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 std::collections::{HashMap, VecDeque};
29 : use std::fs::File;
30 : use std::io::SeekFrom;
31 : use std::ops::Range;
32 : use std::os::unix::prelude::FileExt;
33 : use std::str::FromStr;
34 : use std::sync::Arc;
35 :
36 : use anyhow::{Context, Result, bail, ensure};
37 : use bytes::Bytes;
38 : use camino::{Utf8Path, Utf8PathBuf};
39 : use hex;
40 : use itertools::Itertools;
41 : use pageserver_api::config::MaxVectoredReadBytes;
42 : use pageserver_api::key::{DBDIR_KEY, KEY_SIZE, Key};
43 : use pageserver_api::keyspace::KeySpace;
44 : use pageserver_api::shard::{ShardIdentity, TenantShardId};
45 : use pageserver_api::value::Value;
46 : use rand::Rng;
47 : use rand::distributions::Alphanumeric;
48 : use serde::{Deserialize, Serialize};
49 : use tokio::sync::OnceCell;
50 : use tokio_stream::StreamExt;
51 : use tracing::*;
52 : use utils::bin_ser::BeSer;
53 : use utils::id::{TenantId, TimelineId};
54 : use utils::lsn::Lsn;
55 :
56 : use super::layer_name::ImageLayerName;
57 : use super::{
58 : AsLayerDesc, LayerName, OnDiskValue, OnDiskValueIo, PersistentLayerDesc, ResidentLayer,
59 : ValuesReconstructState,
60 : };
61 : use crate::config::PageServerConf;
62 : use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
63 : use crate::page_cache::{self, FileId, PAGE_SZ};
64 : use crate::tenant::blob_io::BlobWriter;
65 : use crate::tenant::block_io::{BlockBuf, FileBlockReader};
66 : use crate::tenant::disk_btree::{
67 : DiskBtreeBuilder, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
68 : };
69 : use crate::tenant::timeline::GetVectoredError;
70 : use crate::tenant::vectored_blob_io::{
71 : BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
72 : VectoredReadPlanner,
73 : };
74 : use crate::virtual_file::owned_buffers_io::io_buf_ext::IoBufExt;
75 : use crate::virtual_file::{self, IoBufferMut, MaybeFatalIo, VirtualFile};
76 : use crate::{IMAGE_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
77 :
78 : ///
79 : /// Header stored in the beginning of the file
80 : ///
81 : /// After this comes the 'values' part, starting on block 1. After that,
82 : /// the 'index' starts at the block indicated by 'index_start_blk'
83 : ///
84 0 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
85 : pub struct Summary {
86 : /// Magic value to identify this as a neon image file. Always IMAGE_FILE_MAGIC.
87 : pub magic: u16,
88 : pub format_version: u16,
89 :
90 : pub tenant_id: TenantId,
91 : pub timeline_id: TimelineId,
92 : pub key_range: Range<Key>,
93 : pub lsn: Lsn,
94 :
95 : /// Block number where the 'index' part of the file begins.
96 : pub index_start_blk: u32,
97 : /// Block within the 'index', where the B-tree root page is stored
98 : pub index_root_blk: u32,
99 : // the 'values' part starts after the summary header, on block 1.
100 : }
101 :
102 : impl From<&ImageLayer> for Summary {
103 0 : fn from(layer: &ImageLayer) -> Self {
104 0 : Self::expected(
105 0 : layer.desc.tenant_shard_id.tenant_id,
106 0 : layer.desc.timeline_id,
107 0 : layer.desc.key_range.clone(),
108 0 : layer.lsn,
109 0 : )
110 0 : }
111 : }
112 :
113 : impl Summary {
114 276 : pub(super) fn expected(
115 276 : tenant_id: TenantId,
116 276 : timeline_id: TimelineId,
117 276 : key_range: Range<Key>,
118 276 : lsn: Lsn,
119 276 : ) -> Self {
120 276 : Self {
121 276 : magic: IMAGE_FILE_MAGIC,
122 276 : format_version: STORAGE_FORMAT_VERSION,
123 276 : tenant_id,
124 276 : timeline_id,
125 276 : key_range,
126 276 : lsn,
127 276 :
128 276 : index_start_blk: 0,
129 276 : index_root_blk: 0,
130 276 : }
131 276 : }
132 : }
133 :
134 : /// This is used only from `pagectl`. Within pageserver, all layers are
135 : /// [`crate::tenant::storage_layer::Layer`], which can hold an [`ImageLayerInner`].
136 : pub struct ImageLayer {
137 : path: Utf8PathBuf,
138 : pub desc: PersistentLayerDesc,
139 : // This entry contains an image of all pages as of this LSN, should be the same as desc.lsn
140 : pub lsn: Lsn,
141 : inner: OnceCell<ImageLayerInner>,
142 : }
143 :
144 : impl std::fmt::Debug for ImageLayer {
145 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
146 : use super::RangeDisplayDebug;
147 :
148 0 : f.debug_struct("ImageLayer")
149 0 : .field("key_range", &RangeDisplayDebug(&self.desc.key_range))
150 0 : .field("file_size", &self.desc.file_size)
151 0 : .field("lsn", &self.lsn)
152 0 : .field("inner", &self.inner)
153 0 : .finish()
154 0 : }
155 : }
156 :
157 : /// ImageLayer is the in-memory data structure associated with an on-disk image
158 : /// file.
159 : pub struct ImageLayerInner {
160 : // values copied from summary
161 : index_start_blk: u32,
162 : index_root_blk: u32,
163 :
164 : key_range: Range<Key>,
165 : lsn: Lsn,
166 :
167 : file: Arc<VirtualFile>,
168 : file_id: FileId,
169 :
170 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
171 : }
172 :
173 : impl ImageLayerInner {
174 0 : pub(crate) fn layer_dbg_info(&self) -> String {
175 0 : format!(
176 0 : "image {}..{} {}",
177 0 : self.key_range().start,
178 0 : self.key_range().end,
179 0 : self.lsn()
180 0 : )
181 0 : }
182 : }
183 :
184 : impl std::fmt::Debug for ImageLayerInner {
185 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
186 0 : f.debug_struct("ImageLayerInner")
187 0 : .field("index_start_blk", &self.index_start_blk)
188 0 : .field("index_root_blk", &self.index_root_blk)
189 0 : .finish()
190 0 : }
191 : }
192 :
193 : impl ImageLayerInner {
194 0 : pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
195 0 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
196 0 : let tree_reader = DiskBtreeReader::<_, KEY_SIZE>::new(
197 0 : self.index_start_blk,
198 0 : self.index_root_blk,
199 0 : block_reader,
200 0 : );
201 0 :
202 0 : tree_reader.dump(ctx).await?;
203 :
204 0 : tree_reader
205 0 : .visit(
206 0 : &[0u8; KEY_SIZE],
207 0 : VisitDirection::Forwards,
208 0 : |key, value| {
209 0 : println!("key: {} offset {}", hex::encode(key), value);
210 0 : true
211 0 : },
212 0 : ctx,
213 0 : )
214 0 : .await?;
215 :
216 0 : Ok(())
217 0 : }
218 : }
219 :
220 : /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
221 : impl std::fmt::Display for ImageLayer {
222 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
223 0 : write!(f, "{}", self.layer_desc().short_id())
224 0 : }
225 : }
226 :
227 : impl AsLayerDesc for ImageLayer {
228 0 : fn layer_desc(&self) -> &PersistentLayerDesc {
229 0 : &self.desc
230 0 : }
231 : }
232 :
233 : impl ImageLayer {
234 0 : pub async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
235 0 : self.desc.dump();
236 0 :
237 0 : if !verbose {
238 0 : return Ok(());
239 0 : }
240 :
241 0 : let inner = self.load(ctx).await?;
242 :
243 0 : inner.dump(ctx).await?;
244 :
245 0 : Ok(())
246 0 : }
247 :
248 1204 : fn temp_path_for(
249 1204 : conf: &PageServerConf,
250 1204 : timeline_id: TimelineId,
251 1204 : tenant_shard_id: TenantShardId,
252 1204 : fname: &ImageLayerName,
253 1204 : ) -> Utf8PathBuf {
254 1204 : let rand_string: String = rand::thread_rng()
255 1204 : .sample_iter(&Alphanumeric)
256 1204 : .take(8)
257 1204 : .map(char::from)
258 1204 : .collect();
259 1204 :
260 1204 : conf.timeline_path(&tenant_shard_id, &timeline_id)
261 1204 : .join(format!("{fname}.{rand_string}.{TEMP_FILE_SUFFIX}"))
262 1204 : }
263 :
264 : ///
265 : /// Open the underlying file and read the metadata into memory, if it's
266 : /// not loaded already.
267 : ///
268 0 : async fn load(&self, ctx: &RequestContext) -> Result<&ImageLayerInner> {
269 0 : self.inner
270 0 : .get_or_try_init(|| self.load_inner(ctx))
271 0 : .await
272 0 : .with_context(|| format!("Failed to load image layer {}", self.path()))
273 0 : }
274 :
275 0 : async fn load_inner(&self, ctx: &RequestContext) -> Result<ImageLayerInner> {
276 0 : let path = self.path();
277 :
278 0 : let loaded =
279 0 : ImageLayerInner::load(&path, self.desc.image_layer_lsn(), None, None, ctx).await?;
280 :
281 : // not production code
282 0 : let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
283 0 : let expected_layer_name = self.layer_desc().layer_name();
284 0 :
285 0 : if actual_layer_name != expected_layer_name {
286 0 : println!("warning: filename does not match what is expected from in-file summary");
287 0 : println!("actual: {:?}", actual_layer_name.to_string());
288 0 : println!("expected: {:?}", expected_layer_name.to_string());
289 0 : }
290 :
291 0 : Ok(loaded)
292 0 : }
293 :
294 : /// Create an ImageLayer struct representing an existing file on disk.
295 : ///
296 : /// This variant is only used for debugging purposes, by the 'pagectl' binary.
297 0 : pub fn new_for_path(path: &Utf8Path, file: File) -> Result<ImageLayer> {
298 0 : let mut summary_buf = vec![0; PAGE_SZ];
299 0 : file.read_exact_at(&mut summary_buf, 0)?;
300 0 : let summary = Summary::des_prefix(&summary_buf)?;
301 0 : let metadata = file
302 0 : .metadata()
303 0 : .context("get file metadata to determine size")?;
304 :
305 : // This function is never used for constructing layers in a running pageserver,
306 : // so it does not need an accurate TenantShardId.
307 0 : let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
308 0 :
309 0 : Ok(ImageLayer {
310 0 : path: path.to_path_buf(),
311 0 : desc: PersistentLayerDesc::new_img(
312 0 : tenant_shard_id,
313 0 : summary.timeline_id,
314 0 : summary.key_range,
315 0 : summary.lsn,
316 0 : metadata.len(),
317 0 : ), // Now we assume image layer ALWAYS covers the full range. This may change in the future.
318 0 : lsn: summary.lsn,
319 0 : inner: OnceCell::new(),
320 0 : })
321 0 : }
322 :
323 0 : fn path(&self) -> Utf8PathBuf {
324 0 : self.path.clone()
325 0 : }
326 : }
327 :
328 : #[derive(thiserror::Error, Debug)]
329 : pub enum RewriteSummaryError {
330 : #[error("magic mismatch")]
331 : MagicMismatch,
332 : #[error(transparent)]
333 : Other(#[from] anyhow::Error),
334 : }
335 :
336 : impl From<std::io::Error> for RewriteSummaryError {
337 0 : fn from(e: std::io::Error) -> Self {
338 0 : Self::Other(anyhow::anyhow!(e))
339 0 : }
340 : }
341 :
342 : impl ImageLayer {
343 0 : pub async fn rewrite_summary<F>(
344 0 : path: &Utf8Path,
345 0 : rewrite: F,
346 0 : ctx: &RequestContext,
347 0 : ) -> Result<(), RewriteSummaryError>
348 0 : where
349 0 : F: Fn(Summary) -> Summary,
350 0 : {
351 0 : let mut file = VirtualFile::open_with_options(
352 0 : path,
353 0 : virtual_file::OpenOptions::new().read(true).write(true),
354 0 : ctx,
355 0 : )
356 0 : .await
357 0 : .with_context(|| format!("Failed to open file '{}'", path))?;
358 0 : let file_id = page_cache::next_file_id();
359 0 : let block_reader = FileBlockReader::new(&file, file_id);
360 0 : let summary_blk = block_reader.read_blk(0, ctx).await?;
361 0 : let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
362 0 : if actual_summary.magic != IMAGE_FILE_MAGIC {
363 0 : return Err(RewriteSummaryError::MagicMismatch);
364 0 : }
365 0 :
366 0 : let new_summary = rewrite(actual_summary);
367 0 :
368 0 : let mut buf = Vec::with_capacity(PAGE_SZ);
369 0 : // TODO: could use smallvec here but it's a pain with Slice<T>
370 0 : Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
371 0 : file.seek(SeekFrom::Start(0)).await?;
372 0 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
373 0 : res?;
374 0 : Ok(())
375 0 : }
376 : }
377 :
378 : impl ImageLayerInner {
379 264 : pub(crate) fn key_range(&self) -> &Range<Key> {
380 264 : &self.key_range
381 264 : }
382 :
383 264 : pub(crate) fn lsn(&self) -> Lsn {
384 264 : self.lsn
385 264 : }
386 :
387 276 : pub(super) async fn load(
388 276 : path: &Utf8Path,
389 276 : lsn: Lsn,
390 276 : summary: Option<Summary>,
391 276 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
392 276 : ctx: &RequestContext,
393 276 : ) -> anyhow::Result<Self> {
394 276 : let file = Arc::new(
395 276 : VirtualFile::open_v2(path, ctx)
396 276 : .await
397 276 : .context("open layer file")?,
398 : );
399 276 : let file_id = page_cache::next_file_id();
400 276 : let block_reader = FileBlockReader::new(&file, file_id);
401 276 : let summary_blk = block_reader
402 276 : .read_blk(0, ctx)
403 276 : .await
404 276 : .context("read first block")?;
405 :
406 : // length is the only way how this could fail, so it's not actually likely at all unless
407 : // read_blk returns wrong sized block.
408 : //
409 : // TODO: confirm and make this into assertion
410 276 : let actual_summary =
411 276 : Summary::des_prefix(summary_blk.as_ref()).context("deserialize first block")?;
412 :
413 276 : if let Some(mut expected_summary) = summary {
414 : // production code path
415 276 : expected_summary.index_start_blk = actual_summary.index_start_blk;
416 276 : expected_summary.index_root_blk = actual_summary.index_root_blk;
417 276 : // mask out the timeline_id, but still require the layers to be from the same tenant
418 276 : expected_summary.timeline_id = actual_summary.timeline_id;
419 276 :
420 276 : if actual_summary != expected_summary {
421 0 : bail!(
422 0 : "in-file summary does not match expected summary. actual = {:?} expected = {:?}",
423 0 : actual_summary,
424 0 : expected_summary
425 0 : );
426 276 : }
427 0 : }
428 :
429 276 : Ok(ImageLayerInner {
430 276 : index_start_blk: actual_summary.index_start_blk,
431 276 : index_root_blk: actual_summary.index_root_blk,
432 276 : lsn,
433 276 : file,
434 276 : file_id,
435 276 : max_vectored_read_bytes,
436 276 : key_range: actual_summary.key_range,
437 276 : })
438 276 : }
439 :
440 : // Look up the keys in the provided keyspace and update
441 : // the reconstruct state with whatever is found.
442 45124 : pub(super) async fn get_values_reconstruct_data(
443 45124 : &self,
444 45124 : this: ResidentLayer,
445 45124 : keyspace: KeySpace,
446 45124 : reconstruct_state: &mut ValuesReconstructState,
447 45124 : ctx: &RequestContext,
448 45124 : ) -> Result<(), GetVectoredError> {
449 45124 : let reads = self
450 45124 : .plan_reads(keyspace, None, ctx)
451 45124 : .await
452 45124 : .map_err(GetVectoredError::Other)?;
453 :
454 45124 : self.do_reads_and_update_state(this, reads, reconstruct_state, ctx)
455 45124 : .await;
456 :
457 45124 : reconstruct_state.on_image_layer_visited(&self.key_range);
458 45124 :
459 45124 : Ok(())
460 45124 : }
461 :
462 : /// Traverse the layer's index to build read operations on the overlap of the input keyspace
463 : /// and the keys in this layer.
464 : ///
465 : /// If shard_identity is provided, it will be used to filter keys down to those stored on
466 : /// this shard.
467 45140 : async fn plan_reads(
468 45140 : &self,
469 45140 : keyspace: KeySpace,
470 45140 : shard_identity: Option<&ShardIdentity>,
471 45140 : ctx: &RequestContext,
472 45140 : ) -> anyhow::Result<Vec<VectoredRead>> {
473 45140 : let mut planner = VectoredReadPlanner::new(
474 45140 : self.max_vectored_read_bytes
475 45140 : .expect("Layer is loaded with max vectored bytes config")
476 45140 : .0
477 45140 : .into(),
478 45140 : );
479 45140 :
480 45140 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
481 45140 : let tree_reader =
482 45140 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
483 45140 :
484 45140 : let ctx = RequestContextBuilder::extend(ctx)
485 45140 : .page_content_kind(PageContentKind::ImageLayerBtreeNode)
486 45140 : .build();
487 :
488 45176 : for range in keyspace.ranges.iter() {
489 45176 : let mut range_end_handled = false;
490 45176 : let mut search_key: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
491 45176 : range.start.write_to_byte_slice(&mut search_key);
492 45176 :
493 45176 : let index_stream = tree_reader.clone().into_stream(&search_key, &ctx);
494 45176 : let mut index_stream = std::pin::pin!(index_stream);
495 :
496 2251528 : while let Some(index_entry) = index_stream.next().await {
497 2250900 : let (raw_key, offset) = index_entry?;
498 :
499 2250900 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
500 2250900 : assert!(key >= range.start);
501 :
502 2250900 : let flag = if let Some(shard_identity) = shard_identity {
503 2097152 : if shard_identity.is_key_disposable(&key) {
504 1572864 : BlobFlag::Ignore
505 : } else {
506 524288 : BlobFlag::None
507 : }
508 : } else {
509 153748 : BlobFlag::None
510 : };
511 :
512 2250900 : if key >= range.end {
513 44548 : planner.handle_range_end(offset);
514 44548 : range_end_handled = true;
515 44548 : break;
516 2206352 : } else {
517 2206352 : planner.handle(key, self.lsn, offset, flag);
518 2206352 : }
519 : }
520 :
521 45176 : if !range_end_handled {
522 628 : let payload_end = self.index_start_blk as u64 * PAGE_SZ as u64;
523 628 : planner.handle_range_end(payload_end);
524 44548 : }
525 : }
526 :
527 45140 : Ok(planner.finish())
528 45140 : }
529 :
530 : /// Given a key range, select the parts of that range that should be retained by the ShardIdentity,
531 : /// then execute vectored GET operations, passing the results of all read keys into the writer.
532 16 : pub(super) async fn filter(
533 16 : &self,
534 16 : shard_identity: &ShardIdentity,
535 16 : writer: &mut ImageLayerWriter,
536 16 : ctx: &RequestContext,
537 16 : ) -> anyhow::Result<usize> {
538 : // Fragment the range into the regions owned by this ShardIdentity
539 16 : let plan = self
540 16 : .plan_reads(
541 16 : KeySpace {
542 16 : // If asked for the total key space, plan_reads will give us all the keys in the layer
543 16 : ranges: vec![Key::MIN..Key::MAX],
544 16 : },
545 16 : Some(shard_identity),
546 16 : ctx,
547 16 : )
548 16 : .await?;
549 :
550 16 : let vectored_blob_reader = VectoredBlobReader::new(&self.file);
551 16 : let mut key_count = 0;
552 32 : for read in plan.into_iter() {
553 32 : let buf_size = read.size();
554 32 :
555 32 : let buf = IoBufferMut::with_capacity(buf_size);
556 32 : let blobs_buf = vectored_blob_reader.read_blobs(&read, buf, ctx).await?;
557 :
558 32 : let view = BufView::new_slice(&blobs_buf.buf);
559 :
560 524288 : for meta in blobs_buf.blobs.iter() {
561 524288 : let img_buf = meta.read(&view).await?;
562 :
563 524288 : key_count += 1;
564 524288 : writer
565 524288 : .put_image(meta.meta.key, img_buf.into_bytes(), ctx)
566 524288 : .await
567 524288 : .context(format!("Storing key {}", meta.meta.key))?;
568 : }
569 : }
570 :
571 16 : Ok(key_count)
572 16 : }
573 :
574 45124 : async fn do_reads_and_update_state(
575 45124 : &self,
576 45124 : this: ResidentLayer,
577 45124 : reads: Vec<VectoredRead>,
578 45124 : reconstruct_state: &mut ValuesReconstructState,
579 45124 : ctx: &RequestContext,
580 45124 : ) {
581 45124 : let max_vectored_read_bytes = self
582 45124 : .max_vectored_read_bytes
583 45124 : .expect("Layer is loaded with max vectored bytes config")
584 45124 : .0
585 45124 : .into();
586 :
587 45124 : for read in reads.into_iter() {
588 45048 : let mut ios: HashMap<(Key, Lsn), OnDiskValueIo> = Default::default();
589 109200 : for (_, blob_meta) in read.blobs_at.as_slice() {
590 109200 : let io = reconstruct_state.update_key(&blob_meta.key, blob_meta.lsn, true);
591 109200 : ios.insert((blob_meta.key, blob_meta.lsn), io);
592 109200 : }
593 :
594 45048 : let buf_size = read.size();
595 45048 :
596 45048 : if buf_size > max_vectored_read_bytes {
597 : // If the read is oversized, it should only contain one key.
598 0 : let offenders = read
599 0 : .blobs_at
600 0 : .as_slice()
601 0 : .iter()
602 0 : .filter_map(|(_, blob_meta)| {
603 0 : if blob_meta.key.is_rel_dir_key()
604 0 : || blob_meta.key == DBDIR_KEY
605 0 : || blob_meta.key.is_aux_file_key()
606 : {
607 : // The size of values for these keys is unbounded and can
608 : // grow very large in pathological cases.
609 0 : None
610 : } else {
611 0 : Some(format!("{}@{}", blob_meta.key, blob_meta.lsn))
612 : }
613 0 : })
614 0 : .join(", ");
615 0 :
616 0 : if !offenders.is_empty() {
617 0 : tracing::warn!(
618 0 : "Oversized vectored read ({} > {}) for keys {}",
619 : buf_size,
620 : max_vectored_read_bytes,
621 : offenders
622 : );
623 0 : }
624 45048 : }
625 :
626 45048 : let read_extend_residency = this.clone();
627 45048 : let read_from = self.file.clone();
628 45048 : let read_ctx = ctx.attached_child();
629 45048 : reconstruct_state
630 45048 : .spawn_io(async move {
631 45048 : let buf = IoBufferMut::with_capacity(buf_size);
632 45048 : let vectored_blob_reader = VectoredBlobReader::new(&read_from);
633 45048 : let res = vectored_blob_reader.read_blobs(&read, buf, &read_ctx).await;
634 :
635 45048 : match res {
636 45048 : Ok(blobs_buf) => {
637 45048 : let view = BufView::new_slice(&blobs_buf.buf);
638 109200 : for meta in blobs_buf.blobs.iter() {
639 109200 : let io: OnDiskValueIo =
640 109200 : ios.remove(&(meta.meta.key, meta.meta.lsn)).unwrap();
641 109200 : let img_buf = meta.read(&view).await;
642 :
643 109200 : let img_buf = match img_buf {
644 109200 : Ok(img_buf) => img_buf,
645 0 : Err(e) => {
646 0 : io.complete(Err(e));
647 0 : continue;
648 : }
649 : };
650 :
651 109200 : io.complete(Ok(OnDiskValue::RawImage(img_buf.into_bytes())));
652 : }
653 :
654 45048 : assert!(ios.is_empty());
655 : }
656 0 : Err(err) => {
657 0 : for (_, io) in ios {
658 0 : io.complete(Err(std::io::Error::new(
659 0 : err.kind(),
660 0 : "vec read failed",
661 0 : )));
662 0 : }
663 : }
664 : }
665 :
666 : // keep layer resident until this IO is done; this spawned IO future generally outlives the
667 : // call to `self` / the `Arc<DownloadedLayer>` / the `ResidentLayer` that guarantees residency
668 45048 : drop(read_extend_residency);
669 45048 : })
670 45048 : .await;
671 : }
672 45124 : }
673 :
674 244 : pub(crate) fn iter<'a>(&'a self, ctx: &'a RequestContext) -> ImageLayerIterator<'a> {
675 244 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
676 244 : let tree_reader =
677 244 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
678 244 : ImageLayerIterator {
679 244 : image_layer: self,
680 244 : ctx,
681 244 : index_iter: tree_reader.iter(&[0; KEY_SIZE], ctx),
682 244 : key_values_batch: VecDeque::new(),
683 244 : is_end: false,
684 244 : planner: StreamingVectoredReadPlanner::new(
685 244 : 1024 * 8192, // The default value. Unit tests might use a different value. 1024 * 8K = 8MB buffer.
686 244 : 1024, // The default value. Unit tests might use a different value
687 244 : ),
688 244 : }
689 244 : }
690 :
691 : /// NB: not super efficient, but not terrible either. Should prob be an iterator.
692 : //
693 : // We're reusing the index traversal logical in plan_reads; would be nice to
694 : // factor that out.
695 0 : pub(crate) async fn load_keys(&self, ctx: &RequestContext) -> anyhow::Result<Vec<Key>> {
696 0 : let plan = self
697 0 : .plan_reads(KeySpace::single(self.key_range.clone()), None, ctx)
698 0 : .await?;
699 0 : Ok(plan
700 0 : .into_iter()
701 0 : .flat_map(|read| read.blobs_at)
702 0 : .map(|(_, blob_meta)| blob_meta.key)
703 0 : .collect())
704 0 : }
705 : }
706 :
707 : /// A builder object for constructing a new image layer.
708 : ///
709 : /// Usage:
710 : ///
711 : /// 1. Create the ImageLayerWriter by calling ImageLayerWriter::new(...)
712 : ///
713 : /// 2. Write the contents by calling `put_page_image` for every key-value
714 : /// pair in the key range.
715 : ///
716 : /// 3. Call `finish`.
717 : ///
718 : struct ImageLayerWriterInner {
719 : conf: &'static PageServerConf,
720 : path: Utf8PathBuf,
721 : timeline_id: TimelineId,
722 : tenant_shard_id: TenantShardId,
723 : key_range: Range<Key>,
724 : lsn: Lsn,
725 :
726 : // Total uncompressed bytes passed into put_image
727 : uncompressed_bytes: u64,
728 :
729 : // Like `uncompressed_bytes`,
730 : // but only of images we might consider for compression
731 : uncompressed_bytes_eligible: u64,
732 :
733 : // Like `uncompressed_bytes`, but only of images
734 : // where we have chosen their compressed form
735 : uncompressed_bytes_chosen: u64,
736 :
737 : // Number of keys in the layer.
738 : num_keys: usize,
739 :
740 : blob_writer: BlobWriter<false>,
741 : tree: DiskBtreeBuilder<BlockBuf, KEY_SIZE>,
742 :
743 : #[cfg(feature = "testing")]
744 : last_written_key: Key,
745 : }
746 :
747 : impl ImageLayerWriterInner {
748 : ///
749 : /// Start building a new image layer.
750 : ///
751 1204 : async fn new(
752 1204 : conf: &'static PageServerConf,
753 1204 : timeline_id: TimelineId,
754 1204 : tenant_shard_id: TenantShardId,
755 1204 : key_range: &Range<Key>,
756 1204 : lsn: Lsn,
757 1204 : ctx: &RequestContext,
758 1204 : ) -> anyhow::Result<Self> {
759 1204 : // Create the file initially with a temporary filename.
760 1204 : // We'll atomically rename it to the final name when we're done.
761 1204 : let path = ImageLayer::temp_path_for(
762 1204 : conf,
763 1204 : timeline_id,
764 1204 : tenant_shard_id,
765 1204 : &ImageLayerName {
766 1204 : key_range: key_range.clone(),
767 1204 : lsn,
768 1204 : },
769 1204 : );
770 1204 : trace!("creating image layer {}", path);
771 1204 : let mut file = {
772 1204 : VirtualFile::open_with_options(
773 1204 : &path,
774 1204 : virtual_file::OpenOptions::new()
775 1204 : .write(true)
776 1204 : .create_new(true),
777 1204 : ctx,
778 1204 : )
779 1204 : .await?
780 : };
781 : // make room for the header block
782 1204 : file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
783 1204 : let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
784 1204 :
785 1204 : // Initialize the b-tree index builder
786 1204 : let block_buf = BlockBuf::new();
787 1204 : let tree_builder = DiskBtreeBuilder::new(block_buf);
788 1204 :
789 1204 : let writer = Self {
790 1204 : conf,
791 1204 : path,
792 1204 : timeline_id,
793 1204 : tenant_shard_id,
794 1204 : key_range: key_range.clone(),
795 1204 : lsn,
796 1204 : tree: tree_builder,
797 1204 : blob_writer,
798 1204 : uncompressed_bytes: 0,
799 1204 : uncompressed_bytes_eligible: 0,
800 1204 : uncompressed_bytes_chosen: 0,
801 1204 : num_keys: 0,
802 1204 : #[cfg(feature = "testing")]
803 1204 : last_written_key: Key::MIN,
804 1204 : };
805 1204 :
806 1204 : Ok(writer)
807 1204 : }
808 :
809 : ///
810 : /// Write next value to the file.
811 : ///
812 : /// The page versions must be appended in blknum order.
813 : ///
814 1093372 : async fn put_image(
815 1093372 : &mut self,
816 1093372 : key: Key,
817 1093372 : img: Bytes,
818 1093372 : ctx: &RequestContext,
819 1093372 : ) -> anyhow::Result<()> {
820 1093372 : ensure!(self.key_range.contains(&key));
821 1093372 : let compression = self.conf.image_compression;
822 1093372 : let uncompressed_len = img.len() as u64;
823 1093372 : self.uncompressed_bytes += uncompressed_len;
824 1093372 : self.num_keys += 1;
825 1093372 : let (_img, res) = self
826 1093372 : .blob_writer
827 1093372 : .write_blob_maybe_compressed(img.slice_len(), ctx, compression)
828 1093372 : .await;
829 : // TODO: re-use the buffer for `img` further upstack
830 1093372 : let (off, compression_info) = res?;
831 1093372 : if compression_info.compressed_size.is_some() {
832 16004 : // The image has been considered for compression at least
833 16004 : self.uncompressed_bytes_eligible += uncompressed_len;
834 1077368 : }
835 1093372 : if compression_info.written_compressed {
836 0 : // The image has been compressed
837 0 : self.uncompressed_bytes_chosen += uncompressed_len;
838 1093372 : }
839 :
840 1093372 : let mut keybuf: [u8; KEY_SIZE] = [0u8; KEY_SIZE];
841 1093372 : key.write_to_byte_slice(&mut keybuf);
842 1093372 : self.tree.append(&keybuf, off)?;
843 :
844 : #[cfg(feature = "testing")]
845 1093372 : {
846 1093372 : self.last_written_key = key;
847 1093372 : }
848 1093372 :
849 1093372 : Ok(())
850 1093372 : }
851 :
852 : ///
853 : /// Finish writing the image layer.
854 : ///
855 732 : async fn finish(
856 732 : self,
857 732 : ctx: &RequestContext,
858 732 : end_key: Option<Key>,
859 732 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
860 732 : let temp_path = self.path.clone();
861 732 : let result = self.finish0(ctx, end_key).await;
862 732 : if let Err(ref e) = result {
863 0 : tracing::info!(%temp_path, "cleaning up temporary file after error during writing: {e}");
864 0 : if let Err(e) = std::fs::remove_file(&temp_path) {
865 0 : tracing::warn!(error=%e, %temp_path, "error cleaning up temporary layer file after error during writing");
866 0 : }
867 732 : }
868 732 : result
869 732 : }
870 :
871 : ///
872 : /// Finish writing the image layer.
873 : ///
874 732 : async fn finish0(
875 732 : self,
876 732 : ctx: &RequestContext,
877 732 : end_key: Option<Key>,
878 732 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
879 732 : let index_start_blk = self.blob_writer.size().div_ceil(PAGE_SZ as u64) as u32;
880 732 :
881 732 : // Calculate compression ratio
882 732 : let compressed_size = self.blob_writer.size() - PAGE_SZ as u64; // Subtract PAGE_SZ for header
883 732 : crate::metrics::COMPRESSION_IMAGE_INPUT_BYTES.inc_by(self.uncompressed_bytes);
884 732 : crate::metrics::COMPRESSION_IMAGE_INPUT_BYTES_CONSIDERED
885 732 : .inc_by(self.uncompressed_bytes_eligible);
886 732 : crate::metrics::COMPRESSION_IMAGE_INPUT_BYTES_CHOSEN.inc_by(self.uncompressed_bytes_chosen);
887 732 : crate::metrics::COMPRESSION_IMAGE_OUTPUT_BYTES.inc_by(compressed_size);
888 732 :
889 732 : let mut file = self.blob_writer.into_inner();
890 732 :
891 732 : // Write out the index
892 732 : file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
893 732 : .await?;
894 732 : let (index_root_blk, block_buf) = self.tree.finish()?;
895 2376 : for buf in block_buf.blocks {
896 1644 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
897 1644 : res?;
898 : }
899 :
900 732 : let final_key_range = if let Some(end_key) = end_key {
901 584 : self.key_range.start..end_key
902 : } else {
903 148 : self.key_range.clone()
904 : };
905 :
906 : // Fill in the summary on blk 0
907 732 : let summary = Summary {
908 732 : magic: IMAGE_FILE_MAGIC,
909 732 : format_version: STORAGE_FORMAT_VERSION,
910 732 : tenant_id: self.tenant_shard_id.tenant_id,
911 732 : timeline_id: self.timeline_id,
912 732 : key_range: final_key_range.clone(),
913 732 : lsn: self.lsn,
914 732 : index_start_blk,
915 732 : index_root_blk,
916 732 : };
917 732 :
918 732 : let mut buf = Vec::with_capacity(PAGE_SZ);
919 732 : // TODO: could use smallvec here but it's a pain with Slice<T>
920 732 : Summary::ser_into(&summary, &mut buf)?;
921 732 : file.seek(SeekFrom::Start(0)).await?;
922 732 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
923 732 : res?;
924 :
925 732 : let metadata = file
926 732 : .metadata()
927 732 : .await
928 732 : .context("get metadata to determine file size")?;
929 :
930 732 : let desc = PersistentLayerDesc::new_img(
931 732 : self.tenant_shard_id,
932 732 : self.timeline_id,
933 732 : final_key_range,
934 732 : self.lsn,
935 732 : metadata.len(),
936 732 : );
937 :
938 : #[cfg(feature = "testing")]
939 732 : if let Some(end_key) = end_key {
940 584 : assert!(
941 584 : self.last_written_key < end_key,
942 0 : "written key violates end_key range"
943 : );
944 148 : }
945 :
946 : // Note: Because we open the file in write-only mode, we cannot
947 : // reuse the same VirtualFile for reading later. That's why we don't
948 : // set inner.file here. The first read will have to re-open it.
949 :
950 : // fsync the file
951 732 : file.sync_all()
952 732 : .await
953 732 : .maybe_fatal_err("image_layer sync_all")?;
954 :
955 732 : trace!("created image layer {}", self.path);
956 :
957 732 : Ok((desc, self.path))
958 732 : }
959 : }
960 :
961 : /// A builder object for constructing a new image layer.
962 : ///
963 : /// Usage:
964 : ///
965 : /// 1. Create the ImageLayerWriter by calling ImageLayerWriter::new(...)
966 : ///
967 : /// 2. Write the contents by calling `put_page_image` for every key-value
968 : /// pair in the key range.
969 : ///
970 : /// 3. Call `finish`.
971 : ///
972 : /// # Note
973 : ///
974 : /// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
975 : /// possible for the writer to drop before `finish` is actually called. So this
976 : /// could lead to odd temporary files in the directory, exhausting file system.
977 : /// This structure wraps `ImageLayerWriterInner` and also contains `Drop`
978 : /// implementation that cleans up the temporary file in failure. It's not
979 : /// possible to do this directly in `ImageLayerWriterInner` since `finish` moves
980 : /// out some fields, making it impossible to implement `Drop`.
981 : ///
982 : #[must_use]
983 : pub struct ImageLayerWriter {
984 : inner: Option<ImageLayerWriterInner>,
985 : }
986 :
987 : impl ImageLayerWriter {
988 : ///
989 : /// Start building a new image layer.
990 : ///
991 1204 : pub async fn new(
992 1204 : conf: &'static PageServerConf,
993 1204 : timeline_id: TimelineId,
994 1204 : tenant_shard_id: TenantShardId,
995 1204 : key_range: &Range<Key>,
996 1204 : lsn: Lsn,
997 1204 : ctx: &RequestContext,
998 1204 : ) -> anyhow::Result<ImageLayerWriter> {
999 1204 : Ok(Self {
1000 1204 : inner: Some(
1001 1204 : ImageLayerWriterInner::new(conf, timeline_id, tenant_shard_id, key_range, lsn, ctx)
1002 1204 : .await?,
1003 : ),
1004 : })
1005 1204 : }
1006 :
1007 : ///
1008 : /// Write next value to the file.
1009 : ///
1010 : /// The page versions must be appended in blknum order.
1011 : ///
1012 1093372 : pub async fn put_image(
1013 1093372 : &mut self,
1014 1093372 : key: Key,
1015 1093372 : img: Bytes,
1016 1093372 : ctx: &RequestContext,
1017 1093372 : ) -> anyhow::Result<()> {
1018 1093372 : self.inner.as_mut().unwrap().put_image(key, img, ctx).await
1019 1093372 : }
1020 :
1021 : /// Estimated size of the image layer.
1022 17080 : pub(crate) fn estimated_size(&self) -> u64 {
1023 17080 : let inner = self.inner.as_ref().unwrap();
1024 17080 : inner.blob_writer.size() + inner.tree.borrow_writer().size() + PAGE_SZ as u64
1025 17080 : }
1026 :
1027 17272 : pub(crate) fn num_keys(&self) -> usize {
1028 17272 : self.inner.as_ref().unwrap().num_keys
1029 17272 : }
1030 :
1031 : ///
1032 : /// Finish writing the image layer.
1033 : ///
1034 148 : pub(crate) async fn finish(
1035 148 : mut self,
1036 148 : ctx: &RequestContext,
1037 148 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
1038 148 : self.inner.take().unwrap().finish(ctx, None).await
1039 148 : }
1040 :
1041 : /// Finish writing the image layer with an end key, used in [`super::batch_split_writer::SplitImageLayerWriter`]. The end key determines the end of the image layer's covered range and is exclusive.
1042 584 : pub(super) async fn finish_with_end_key(
1043 584 : mut self,
1044 584 : end_key: Key,
1045 584 : ctx: &RequestContext,
1046 584 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
1047 584 : self.inner.take().unwrap().finish(ctx, Some(end_key)).await
1048 584 : }
1049 : }
1050 :
1051 : impl Drop for ImageLayerWriter {
1052 1204 : fn drop(&mut self) {
1053 1204 : if let Some(inner) = self.inner.take() {
1054 472 : inner.blob_writer.into_inner().remove();
1055 732 : }
1056 1204 : }
1057 : }
1058 :
1059 : pub struct ImageLayerIterator<'a> {
1060 : image_layer: &'a ImageLayerInner,
1061 : ctx: &'a RequestContext,
1062 : planner: StreamingVectoredReadPlanner,
1063 : index_iter: DiskBtreeIterator<'a>,
1064 : key_values_batch: VecDeque<(Key, Lsn, Value)>,
1065 : is_end: bool,
1066 : }
1067 :
1068 : impl ImageLayerIterator<'_> {
1069 0 : pub(crate) fn layer_dbg_info(&self) -> String {
1070 0 : self.image_layer.layer_dbg_info()
1071 0 : }
1072 :
1073 : /// Retrieve a batch of key-value pairs into the iterator buffer.
1074 38252 : async fn next_batch(&mut self) -> anyhow::Result<()> {
1075 38252 : assert!(self.key_values_batch.is_empty());
1076 38252 : assert!(!self.is_end);
1077 :
1078 38252 : let plan = loop {
1079 58072 : if let Some(res) = self.index_iter.next().await {
1080 57884 : let (raw_key, offset) = res?;
1081 57884 : if let Some(batch_plan) = self.planner.handle(
1082 57884 : Key::from_slice(&raw_key[..KEY_SIZE]),
1083 57884 : self.image_layer.lsn,
1084 57884 : offset,
1085 57884 : true,
1086 57884 : ) {
1087 38064 : break batch_plan;
1088 19820 : }
1089 : } else {
1090 188 : self.is_end = true;
1091 188 : let payload_end = self.image_layer.index_start_blk as u64 * PAGE_SZ as u64;
1092 188 : if let Some(item) = self.planner.handle_range_end(payload_end) {
1093 188 : break item;
1094 : } else {
1095 0 : return Ok(()); // TODO: a test case on empty iterator
1096 : }
1097 : }
1098 : };
1099 38252 : let vectored_blob_reader = VectoredBlobReader::new(&self.image_layer.file);
1100 38252 : let mut next_batch = std::collections::VecDeque::new();
1101 38252 : let buf_size = plan.size();
1102 38252 : let buf = IoBufferMut::with_capacity(buf_size);
1103 38252 : let blobs_buf = vectored_blob_reader
1104 38252 : .read_blobs(&plan, buf, self.ctx)
1105 38252 : .await?;
1106 38252 : let view = BufView::new_slice(&blobs_buf.buf);
1107 57828 : for meta in blobs_buf.blobs.iter() {
1108 57828 : let img_buf = meta.read(&view).await?;
1109 57828 : next_batch.push_back((
1110 57828 : meta.meta.key,
1111 57828 : self.image_layer.lsn,
1112 57828 : Value::Image(img_buf.into_bytes()),
1113 57828 : ));
1114 : }
1115 38252 : self.key_values_batch = next_batch;
1116 38252 : Ok(())
1117 38252 : }
1118 :
1119 57608 : pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
1120 57608 : if self.key_values_batch.is_empty() {
1121 38404 : if self.is_end {
1122 320 : return Ok(None);
1123 38084 : }
1124 38084 : self.next_batch().await?;
1125 19204 : }
1126 57288 : Ok(Some(
1127 57288 : self.key_values_batch
1128 57288 : .pop_front()
1129 57288 : .expect("should not be empty"),
1130 57288 : ))
1131 57608 : }
1132 : }
1133 :
1134 : #[cfg(test)]
1135 : mod test {
1136 : use std::sync::Arc;
1137 : use std::time::Duration;
1138 :
1139 : use bytes::Bytes;
1140 : use itertools::Itertools;
1141 : use pageserver_api::key::Key;
1142 : use pageserver_api::shard::{ShardCount, ShardIdentity, ShardNumber, ShardStripeSize};
1143 : use pageserver_api::value::Value;
1144 : use utils::generation::Generation;
1145 : use utils::id::{TenantId, TimelineId};
1146 : use utils::lsn::Lsn;
1147 :
1148 : use super::{ImageLayerIterator, ImageLayerWriter};
1149 : use crate::DEFAULT_PG_VERSION;
1150 : use crate::context::RequestContext;
1151 : use crate::tenant::config::TenantConf;
1152 : use crate::tenant::harness::{TIMELINE_ID, TenantHarness};
1153 : use crate::tenant::storage_layer::{Layer, ResidentLayer};
1154 : use crate::tenant::vectored_blob_io::StreamingVectoredReadPlanner;
1155 : use crate::tenant::{Tenant, Timeline};
1156 :
1157 : #[tokio::test]
1158 4 : async fn image_layer_rewrite() {
1159 4 : let tenant_conf = TenantConf {
1160 4 : gc_period: Duration::ZERO,
1161 4 : compaction_period: Duration::ZERO,
1162 4 : ..TenantConf::default()
1163 4 : };
1164 4 : let tenant_id = TenantId::generate();
1165 4 : let mut gen_ = Generation::new(0xdead0001);
1166 20 : let mut get_next_gen = || {
1167 20 : let ret = gen_;
1168 20 : gen_ = gen_.next();
1169 20 : ret
1170 20 : };
1171 4 : // The LSN at which we will create an image layer to filter
1172 4 : let lsn = Lsn(0xdeadbeef0000);
1173 4 : let timeline_id = TimelineId::generate();
1174 4 :
1175 4 : //
1176 4 : // Create an unsharded parent with a layer.
1177 4 : //
1178 4 :
1179 4 : let harness = TenantHarness::create_custom(
1180 4 : "test_image_layer_rewrite--parent",
1181 4 : tenant_conf.clone(),
1182 4 : tenant_id,
1183 4 : ShardIdentity::unsharded(),
1184 4 : get_next_gen(),
1185 4 : )
1186 4 : .await
1187 4 : .unwrap();
1188 4 : let (tenant, ctx) = harness.load().await;
1189 4 : let timeline = tenant
1190 4 : .create_test_timeline(timeline_id, lsn, DEFAULT_PG_VERSION, &ctx)
1191 4 : .await
1192 4 : .unwrap();
1193 4 :
1194 4 : // This key range contains several 0x8000 page stripes, only one of which belongs to shard zero
1195 4 : let input_start = Key::from_hex("000000067f00000001000000ae0000000000").unwrap();
1196 4 : let input_end = Key::from_hex("000000067f00000001000000ae0000020000").unwrap();
1197 4 : let range = input_start..input_end;
1198 4 :
1199 4 : // Build an image layer to filter
1200 4 : let resident = {
1201 4 : let mut writer = ImageLayerWriter::new(
1202 4 : harness.conf,
1203 4 : timeline_id,
1204 4 : harness.tenant_shard_id,
1205 4 : &range,
1206 4 : lsn,
1207 4 : &ctx,
1208 4 : )
1209 4 : .await
1210 4 : .unwrap();
1211 4 :
1212 4 : let foo_img = Bytes::from_static(&[1, 2, 3, 4]);
1213 4 : let mut key = range.start;
1214 524292 : while key < range.end {
1215 524288 : writer.put_image(key, foo_img.clone(), &ctx).await.unwrap();
1216 524288 :
1217 524288 : key = key.next();
1218 4 : }
1219 4 : let (desc, path) = writer.finish(&ctx).await.unwrap();
1220 4 : Layer::finish_creating(tenant.conf, &timeline, desc, &path).unwrap()
1221 4 : };
1222 4 : let original_size = resident.metadata().file_size;
1223 4 :
1224 4 : //
1225 4 : // Create child shards and do the rewrite, exercising filter().
1226 4 : // TODO: abstraction in TenantHarness for splits.
1227 4 : //
1228 4 :
1229 4 : // Filter for various shards: this exercises cases like values at start of key range, end of key
1230 4 : // range, middle of key range.
1231 4 : let shard_count = ShardCount::new(4);
1232 16 : for shard_number in 0..shard_count.count() {
1233 4 : //
1234 4 : // mimic the shard split
1235 4 : //
1236 16 : let shard_identity = ShardIdentity::new(
1237 16 : ShardNumber(shard_number),
1238 16 : shard_count,
1239 16 : ShardStripeSize(0x8000),
1240 16 : )
1241 16 : .unwrap();
1242 16 : let harness = TenantHarness::create_custom(
1243 16 : Box::leak(Box::new(format!(
1244 16 : "test_image_layer_rewrite--child{}",
1245 16 : shard_identity.shard_slug()
1246 16 : ))),
1247 16 : tenant_conf.clone(),
1248 16 : tenant_id,
1249 16 : shard_identity,
1250 16 : // NB: in reality, the shards would each fork off their own gen number sequence from the parent.
1251 16 : // But here, all we care about is that the gen number is unique.
1252 16 : get_next_gen(),
1253 16 : )
1254 16 : .await
1255 16 : .unwrap();
1256 16 : let (tenant, ctx) = harness.load().await;
1257 16 : let timeline = tenant
1258 16 : .create_test_timeline(timeline_id, lsn, DEFAULT_PG_VERSION, &ctx)
1259 16 : .await
1260 16 : .unwrap();
1261 4 :
1262 4 : //
1263 4 : // use filter() and make assertions
1264 4 : //
1265 4 :
1266 16 : let mut filtered_writer = ImageLayerWriter::new(
1267 16 : harness.conf,
1268 16 : timeline_id,
1269 16 : harness.tenant_shard_id,
1270 16 : &range,
1271 16 : lsn,
1272 16 : &ctx,
1273 16 : )
1274 16 : .await
1275 16 : .unwrap();
1276 4 :
1277 16 : let wrote_keys = resident
1278 16 : .filter(&shard_identity, &mut filtered_writer, &ctx)
1279 16 : .await
1280 16 : .unwrap();
1281 16 : let replacement = if wrote_keys > 0 {
1282 12 : let (desc, path) = filtered_writer.finish(&ctx).await.unwrap();
1283 12 : let resident = Layer::finish_creating(tenant.conf, &timeline, desc, &path).unwrap();
1284 12 : Some(resident)
1285 4 : } else {
1286 4 : None
1287 4 : };
1288 4 :
1289 4 : // This exact size and those below will need updating as/when the layer encoding changes, but
1290 4 : // should be deterministic for a given version of the format, as we used no randomness generating the input.
1291 16 : assert_eq!(original_size, 1597440);
1292 4 :
1293 16 : match shard_number {
1294 4 : 0 => {
1295 4 : // We should have written out just one stripe for our shard identity
1296 4 : assert_eq!(wrote_keys, 0x8000);
1297 4 : let replacement = replacement.unwrap();
1298 4 :
1299 4 : // We should have dropped some of the data
1300 4 : assert!(replacement.metadata().file_size < original_size);
1301 4 : assert!(replacement.metadata().file_size > 0);
1302 4 :
1303 4 : // Assert that we dropped ~3/4 of the data.
1304 4 : assert_eq!(replacement.metadata().file_size, 417792);
1305 4 : }
1306 4 : 1 => {
1307 4 : // Shard 1 has no keys in our input range
1308 4 : assert_eq!(wrote_keys, 0x0);
1309 4 : assert!(replacement.is_none());
1310 4 : }
1311 4 : 2 => {
1312 4 : // Shard 2 has one stripes in the input range
1313 4 : assert_eq!(wrote_keys, 0x8000);
1314 4 : let replacement = replacement.unwrap();
1315 4 : assert!(replacement.metadata().file_size < original_size);
1316 4 : assert!(replacement.metadata().file_size > 0);
1317 4 : assert_eq!(replacement.metadata().file_size, 417792);
1318 4 : }
1319 4 : 3 => {
1320 4 : // Shard 3 has two stripes in the input range
1321 4 : assert_eq!(wrote_keys, 0x10000);
1322 4 : let replacement = replacement.unwrap();
1323 4 : assert!(replacement.metadata().file_size < original_size);
1324 4 : assert!(replacement.metadata().file_size > 0);
1325 4 : assert_eq!(replacement.metadata().file_size, 811008);
1326 4 : }
1327 4 : _ => unreachable!(),
1328 4 : }
1329 4 : }
1330 4 : }
1331 :
1332 4 : async fn produce_image_layer(
1333 4 : tenant: &Tenant,
1334 4 : tline: &Arc<Timeline>,
1335 4 : mut images: Vec<(Key, Bytes)>,
1336 4 : lsn: Lsn,
1337 4 : ctx: &RequestContext,
1338 4 : ) -> anyhow::Result<ResidentLayer> {
1339 4 : images.sort();
1340 4 : let (key_start, _) = images.first().unwrap();
1341 4 : let (key_last, _) = images.last().unwrap();
1342 4 : let key_end = key_last.next();
1343 4 : let key_range = *key_start..key_end;
1344 4 : let mut writer = ImageLayerWriter::new(
1345 4 : tenant.conf,
1346 4 : tline.timeline_id,
1347 4 : tenant.tenant_shard_id,
1348 4 : &key_range,
1349 4 : lsn,
1350 4 : ctx,
1351 4 : )
1352 4 : .await?;
1353 :
1354 4004 : for (key, img) in images {
1355 4000 : writer.put_image(key, img, ctx).await?;
1356 : }
1357 4 : let (desc, path) = writer.finish(ctx).await?;
1358 4 : let img_layer = Layer::finish_creating(tenant.conf, tline, desc, &path)?;
1359 :
1360 4 : Ok::<_, anyhow::Error>(img_layer)
1361 4 : }
1362 :
1363 56 : async fn assert_img_iter_equal(
1364 56 : img_iter: &mut ImageLayerIterator<'_>,
1365 56 : expect: &[(Key, Bytes)],
1366 56 : expect_lsn: Lsn,
1367 56 : ) {
1368 56 : let mut expect_iter = expect.iter();
1369 : loop {
1370 56056 : let o1 = img_iter.next().await.unwrap();
1371 56056 : let o2 = expect_iter.next();
1372 56056 : match (o1, o2) {
1373 56 : (None, None) => break,
1374 56000 : (Some((k1, l1, v1)), Some((k2, i2))) => {
1375 56000 : let Value::Image(i1) = v1 else {
1376 0 : panic!("expect Value::Image")
1377 : };
1378 56000 : assert_eq!(&k1, k2);
1379 56000 : assert_eq!(l1, expect_lsn);
1380 56000 : assert_eq!(&i1, i2);
1381 : }
1382 0 : (o1, o2) => panic!("iterators length mismatch: {:?}, {:?}", o1, o2),
1383 : }
1384 : }
1385 56 : }
1386 :
1387 : #[tokio::test]
1388 4 : async fn image_layer_iterator() {
1389 4 : let harness = TenantHarness::create("image_layer_iterator").await.unwrap();
1390 4 : let (tenant, ctx) = harness.load().await;
1391 4 :
1392 4 : let tline = tenant
1393 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
1394 4 : .await
1395 4 : .unwrap();
1396 4 :
1397 4000 : fn get_key(id: u32) -> Key {
1398 4000 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
1399 4000 : key.field6 = id;
1400 4000 : key
1401 4000 : }
1402 4 : const N: usize = 1000;
1403 4 : let test_imgs = (0..N)
1404 4000 : .map(|idx| (get_key(idx as u32), Bytes::from(format!("img{idx:05}"))))
1405 4 : .collect_vec();
1406 4 : let resident_layer =
1407 4 : produce_image_layer(&tenant, &tline, test_imgs.clone(), Lsn(0x10), &ctx)
1408 4 : .await
1409 4 : .unwrap();
1410 4 : let img_layer = resident_layer.get_as_image(&ctx).await.unwrap();
1411 12 : for max_read_size in [1, 1024] {
1412 64 : for batch_size in [1, 2, 4, 8, 3, 7, 13] {
1413 56 : println!("running with batch_size={batch_size} max_read_size={max_read_size}");
1414 56 : // Test if the batch size is correctly determined
1415 56 : let mut iter = img_layer.iter(&ctx);
1416 56 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
1417 56 : let mut num_items = 0;
1418 224 : for _ in 0..3 {
1419 168 : iter.next_batch().await.unwrap();
1420 168 : num_items += iter.key_values_batch.len();
1421 168 : if max_read_size == 1 {
1422 4 : // every key should be a batch b/c the value is larger than max_read_size
1423 84 : assert_eq!(iter.key_values_batch.len(), 1);
1424 4 : } else {
1425 84 : assert!(iter.key_values_batch.len() <= batch_size);
1426 4 : }
1427 168 : if num_items >= N {
1428 4 : break;
1429 168 : }
1430 168 : iter.key_values_batch.clear();
1431 4 : }
1432 4 : // Test if the result is correct
1433 56 : let mut iter = img_layer.iter(&ctx);
1434 56 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
1435 56 : assert_img_iter_equal(&mut iter, &test_imgs, Lsn(0x10)).await;
1436 4 : }
1437 4 : }
1438 4 : }
1439 : }
|