 : key_range: &Range<Key>,
918 244 : lsn: Lsn,
919 244 : ctx: &RequestContext,
920 244 : ) -> anyhow::Result<ImageLayerWriter> {
921 244 : Ok(Self {
922 244 : inner: Some(
923 244 : ImageLayerWriterInner::new(conf, timeline_id, tenant_shard_id, key_range, lsn, ctx)
924 191 : .await?,
925 : ),
926 : })
927 244 : }
928 :
929 : ///
930 : /// Write next value to the file.
931 : ///
932 : /// The page versions must be appended in blknum order.
933 : ///
934 537711 : pub async fn put_image(
935 537711 : &mut self,
936 537711 : key: Key,
937 537711 : img: Bytes,
938 537711 : ctx: &RequestContext,
939 537711 : ) -> anyhow::Result<()> {
940 546204 : self.inner.as_mut().unwrap().put_image(key, img, ctx).await
941 537711 : }
942 :
943 : ///
944 : /// Finish writing the image layer.
945 : ///
946 238 : pub(crate) async fn finish(
947 238 : mut self,
948 238 : timeline: &Arc<Timeline>,
949 238 : ctx: &RequestContext,
950 238 : ) -> anyhow::Result<super::ResidentLayer> {
951 720 : self.inner.take().unwrap().finish(timeline, ctx).await
952 238 : }
953 : }
954 :
955 : impl Drop for ImageLayerWriter {
956 244 : fn drop(&mut self) {
957 244 : if let Some(inner) = self.inner.take() {
958 6 : inner.blob_writer.into_inner().remove();
959 238 : }
960 244 : }
961 : }
962 :
963 : #[cfg(test)]
964 : pub struct ImageLayerIterator<'a> {
965 : image_layer: &'a ImageLayerInner,
966 : ctx: &'a RequestContext,
967 : planner: crate::tenant::vectored_blob_io::StreamingVectoredReadPlanner,
968 : index_iter: crate::tenant::disk_btree::DiskBtreeIterator<'a>,
969 : key_values_batch: std::collections::VecDeque<(Key, Lsn, Value)>,
970 : is_end: bool,
971 : }
972 :
973 : #[cfg(test)]
974 : impl<'a> ImageLayerIterator<'a> {
975 : /// Retrieve a batch of key-value pairs into the iterator buffer.
976 18942 : async fn next_batch(&mut self) -> anyhow::Result<()> {
977 18942 : assert!(self.key_values_batch.is_empty());
978 18942 : assert!(!self.is_end);
979 :
980 18942 : let plan = loop {
981 28326 : if let Some(res) = self.index_iter.next().await {
982 28298 : let (raw_key, offset) = res?;
983 28298 : if let Some(batch_plan) = self.planner.handle(
984 28298 : Key::from_slice(&raw_key[..KEY_SIZE]),
985 28298 : self.image_layer.lsn,
986 28298 : offset,
987 28298 : BlobFlag::None,
988 28298 : ) {
989 18914 : break batch_plan;
990 9384 : }
991 : } else {
992 28 : self.is_end = true;
993 28 : let payload_end = self.image_layer.index_start_blk as u64 * PAGE_SZ as u64;
994 28 : break self.planner.handle_range_end(payload_end);
995 : }
996 : };
997 18942 : let vectored_blob_reader = VectoredBlobReader::new(&self.image_layer.file);
998 18942 : let mut next_batch = std::collections::VecDeque::new();
999 18942 : let buf_size = plan.size();
1000 18942 : let buf = BytesMut::with_capacity(buf_size);
1001 18942 : let blobs_buf = vectored_blob_reader
1002 18942 : .read_blobs(&plan, buf, self.ctx)
1003 9619 : .await?;
1004 18942 : let frozen_buf: Bytes = blobs_buf.buf.freeze();
1005 28270 : for meta in blobs_buf.blobs.iter() {
1006 28270 : let img_buf = frozen_buf.slice(meta.start..meta.end);
1007 28270 : next_batch.push_back((meta.meta.key, self.image_layer.lsn, Value::Image(img_buf)));
1008 28270 : }
1009 18942 : self.key_values_batch = next_batch;
1010 18942 : Ok(())
1011 18942 : }
1012 :
1013 28028 : pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
1014 28028 : if self.key_values_batch.is_empty() {
1015 18886 : if self.is_end {
1016 28 : return Ok(None);
1017 18858 : }
1018 18858 : self.next_batch().await?;
1019 9142 : }
1020 28000 : Ok(Some(
1021 28000 : self.key_values_batch
1022 28000 : .pop_front()
1023 28000 : .expect("should not be empty"),
1024 28000 : ))
1025 28028 : }
1026 : }
1027 :
1028 : #[cfg(test)]
1029 : mod test {
1030 : use std::{sync::Arc, time::Duration};
1031 :
1032 : use bytes::Bytes;
1033 : use itertools::Itertools;
1034 : use pageserver_api::{
1035 : key::Key,
1036 : shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize},
1037 : };
1038 : use utils::{
1039 : generation::Generation,
1040 : id::{TenantId, TimelineId},
1041 : lsn::Lsn,
1042 : };
1043 :
1044 : use crate::{
1045 : context::RequestContext,
1046 : repository::Value,
1047 : tenant::{
1048 : config::TenantConf,
1049 : harness::{TenantHarness, TIMELINE_ID},
1050 : storage_layer::ResidentLayer,
1051 : vectored_blob_io::StreamingVectoredReadPlanner,
1052 : Tenant, Timeline,
1053 : },
1054 : DEFAULT_PG_VERSION,
1055 : };
1056 :
1057 : use super::{ImageLayerIterator, ImageLayerWriter};
1058 :
1059 : #[tokio::test]
1060 2 : async fn image_layer_rewrite() {
1061 2 : let tenant_conf = TenantConf {
1062 2 : gc_period: Duration::ZERO,
1063 2 : compaction_period: Duration::ZERO,
1064 2 : ..TenantConf::default()
1065 2 : };
1066 2 : let tenant_id = TenantId::generate();
1067 2 : let mut gen = Generation::new(0xdead0001);
1068 10 : let mut get_next_gen = || {
1069 10 : let ret = gen;
1070 10 : gen = gen.next();
1071 10 : ret
1072 10 : };
1073 2 : // The LSN at which we will create an image layer to filter
1074 2 : let lsn = Lsn(0xdeadbeef0000);
1075 2 : let timeline_id = TimelineId::generate();
1076 2 :
1077 2 : //
1078 2 : // Create an unsharded parent with a layer.
1079 2 : //
1080 2 :
1081 2 : let harness = TenantHarness::create_custom(
1082 2 : "test_image_layer_rewrite--parent",
1083 2 : tenant_conf.clone(),
1084 2 : tenant_id,
1085 2 : ShardIdentity::unsharded(),
1086 2 : get_next_gen(),
1087 2 : )
1088 2 : .unwrap();
1089 8 : let (tenant, ctx) = harness.load().await;
1090 2 : let timeline = tenant
1091 2 : .create_test_timeline(timeline_id, lsn, DEFAULT_PG_VERSION, &ctx)
1092 6 : .await
1093 2 : .unwrap();
1094 2 :
1095 2 : // This key range contains several 0x8000 page stripes, only one of which belongs to shard zero
1096 2 : let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
1097 2 : let input_end = Key::from_hex("000000067f00000001000000ae0000020000").unwrap();
1098 2 : let range = input_start..input_end;
1099 2 :
1100 2 : // Build an image layer to filter
1101 2 : let resident = {
1102 2 : let mut writer = ImageLayerWriter::new(
1103 2 : harness.conf,
1104 2 : timeline_id,
1105 2 : harness.tenant_shard_id,
1106 2 : &range,
1107 2 : lsn,
1108 2 : &ctx,
1109 2 : )
1110 2 : .await
1111 2 : .unwrap();
1112 2 :
1113 2 : let foo_img = Bytes::from_static(&[1, 2, 3, 4]);
1114 2 : let mut key = range.start;
1115 262146 : while key < range.end {
1116 266239 : writer.put_image(key, foo_img.clone(), &ctx).await.unwrap();
1117 262144 :
1118 262144 : key = key.next();
1119 2 : }
1120 119 : writer.finish(&timeline, &ctx).await.unwrap()
1121 2 : };
1122 2 : let original_size = resident.metadata().file_size;
1123 2 :
1124 2 : //
1125 2 : // Create child shards and do the rewrite, exercising filter().
1126 2 : // TODO: abstraction in TenantHarness for splits.
1127 2 : //
1128 2 :
1129 2 : // Filter for various shards: this exercises cases like values at start of key range, end of key
1130 2 : // range, middle of key range.
1131 2 : let shard_count = ShardCount::new(4);
1132 8 : for shard_number in 0..shard_count.count() {
1133 2 : //
1134 2 : // mimic the shard split
1135 2 : //
1136 8 : let shard_identity = ShardIdentity::new(
1137 8 : ShardNumber(shard_number),
1138 8 : shard_count,
1139 8 : ShardStripeSize(0x8000),
1140 8 : )
1141 8 : .unwrap();
1142 8 : let harness = TenantHarness::create_custom(
1143 8 : Box::leak(Box::new(format!(
1144 8 : "test_image_layer_rewrite--child{}",
1145 8 : shard_identity.shard_slug()
1146 8 : ))),
1147 8 : tenant_conf.clone(),
1148 8 : tenant_id,
1149 8 : shard_identity,
1150 8 : // NB: in reality, the shards would each fork off their own gen number sequence from the parent.
1151 8 : // But here, all we care about is that the gen number is unique.
1152 8 : get_next_gen(),
1153 8 : )
1154 8 : .unwrap();
1155 32 : let (tenant, ctx) = harness.load().await;
1156 8 : let timeline = tenant
1157 8 : .create_test_timeline(timeline_id, lsn, DEFAULT_PG_VERSION, &ctx)
1158 20 : .await
1159 8 : .unwrap();
1160 2 :
1161 2 : //
1162 2 : // use filter() and make assertions
1163 2 : //
1164 2 :
1165 8 : let mut filtered_writer = ImageLayerWriter::new(
1166 8 : harness.conf,
1167 8 : timeline_id,
1168 8 : harness.tenant_shard_id,
1169 8 : &range,
1170 8 : lsn,
1171 8 : &ctx,
1172 8 : )
1173 4 : .await
1174 8 : .unwrap();
1175 2 :
1176 8 : let wrote_keys = resident
1177 8 : .filter(&shard_identity, &mut filtered_writer, &ctx)
1178 266719 : .await
1179 8 : .unwrap();
1180 8 : let replacement = if wrote_keys > 0 {
1181 129 : Some(filtered_writer.finish(&timeline, &ctx).await.unwrap())
1182 2 : } else {
1183 2 : None
1184 2 : };
1185 2 :
1186 2 : // This exact size and those below will need updating as/when the layer encoding changes, but
1187 2 : // should be deterministic for a given version of the format, as we used no randomness generating the input.
1188 8 : assert_eq!(original_size, 1597440);
1189 2 :
1190 8 : match shard_number {
1191 2 : 0 => {
1192 2 : // We should have written out just one stripe for our shard identity
1193 2 : assert_eq!(wrote_keys, 0x8000);
1194 2 : let replacement = replacement.unwrap();
1195 2 :
1196 2 : // We should have dropped some of the data
1197 2 : assert!(replacement.metadata().file_size < original_size);
1198 2 : assert!(replacement.metadata().file_size > 0);
1199 2 :
1200 2 : // Assert that we dropped ~3/4 of the data.
1201 2 : assert_eq!(replacement.metadata().file_size, 417792);
1202 2 : }
1203 2 : 1 => {
1204 2 : // Shard 1 has no keys in our input range
1205 2 : assert_eq!(wrote_keys, 0x0);
1206 2 : assert!(replacement.is_none());
1207 2 : }
1208 2 : 2 => {
1209 2 : // Shard 2 has one stripes in the input range
1210 2 : assert_eq!(wrote_keys, 0x8000);
1211 2 : let replacement = replacement.unwrap();
1212 2 : assert!(replacement.metadata().file_size < original_size);
1213 2 : assert!(replacement.metadata().file_size > 0);
1214 2 : assert_eq!(replacement.metadata().file_size, 417792);
1215 2 : }
1216 2 : 3 => {
1217 2 : // Shard 3 has two stripes in the input range
1218 2 : assert_eq!(wrote_keys, 0x10000);
1219 2 : let replacement = replacement.unwrap();
1220 2 : assert!(replacement.metadata().file_size < original_size);
1221 2 : assert!(replacement.metadata().file_size > 0);
1222 2 : assert_eq!(replacement.metadata().file_size, 811008);
1223 2 : }
1224 2 : _ => unreachable!(),
1225 2 : }
1226 2 : }
1227 2 : }
1228 :
1229 2 : async fn produce_image_layer(
1230 2 : tenant: &Tenant,
1231 2 : tline: &Arc<Timeline>,
1232 2 : mut images: Vec<(Key, Bytes)>,
1233 2 : lsn: Lsn,
1234 2 : ctx: &RequestContext,
1235 2 : ) -> anyhow::Result<ResidentLayer> {
1236 2 : images.sort();
1237 2 : let (key_start, _) = images.first().unwrap();
1238 2 : let (key_last, _) = images.last().unwrap();
1239 2 : let key_end = key_last.next();
1240 2 : let key_range = *key_start..key_end;
1241 2 : let mut writer = ImageLayerWriter::new(
1242 2 : tenant.conf,
1243 2 : tline.timeline_id,
1244 2 : tenant.tenant_shard_id,
1245 2 : &key_range,
1246 2 : lsn,
1247 2 : ctx,
1248 2 : )
1249 1 : .await?;
1250 :
1251 2002 : for (key, img) in images {
1252 2031 : writer.put_image(key, img, ctx).await?;
1253 : }
1254 4 : let img_layer = writer.finish(tline, ctx).await?;
1255 :
1256 2 : Ok::<_, anyhow::Error>(img_layer)
1257 2 : }
1258 :
1259 28 : async fn assert_img_iter_equal(
1260 28 : img_iter: &mut ImageLayerIterator<'_>,
1261 28 : expect: &[(Key, Bytes)],
1262 28 : expect_lsn: Lsn,
1263 28 : ) {
1264 28 : let mut expect_iter = expect.iter();
1265 : loop {
1266 28028 : let o1 = img_iter.next().await.unwrap();
1267 28028 : let o2 = expect_iter.next();
1268 28028 : match (o1, o2) {
1269 28 : (None, None) => break,
1270 28000 : (Some((k1, l1, v1)), Some((k2, i2))) => {
1271 28000 : let Value::Image(i1) = v1 else {
1272 0 : panic!("expect Value::Image")
1273 : };
1274 28000 : assert_eq!(&k1, k2);
1275 28000 : assert_eq!(l1, expect_lsn);
1276 28000 : assert_eq!(&i1, i2);
1277 : }
1278 0 : (o1, o2) => panic!("iterators length mismatch: {:?}, {:?}", o1, o2),
1279 : }
1280 : }
1281 28 : }
1282 :
1283 : #[tokio::test]
1284 2 : async fn image_layer_iterator() {
1285 2 : let harness = TenantHarness::create("image_layer_iterator").unwrap();
1286 8 : let (tenant, ctx) = harness.load().await;
1287 2 :
1288 2 : let tline = tenant
1289 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
1290 6 : .await
1291 2 : .unwrap();
1292 2 :
1293 2000 : fn get_key(id: u32) -> Key {
1294 2000 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
1295 2000 : key.field6 = id;
1296 2000 : key
1297 2000 : }
1298 2 : const N: usize = 1000;
1299 2 : let test_imgs = (0..N)
1300 2000 : .map(|idx| (get_key(idx as u32), Bytes::from(format!("img{idx:05}"))))
1301 2 : .collect_vec();
1302 2 : let resident_layer =
1303 2 : produce_image_layer(&tenant, &tline, test_imgs.clone(), Lsn(0x10), &ctx)
1304 2036 : .await
1305 2 : .unwrap();
1306 2 : let img_layer = resident_layer.get_as_image(&ctx).await.unwrap();
1307 6 : for max_read_size in [1, 1024] {
1308 32 : for batch_size in [1, 2, 4, 8, 3, 7, 13] {
1309 28 : println!("running with batch_size={batch_size} max_read_size={max_read_size}");
1310 28 : // Test if the batch size is correctly determined
1311 28 : let mut iter = img_layer.iter(&ctx);
1312 28 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
1313 28 : let mut num_items = 0;
1314 112 : for _ in 0..3 {
1315 84 : iter.next_batch().await.unwrap();
1316 84 : num_items += iter.key_values_batch.len();
1317 84 : if max_read_size == 1 {
1318 2 : // every key should be a batch b/c the value is larger than max_read_size
1319 42 : assert_eq!(iter.key_values_batch.len(), 1);
1320 2 : } else {
1321 42 : assert_eq!(iter.key_values_batch.len(), batch_size);
1322 2 : }
1323 84 : if num_items >= N {
1324 2 : break;
1325 84 : }
1326 84 : iter.key_values_batch.clear();
1327 2 : }
1328 2 : // Test if the result is correct
1329 28 : let mut iter = img_layer.iter(&ctx);
1330 28 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
1331 9576 : assert_img_iter_equal(&mut iter, &test_imgs, Lsn(0x10)).await;
1332 2 : }
1333 2 : }
1334 2 : }
1335 : }
|
