Line data Source code
1 : //! An ImageLayer represents an image or a snapshot of a key-range at
2 : //! one particular LSN.
3 : //!
4 : //! It contains an image of all key-value pairs in its key-range. Any key
5 : //! that falls into the image layer's range but does not exist in the layer,
6 : //! does not exist.
7 : //!
8 : //! An image layer is stored in a file on disk. The file is stored in
9 : //! timelines/<timeline_id> directory. Currently, there are no
10 : //! subdirectories, and each image layer file is named like this:
11 : //!
12 : //! ```text
13 : //! <key start>-<key end>__<LSN>
14 : //! ```
15 : //!
16 : //! For example:
17 : //!
18 : //! ```text
19 : //! 000000067F000032BE0000400000000070B6-000000067F000032BE0000400000000080B6__00000000346BC568
20 : //! ```
21 : //!
22 : //! Every image layer file consists of three parts: "summary",
23 : //! "index", and "values". The summary is a fixed size header at the
24 : //! beginning of the file, and it contains basic information about the
25 : //! layer, and offsets to the other parts. The "index" is a B-tree,
26 : //! mapping from Key to an offset in the "values" part. The
27 : //! actual page images are stored in the "values" part.
28 : use crate::config::PageServerConf;
29 : use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
30 : use crate::page_cache::{self, FileId, PAGE_SZ};
31 : use crate::repository::{Key, Value, KEY_SIZE};
32 : use crate::tenant::blob_io::BlobWriter;
33 : use crate::tenant::block_io::{BlockBuf, FileBlockReader};
34 : use crate::tenant::disk_btree::{
35 : DiskBtreeBuilder, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
36 : };
37 : use crate::tenant::timeline::GetVectoredError;
38 : use crate::tenant::vectored_blob_io::{
39 : BlobFlag, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead, VectoredReadPlanner,
40 : };
41 : use crate::tenant::{PageReconstructError, Timeline};
42 : use crate::virtual_file::owned_buffers_io::io_buf_ext::IoBufExt;
43 : use crate::virtual_file::{self, VirtualFile};
44 : use crate::{IMAGE_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
45 : use anyhow::{anyhow, bail, ensure, Context, Result};
46 : use bytes::{Bytes, BytesMut};
47 : use camino::{Utf8Path, Utf8PathBuf};
48 : use hex;
49 : use itertools::Itertools;
50 : use pageserver_api::config::MaxVectoredReadBytes;
51 : use pageserver_api::keyspace::KeySpace;
52 : use pageserver_api::shard::{ShardIdentity, TenantShardId};
53 : use rand::{distributions::Alphanumeric, Rng};
54 : use serde::{Deserialize, Serialize};
55 : use std::collections::VecDeque;
56 : use std::fs::File;
57 : use std::io::SeekFrom;
58 : use std::ops::Range;
59 : use std::os::unix::prelude::FileExt;
60 : use std::str::FromStr;
61 : use std::sync::Arc;
62 : use tokio::sync::OnceCell;
63 : use tokio_stream::StreamExt;
64 : use tracing::*;
65 :
66 : use utils::{
67 : bin_ser::BeSer,
68 : id::{TenantId, TimelineId},
69 : lsn::Lsn,
70 : };
71 :
72 : use super::layer_name::ImageLayerName;
73 : use super::{
74 : AsLayerDesc, Layer, LayerName, PersistentLayerDesc, ResidentLayer, ValuesReconstructState,
75 : };
76 :
77 : ///
78 : /// Header stored in the beginning of the file
79 : ///
80 : /// After this comes the 'values' part, starting on block 1. After that,
81 : /// the 'index' starts at the block indicated by 'index_start_blk'
82 : ///
83 306 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
84 : pub struct Summary {
85 : /// Magic value to identify this as a neon image file. Always IMAGE_FILE_MAGIC.
86 : pub magic: u16,
87 : pub format_version: u16,
88 :
89 : pub tenant_id: TenantId,
90 : pub timeline_id: TimelineId,
91 : pub key_range: Range<Key>,
92 : pub lsn: Lsn,
93 :
94 : /// Block number where the 'index' part of the file begins.
95 : pub index_start_blk: u32,
96 : /// Block within the 'index', where the B-tree root page is stored
97 : pub index_root_blk: u32,
98 : // the 'values' part starts after the summary header, on block 1.
99 : }
100 :
101 : impl From<&ImageLayer> for Summary {
102 0 : fn from(layer: &ImageLayer) -> Self {
103 0 : Self::expected(
104 0 : layer.desc.tenant_shard_id.tenant_id,
105 0 : layer.desc.timeline_id,
106 0 : layer.desc.key_range.clone(),
107 0 : layer.lsn,
108 0 : )
109 0 : }
110 : }
111 :
112 : impl Summary {
113 306 : pub(super) fn expected(
114 306 : tenant_id: TenantId,
115 306 : timeline_id: TimelineId,
116 306 : key_range: Range<Key>,
117 306 : lsn: Lsn,
118 306 : ) -> Self {
119 306 : Self {
120 306 : magic: IMAGE_FILE_MAGIC,
121 306 : format_version: STORAGE_FORMAT_VERSION,
122 306 : tenant_id,
123 306 : timeline_id,
124 306 : key_range,
125 306 : lsn,
126 306 :
127 306 : index_start_blk: 0,
128 306 : index_root_blk: 0,
129 306 : }
130 306 : }
131 : }
132 :
133 : /// This is used only from `pagectl`. Within pageserver, all layers are
134 : /// [`crate::tenant::storage_layer::Layer`], which can hold an [`ImageLayerInner`].
135 : pub struct ImageLayer {
136 : path: Utf8PathBuf,
137 : pub desc: PersistentLayerDesc,
138 : // This entry contains an image of all pages as of this LSN, should be the same as desc.lsn
139 : pub lsn: Lsn,
140 : inner: OnceCell<ImageLayerInner>,
141 : }
142 :
143 : impl std::fmt::Debug for ImageLayer {
144 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
145 : use super::RangeDisplayDebug;
146 :
147 0 : f.debug_struct("ImageLayer")
148 0 : .field("key_range", &RangeDisplayDebug(&self.desc.key_range))
149 0 : .field("file_size", &self.desc.file_size)
150 0 : .field("lsn", &self.lsn)
151 0 : .field("inner", &self.inner)
152 0 : .finish()
153 0 : }
154 : }
155 :
156 : /// ImageLayer is the in-memory data structure associated with an on-disk image
157 : /// file.
158 : pub struct ImageLayerInner {
159 : // values copied from summary
160 : index_start_blk: u32,
161 : index_root_blk: u32,
162 :
163 : key_range: Range<Key>,
164 : lsn: Lsn,
165 :
166 : file: VirtualFile,
167 : file_id: FileId,
168 :
169 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
170 : }
171 :
172 : impl ImageLayerInner {
173 0 : pub(crate) fn layer_dbg_info(&self) -> String {
174 0 : format!(
175 0 : "image {}..{} {}",
176 0 : self.key_range().start,
177 0 : self.key_range().end,
178 0 : self.lsn()
179 0 : )
180 0 : }
181 : }
182 :
183 : impl std::fmt::Debug for ImageLayerInner {
184 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
185 0 : f.debug_struct("ImageLayerInner")
186 0 : .field("index_start_blk", &self.index_start_blk)
187 0 : .field("index_root_blk", &self.index_root_blk)
188 0 : .finish()
189 0 : }
190 : }
191 :
192 : impl ImageLayerInner {
193 0 : pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
194 0 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
195 0 : let tree_reader = DiskBtreeReader::<_, KEY_SIZE>::new(
196 0 : self.index_start_blk,
197 0 : self.index_root_blk,
198 0 : block_reader,
199 0 : );
200 0 :
201 0 : tree_reader.dump().await?;
202 :
203 0 : tree_reader
204 0 : .visit(
205 0 : &[0u8; KEY_SIZE],
206 0 : VisitDirection::Forwards,
207 0 : |key, value| {
208 0 : println!("key: {} offset {}", hex::encode(key), value);
209 0 : true
210 0 : },
211 0 : ctx,
212 0 : )
213 0 : .await?;
214 :
215 0 : Ok(())
216 0 : }
217 : }
218 :
219 : /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
220 : impl std::fmt::Display for ImageLayer {
221 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
222 0 : write!(f, "{}", self.layer_desc().short_id())
223 0 : }
224 : }
225 :
226 : impl AsLayerDesc for ImageLayer {
227 0 : fn layer_desc(&self) -> &PersistentLayerDesc {
228 0 : &self.desc
229 0 : }
230 : }
231 :
232 : impl ImageLayer {
233 0 : pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
234 0 : self.desc.dump();
235 0 :
236 0 : if !verbose {
237 0 : return Ok(());
238 0 : }
239 :
240 0 : let inner = self.load(ctx).await?;
241 :
242 0 : inner.dump(ctx).await?;
243 :
244 0 : Ok(())
245 0 : }
246 :
247 1470 : fn temp_path_for(
248 1470 : conf: &PageServerConf,
249 1470 : timeline_id: TimelineId,
250 1470 : tenant_shard_id: TenantShardId,
251 1470 : fname: &ImageLayerName,
252 1470 : ) -> Utf8PathBuf {
253 1470 : let rand_string: String = rand::thread_rng()
254 1470 : .sample_iter(&Alphanumeric)
255 1470 : .take(8)
256 1470 : .map(char::from)
257 1470 : .collect();
258 1470 :
259 1470 : conf.timeline_path(&tenant_shard_id, &timeline_id)
260 1470 : .join(format!("{fname}.{rand_string}.{TEMP_FILE_SUFFIX}"))
261 1470 : }
262 :
263 : ///
264 : /// Open the underlying file and read the metadata into memory, if it's
265 : /// not loaded already.
266 : ///
267 0 : async fn load(&self, ctx: &RequestContext) -> Result<&ImageLayerInner> {
268 0 : self.inner
269 0 : .get_or_try_init(|| self.load_inner(ctx))
270 0 : .await
271 0 : .with_context(|| format!("Failed to load image layer {}", self.path()))
272 0 : }
273 :
274 0 : async fn load_inner(&self, ctx: &RequestContext) -> Result<ImageLayerInner> {
275 0 : let path = self.path();
276 :
277 0 : let loaded =
278 0 : ImageLayerInner::load(&path, self.desc.image_layer_lsn(), None, None, ctx).await?;
279 :
280 : // not production code
281 0 : let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
282 0 : let expected_layer_name = self.layer_desc().layer_name();
283 0 :
284 0 : if actual_layer_name != expected_layer_name {
285 0 : println!("warning: filename does not match what is expected from in-file summary");
286 0 : println!("actual: {:?}", actual_layer_name.to_string());
287 0 : println!("expected: {:?}", expected_layer_name.to_string());
288 0 : }
289 :
290 0 : Ok(loaded)
291 0 : }
292 :
293 : /// Create an ImageLayer struct representing an existing file on disk.
294 : ///
295 : /// This variant is only used for debugging purposes, by the 'pagectl' binary.
296 0 : pub fn new_for_path(path: &Utf8Path, file: File) -> Result<ImageLayer> {
297 0 : let mut summary_buf = vec![0; PAGE_SZ];
298 0 : file.read_exact_at(&mut summary_buf, 0)?;
299 0 : let summary = Summary::des_prefix(&summary_buf)?;
300 0 : let metadata = file
301 0 : .metadata()
302 0 : .context("get file metadata to determine size")?;
303 :
304 : // This function is never used for constructing layers in a running pageserver,
305 : // so it does not need an accurate TenantShardId.
306 0 : let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
307 0 :
308 0 : Ok(ImageLayer {
309 0 : path: path.to_path_buf(),
310 0 : desc: PersistentLayerDesc::new_img(
311 0 : tenant_shard_id,
312 0 : summary.timeline_id,
313 0 : summary.key_range,
314 0 : summary.lsn,
315 0 : metadata.len(),
316 0 : ), // Now we assume image layer ALWAYS covers the full range. This may change in the future.
317 0 : lsn: summary.lsn,
318 0 : inner: OnceCell::new(),
319 0 : })
320 0 : }
321 :
322 0 : fn path(&self) -> Utf8PathBuf {
323 0 : self.path.clone()
324 0 : }
325 : }
326 :
327 0 : #[derive(thiserror::Error, Debug)]
328 : pub enum RewriteSummaryError {
329 : #[error("magic mismatch")]
330 : MagicMismatch,
331 : #[error(transparent)]
332 : Other(#[from] anyhow::Error),
333 : }
334 :
335 : impl From<std::io::Error> for RewriteSummaryError {
336 0 : fn from(e: std::io::Error) -> Self {
337 0 : Self::Other(anyhow::anyhow!(e))
338 0 : }
339 : }
340 :
341 : impl ImageLayer {
342 0 : pub async fn rewrite_summary<F>(
343 0 : path: &Utf8Path,
344 0 : rewrite: F,
345 0 : ctx: &RequestContext,
346 0 : ) -> Result<(), RewriteSummaryError>
347 0 : where
348 0 : F: Fn(Summary) -> Summary,
349 0 : {
350 0 : let mut file = VirtualFile::open_with_options(
351 0 : path,
352 0 : virtual_file::OpenOptions::new().read(true).write(true),
353 0 : ctx,
354 0 : )
355 0 : .await
356 0 : .with_context(|| format!("Failed to open file '{}'", path))?;
357 0 : let file_id = page_cache::next_file_id();
358 0 : let block_reader = FileBlockReader::new(&file, file_id);
359 0 : let summary_blk = block_reader.read_blk(0, ctx).await?;
360 0 : let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
361 0 : if actual_summary.magic != IMAGE_FILE_MAGIC {
362 0 : return Err(RewriteSummaryError::MagicMismatch);
363 0 : }
364 0 :
365 0 : let new_summary = rewrite(actual_summary);
366 0 :
367 0 : let mut buf = Vec::with_capacity(PAGE_SZ);
368 0 : // TODO: could use smallvec here but it's a pain with Slice<T>
369 0 : Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
370 0 : file.seek(SeekFrom::Start(0)).await?;
371 0 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
372 0 : res?;
373 0 : Ok(())
374 0 : }
375 : }
376 :
377 : impl ImageLayerInner {
378 108 : pub(crate) fn key_range(&self) -> &Range<Key> {
379 108 : &self.key_range
380 108 : }
381 :
382 108 : pub(crate) fn lsn(&self) -> Lsn {
383 108 : self.lsn
384 108 : }
385 :
386 306 : pub(super) async fn load(
387 306 : path: &Utf8Path,
388 306 : lsn: Lsn,
389 306 : summary: Option<Summary>,
390 306 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
391 306 : ctx: &RequestContext,
392 306 : ) -> anyhow::Result<Self> {
393 306 : let file = VirtualFile::open(path, ctx)
394 153 : .await
395 306 : .context("open layer file")?;
396 306 : let file_id = page_cache::next_file_id();
397 306 : let block_reader = FileBlockReader::new(&file, file_id);
398 306 : let summary_blk = block_reader
399 306 : .read_blk(0, ctx)
400 157 : .await
401 306 : .context("read first block")?;
402 :
403 : // length is the only way how this could fail, so it's not actually likely at all unless
404 : // read_blk returns wrong sized block.
405 : //
406 : // TODO: confirm and make this into assertion
407 306 : let actual_summary =
408 306 : Summary::des_prefix(summary_blk.as_ref()).context("deserialize first block")?;
409 :
410 306 : if let Some(mut expected_summary) = summary {
411 : // production code path
412 306 : expected_summary.index_start_blk = actual_summary.index_start_blk;
413 306 : expected_summary.index_root_blk = actual_summary.index_root_blk;
414 306 : // mask out the timeline_id, but still require the layers to be from the same tenant
415 306 : expected_summary.timeline_id = actual_summary.timeline_id;
416 306 :
417 306 : if actual_summary != expected_summary {
418 0 : bail!(
419 0 : "in-file summary does not match expected summary. actual = {:?} expected = {:?}",
420 0 : actual_summary,
421 0 : expected_summary
422 0 : );
423 306 : }
424 0 : }
425 :
426 306 : Ok(ImageLayerInner {
427 306 : index_start_blk: actual_summary.index_start_blk,
428 306 : index_root_blk: actual_summary.index_root_blk,
429 306 : lsn,
430 306 : file,
431 306 : file_id,
432 306 : max_vectored_read_bytes,
433 306 : key_range: actual_summary.key_range,
434 306 : })
435 306 : }
436 :
437 : // Look up the keys in the provided keyspace and update
438 : // the reconstruct state with whatever is found.
439 24061 : pub(super) async fn get_values_reconstruct_data(
440 24061 : &self,
441 24061 : keyspace: KeySpace,
442 24061 : reconstruct_state: &mut ValuesReconstructState,
443 24061 : ctx: &RequestContext,
444 24061 : ) -> Result<(), GetVectoredError> {
445 24061 : let reads = self
446 24061 : .plan_reads(keyspace, None, ctx)
447 2128 : .await
448 24061 : .map_err(GetVectoredError::Other)?;
449 :
450 24061 : self.do_reads_and_update_state(reads, reconstruct_state, ctx)
451 14111 : .await;
452 :
453 24061 : reconstruct_state.on_image_layer_visited(&self.key_range);
454 24061 :
455 24061 : Ok(())
456 24061 : }
457 :
458 : /// Traverse the layer's index to build read operations on the overlap of the input keyspace
459 : /// and the keys in this layer.
460 : ///
461 : /// If shard_identity is provided, it will be used to filter keys down to those stored on
462 : /// this shard.
463 24085 : async fn plan_reads(
464 24085 : &self,
465 24085 : keyspace: KeySpace,
466 24085 : shard_identity: Option<&ShardIdentity>,
467 24085 : ctx: &RequestContext,
468 24085 : ) -> anyhow::Result<Vec<VectoredRead>> {
469 24085 : let mut planner = VectoredReadPlanner::new(
470 24085 : self.max_vectored_read_bytes
471 24085 : .expect("Layer is loaded with max vectored bytes config")
472 24085 : .0
473 24085 : .into(),
474 24085 : );
475 24085 :
476 24085 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
477 24085 : let tree_reader =
478 24085 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
479 24085 :
480 24085 : let ctx = RequestContextBuilder::extend(ctx)
481 24085 : .page_content_kind(PageContentKind::ImageLayerBtreeNode)
482 24085 : .build();
483 :
484 89115 : for range in keyspace.ranges.iter() {
485 89115 : let mut range_end_handled = false;
486 89115 : let mut search_key: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
487 89115 : range.start.write_to_byte_slice(&mut search_key);
488 89115 :
489 89115 : let index_stream = tree_reader.clone().into_stream(&search_key, &ctx);
490 89115 : let mut index_stream = std::pin::pin!(index_stream);
491 :
492 3289835 : while let Some(index_entry) = index_stream.next().await {
493 3289236 : let (raw_key, offset) = index_entry?;
494 :
495 3289236 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
496 3289236 : assert!(key >= range.start);
497 :
498 3289236 : let flag = if let Some(shard_identity) = shard_identity {
499 3145728 : if shard_identity.is_key_disposable(&key) {
500 2359296 : BlobFlag::Ignore
501 : } else {
502 786432 : BlobFlag::None
503 : }
504 : } else {
505 143508 : BlobFlag::None
506 : };
507 :
508 3289236 : if key >= range.end {
509 88516 : planner.handle_range_end(offset);
510 88516 : range_end_handled = true;
511 88516 : break;
512 3200720 : } else {
513 3200720 : planner.handle(key, self.lsn, offset, flag);
514 3200720 : }
515 : }
516 :
517 89115 : if !range_end_handled {
518 599 : let payload_end = self.index_start_blk as u64 * PAGE_SZ as u64;
519 599 : planner.handle_range_end(payload_end);
520 88516 : }
521 : }
522 :
523 24085 : Ok(planner.finish())
524 24085 : }
525 :
526 : /// Given a key range, select the parts of that range that should be retained by the ShardIdentity,
527 : /// then execute vectored GET operations, passing the results of all read keys into the writer.
528 24 : pub(super) async fn filter(
529 24 : &self,
530 24 : shard_identity: &ShardIdentity,
531 24 : writer: &mut ImageLayerWriter,
532 24 : ctx: &RequestContext,
533 24 : ) -> anyhow::Result<usize> {
534 : // Fragment the range into the regions owned by this ShardIdentity
535 24 : let plan = self
536 24 : .plan_reads(
537 24 : KeySpace {
538 24 : // If asked for the total key space, plan_reads will give us all the keys in the layer
539 24 : ranges: vec![Key::MIN..Key::MAX],
540 24 : },
541 24 : Some(shard_identity),
542 24 : ctx,
543 24 : )
544 1407 : .await?;
545 :
546 24 : let vectored_blob_reader = VectoredBlobReader::new(&self.file);
547 24 : let mut key_count = 0;
548 48 : for read in plan.into_iter() {
549 48 : let buf_size = read.size();
550 48 :
551 48 : let buf = BytesMut::with_capacity(buf_size);
552 48 : let blobs_buf = vectored_blob_reader.read_blobs(&read, buf, ctx).await?;
553 :
554 48 : let frozen_buf = blobs_buf.buf.freeze();
555 :
556 786432 : for meta in blobs_buf.blobs.iter() {
557 786432 : let img_buf = frozen_buf.slice(meta.start..meta.end);
558 786432 :
559 786432 : key_count += 1;
560 786432 : writer
561 786432 : .put_image(meta.meta.key, img_buf, ctx)
562 798720 : .await
563 786432 : .context(format!("Storing key {}", meta.meta.key))?;
564 : }
565 : }
566 :
567 24 : Ok(key_count)
568 24 : }
569 :
570 24061 : async fn do_reads_and_update_state(
571 24061 : &self,
572 24061 : reads: Vec<VectoredRead>,
573 24061 : reconstruct_state: &mut ValuesReconstructState,
574 24061 : ctx: &RequestContext,
575 24061 : ) {
576 24061 : let max_vectored_read_bytes = self
577 24061 : .max_vectored_read_bytes
578 24061 : .expect("Layer is loaded with max vectored bytes config")
579 24061 : .0
580 24061 : .into();
581 24061 :
582 24061 : let vectored_blob_reader = VectoredBlobReader::new(&self.file);
583 27665 : for read in reads.into_iter() {
584 27665 : let buf_size = read.size();
585 27665 :
586 27665 : if buf_size > max_vectored_read_bytes {
587 : // If the read is oversized, it should only contain one key.
588 0 : let offenders = read
589 0 : .blobs_at
590 0 : .as_slice()
591 0 : .iter()
592 0 : .map(|(_, blob_meta)| format!("{}@{}", blob_meta.key, blob_meta.lsn))
593 0 : .join(", ");
594 0 : tracing::warn!(
595 0 : "Oversized vectored read ({} > {}) for keys {}",
596 : buf_size,
597 : max_vectored_read_bytes,
598 : offenders
599 : );
600 27665 : }
601 :
602 27665 : let buf = BytesMut::with_capacity(buf_size);
603 27665 : let res = vectored_blob_reader.read_blobs(&read, buf, ctx).await;
604 :
605 27665 : match res {
606 27665 : Ok(blobs_buf) => {
607 27665 : let frozen_buf = blobs_buf.buf.freeze();
608 :
609 54992 : for meta in blobs_buf.blobs.iter() {
610 54992 : let img_buf = frozen_buf.slice(meta.start..meta.end);
611 54992 : reconstruct_state.update_key(
612 54992 : &meta.meta.key,
613 54992 : self.lsn,
614 54992 : Value::Image(img_buf),
615 54992 : );
616 54992 : }
617 : }
618 0 : Err(err) => {
619 0 : let kind = err.kind();
620 0 : for (_, blob_meta) in read.blobs_at.as_slice() {
621 0 : reconstruct_state.on_key_error(
622 0 : blob_meta.key,
623 0 : PageReconstructError::from(anyhow!(
624 0 : "Failed to read blobs from virtual file {}: {}",
625 0 : self.file.path,
626 0 : kind
627 0 : )),
628 0 : );
629 0 : }
630 : }
631 : };
632 : }
633 24061 : }
634 :
635 276 : pub(crate) fn iter<'a>(&'a self, ctx: &'a RequestContext) -> ImageLayerIterator<'a> {
636 276 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
637 276 : let tree_reader =
638 276 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
639 276 : ImageLayerIterator {
640 276 : image_layer: self,
641 276 : ctx,
642 276 : index_iter: tree_reader.iter(&[0; KEY_SIZE], ctx),
643 276 : key_values_batch: VecDeque::new(),
644 276 : is_end: false,
645 276 : planner: StreamingVectoredReadPlanner::new(
646 276 : 1024 * 8192, // The default value. Unit tests might use a different value. 1024 * 8K = 8MB buffer.
647 276 : 1024, // The default value. Unit tests might use a different value
648 276 : ),
649 276 : }
650 276 : }
651 : }
652 :
653 : /// A builder object for constructing a new image layer.
654 : ///
655 : /// Usage:
656 : ///
657 : /// 1. Create the ImageLayerWriter by calling ImageLayerWriter::new(...)
658 : ///
659 : /// 2. Write the contents by calling `put_page_image` for every key-value
660 : /// pair in the key range.
661 : ///
662 : /// 3. Call `finish`.
663 : ///
664 : struct ImageLayerWriterInner {
665 : conf: &'static PageServerConf,
666 : path: Utf8PathBuf,
667 : timeline_id: TimelineId,
668 : tenant_shard_id: TenantShardId,
669 : key_range: Range<Key>,
670 : lsn: Lsn,
671 :
672 : // Total uncompressed bytes passed into put_image
673 : uncompressed_bytes: u64,
674 :
675 : // Like `uncompressed_bytes`,
676 : // but only of images we might consider for compression
677 : uncompressed_bytes_eligible: u64,
678 :
679 : // Like `uncompressed_bytes`, but only of images
680 : // where we have chosen their compressed form
681 : uncompressed_bytes_chosen: u64,
682 :
683 : // Number of keys in the layer.
684 : num_keys: usize,
685 :
686 : blob_writer: BlobWriter<false>,
687 : tree: DiskBtreeBuilder<BlockBuf, KEY_SIZE>,
688 :
689 : #[cfg(feature = "testing")]
690 : last_written_key: Key,
691 : }
692 :
693 : impl ImageLayerWriterInner {
694 : ///
695 : /// Start building a new image layer.
696 : ///
697 1470 : async fn new(
698 1470 : conf: &'static PageServerConf,
699 1470 : timeline_id: TimelineId,
700 1470 : tenant_shard_id: TenantShardId,
701 1470 : key_range: &Range<Key>,
702 1470 : lsn: Lsn,
703 1470 : ctx: &RequestContext,
704 1470 : ) -> anyhow::Result<Self> {
705 1470 : // Create the file initially with a temporary filename.
706 1470 : // We'll atomically rename it to the final name when we're done.
707 1470 : let path = ImageLayer::temp_path_for(
708 1470 : conf,
709 1470 : timeline_id,
710 1470 : tenant_shard_id,
711 1470 : &ImageLayerName {
712 1470 : key_range: key_range.clone(),
713 1470 : lsn,
714 1470 : },
715 1470 : );
716 1470 : trace!("creating image layer {}", path);
717 1470 : let mut file = {
718 1470 : VirtualFile::open_with_options(
719 1470 : &path,
720 1470 : virtual_file::OpenOptions::new()
721 1470 : .write(true)
722 1470 : .create_new(true),
723 1470 : ctx,
724 1470 : )
725 982 : .await?
726 : };
727 : // make room for the header block
728 1470 : file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
729 1470 : let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
730 1470 :
731 1470 : // Initialize the b-tree index builder
732 1470 : let block_buf = BlockBuf::new();
733 1470 : let tree_builder = DiskBtreeBuilder::new(block_buf);
734 1470 :
735 1470 : let writer = Self {
736 1470 : conf,
737 1470 : path,
738 1470 : timeline_id,
739 1470 : tenant_shard_id,
740 1470 : key_range: key_range.clone(),
741 1470 : lsn,
742 1470 : tree: tree_builder,
743 1470 : blob_writer,
744 1470 : uncompressed_bytes: 0,
745 1470 : uncompressed_bytes_eligible: 0,
746 1470 : uncompressed_bytes_chosen: 0,
747 1470 : num_keys: 0,
748 1470 : #[cfg(feature = "testing")]
749 1470 : last_written_key: Key::MIN,
750 1470 : };
751 1470 :
752 1470 : Ok(writer)
753 1470 : }
754 :
755 : ///
756 : /// Write next value to the file.
757 : ///
758 : /// The page versions must be appended in blknum order.
759 : ///
760 1638450 : async fn put_image(
761 1638450 : &mut self,
762 1638450 : key: Key,
763 1638450 : img: Bytes,
764 1638450 : ctx: &RequestContext,
765 1638450 : ) -> anyhow::Result<()> {
766 1638450 : ensure!(self.key_range.contains(&key));
767 1638450 : let compression = self.conf.image_compression;
768 1638450 : let uncompressed_len = img.len() as u64;
769 1638450 : self.uncompressed_bytes += uncompressed_len;
770 1638450 : self.num_keys += 1;
771 1638450 : let (_img, res) = self
772 1638450 : .blob_writer
773 1638450 : .write_blob_maybe_compressed(img.slice_len(), ctx, compression)
774 1664353 : .await;
775 : // TODO: re-use the buffer for `img` further upstack
776 1638450 : let (off, compression_info) = res?;
777 1638450 : if compression_info.compressed_size.is_some() {
778 24006 : // The image has been considered for compression at least
779 24006 : self.uncompressed_bytes_eligible += uncompressed_len;
780 1614444 : }
781 1638450 : if compression_info.written_compressed {
782 0 : // The image has been compressed
783 0 : self.uncompressed_bytes_chosen += uncompressed_len;
784 1638450 : }
785 :
786 1638450 : let mut keybuf: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
787 1638450 : key.write_to_byte_slice(&mut keybuf);
788 1638450 : self.tree.append(&keybuf, off)?;
789 :
790 : #[cfg(feature = "testing")]
791 1638450 : {
792 1638450 : self.last_written_key = key;
793 1638450 : }
794 1638450 :
795 1638450 : Ok(())
796 1638450 : }
797 :
798 : ///
799 : /// Finish writing the image layer.
800 : ///
801 900 : async fn finish(
802 900 : self,
803 900 : timeline: &Arc<Timeline>,
804 900 : ctx: &RequestContext,
805 900 : end_key: Option<Key>,
806 900 : ) -> anyhow::Result<ResidentLayer> {
807 900 : let index_start_blk =
808 900 : ((self.blob_writer.size() + PAGE_SZ as u64 - 1) / PAGE_SZ as u64) as u32;
809 900 :
810 900 : // Calculate compression ratio
811 900 : let compressed_size = self.blob_writer.size() - PAGE_SZ as u64; // Subtract PAGE_SZ for header
812 900 : crate::metrics::COMPRESSION_IMAGE_INPUT_BYTES.inc_by(self.uncompressed_bytes);
813 900 : crate::metrics::COMPRESSION_IMAGE_INPUT_BYTES_CONSIDERED
814 900 : .inc_by(self.uncompressed_bytes_eligible);
815 900 : crate::metrics::COMPRESSION_IMAGE_INPUT_BYTES_CHOSEN.inc_by(self.uncompressed_bytes_chosen);
816 900 : crate::metrics::COMPRESSION_IMAGE_OUTPUT_BYTES.inc_by(compressed_size);
817 900 :
818 900 : let mut file = self.blob_writer.into_inner();
819 900 :
820 900 : // Write out the index
821 900 : file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
822 0 : .await?;
823 900 : let (index_root_blk, block_buf) = self.tree.finish()?;
824 3168 : for buf in block_buf.blocks {
825 2268 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
826 2268 : res?;
827 : }
828 :
829 900 : let final_key_range = if let Some(end_key) = end_key {
830 102 : self.key_range.start..end_key
831 : } else {
832 798 : self.key_range.clone()
833 : };
834 :
835 : // Fill in the summary on blk 0
836 900 : let summary = Summary {
837 900 : magic: IMAGE_FILE_MAGIC,
838 900 : format_version: STORAGE_FORMAT_VERSION,
839 900 : tenant_id: self.tenant_shard_id.tenant_id,
840 900 : timeline_id: self.timeline_id,
841 900 : key_range: final_key_range.clone(),
842 900 : lsn: self.lsn,
843 900 : index_start_blk,
844 900 : index_root_blk,
845 900 : };
846 900 :
847 900 : let mut buf = Vec::with_capacity(PAGE_SZ);
848 900 : // TODO: could use smallvec here but it's a pain with Slice<T>
849 900 : Summary::ser_into(&summary, &mut buf)?;
850 900 : file.seek(SeekFrom::Start(0)).await?;
851 900 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
852 900 : res?;
853 :
854 900 : let metadata = file
855 900 : .metadata()
856 455 : .await
857 900 : .context("get metadata to determine file size")?;
858 :
859 900 : let desc = PersistentLayerDesc::new_img(
860 900 : self.tenant_shard_id,
861 900 : self.timeline_id,
862 900 : final_key_range,
863 900 : self.lsn,
864 900 : metadata.len(),
865 900 : );
866 :
867 : #[cfg(feature = "testing")]
868 900 : if let Some(end_key) = end_key {
869 102 : assert!(
870 102 : self.last_written_key < end_key,
871 0 : "written key violates end_key range"
872 : );
873 798 : }
874 :
875 : // Note: Because we open the file in write-only mode, we cannot
876 : // reuse the same VirtualFile for reading later. That's why we don't
877 : // set inner.file here. The first read will have to re-open it.
878 :
879 : // fsync the file
880 900 : file.sync_all().await?;
881 :
882 : // FIXME: why not carry the virtualfile here, it supports renaming?
883 900 : let layer = Layer::finish_creating(self.conf, timeline, desc, &self.path)?;
884 :
885 900 : info!("created image layer {}", layer.local_path());
886 :
887 900 : Ok(layer)
888 900 : }
889 : }
890 :
891 : /// A builder object for constructing a new image layer.
892 : ///
893 : /// Usage:
894 : ///
895 : /// 1. Create the ImageLayerWriter by calling ImageLayerWriter::new(...)
896 : ///
897 : /// 2. Write the contents by calling `put_page_image` for every key-value
898 : /// pair in the key range.
899 : ///
900 : /// 3. Call `finish`.
901 : ///
902 : /// # Note
903 : ///
904 : /// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
905 : /// possible for the writer to drop before `finish` is actually called. So this
906 : /// could lead to odd temporary files in the directory, exhausting file system.
907 : /// This structure wraps `ImageLayerWriterInner` and also contains `Drop`
908 : /// implementation that cleans up the temporary file in failure. It's not
909 : /// possible to do this directly in `ImageLayerWriterInner` since `finish` moves
910 : /// out some fields, making it impossible to implement `Drop`.
911 : ///
912 : #[must_use]
913 : pub struct ImageLayerWriter {
914 : inner: Option<ImageLayerWriterInner>,
915 : }
916 :
917 : impl ImageLayerWriter {
918 : ///
919 : /// Start building a new image layer.
920 : ///
921 1470 : pub async fn new(
922 1470 : conf: &'static PageServerConf,
923 1470 : timeline_id: TimelineId,
924 1470 : tenant_shard_id: TenantShardId,
925 1470 : key_range: &Range<Key>,
926 1470 : lsn: Lsn,
927 1470 : ctx: &RequestContext,
928 1470 : ) -> anyhow::Result<ImageLayerWriter> {
929 1470 : Ok(Self {
930 1470 : inner: Some(
931 1470 : ImageLayerWriterInner::new(conf, timeline_id, tenant_shard_id, key_range, lsn, ctx)
932 982 : .await?,
933 : ),
934 : })
935 1470 : }
936 :
937 : ///
938 : /// Write next value to the file.
939 : ///
940 : /// The page versions must be appended in blknum order.
941 : ///
942 1638450 : pub async fn put_image(
943 1638450 : &mut self,
944 1638450 : key: Key,
945 1638450 : img: Bytes,
946 1638450 : ctx: &RequestContext,
947 1638450 : ) -> anyhow::Result<()> {
948 1664353 : self.inner.as_mut().unwrap().put_image(key, img, ctx).await
949 1638450 : }
950 :
951 : /// Estimated size of the image layer.
952 25146 : pub(crate) fn estimated_size(&self) -> u64 {
953 25146 : let inner = self.inner.as_ref().unwrap();
954 25146 : inner.blob_writer.size() + inner.tree.borrow_writer().size() + PAGE_SZ as u64
955 25146 : }
956 :
957 25326 : pub(crate) fn num_keys(&self) -> usize {
958 25326 : self.inner.as_ref().unwrap().num_keys
959 25326 : }
960 :
961 : ///
962 : /// Finish writing the image layer.
963 : ///
964 798 : pub(crate) async fn finish(
965 798 : mut self,
966 798 : timeline: &Arc<Timeline>,
967 798 : ctx: &RequestContext,
968 798 : ) -> anyhow::Result<super::ResidentLayer> {
969 2311 : self.inner.take().unwrap().finish(timeline, ctx, None).await
970 798 : }
971 :
972 : /// Finish writing the image layer with an end key, used in [`super::split_writer::SplitImageLayerWriter`]. The end key determines the end of the image layer's covered range and is exclusive.
973 102 : pub(super) async fn finish_with_end_key(
974 102 : mut self,
975 102 : timeline: &Arc<Timeline>,
976 102 : end_key: Key,
977 102 : ctx: &RequestContext,
978 102 : ) -> anyhow::Result<super::ResidentLayer> {
979 102 : self.inner
980 102 : .take()
981 102 : .unwrap()
982 102 : .finish(timeline, ctx, Some(end_key))
983 210 : .await
984 102 : }
985 : }
986 :
987 : impl Drop for ImageLayerWriter {
988 1470 : fn drop(&mut self) {
989 1470 : if let Some(inner) = self.inner.take() {
990 570 : inner.blob_writer.into_inner().remove();
991 900 : }
992 1470 : }
993 : }
994 :
995 : pub struct ImageLayerIterator<'a> {
996 : image_layer: &'a ImageLayerInner,
997 : ctx: &'a RequestContext,
998 : planner: StreamingVectoredReadPlanner,
999 : index_iter: DiskBtreeIterator<'a>,
1000 : key_values_batch: VecDeque<(Key, Lsn, Value)>,
1001 : is_end: bool,
1002 : }
1003 :
1004 : impl<'a> ImageLayerIterator<'a> {
1005 0 : pub(crate) fn layer_dbg_info(&self) -> String {
1006 0 : self.image_layer.layer_dbg_info()
1007 0 : }
1008 :
1009 : /// Retrieve a batch of key-value pairs into the iterator buffer.
1010 57052 : async fn next_batch(&mut self) -> anyhow::Result<()> {
1011 57052 : assert!(self.key_values_batch.is_empty());
1012 57052 : assert!(!self.is_end);
1013 :
1014 57052 : let plan = loop {
1015 86310 : if let Some(res) = self.index_iter.next().await {
1016 86118 : let (raw_key, offset) = res?;
1017 86118 : if let Some(batch_plan) = self.planner.handle(
1018 86118 : Key::from_slice(&raw_key[..KEY_SIZE]),
1019 86118 : self.image_layer.lsn,
1020 86118 : offset,
1021 86118 : ) {
1022 56860 : break batch_plan;
1023 29258 : }
1024 : } else {
1025 192 : self.is_end = true;
1026 192 : let payload_end = self.image_layer.index_start_blk as u64 * PAGE_SZ as u64;
1027 192 : if let Some(item) = self.planner.handle_range_end(payload_end) {
1028 192 : break item;
1029 : } else {
1030 0 : return Ok(()); // TODO: a test case on empty iterator
1031 : }
1032 : }
1033 : };
1034 57052 : let vectored_blob_reader = VectoredBlobReader::new(&self.image_layer.file);
1035 57052 : let mut next_batch = std::collections::VecDeque::new();
1036 57052 : let buf_size = plan.size();
1037 57052 : let buf = BytesMut::with_capacity(buf_size);
1038 57052 : let blobs_buf = vectored_blob_reader
1039 57052 : .read_blobs(&plan, buf, self.ctx)
1040 28970 : .await?;
1041 57052 : let frozen_buf: Bytes = blobs_buf.buf.freeze();
1042 86034 : for meta in blobs_buf.blobs.iter() {
1043 86034 : let img_buf = frozen_buf.slice(meta.start..meta.end);
1044 86034 : next_batch.push_back((meta.meta.key, self.image_layer.lsn, Value::Image(img_buf)));
1045 86034 : }
1046 57052 : self.key_values_batch = next_batch;
1047 57052 : Ok(())
1048 57052 : }
1049 :
1050 85524 : pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
1051 85524 : if self.key_values_batch.is_empty() {
1052 57100 : if self.is_end {
1053 300 : return Ok(None);
1054 56800 : }
1055 56800 : self.next_batch().await?;
1056 28424 : }
1057 85224 : Ok(Some(
1058 85224 : self.key_values_batch
1059 85224 : .pop_front()
1060 85224 : .expect("should not be empty"),
1061 85224 : ))
1062 85524 : }
1063 : }
1064 :
1065 : #[cfg(test)]
1066 : mod test {
1067 : use std::{sync::Arc, time::Duration};
1068 :
1069 : use bytes::Bytes;
1070 : use itertools::Itertools;
1071 : use pageserver_api::{
1072 : key::Key,
1073 : shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize},
1074 : };
1075 : use utils::{
1076 : generation::Generation,
1077 : id::{TenantId, TimelineId},
1078 : lsn::Lsn,
1079 : };
1080 :
1081 : use crate::{
1082 : context::RequestContext,
1083 : repository::Value,
1084 : tenant::{
1085 : config::TenantConf,
1086 : harness::{TenantHarness, TIMELINE_ID},
1087 : storage_layer::ResidentLayer,
1088 : vectored_blob_io::StreamingVectoredReadPlanner,
1089 : Tenant, Timeline,
1090 : },
1091 : DEFAULT_PG_VERSION,
1092 : };
1093 :
1094 : use super::{ImageLayerIterator, ImageLayerWriter};
1095 :
1096 : #[tokio::test]
1097 6 : async fn image_layer_rewrite() {
1098 6 : let tenant_conf = TenantConf {
1099 6 : gc_period: Duration::ZERO,
1100 6 : compaction_period: Duration::ZERO,
1101 6 : ..TenantConf::default()
1102 6 : };
1103 6 : let tenant_id = TenantId::generate();
1104 6 : let mut gen = Generation::new(0xdead0001);
1105 30 : let mut get_next_gen = || {
1106 30 : let ret = gen;
1107 30 : gen = gen.next();
1108 30 : ret
1109 30 : };
1110 6 : // The LSN at which we will create an image layer to filter
1111 6 : let lsn = Lsn(0xdeadbeef0000);
1112 6 : let timeline_id = TimelineId::generate();
1113 6 :
1114 6 : //
1115 6 : // Create an unsharded parent with a layer.
1116 6 : //
1117 6 :
1118 6 : let harness = TenantHarness::create_custom(
1119 6 : "test_image_layer_rewrite--parent",
1120 6 : tenant_conf.clone(),
1121 6 : tenant_id,
1122 6 : ShardIdentity::unsharded(),
1123 6 : get_next_gen(),
1124 6 : )
1125 6 : .await
1126 6 : .unwrap();
1127 24 : let (tenant, ctx) = harness.load().await;
1128 6 : let timeline = tenant
1129 6 : .create_test_timeline(timeline_id, lsn, DEFAULT_PG_VERSION, &ctx)
1130 12 : .await
1131 6 : .unwrap();
1132 6 :
1133 6 : // This key range contains several 0x8000 page stripes, only one of which belongs to shard zero
1134 6 : let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
1135 6 : let input_end = Key::from_hex("000000067f00000001000000ae0000020000").unwrap();
1136 6 : let range = input_start..input_end;
1137 6 :
1138 6 : // Build an image layer to filter
1139 6 : let resident = {
1140 6 : let mut writer = ImageLayerWriter::new(
1141 6 : harness.conf,
1142 6 : timeline_id,
1143 6 : harness.tenant_shard_id,
1144 6 : &range,
1145 6 : lsn,
1146 6 : &ctx,
1147 6 : )
1148 6 : .await
1149 6 : .unwrap();
1150 6 :
1151 6 : let foo_img = Bytes::from_static(&[1, 2, 3, 4]);
1152 6 : let mut key = range.start;
1153 786438 : while key < range.end {
1154 798717 : writer.put_image(key, foo_img.clone(), &ctx).await.unwrap();
1155 786432 :
1156 786432 : key = key.next();
1157 6 : }
1158 357 : writer.finish(&timeline, &ctx).await.unwrap()
1159 6 : };
1160 6 : let original_size = resident.metadata().file_size;
1161 6 :
1162 6 : //
1163 6 : // Create child shards and do the rewrite, exercising filter().
1164 6 : // TODO: abstraction in TenantHarness for splits.
1165 6 : //
1166 6 :
1167 6 : // Filter for various shards: this exercises cases like values at start of key range, end of key
1168 6 : // range, middle of key range.
1169 6 : let shard_count = ShardCount::new(4);
1170 24 : for shard_number in 0..shard_count.count() {
1171 6 : //
1172 6 : // mimic the shard split
1173 6 : //
1174 24 : let shard_identity = ShardIdentity::new(
1175 24 : ShardNumber(shard_number),
1176 24 : shard_count,
1177 24 : ShardStripeSize(0x8000),
1178 24 : )
1179 24 : .unwrap();
1180 24 : let harness = TenantHarness::create_custom(
1181 24 : Box::leak(Box::new(format!(
1182 24 : "test_image_layer_rewrite--child{}",
1183 24 : shard_identity.shard_slug()
1184 24 : ))),
1185 24 : tenant_conf.clone(),
1186 24 : tenant_id,
1187 24 : shard_identity,
1188 24 : // NB: in reality, the shards would each fork off their own gen number sequence from the parent.
1189 24 : // But here, all we care about is that the gen number is unique.
1190 24 : get_next_gen(),
1191 24 : )
1192 6 : .await
1193 24 : .unwrap();
1194 96 : let (tenant, ctx) = harness.load().await;
1195 24 : let timeline = tenant
1196 24 : .create_test_timeline(timeline_id, lsn, DEFAULT_PG_VERSION, &ctx)
1197 36 : .await
1198 24 : .unwrap();
1199 6 :
1200 6 : //
1201 6 : // use filter() and make assertions
1202 6 : //
1203 6 :
1204 24 : let mut filtered_writer = ImageLayerWriter::new(
1205 24 : harness.conf,
1206 24 : timeline_id,
1207 24 : harness.tenant_shard_id,
1208 24 : &range,
1209 24 : lsn,
1210 24 : &ctx,
1211 24 : )
1212 12 : .await
1213 24 : .unwrap();
1214 6 :
1215 24 : let wrote_keys = resident
1216 24 : .filter(&shard_identity, &mut filtered_writer, &ctx)
1217 800157 : .await
1218 24 : .unwrap();
1219 24 : let replacement = if wrote_keys > 0 {
1220 387 : Some(filtered_writer.finish(&timeline, &ctx).await.unwrap())
1221 6 : } else {
1222 6 : None
1223 6 : };
1224 6 :
1225 6 : // This exact size and those below will need updating as/when the layer encoding changes, but
1226 6 : // should be deterministic for a given version of the format, as we used no randomness generating the input.
1227 24 : assert_eq!(original_size, 1597440);
1228 6 :
1229 24 : match shard_number {
1230 6 : 0 => {
1231 6 : // We should have written out just one stripe for our shard identity
1232 6 : assert_eq!(wrote_keys, 0x8000);
1233 6 : let replacement = replacement.unwrap();
1234 6 :
1235 6 : // We should have dropped some of the data
1236 6 : assert!(replacement.metadata().file_size < original_size);
1237 6 : assert!(replacement.metadata().file_size > 0);
1238 6 :
1239 6 : // Assert that we dropped ~3/4 of the data.
1240 6 : assert_eq!(replacement.metadata().file_size, 417792);
1241 6 : }
1242 6 : 1 => {
1243 6 : // Shard 1 has no keys in our input range
1244 6 : assert_eq!(wrote_keys, 0x0);
1245 6 : assert!(replacement.is_none());
1246 6 : }
1247 6 : 2 => {
1248 6 : // Shard 2 has one stripes in the input range
1249 6 : assert_eq!(wrote_keys, 0x8000);
1250 6 : let replacement = replacement.unwrap();
1251 6 : assert!(replacement.metadata().file_size < original_size);
1252 6 : assert!(replacement.metadata().file_size > 0);
1253 6 : assert_eq!(replacement.metadata().file_size, 417792);
1254 6 : }
1255 6 : 3 => {
1256 6 : // Shard 3 has two stripes in the input range
1257 6 : assert_eq!(wrote_keys, 0x10000);
1258 6 : let replacement = replacement.unwrap();
1259 6 : assert!(replacement.metadata().file_size < original_size);
1260 6 : assert!(replacement.metadata().file_size > 0);
1261 6 : assert_eq!(replacement.metadata().file_size, 811008);
1262 6 : }
1263 6 : _ => unreachable!(),
1264 6 : }
1265 6 : }
1266 6 : }
1267 :
1268 6 : async fn produce_image_layer(
1269 6 : tenant: &Tenant,
1270 6 : tline: &Arc<Timeline>,
1271 6 : mut images: Vec<(Key, Bytes)>,
1272 6 : lsn: Lsn,
1273 6 : ctx: &RequestContext,
1274 6 : ) -> anyhow::Result<ResidentLayer> {
1275 6 : images.sort();
1276 6 : let (key_start, _) = images.first().unwrap();
1277 6 : let (key_last, _) = images.last().unwrap();
1278 6 : let key_end = key_last.next();
1279 6 : let key_range = *key_start..key_end;
1280 6 : let mut writer = ImageLayerWriter::new(
1281 6 : tenant.conf,
1282 6 : tline.timeline_id,
1283 6 : tenant.tenant_shard_id,
1284 6 : &key_range,
1285 6 : lsn,
1286 6 : ctx,
1287 6 : )
1288 3 : .await?;
1289 :
1290 6006 : for (key, img) in images {
1291 6093 : writer.put_image(key, img, ctx).await?;
1292 : }
1293 12 : let img_layer = writer.finish(tline, ctx).await?;
1294 :
1295 6 : Ok::<_, anyhow::Error>(img_layer)
1296 6 : }
1297 :
1298 84 : async fn assert_img_iter_equal(
1299 84 : img_iter: &mut ImageLayerIterator<'_>,
1300 84 : expect: &[(Key, Bytes)],
1301 84 : expect_lsn: Lsn,
1302 84 : ) {
1303 84 : let mut expect_iter = expect.iter();
1304 : loop {
1305 84084 : let o1 = img_iter.next().await.unwrap();
1306 84084 : let o2 = expect_iter.next();
1307 84084 : match (o1, o2) {
1308 84 : (None, None) => break,
1309 84000 : (Some((k1, l1, v1)), Some((k2, i2))) => {
1310 84000 : let Value::Image(i1) = v1 else {
1311 0 : panic!("expect Value::Image")
1312 : };
1313 84000 : assert_eq!(&k1, k2);
1314 84000 : assert_eq!(l1, expect_lsn);
1315 84000 : assert_eq!(&i1, i2);
1316 : }
1317 0 : (o1, o2) => panic!("iterators length mismatch: {:?}, {:?}", o1, o2),
1318 : }
1319 : }
1320 84 : }
1321 :
1322 : #[tokio::test]
1323 6 : async fn image_layer_iterator() {
1324 6 : let harness = TenantHarness::create("image_layer_iterator").await.unwrap();
1325 24 : let (tenant, ctx) = harness.load().await;
1326 6 :
1327 6 : let tline = tenant
1328 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
1329 12 : .await
1330 6 : .unwrap();
1331 6 :
1332 6000 : fn get_key(id: u32) -> Key {
1333 6000 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
1334 6000 : key.field6 = id;
1335 6000 : key
1336 6000 : }
1337 6 : const N: usize = 1000;
1338 6 : let test_imgs = (0..N)
1339 6000 : .map(|idx| (get_key(idx as u32), Bytes::from(format!("img{idx:05}"))))
1340 6 : .collect_vec();
1341 6 : let resident_layer =
1342 6 : produce_image_layer(&tenant, &tline, test_imgs.clone(), Lsn(0x10), &ctx)
1343 6108 : .await
1344 6 : .unwrap();
1345 6 : let img_layer = resident_layer.get_as_image(&ctx).await.unwrap();
1346 18 : for max_read_size in [1, 1024] {
1347 96 : for batch_size in [1, 2, 4, 8, 3, 7, 13] {
1348 84 : println!("running with batch_size={batch_size} max_read_size={max_read_size}");
1349 84 : // Test if the batch size is correctly determined
1350 84 : let mut iter = img_layer.iter(&ctx);
1351 84 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
1352 84 : let mut num_items = 0;
1353 336 : for _ in 0..3 {
1354 252 : iter.next_batch().await.unwrap();
1355 252 : num_items += iter.key_values_batch.len();
1356 252 : if max_read_size == 1 {
1357 6 : // every key should be a batch b/c the value is larger than max_read_size
1358 126 : assert_eq!(iter.key_values_batch.len(), 1);
1359 6 : } else {
1360 126 : assert!(iter.key_values_batch.len() <= batch_size);
1361 6 : }
1362 252 : if num_items >= N {
1363 6 : break;
1364 252 : }
1365 252 : iter.key_values_batch.clear();
1366 6 : }
1367 6 : // Test if the result is correct
1368 84 : let mut iter = img_layer.iter(&ctx);
1369 84 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
1370 28787 : assert_img_iter_equal(&mut iter, &test_imgs, Lsn(0x10)).await;
1371 6 : }
1372 6 : }
1373 6 : }
1374 : }
|
