Line data Source code
1 : //! A DeltaLayer represents a collection of WAL records or page images in a range of
2 : //! LSNs, and in a range of Keys. It is stored on a file on disk.
3 : //!
4 : //! Usually a delta layer only contains differences, in the form of WAL records
5 : //! against a base LSN. However, if a relation extended or a whole new relation
6 : //! is created, there would be no base for the new pages. The entries for them
7 : //! must be page images or WAL records with the 'will_init' flag set, so that
8 : //! they can be replayed without referring to an older page version.
9 : //!
10 : //! The delta files are stored in `timelines/<timeline_id>` directory. Currently,
11 : //! there are no subdirectories, and each delta file is named like this:
12 : //!
13 : //! ```text
14 : //! <key start>-<key end>__<start LSN>-<end LSN>
15 : //! ```
16 : //!
17 : //! For example:
18 : //!
19 : //! ```text
20 : //! 000000067F000032BE0000400000000020B6-000000067F000032BE0000400000000030B6__000000578C6B29-0000000057A50051
21 : //! ```
22 : //!
23 : //! Every delta file consists of three parts: "summary", "values", and
24 : //! "index". The summary is a fixed size header at the beginning of the file,
25 : //! and it contains basic information about the layer, and offsets to the other
26 : //! parts. The "index" is a B-tree, mapping from Key and LSN to an offset in the
27 : //! "values" part. The actual page images and WAL records are stored in the
28 : //! "values" part.
29 : //!
30 : use crate::config::PageServerConf;
31 : use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
32 : use crate::page_cache::{self, FileId, PAGE_SZ};
33 : use crate::repository::{Key, Value, KEY_SIZE};
34 : use crate::tenant::blob_io::BlobWriter;
35 : use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockLease, BlockReader, FileBlockReader};
36 : use crate::tenant::disk_btree::{
37 : DiskBtreeBuilder, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
38 : };
39 : use crate::tenant::storage_layer::{Layer, ValueReconstructResult, ValueReconstructState};
40 : use crate::tenant::timeline::GetVectoredError;
41 : use crate::tenant::vectored_blob_io::{
42 : BlobFlag, MaxVectoredReadBytes, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
43 : VectoredReadPlanner,
44 : };
45 : use crate::tenant::{PageReconstructError, Timeline};
46 : use crate::virtual_file::{self, VirtualFile};
47 : use crate::{walrecord, TEMP_FILE_SUFFIX};
48 : use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
49 : use anyhow::{anyhow, bail, ensure, Context, Result};
50 : use bytes::BytesMut;
51 : use camino::{Utf8Path, Utf8PathBuf};
52 : use futures::StreamExt;
53 : use itertools::Itertools;
54 : use pageserver_api::keyspace::KeySpace;
55 : use pageserver_api::models::ImageCompressionAlgorithm;
56 : use pageserver_api::shard::TenantShardId;
57 : use rand::{distributions::Alphanumeric, Rng};
58 : use serde::{Deserialize, Serialize};
59 : use std::collections::VecDeque;
60 : use std::fs::File;
61 : use std::io::SeekFrom;
62 : use std::ops::Range;
63 : use std::os::unix::fs::FileExt;
64 : use std::str::FromStr;
65 : use std::sync::Arc;
66 : use tokio::sync::OnceCell;
67 : use tracing::*;
68 :
69 : use utils::{
70 : bin_ser::BeSer,
71 : id::{TenantId, TimelineId},
72 : lsn::Lsn,
73 : };
74 :
75 : use super::{
76 : AsLayerDesc, LayerAccessStats, LayerName, PersistentLayerDesc, ResidentLayer,
77 : ValuesReconstructState,
78 : };
79 :
80 : ///
81 : /// Header stored in the beginning of the file
82 : ///
83 : /// After this comes the 'values' part, starting on block 1. After that,
84 : /// the 'index' starts at the block indicated by 'index_start_blk'
85 : ///
86 1050 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
87 : pub struct Summary {
88 : /// Magic value to identify this as a neon delta file. Always DELTA_FILE_MAGIC.
89 : pub magic: u16,
90 : pub format_version: u16,
91 :
92 : pub tenant_id: TenantId,
93 : pub timeline_id: TimelineId,
94 : pub key_range: Range<Key>,
95 : pub lsn_range: Range<Lsn>,
96 :
97 : /// Block number where the 'index' part of the file begins.
98 : pub index_start_blk: u32,
99 : /// Block within the 'index', where the B-tree root page is stored
100 : pub index_root_blk: u32,
101 : }
102 :
103 : impl From<&DeltaLayer> for Summary {
104 0 : fn from(layer: &DeltaLayer) -> Self {
105 0 : Self::expected(
106 0 : layer.desc.tenant_shard_id.tenant_id,
107 0 : layer.desc.timeline_id,
108 0 : layer.desc.key_range.clone(),
109 0 : layer.desc.lsn_range.clone(),
110 0 : )
111 0 : }
112 : }
113 :
114 : impl Summary {
115 1050 : pub(super) fn expected(
116 1050 : tenant_id: TenantId,
117 1050 : timeline_id: TimelineId,
118 1050 : keys: Range<Key>,
119 1050 : lsns: Range<Lsn>,
120 1050 : ) -> Self {
121 1050 : Self {
122 1050 : magic: DELTA_FILE_MAGIC,
123 1050 : format_version: STORAGE_FORMAT_VERSION,
124 1050 :
125 1050 : tenant_id,
126 1050 : timeline_id,
127 1050 : key_range: keys,
128 1050 : lsn_range: lsns,
129 1050 :
130 1050 : index_start_blk: 0,
131 1050 : index_root_blk: 0,
132 1050 : }
133 1050 : }
134 : }
135 :
136 : // Flag indicating that this version initialize the page
137 : const WILL_INIT: u64 = 1;
138 :
139 : /// Struct representing reference to BLOB in layers. Reference contains BLOB
140 : /// offset, and for WAL records it also contains `will_init` flag. The flag
141 : /// helps to determine the range of records that needs to be applied, without
142 : /// reading/deserializing records themselves.
143 0 : #[derive(Debug, Serialize, Deserialize, Copy, Clone)]
144 : pub struct BlobRef(pub u64);
145 :
146 : impl BlobRef {
147 510532 : pub fn will_init(&self) -> bool {
148 510532 : (self.0 & WILL_INIT) != 0
149 510532 : }
150 :
151 6810204 : pub fn pos(&self) -> u64 {
152 6810204 : self.0 >> 1
153 6810204 : }
154 :
155 6458996 : pub fn new(pos: u64, will_init: bool) -> BlobRef {
156 6458996 : let mut blob_ref = pos << 1;
157 6458996 : if will_init {
158 6457656 : blob_ref |= WILL_INIT;
159 6457656 : }
160 6458996 : BlobRef(blob_ref)
161 6458996 : }
162 : }
163 :
164 : pub const DELTA_KEY_SIZE: usize = KEY_SIZE + 8;
165 : struct DeltaKey([u8; DELTA_KEY_SIZE]);
166 :
167 : /// This is the key of the B-tree index stored in the delta layer. It consists
168 : /// of the serialized representation of a Key and LSN.
169 : impl DeltaKey {
170 2064198 : fn from_slice(buf: &[u8]) -> Self {
171 2064198 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
172 2064198 : bytes.copy_from_slice(buf);
173 2064198 : DeltaKey(bytes)
174 2064198 : }
175 :
176 6706059 : fn from_key_lsn(key: &Key, lsn: Lsn) -> Self {
177 6706059 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
178 6706059 : key.write_to_byte_slice(&mut bytes[0..KEY_SIZE]);
179 6706059 : bytes[KEY_SIZE..].copy_from_slice(&u64::to_be_bytes(lsn.0));
180 6706059 : DeltaKey(bytes)
181 6706059 : }
182 :
183 2064198 : fn key(&self) -> Key {
184 2064198 : Key::from_slice(&self.0)
185 2064198 : }
186 :
187 2064198 : fn lsn(&self) -> Lsn {
188 2064198 : Lsn(u64::from_be_bytes(self.0[KEY_SIZE..].try_into().unwrap()))
189 2064198 : }
190 :
191 2681908 : fn extract_lsn_from_buf(buf: &[u8]) -> Lsn {
192 2681908 : let mut lsn_buf = [0u8; 8];
193 2681908 : lsn_buf.copy_from_slice(&buf[KEY_SIZE..]);
194 2681908 : Lsn(u64::from_be_bytes(lsn_buf))
195 2681908 : }
196 : }
197 :
198 : /// This is used only from `pagectl`. Within pageserver, all layers are
199 : /// [`crate::tenant::storage_layer::Layer`], which can hold a [`DeltaLayerInner`].
200 : pub struct DeltaLayer {
201 : path: Utf8PathBuf,
202 : pub desc: PersistentLayerDesc,
203 : access_stats: LayerAccessStats,
204 : inner: OnceCell<Arc<DeltaLayerInner>>,
205 : }
206 :
207 : impl std::fmt::Debug for DeltaLayer {
208 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
209 0 : use super::RangeDisplayDebug;
210 0 :
211 0 : f.debug_struct("DeltaLayer")
212 0 : .field("key_range", &RangeDisplayDebug(&self.desc.key_range))
213 0 : .field("lsn_range", &self.desc.lsn_range)
214 0 : .field("file_size", &self.desc.file_size)
215 0 : .field("inner", &self.inner)
216 0 : .finish()
217 0 : }
218 : }
219 :
220 : /// `DeltaLayerInner` is the in-memory data structure associated with an on-disk delta
221 : /// file.
222 : pub struct DeltaLayerInner {
223 : // values copied from summary
224 : index_start_blk: u32,
225 : index_root_blk: u32,
226 :
227 : file: VirtualFile,
228 : file_id: FileId,
229 :
230 : #[allow(dead_code)]
231 : layer_key_range: Range<Key>,
232 : #[allow(dead_code)]
233 : layer_lsn_range: Range<Lsn>,
234 :
235 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
236 : }
237 :
238 : impl std::fmt::Debug for DeltaLayerInner {
239 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
240 0 : f.debug_struct("DeltaLayerInner")
241 0 : .field("index_start_blk", &self.index_start_blk)
242 0 : .field("index_root_blk", &self.index_root_blk)
243 0 : .finish()
244 0 : }
245 : }
246 :
247 : /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
248 : impl std::fmt::Display for DeltaLayer {
249 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
250 0 : write!(f, "{}", self.layer_desc().short_id())
251 0 : }
252 : }
253 :
254 : impl AsLayerDesc for DeltaLayer {
255 0 : fn layer_desc(&self) -> &PersistentLayerDesc {
256 0 : &self.desc
257 0 : }
258 : }
259 :
260 : impl DeltaLayer {
261 0 : pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
262 0 : self.desc.dump();
263 0 :
264 0 : if !verbose {
265 0 : return Ok(());
266 0 : }
267 :
268 0 : let inner = self.load(ctx).await?;
269 :
270 0 : inner.dump(ctx).await
271 0 : }
272 :
273 1378 : fn temp_path_for(
274 1378 : conf: &PageServerConf,
275 1378 : tenant_shard_id: &TenantShardId,
276 1378 : timeline_id: &TimelineId,
277 1378 : key_start: Key,
278 1378 : lsn_range: &Range<Lsn>,
279 1378 : ) -> Utf8PathBuf {
280 1378 : let rand_string: String = rand::thread_rng()
281 1378 : .sample_iter(&Alphanumeric)
282 1378 : .take(8)
283 1378 : .map(char::from)
284 1378 : .collect();
285 1378 :
286 1378 : conf.timeline_path(tenant_shard_id, timeline_id)
287 1378 : .join(format!(
288 1378 : "{}-XXX__{:016X}-{:016X}.{}.{}",
289 1378 : key_start,
290 1378 : u64::from(lsn_range.start),
291 1378 : u64::from(lsn_range.end),
292 1378 : rand_string,
293 1378 : TEMP_FILE_SUFFIX,
294 1378 : ))
295 1378 : }
296 :
297 : ///
298 : /// Open the underlying file and read the metadata into memory, if it's
299 : /// not loaded already.
300 : ///
301 0 : async fn load(&self, ctx: &RequestContext) -> Result<&Arc<DeltaLayerInner>> {
302 0 : self.access_stats.record_access(ctx);
303 0 : // Quick exit if already loaded
304 0 : self.inner
305 0 : .get_or_try_init(|| self.load_inner(ctx))
306 0 : .await
307 0 : .with_context(|| format!("Failed to load delta layer {}", self.path()))
308 0 : }
309 :
310 0 : async fn load_inner(&self, ctx: &RequestContext) -> anyhow::Result<Arc<DeltaLayerInner>> {
311 0 : let path = self.path();
312 :
313 0 : let loaded = DeltaLayerInner::load(&path, None, None, ctx).await?;
314 :
315 : // not production code
316 0 : let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
317 0 : let expected_layer_name = self.layer_desc().layer_name();
318 0 :
319 0 : if actual_layer_name != expected_layer_name {
320 0 : println!("warning: filename does not match what is expected from in-file summary");
321 0 : println!("actual: {:?}", actual_layer_name.to_string());
322 0 : println!("expected: {:?}", expected_layer_name.to_string());
323 0 : }
324 :
325 0 : Ok(Arc::new(loaded))
326 0 : }
327 :
328 : /// Create a DeltaLayer struct representing an existing file on disk.
329 : ///
330 : /// This variant is only used for debugging purposes, by the 'pagectl' binary.
331 0 : pub fn new_for_path(path: &Utf8Path, file: File) -> Result<Self> {
332 0 : let mut summary_buf = vec![0; PAGE_SZ];
333 0 : file.read_exact_at(&mut summary_buf, 0)?;
334 0 : let summary = Summary::des_prefix(&summary_buf)?;
335 :
336 0 : let metadata = file
337 0 : .metadata()
338 0 : .context("get file metadata to determine size")?;
339 :
340 : // This function is never used for constructing layers in a running pageserver,
341 : // so it does not need an accurate TenantShardId.
342 0 : let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
343 0 :
344 0 : Ok(DeltaLayer {
345 0 : path: path.to_path_buf(),
346 0 : desc: PersistentLayerDesc::new_delta(
347 0 : tenant_shard_id,
348 0 : summary.timeline_id,
349 0 : summary.key_range,
350 0 : summary.lsn_range,
351 0 : metadata.len(),
352 0 : ),
353 0 : access_stats: Default::default(),
354 0 : inner: OnceCell::new(),
355 0 : })
356 0 : }
357 :
358 : /// Path to the layer file in pageserver workdir.
359 0 : fn path(&self) -> Utf8PathBuf {
360 0 : self.path.clone()
361 0 : }
362 : }
363 :
364 : /// A builder object for constructing a new delta layer.
365 : ///
366 : /// Usage:
367 : ///
368 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
369 : ///
370 : /// 2. Write the contents by calling `put_value` for every page
371 : /// version to store in the layer.
372 : ///
373 : /// 3. Call `finish`.
374 : ///
375 : struct DeltaLayerWriterInner {
376 : conf: &'static PageServerConf,
377 : pub path: Utf8PathBuf,
378 : timeline_id: TimelineId,
379 : tenant_shard_id: TenantShardId,
380 :
381 : key_start: Key,
382 : lsn_range: Range<Lsn>,
383 :
384 : tree: DiskBtreeBuilder<BlockBuf, DELTA_KEY_SIZE>,
385 :
386 : blob_writer: BlobWriter<true>,
387 : }
388 :
389 : impl DeltaLayerWriterInner {
390 : ///
391 : /// Start building a new delta layer.
392 : ///
393 1378 : async fn new(
394 1378 : conf: &'static PageServerConf,
395 1378 : timeline_id: TimelineId,
396 1378 : tenant_shard_id: TenantShardId,
397 1378 : key_start: Key,
398 1378 : lsn_range: Range<Lsn>,
399 1378 : ctx: &RequestContext,
400 1378 : ) -> anyhow::Result<Self> {
401 1378 : // Create the file initially with a temporary filename. We don't know
402 1378 : // the end key yet, so we cannot form the final filename yet. We will
403 1378 : // rename it when we're done.
404 1378 : //
405 1378 : // Note: This overwrites any existing file. There shouldn't be any.
406 1378 : // FIXME: throw an error instead?
407 1378 : let path =
408 1378 : DeltaLayer::temp_path_for(conf, &tenant_shard_id, &timeline_id, key_start, &lsn_range);
409 :
410 1378 : let mut file = VirtualFile::create(&path, ctx).await?;
411 : // make room for the header block
412 1378 : file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
413 1378 : let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
414 1378 :
415 1378 : // Initialize the b-tree index builder
416 1378 : let block_buf = BlockBuf::new();
417 1378 : let tree_builder = DiskBtreeBuilder::new(block_buf);
418 1378 :
419 1378 : Ok(Self {
420 1378 : conf,
421 1378 : path,
422 1378 : timeline_id,
423 1378 : tenant_shard_id,
424 1378 : key_start,
425 1378 : lsn_range,
426 1378 : tree: tree_builder,
427 1378 : blob_writer,
428 1378 : })
429 1378 : }
430 :
431 : ///
432 : /// Append a key-value pair to the file.
433 : ///
434 : /// The values must be appended in key, lsn order.
435 : ///
436 2072230 : async fn put_value(
437 2072230 : &mut self,
438 2072230 : key: Key,
439 2072230 : lsn: Lsn,
440 2072230 : val: Value,
441 2072230 : ctx: &RequestContext,
442 2072230 : ) -> anyhow::Result<()> {
443 2072230 : let (_, res) = self
444 2072230 : .put_value_bytes(key, lsn, Value::ser(&val)?, val.will_init(), ctx)
445 1217 : .await;
446 2072230 : res
447 2072230 : }
448 :
449 6458892 : async fn put_value_bytes(
450 6458892 : &mut self,
451 6458892 : key: Key,
452 6458892 : lsn: Lsn,
453 6458892 : val: Vec<u8>,
454 6458892 : will_init: bool,
455 6458892 : ctx: &RequestContext,
456 6458892 : ) -> (Vec<u8>, anyhow::Result<()>) {
457 6458892 : assert!(
458 6458892 : self.lsn_range.start <= lsn,
459 0 : "lsn_start={}, lsn={}",
460 : self.lsn_range.start,
461 : lsn
462 : );
463 : // We don't want to use compression in delta layer creation
464 6458892 : let compression = ImageCompressionAlgorithm::Disabled;
465 6458892 : let (val, res) = self
466 6458892 : .blob_writer
467 6458892 : .write_blob_maybe_compressed(val, ctx, compression)
468 4164 : .await;
469 6458892 : let off = match res {
470 6458892 : Ok((off, _)) => off,
471 0 : Err(e) => return (val, Err(anyhow::anyhow!(e))),
472 : };
473 :
474 6458892 : let blob_ref = BlobRef::new(off, will_init);
475 6458892 :
476 6458892 : let delta_key = DeltaKey::from_key_lsn(&key, lsn);
477 6458892 : let res = self.tree.append(&delta_key.0, blob_ref.0);
478 6458892 : (val, res.map_err(|e| anyhow::anyhow!(e)))
479 6458892 : }
480 :
481 2023972 : fn size(&self) -> u64 {
482 2023972 : self.blob_writer.size() + self.tree.borrow_writer().size()
483 2023972 : }
484 :
485 : ///
486 : /// Finish writing the delta layer.
487 : ///
488 1378 : async fn finish(
489 1378 : self,
490 1378 : key_end: Key,
491 1378 : timeline: &Arc<Timeline>,
492 1378 : ctx: &RequestContext,
493 1378 : ) -> anyhow::Result<ResidentLayer> {
494 1378 : let temp_path = self.path.clone();
495 9633 : let result = self.finish0(key_end, timeline, ctx).await;
496 1378 : if result.is_err() {
497 0 : tracing::info!(%temp_path, "cleaning up temporary file after error during writing");
498 0 : if let Err(e) = std::fs::remove_file(&temp_path) {
499 0 : tracing::warn!(error=%e, %temp_path, "error cleaning up temporary layer file after error during writing");
500 0 : }
501 1378 : }
502 1378 : result
503 1378 : }
504 :
505 1378 : async fn finish0(
506 1378 : self,
507 1378 : key_end: Key,
508 1378 : timeline: &Arc<Timeline>,
509 1378 : ctx: &RequestContext,
510 1378 : ) -> anyhow::Result<ResidentLayer> {
511 1378 : let index_start_blk =
512 1378 : ((self.blob_writer.size() + PAGE_SZ as u64 - 1) / PAGE_SZ as u64) as u32;
513 :
514 1378 : let mut file = self.blob_writer.into_inner(ctx).await?;
515 :
516 : // Write out the index
517 1378 : let (index_root_blk, block_buf) = self.tree.finish()?;
518 1378 : file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
519 0 : .await?;
520 14973 : for buf in block_buf.blocks {
521 13595 : let (_buf, res) = file.write_all(buf, ctx).await;
522 13595 : res?;
523 : }
524 1378 : assert!(self.lsn_range.start < self.lsn_range.end);
525 : // Fill in the summary on blk 0
526 1378 : let summary = Summary {
527 1378 : magic: DELTA_FILE_MAGIC,
528 1378 : format_version: STORAGE_FORMAT_VERSION,
529 1378 : tenant_id: self.tenant_shard_id.tenant_id,
530 1378 : timeline_id: self.timeline_id,
531 1378 : key_range: self.key_start..key_end,
532 1378 : lsn_range: self.lsn_range.clone(),
533 1378 : index_start_blk,
534 1378 : index_root_blk,
535 1378 : };
536 1378 :
537 1378 : let mut buf = Vec::with_capacity(PAGE_SZ);
538 1378 : // TODO: could use smallvec here but it's a pain with Slice<T>
539 1378 : Summary::ser_into(&summary, &mut buf)?;
540 1378 : file.seek(SeekFrom::Start(0)).await?;
541 1378 : let (_buf, res) = file.write_all(buf, ctx).await;
542 1378 : res?;
543 :
544 1378 : let metadata = file
545 1378 : .metadata()
546 690 : .await
547 1378 : .context("get file metadata to determine size")?;
548 :
549 : // 5GB limit for objects without multipart upload (which we don't want to use)
550 : // Make it a little bit below to account for differing GB units
551 : // https://docs.aws.amazon.com/AmazonS3/latest/userguide/upload-objects.html
552 : const S3_UPLOAD_LIMIT: u64 = 4_500_000_000;
553 1378 : ensure!(
554 1378 : metadata.len() <= S3_UPLOAD_LIMIT,
555 0 : "Created delta layer file at {} of size {} above limit {S3_UPLOAD_LIMIT}!",
556 0 : file.path,
557 0 : metadata.len()
558 : );
559 :
560 : // Note: Because we opened the file in write-only mode, we cannot
561 : // reuse the same VirtualFile for reading later. That's why we don't
562 : // set inner.file here. The first read will have to re-open it.
563 :
564 1378 : let desc = PersistentLayerDesc::new_delta(
565 1378 : self.tenant_shard_id,
566 1378 : self.timeline_id,
567 1378 : self.key_start..key_end,
568 1378 : self.lsn_range.clone(),
569 1378 : metadata.len(),
570 1378 : );
571 1378 :
572 1378 : // fsync the file
573 1378 : file.sync_all().await?;
574 :
575 1378 : let layer = Layer::finish_creating(self.conf, timeline, desc, &self.path)?;
576 :
577 1378 : trace!("created delta layer {}", layer.local_path());
578 :
579 1378 : Ok(layer)
580 1378 : }
581 : }
582 :
583 : /// A builder object for constructing a new delta layer.
584 : ///
585 : /// Usage:
586 : ///
587 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
588 : ///
589 : /// 2. Write the contents by calling `put_value` for every page
590 : /// version to store in the layer.
591 : ///
592 : /// 3. Call `finish`.
593 : ///
594 : /// # Note
595 : ///
596 : /// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
597 : /// possible for the writer to drop before `finish` is actually called. So this
598 : /// could lead to odd temporary files in the directory, exhausting file system.
599 : /// This structure wraps `DeltaLayerWriterInner` and also contains `Drop`
600 : /// implementation that cleans up the temporary file in failure. It's not
601 : /// possible to do this directly in `DeltaLayerWriterInner` since `finish` moves
602 : /// out some fields, making it impossible to implement `Drop`.
603 : ///
604 : #[must_use]
605 : pub struct DeltaLayerWriter {
606 : inner: Option<DeltaLayerWriterInner>,
607 : }
608 :
609 : impl DeltaLayerWriter {
610 : ///
611 : /// Start building a new delta layer.
612 : ///
613 1378 : pub async fn new(
614 1378 : conf: &'static PageServerConf,
615 1378 : timeline_id: TimelineId,
616 1378 : tenant_shard_id: TenantShardId,
617 1378 : key_start: Key,
618 1378 : lsn_range: Range<Lsn>,
619 1378 : ctx: &RequestContext,
620 1378 : ) -> anyhow::Result<Self> {
621 1378 : Ok(Self {
622 1378 : inner: Some(
623 1378 : DeltaLayerWriterInner::new(
624 1378 : conf,
625 1378 : timeline_id,
626 1378 : tenant_shard_id,
627 1378 : key_start,
628 1378 : lsn_range,
629 1378 : ctx,
630 1378 : )
631 696 : .await?,
632 : ),
633 : })
634 1378 : }
635 :
636 : ///
637 : /// Append a key-value pair to the file.
638 : ///
639 : /// The values must be appended in key, lsn order.
640 : ///
641 2072230 : pub async fn put_value(
642 2072230 : &mut self,
643 2072230 : key: Key,
644 2072230 : lsn: Lsn,
645 2072230 : val: Value,
646 2072230 : ctx: &RequestContext,
647 2072230 : ) -> anyhow::Result<()> {
648 2072230 : self.inner
649 2072230 : .as_mut()
650 2072230 : .unwrap()
651 2072230 : .put_value(key, lsn, val, ctx)
652 1217 : .await
653 2072230 : }
654 :
655 4386662 : pub async fn put_value_bytes(
656 4386662 : &mut self,
657 4386662 : key: Key,
658 4386662 : lsn: Lsn,
659 4386662 : val: Vec<u8>,
660 4386662 : will_init: bool,
661 4386662 : ctx: &RequestContext,
662 4386662 : ) -> (Vec<u8>, anyhow::Result<()>) {
663 4386662 : self.inner
664 4386662 : .as_mut()
665 4386662 : .unwrap()
666 4386662 : .put_value_bytes(key, lsn, val, will_init, ctx)
667 2947 : .await
668 4386662 : }
669 :
670 2023972 : pub fn size(&self) -> u64 {
671 2023972 : self.inner.as_ref().unwrap().size()
672 2023972 : }
673 :
674 : ///
675 : /// Finish writing the delta layer.
676 : ///
677 1378 : pub(crate) async fn finish(
678 1378 : mut self,
679 1378 : key_end: Key,
680 1378 : timeline: &Arc<Timeline>,
681 1378 : ctx: &RequestContext,
682 1378 : ) -> anyhow::Result<ResidentLayer> {
683 1378 : self.inner
684 1378 : .take()
685 1378 : .unwrap()
686 1378 : .finish(key_end, timeline, ctx)
687 9633 : .await
688 1378 : }
689 : }
690 :
691 : impl Drop for DeltaLayerWriter {
692 1378 : fn drop(&mut self) {
693 1378 : if let Some(inner) = self.inner.take() {
694 0 : // We want to remove the virtual file here, so it's fine to not
695 0 : // having completely flushed unwritten data.
696 0 : let vfile = inner.blob_writer.into_inner_no_flush();
697 0 : vfile.remove();
698 1378 : }
699 1378 : }
700 : }
701 :
702 0 : #[derive(thiserror::Error, Debug)]
703 : pub enum RewriteSummaryError {
704 : #[error("magic mismatch")]
705 : MagicMismatch,
706 : #[error(transparent)]
707 : Other(#[from] anyhow::Error),
708 : }
709 :
710 : impl From<std::io::Error> for RewriteSummaryError {
711 0 : fn from(e: std::io::Error) -> Self {
712 0 : Self::Other(anyhow::anyhow!(e))
713 0 : }
714 : }
715 :
716 : impl DeltaLayer {
717 0 : pub async fn rewrite_summary<F>(
718 0 : path: &Utf8Path,
719 0 : rewrite: F,
720 0 : ctx: &RequestContext,
721 0 : ) -> Result<(), RewriteSummaryError>
722 0 : where
723 0 : F: Fn(Summary) -> Summary,
724 0 : {
725 0 : let mut file = VirtualFile::open_with_options(
726 0 : path,
727 0 : virtual_file::OpenOptions::new().read(true).write(true),
728 0 : ctx,
729 0 : )
730 0 : .await
731 0 : .with_context(|| format!("Failed to open file '{}'", path))?;
732 0 : let file_id = page_cache::next_file_id();
733 0 : let block_reader = FileBlockReader::new(&file, file_id);
734 0 : let summary_blk = block_reader.read_blk(0, ctx).await?;
735 0 : let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
736 0 : if actual_summary.magic != DELTA_FILE_MAGIC {
737 0 : return Err(RewriteSummaryError::MagicMismatch);
738 0 : }
739 0 :
740 0 : let new_summary = rewrite(actual_summary);
741 0 :
742 0 : let mut buf = Vec::with_capacity(PAGE_SZ);
743 0 : // TODO: could use smallvec here, but it's a pain with Slice<T>
744 0 : Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
745 0 : file.seek(SeekFrom::Start(0)).await?;
746 0 : let (_buf, res) = file.write_all(buf, ctx).await;
747 0 : res?;
748 0 : Ok(())
749 0 : }
750 : }
751 :
752 : impl DeltaLayerInner {
753 478 : pub(crate) fn key_range(&self) -> &Range<Key> {
754 478 : &self.layer_key_range
755 478 : }
756 :
757 478 : pub(crate) fn lsn_range(&self) -> &Range<Lsn> {
758 478 : &self.layer_lsn_range
759 478 : }
760 :
761 1050 : pub(super) async fn load(
762 1050 : path: &Utf8Path,
763 1050 : summary: Option<Summary>,
764 1050 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
765 1050 : ctx: &RequestContext,
766 1050 : ) -> anyhow::Result<Self> {
767 1050 : let file = VirtualFile::open(path, ctx)
768 525 : .await
769 1050 : .context("open layer file")?;
770 :
771 1050 : let file_id = page_cache::next_file_id();
772 1050 :
773 1050 : let block_reader = FileBlockReader::new(&file, file_id);
774 :
775 1050 : let summary_blk = block_reader
776 1050 : .read_blk(0, ctx)
777 529 : .await
778 1050 : .context("read first block")?;
779 :
780 : // TODO: this should be an assertion instead; see ImageLayerInner::load
781 1050 : let actual_summary =
782 1050 : Summary::des_prefix(summary_blk.as_ref()).context("deserialize first block")?;
783 :
784 1050 : if let Some(mut expected_summary) = summary {
785 : // production code path
786 1050 : expected_summary.index_start_blk = actual_summary.index_start_blk;
787 1050 : expected_summary.index_root_blk = actual_summary.index_root_blk;
788 1050 : // mask out the timeline_id, but still require the layers to be from the same tenant
789 1050 : expected_summary.timeline_id = actual_summary.timeline_id;
790 1050 :
791 1050 : if actual_summary != expected_summary {
792 0 : bail!(
793 0 : "in-file summary does not match expected summary. actual = {:?} expected = {:?}",
794 0 : actual_summary,
795 0 : expected_summary
796 0 : );
797 1050 : }
798 0 : }
799 :
800 1050 : Ok(DeltaLayerInner {
801 1050 : file,
802 1050 : file_id,
803 1050 : index_start_blk: actual_summary.index_start_blk,
804 1050 : index_root_blk: actual_summary.index_root_blk,
805 1050 : max_vectored_read_bytes,
806 1050 : layer_key_range: actual_summary.key_range,
807 1050 : layer_lsn_range: actual_summary.lsn_range,
808 1050 : })
809 1050 : }
810 :
811 320 : pub(super) async fn get_value_reconstruct_data(
812 320 : &self,
813 320 : key: Key,
814 320 : lsn_range: Range<Lsn>,
815 320 : reconstruct_state: &mut ValueReconstructState,
816 320 : ctx: &RequestContext,
817 320 : ) -> anyhow::Result<ValueReconstructResult> {
818 320 : let mut need_image = true;
819 320 : // Scan the page versions backwards, starting from `lsn`.
820 320 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
821 320 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
822 320 : self.index_start_blk,
823 320 : self.index_root_blk,
824 320 : &block_reader,
825 320 : );
826 320 : let search_key = DeltaKey::from_key_lsn(&key, Lsn(lsn_range.end.0 - 1));
827 320 :
828 320 : let mut offsets: Vec<(Lsn, u64)> = Vec::new();
829 320 :
830 320 : tree_reader
831 320 : .visit(
832 320 : &search_key.0,
833 320 : VisitDirection::Backwards,
834 320 : |key, value| {
835 320 : let blob_ref = BlobRef(value);
836 320 : if key[..KEY_SIZE] != search_key.0[..KEY_SIZE] {
837 0 : return false;
838 320 : }
839 320 : let entry_lsn = DeltaKey::extract_lsn_from_buf(key);
840 320 : if entry_lsn < lsn_range.start {
841 0 : return false;
842 320 : }
843 320 : offsets.push((entry_lsn, blob_ref.pos()));
844 320 :
845 320 : !blob_ref.will_init()
846 320 : },
847 320 : &RequestContextBuilder::extend(ctx)
848 320 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
849 320 : .build(),
850 320 : )
851 10 : .await?;
852 :
853 320 : let ctx = &RequestContextBuilder::extend(ctx)
854 320 : .page_content_kind(PageContentKind::DeltaLayerValue)
855 320 : .build();
856 320 :
857 320 : // Ok, 'offsets' now contains the offsets of all the entries we need to read
858 320 : let cursor = block_reader.block_cursor();
859 320 : let mut buf = Vec::new();
860 320 : for (entry_lsn, pos) in offsets {
861 320 : cursor
862 320 : .read_blob_into_buf(pos, &mut buf, ctx)
863 7 : .await
864 320 : .with_context(|| {
865 0 : format!("Failed to read blob from virtual file {}", self.file.path)
866 320 : })?;
867 320 : let val = Value::des(&buf).with_context(|| {
868 0 : format!(
869 0 : "Failed to deserialize file blob from virtual file {}",
870 0 : self.file.path
871 0 : )
872 320 : })?;
873 320 : match val {
874 320 : Value::Image(img) => {
875 320 : reconstruct_state.img = Some((entry_lsn, img));
876 320 : need_image = false;
877 320 : break;
878 : }
879 0 : Value::WalRecord(rec) => {
880 0 : let will_init = rec.will_init();
881 0 : reconstruct_state.records.push((entry_lsn, rec));
882 0 : if will_init {
883 : // This WAL record initializes the page, so no need to go further back
884 0 : need_image = false;
885 0 : break;
886 0 : }
887 : }
888 : }
889 : }
890 :
891 : // If an older page image is needed to reconstruct the page, let the
892 : // caller know.
893 320 : if need_image {
894 0 : Ok(ValueReconstructResult::Continue)
895 : } else {
896 320 : Ok(ValueReconstructResult::Complete)
897 : }
898 320 : }
899 :
900 : // Look up the keys in the provided keyspace and update
901 : // the reconstruct state with whatever is found.
902 : //
903 : // If the key is cached, go no further than the cached Lsn.
904 : //
905 : // Currently, the index is visited for each range, but this
906 : // can be further optimised to visit the index only once.
907 204299 : pub(super) async fn get_values_reconstruct_data(
908 204299 : &self,
909 204299 : keyspace: KeySpace,
910 204299 : lsn_range: Range<Lsn>,
911 204299 : reconstruct_state: &mut ValuesReconstructState,
912 204299 : ctx: &RequestContext,
913 204299 : ) -> Result<(), GetVectoredError> {
914 204299 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
915 204299 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
916 204299 : self.index_start_blk,
917 204299 : self.index_root_blk,
918 204299 : block_reader,
919 204299 : );
920 204299 :
921 204299 : let planner = VectoredReadPlanner::new(
922 204299 : self.max_vectored_read_bytes
923 204299 : .expect("Layer is loaded with max vectored bytes config")
924 204299 : .0
925 204299 : .into(),
926 204299 : );
927 204299 :
928 204299 : let data_end_offset = self.index_start_offset();
929 :
930 204299 : let reads = Self::plan_reads(
931 204299 : &keyspace,
932 204299 : lsn_range.clone(),
933 204299 : data_end_offset,
934 204299 : index_reader,
935 204299 : planner,
936 204299 : reconstruct_state,
937 204299 : ctx,
938 204299 : )
939 21532 : .await
940 204299 : .map_err(GetVectoredError::Other)?;
941 :
942 204299 : self.do_reads_and_update_state(reads, reconstruct_state, ctx)
943 70668 : .await;
944 :
945 204299 : reconstruct_state.on_lsn_advanced(&keyspace, lsn_range.start);
946 204299 :
947 204299 : Ok(())
948 204299 : }
949 :
950 : /// Load all key-values in the delta layer, should be replaced by an iterator-based interface in the future.
951 0 : pub(super) async fn load_key_values(
952 0 : &self,
953 0 : ctx: &RequestContext,
954 0 : ) -> anyhow::Result<Vec<(Key, Lsn, Value)>> {
955 0 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
956 0 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
957 0 : self.index_start_blk,
958 0 : self.index_root_blk,
959 0 : block_reader,
960 0 : );
961 0 : let mut result = Vec::new();
962 0 : let mut stream =
963 0 : Box::pin(self.stream_index_forwards(index_reader, &[0; DELTA_KEY_SIZE], ctx));
964 0 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
965 0 : let cursor = block_reader.block_cursor();
966 0 : let mut buf = Vec::new();
967 0 : while let Some(item) = stream.next().await {
968 0 : let (key, lsn, pos) = item?;
969 : // TODO: dedup code with get_reconstruct_value
970 : // TODO: ctx handling and sharding
971 0 : cursor
972 0 : .read_blob_into_buf(pos.pos(), &mut buf, ctx)
973 0 : .await
974 0 : .with_context(|| {
975 0 : format!("Failed to read blob from virtual file {}", self.file.path)
976 0 : })?;
977 0 : let val = Value::des(&buf).with_context(|| {
978 0 : format!(
979 0 : "Failed to deserialize file blob from virtual file {}",
980 0 : self.file.path
981 0 : )
982 0 : })?;
983 0 : result.push((key, lsn, val));
984 : }
985 0 : Ok(result)
986 0 : }
987 :
988 204501 : async fn plan_reads<Reader>(
989 204501 : keyspace: &KeySpace,
990 204501 : lsn_range: Range<Lsn>,
991 204501 : data_end_offset: u64,
992 204501 : index_reader: DiskBtreeReader<Reader, DELTA_KEY_SIZE>,
993 204501 : mut planner: VectoredReadPlanner,
994 204501 : reconstruct_state: &mut ValuesReconstructState,
995 204501 : ctx: &RequestContext,
996 204501 : ) -> anyhow::Result<Vec<VectoredRead>>
997 204501 : where
998 204501 : Reader: BlockReader + Clone,
999 204501 : {
1000 204501 : let ctx = RequestContextBuilder::extend(ctx)
1001 204501 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
1002 204501 : .build();
1003 :
1004 246813 : for range in keyspace.ranges.iter() {
1005 246813 : let mut range_end_handled = false;
1006 246813 :
1007 246813 : let start_key = DeltaKey::from_key_lsn(&range.start, lsn_range.start);
1008 246813 : let index_stream = index_reader.clone().into_stream(&start_key.0, &ctx);
1009 246813 : let mut index_stream = std::pin::pin!(index_stream);
1010 :
1011 592046 : while let Some(index_entry) = index_stream.next().await {
1012 583078 : let (raw_key, value) = index_entry?;
1013 583078 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
1014 583078 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
1015 583078 : let blob_ref = BlobRef(value);
1016 583078 :
1017 583078 : // Lsns are not monotonically increasing across keys, so we don't assert on them.
1018 583078 : assert!(key >= range.start);
1019 :
1020 583078 : let outside_lsn_range = !lsn_range.contains(&lsn);
1021 583078 : let below_cached_lsn = reconstruct_state.get_cached_lsn(&key) >= Some(lsn);
1022 :
1023 583078 : let flag = {
1024 583078 : if outside_lsn_range || below_cached_lsn {
1025 72866 : BlobFlag::Ignore
1026 510212 : } else if blob_ref.will_init() {
1027 452202 : BlobFlag::ReplaceAll
1028 : } else {
1029 : // Usual path: add blob to the read
1030 58010 : BlobFlag::None
1031 : }
1032 : };
1033 :
1034 583078 : if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
1035 237845 : planner.handle_range_end(blob_ref.pos());
1036 237845 : range_end_handled = true;
1037 237845 : break;
1038 345233 : } else {
1039 345233 : planner.handle(key, lsn, blob_ref.pos(), flag);
1040 345233 : }
1041 : }
1042 :
1043 246813 : if !range_end_handled {
1044 8968 : tracing::debug!("Handling range end fallback at {}", data_end_offset);
1045 8968 : planner.handle_range_end(data_end_offset);
1046 237845 : }
1047 : }
1048 :
1049 204501 : Ok(planner.finish())
1050 204501 : }
1051 :
1052 204499 : fn get_min_read_buffer_size(
1053 204499 : planned_reads: &[VectoredRead],
1054 204499 : read_size_soft_max: usize,
1055 204499 : ) -> usize {
1056 204499 : let Some(largest_read) = planned_reads.iter().max_by_key(|read| read.size()) else {
1057 82444 : return read_size_soft_max;
1058 : };
1059 :
1060 122055 : let largest_read_size = largest_read.size();
1061 122055 : if largest_read_size > read_size_soft_max {
1062 : // If the read is oversized, it should only contain one key.
1063 200 : let offenders = largest_read
1064 200 : .blobs_at
1065 200 : .as_slice()
1066 200 : .iter()
1067 200 : .map(|(_, blob_meta)| format!("{}@{}", blob_meta.key, blob_meta.lsn))
1068 200 : .join(", ");
1069 200 : tracing::warn!(
1070 0 : "Oversized vectored read ({} > {}) for keys {}",
1071 : largest_read_size,
1072 : read_size_soft_max,
1073 : offenders
1074 : );
1075 121855 : }
1076 :
1077 122055 : largest_read_size
1078 204499 : }
1079 :
1080 204299 : async fn do_reads_and_update_state(
1081 204299 : &self,
1082 204299 : reads: Vec<VectoredRead>,
1083 204299 : reconstruct_state: &mut ValuesReconstructState,
1084 204299 : ctx: &RequestContext,
1085 204299 : ) {
1086 204299 : let vectored_blob_reader = VectoredBlobReader::new(&self.file);
1087 204299 : let mut ignore_key_with_err = None;
1088 204299 :
1089 204299 : let max_vectored_read_bytes = self
1090 204299 : .max_vectored_read_bytes
1091 204299 : .expect("Layer is loaded with max vectored bytes config")
1092 204299 : .0
1093 204299 : .into();
1094 204299 : let buf_size = Self::get_min_read_buffer_size(&reads, max_vectored_read_bytes);
1095 204299 : let mut buf = Some(BytesMut::with_capacity(buf_size));
1096 :
1097 : // Note that reads are processed in reverse order (from highest key+lsn).
1098 : // This is the order that `ReconstructState` requires such that it can
1099 : // track when a key is done.
1100 204299 : for read in reads.into_iter().rev() {
1101 138949 : let res = vectored_blob_reader
1102 138949 : .read_blobs(&read, buf.take().expect("Should have a buffer"), ctx)
1103 70668 : .await;
1104 :
1105 138949 : let blobs_buf = match res {
1106 138949 : Ok(blobs_buf) => blobs_buf,
1107 0 : Err(err) => {
1108 0 : let kind = err.kind();
1109 0 : for (_, blob_meta) in read.blobs_at.as_slice() {
1110 0 : reconstruct_state.on_key_error(
1111 0 : blob_meta.key,
1112 0 : PageReconstructError::from(anyhow!(
1113 0 : "Failed to read blobs from virtual file {}: {}",
1114 0 : self.file.path,
1115 0 : kind
1116 0 : )),
1117 0 : );
1118 0 : }
1119 :
1120 : // We have "lost" the buffer since the lower level IO api
1121 : // doesn't return the buffer on error. Allocate a new one.
1122 0 : buf = Some(BytesMut::with_capacity(buf_size));
1123 0 :
1124 0 : continue;
1125 : }
1126 : };
1127 :
1128 149898 : for meta in blobs_buf.blobs.iter().rev() {
1129 149898 : if Some(meta.meta.key) == ignore_key_with_err {
1130 0 : continue;
1131 149898 : }
1132 149898 :
1133 149898 : let value = Value::des(&blobs_buf.buf[meta.start..meta.end]);
1134 149898 : let value = match value {
1135 149898 : Ok(v) => v,
1136 0 : Err(e) => {
1137 0 : reconstruct_state.on_key_error(
1138 0 : meta.meta.key,
1139 0 : PageReconstructError::from(anyhow!(e).context(format!(
1140 0 : "Failed to deserialize blob from virtual file {}",
1141 0 : self.file.path,
1142 0 : ))),
1143 0 : );
1144 0 :
1145 0 : ignore_key_with_err = Some(meta.meta.key);
1146 0 : continue;
1147 : }
1148 : };
1149 :
1150 : // Invariant: once a key reaches [`ValueReconstructSituation::Complete`]
1151 : // state, no further updates shall be made to it. The call below will
1152 : // panic if the invariant is violated.
1153 149898 : reconstruct_state.update_key(&meta.meta.key, meta.meta.lsn, value);
1154 : }
1155 :
1156 138949 : buf = Some(blobs_buf.buf);
1157 : }
1158 204299 : }
1159 :
1160 406 : pub(super) async fn load_keys<'a>(
1161 406 : &'a self,
1162 406 : ctx: &RequestContext,
1163 406 : ) -> Result<Vec<DeltaEntry<'a>>> {
1164 406 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1165 406 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1166 406 : self.index_start_blk,
1167 406 : self.index_root_blk,
1168 406 : block_reader,
1169 406 : );
1170 406 :
1171 406 : let mut all_keys: Vec<DeltaEntry<'_>> = Vec::new();
1172 406 :
1173 406 : tree_reader
1174 406 : .visit(
1175 406 : &[0u8; DELTA_KEY_SIZE],
1176 406 : VisitDirection::Forwards,
1177 2064046 : |key, value| {
1178 2064046 : let delta_key = DeltaKey::from_slice(key);
1179 2064046 : let val_ref = ValueRef {
1180 2064046 : blob_ref: BlobRef(value),
1181 2064046 : layer: self,
1182 2064046 : };
1183 2064046 : let pos = BlobRef(value).pos();
1184 2064046 : if let Some(last) = all_keys.last_mut() {
1185 2063640 : // subtract offset of the current and last entries to get the size
1186 2063640 : // of the value associated with this (key, lsn) tuple
1187 2063640 : let first_pos = last.size;
1188 2063640 : last.size = pos - first_pos;
1189 2063640 : }
1190 2064046 : let entry = DeltaEntry {
1191 2064046 : key: delta_key.key(),
1192 2064046 : lsn: delta_key.lsn(),
1193 2064046 : size: pos,
1194 2064046 : val: val_ref,
1195 2064046 : };
1196 2064046 : all_keys.push(entry);
1197 2064046 : true
1198 2064046 : },
1199 406 : &RequestContextBuilder::extend(ctx)
1200 406 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
1201 406 : .build(),
1202 406 : )
1203 2121 : .await?;
1204 406 : if let Some(last) = all_keys.last_mut() {
1205 406 : // Last key occupies all space till end of value storage,
1206 406 : // which corresponds to beginning of the index
1207 406 : last.size = self.index_start_offset() - last.size;
1208 406 : }
1209 406 : Ok(all_keys)
1210 406 : }
1211 :
1212 : /// Using the given writer, write out a version which has the earlier Lsns than `until`.
1213 : ///
1214 : /// Return the amount of key value records pushed to the writer.
1215 10 : pub(super) async fn copy_prefix(
1216 10 : &self,
1217 10 : writer: &mut DeltaLayerWriter,
1218 10 : until: Lsn,
1219 10 : ctx: &RequestContext,
1220 10 : ) -> anyhow::Result<usize> {
1221 10 : use crate::tenant::vectored_blob_io::{
1222 10 : BlobMeta, VectoredReadBuilder, VectoredReadExtended,
1223 10 : };
1224 10 : use futures::stream::TryStreamExt;
1225 10 :
1226 10 : #[derive(Debug)]
1227 10 : enum Item {
1228 10 : Actual(Key, Lsn, BlobRef),
1229 10 : Sentinel,
1230 10 : }
1231 10 :
1232 10 : impl From<Item> for Option<(Key, Lsn, BlobRef)> {
1233 70 : fn from(value: Item) -> Self {
1234 70 : match value {
1235 60 : Item::Actual(key, lsn, blob) => Some((key, lsn, blob)),
1236 10 : Item::Sentinel => None,
1237 10 : }
1238 70 : }
1239 10 : }
1240 10 :
1241 10 : impl Item {
1242 70 : fn offset(&self) -> Option<BlobRef> {
1243 70 : match self {
1244 60 : Item::Actual(_, _, blob) => Some(*blob),
1245 10 : Item::Sentinel => None,
1246 10 : }
1247 70 : }
1248 10 :
1249 70 : fn is_last(&self) -> bool {
1250 70 : matches!(self, Item::Sentinel)
1251 70 : }
1252 10 : }
1253 10 :
1254 10 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1255 10 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1256 10 : self.index_start_blk,
1257 10 : self.index_root_blk,
1258 10 : block_reader,
1259 10 : );
1260 10 :
1261 10 : let stream = self.stream_index_forwards(tree_reader, &[0u8; DELTA_KEY_SIZE], ctx);
1262 60 : let stream = stream.map_ok(|(key, lsn, pos)| Item::Actual(key, lsn, pos));
1263 10 : // put in a sentinel value for getting the end offset for last item, and not having to
1264 10 : // repeat the whole read part
1265 10 : let stream = stream.chain(futures::stream::once(futures::future::ready(Ok(
1266 10 : Item::Sentinel,
1267 10 : ))));
1268 10 : let mut stream = std::pin::pin!(stream);
1269 10 :
1270 10 : let mut prev: Option<(Key, Lsn, BlobRef)> = None;
1271 10 :
1272 10 : let mut read_builder: Option<VectoredReadBuilder> = None;
1273 10 :
1274 10 : let max_read_size = self
1275 10 : .max_vectored_read_bytes
1276 10 : .map(|x| x.0.get())
1277 10 : .unwrap_or(8192);
1278 10 :
1279 10 : let mut buffer = Some(BytesMut::with_capacity(max_read_size));
1280 10 :
1281 10 : // FIXME: buffering of DeltaLayerWriter
1282 10 : let mut per_blob_copy = Vec::new();
1283 10 :
1284 10 : let mut records = 0;
1285 :
1286 80 : while let Some(item) = stream.try_next().await? {
1287 70 : tracing::debug!(?item, "popped");
1288 70 : let offset = item
1289 70 : .offset()
1290 70 : .unwrap_or(BlobRef::new(self.index_start_offset(), false));
1291 :
1292 70 : let actionable = if let Some((key, lsn, start_offset)) = prev.take() {
1293 60 : let end_offset = offset;
1294 60 :
1295 60 : Some((BlobMeta { key, lsn }, start_offset..end_offset))
1296 : } else {
1297 10 : None
1298 : };
1299 :
1300 70 : let is_last = item.is_last();
1301 70 :
1302 70 : prev = Option::from(item);
1303 70 :
1304 70 : let actionable = actionable.filter(|x| x.0.lsn < until);
1305 :
1306 70 : let builder = if let Some((meta, offsets)) = actionable {
1307 : // extend or create a new builder
1308 32 : if read_builder
1309 32 : .as_mut()
1310 32 : .map(|x| x.extend(offsets.start.pos(), offsets.end.pos(), meta))
1311 32 : .unwrap_or(VectoredReadExtended::No)
1312 32 : == VectoredReadExtended::Yes
1313 : {
1314 16 : None
1315 : } else {
1316 16 : read_builder.replace(VectoredReadBuilder::new(
1317 16 : offsets.start.pos(),
1318 16 : offsets.end.pos(),
1319 16 : meta,
1320 16 : max_read_size,
1321 16 : ))
1322 : }
1323 : } else {
1324 : // nothing to do, except perhaps flush any existing for the last element
1325 38 : None
1326 : };
1327 :
1328 : // flush the possible older builder and also the new one if the item was the last one
1329 70 : let builders = builder.into_iter();
1330 70 : let builders = if is_last {
1331 10 : builders.chain(read_builder.take())
1332 : } else {
1333 60 : builders.chain(None)
1334 : };
1335 :
1336 86 : for builder in builders {
1337 16 : let read = builder.build();
1338 16 :
1339 16 : let reader = VectoredBlobReader::new(&self.file);
1340 16 :
1341 16 : let mut buf = buffer.take().unwrap();
1342 16 :
1343 16 : buf.clear();
1344 16 : buf.reserve(read.size());
1345 16 : let res = reader.read_blobs(&read, buf, ctx).await?;
1346 :
1347 48 : for blob in res.blobs {
1348 32 : let key = blob.meta.key;
1349 32 : let lsn = blob.meta.lsn;
1350 32 : let data = &res.buf[blob.start..blob.end];
1351 32 :
1352 32 : #[cfg(debug_assertions)]
1353 32 : Value::des(data)
1354 32 : .with_context(|| {
1355 0 : format!(
1356 0 : "blob failed to deserialize for {}@{}, {}..{}: {:?}",
1357 0 : blob.meta.key,
1358 0 : blob.meta.lsn,
1359 0 : blob.start,
1360 0 : blob.end,
1361 0 : utils::Hex(data)
1362 0 : )
1363 32 : })
1364 32 : .unwrap();
1365 32 :
1366 32 : // is it an image or will_init walrecord?
1367 32 : // FIXME: this could be handled by threading the BlobRef to the
1368 32 : // VectoredReadBuilder
1369 32 : let will_init = crate::repository::ValueBytes::will_init(data)
1370 32 : .inspect_err(|_e| {
1371 0 : #[cfg(feature = "testing")]
1372 0 : tracing::error!(data=?utils::Hex(data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
1373 32 : })
1374 32 : .unwrap_or(false);
1375 32 :
1376 32 : per_blob_copy.clear();
1377 32 : per_blob_copy.extend_from_slice(data);
1378 :
1379 32 : let (tmp, res) = writer
1380 32 : .put_value_bytes(
1381 32 : key,
1382 32 : lsn,
1383 32 : std::mem::take(&mut per_blob_copy),
1384 32 : will_init,
1385 32 : ctx,
1386 32 : )
1387 4 : .await;
1388 32 : per_blob_copy = tmp;
1389 32 :
1390 32 : res?;
1391 :
1392 32 : records += 1;
1393 : }
1394 :
1395 16 : buffer = Some(res.buf);
1396 : }
1397 : }
1398 :
1399 10 : assert!(
1400 10 : read_builder.is_none(),
1401 0 : "with the sentinel above loop should had handled all"
1402 : );
1403 :
1404 10 : Ok(records)
1405 10 : }
1406 :
1407 4 : pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
1408 4 : println!(
1409 4 : "index_start_blk: {}, root {}",
1410 4 : self.index_start_blk, self.index_root_blk
1411 4 : );
1412 4 :
1413 4 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1414 4 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1415 4 : self.index_start_blk,
1416 4 : self.index_root_blk,
1417 4 : block_reader,
1418 4 : );
1419 4 :
1420 4 : tree_reader.dump().await?;
1421 :
1422 4 : let keys = self.load_keys(ctx).await?;
1423 :
1424 8 : async fn dump_blob(val: &ValueRef<'_>, ctx: &RequestContext) -> anyhow::Result<String> {
1425 8 : let buf = val.load_raw(ctx).await?;
1426 8 : let val = Value::des(&buf)?;
1427 8 : let desc = match val {
1428 8 : Value::Image(img) => {
1429 8 : format!(" img {} bytes", img.len())
1430 : }
1431 0 : Value::WalRecord(rec) => {
1432 0 : let wal_desc = walrecord::describe_wal_record(&rec)?;
1433 0 : format!(
1434 0 : " rec {} bytes will_init: {} {}",
1435 0 : buf.len(),
1436 0 : rec.will_init(),
1437 0 : wal_desc
1438 0 : )
1439 : }
1440 : };
1441 8 : Ok(desc)
1442 8 : }
1443 :
1444 12 : for entry in keys {
1445 8 : let DeltaEntry { key, lsn, val, .. } = entry;
1446 8 : let desc = match dump_blob(&val, ctx).await {
1447 8 : Ok(desc) => desc,
1448 0 : Err(err) => {
1449 0 : format!("ERROR: {err}")
1450 : }
1451 : };
1452 8 : println!(" key {key} at {lsn}: {desc}");
1453 8 :
1454 8 : // Print more details about CHECKPOINT records. Would be nice to print details
1455 8 : // of many other record types too, but these are particularly interesting, as
1456 8 : // have a lot of special processing for them in walingest.rs.
1457 8 : use pageserver_api::key::CHECKPOINT_KEY;
1458 8 : use postgres_ffi::CheckPoint;
1459 8 : if key == CHECKPOINT_KEY {
1460 0 : let val = val.load(ctx).await?;
1461 0 : match val {
1462 0 : Value::Image(img) => {
1463 0 : let checkpoint = CheckPoint::decode(&img)?;
1464 0 : println!(" CHECKPOINT: {:?}", checkpoint);
1465 : }
1466 0 : Value::WalRecord(_rec) => {
1467 0 : println!(" unexpected walrecord value for checkpoint key");
1468 0 : }
1469 : }
1470 8 : }
1471 : }
1472 :
1473 4 : Ok(())
1474 4 : }
1475 :
1476 30 : fn stream_index_forwards<'a, R>(
1477 30 : &'a self,
1478 30 : reader: DiskBtreeReader<R, DELTA_KEY_SIZE>,
1479 30 : start: &'a [u8; DELTA_KEY_SIZE],
1480 30 : ctx: &'a RequestContext,
1481 30 : ) -> impl futures::stream::Stream<
1482 30 : Item = Result<(Key, Lsn, BlobRef), crate::tenant::disk_btree::DiskBtreeError>,
1483 30 : > + 'a
1484 30 : where
1485 30 : R: BlockReader + 'a,
1486 30 : {
1487 30 : use futures::stream::TryStreamExt;
1488 30 : let stream = reader.into_stream(start, ctx);
1489 152 : stream.map_ok(|(key, value)| {
1490 152 : let key = DeltaKey::from_slice(&key);
1491 152 : let (key, lsn) = (key.key(), key.lsn());
1492 152 : let offset = BlobRef(value);
1493 152 :
1494 152 : (key, lsn, offset)
1495 152 : })
1496 30 : }
1497 :
1498 : /// The file offset to the first block of index.
1499 : ///
1500 : /// The file structure is summary, values, and index. We often need this for the size of last blob.
1501 205281 : fn index_start_offset(&self) -> u64 {
1502 205281 : let offset = self.index_start_blk as u64 * PAGE_SZ as u64;
1503 205281 : let bref = BlobRef(offset);
1504 205281 : tracing::debug!(
1505 : index_start_blk = self.index_start_blk,
1506 : offset,
1507 0 : pos = bref.pos(),
1508 0 : "index_start_offset"
1509 : );
1510 205281 : offset
1511 205281 : }
1512 :
1513 534 : pub(crate) fn iter<'a>(&'a self, ctx: &'a RequestContext) -> DeltaLayerIterator<'a> {
1514 534 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1515 534 : let tree_reader =
1516 534 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
1517 534 : DeltaLayerIterator {
1518 534 : delta_layer: self,
1519 534 : ctx,
1520 534 : index_iter: tree_reader.iter(&[0; DELTA_KEY_SIZE], ctx),
1521 534 : key_values_batch: std::collections::VecDeque::new(),
1522 534 : is_end: false,
1523 534 : planner: StreamingVectoredReadPlanner::new(
1524 534 : 1024 * 8192, // The default value. Unit tests might use a different value. 1024 * 8K = 8MB buffer.
1525 534 : 1024, // The default value. Unit tests might use a different value
1526 534 : ),
1527 534 : }
1528 534 : }
1529 : }
1530 :
1531 : /// A set of data associated with a delta layer key and its value
1532 : pub struct DeltaEntry<'a> {
1533 : pub key: Key,
1534 : pub lsn: Lsn,
1535 : /// Size of the stored value
1536 : pub size: u64,
1537 : /// Reference to the on-disk value
1538 : pub val: ValueRef<'a>,
1539 : }
1540 :
1541 : /// Reference to an on-disk value
1542 : pub struct ValueRef<'a> {
1543 : blob_ref: BlobRef,
1544 : layer: &'a DeltaLayerInner,
1545 : }
1546 :
1547 : impl<'a> ValueRef<'a> {
1548 : /// Loads the value from disk
1549 2064038 : pub async fn load(&self, ctx: &RequestContext) -> Result<Value> {
1550 2064038 : let buf = self.load_raw(ctx).await?;
1551 2064038 : let val = Value::des(&buf)?;
1552 2064038 : Ok(val)
1553 2064038 : }
1554 :
1555 2064046 : async fn load_raw(&self, ctx: &RequestContext) -> Result<Vec<u8>> {
1556 2064046 : let reader = BlockCursor::new(crate::tenant::block_io::BlockReaderRef::Adapter(Adapter(
1557 2064046 : self.layer,
1558 2064046 : )));
1559 2064046 : let buf = reader.read_blob(self.blob_ref.pos(), ctx).await?;
1560 2064046 : Ok(buf)
1561 2064046 : }
1562 : }
1563 :
1564 : pub(crate) struct Adapter<T>(T);
1565 :
1566 : impl<T: AsRef<DeltaLayerInner>> Adapter<T> {
1567 2083002 : pub(crate) async fn read_blk(
1568 2083002 : &self,
1569 2083002 : blknum: u32,
1570 2083002 : ctx: &RequestContext,
1571 2083002 : ) -> Result<BlockLease, std::io::Error> {
1572 2083002 : let block_reader = FileBlockReader::new(&self.0.as_ref().file, self.0.as_ref().file_id);
1573 2083002 : block_reader.read_blk(blknum, ctx).await
1574 2083002 : }
1575 : }
1576 :
1577 : impl AsRef<DeltaLayerInner> for DeltaLayerInner {
1578 4166004 : fn as_ref(&self) -> &DeltaLayerInner {
1579 4166004 : self
1580 4166004 : }
1581 : }
1582 :
1583 : impl<'a> pageserver_compaction::interface::CompactionDeltaEntry<'a, Key> for DeltaEntry<'a> {
1584 0 : fn key(&self) -> Key {
1585 0 : self.key
1586 0 : }
1587 0 : fn lsn(&self) -> Lsn {
1588 0 : self.lsn
1589 0 : }
1590 0 : fn size(&self) -> u64 {
1591 0 : self.size
1592 0 : }
1593 : }
1594 :
1595 : pub struct DeltaLayerIterator<'a> {
1596 : delta_layer: &'a DeltaLayerInner,
1597 : ctx: &'a RequestContext,
1598 : planner: StreamingVectoredReadPlanner,
1599 : index_iter: DiskBtreeIterator<'a>,
1600 : key_values_batch: VecDeque<(Key, Lsn, Value)>,
1601 : is_end: bool,
1602 : }
1603 :
1604 : impl<'a> DeltaLayerIterator<'a> {
1605 : /// Retrieve a batch of key-value pairs into the iterator buffer.
1606 21224 : async fn next_batch(&mut self) -> anyhow::Result<()> {
1607 21224 : assert!(self.key_values_batch.is_empty());
1608 21224 : assert!(!self.is_end);
1609 :
1610 21224 : let plan = loop {
1611 2099016 : if let Some(res) = self.index_iter.next().await {
1612 2098510 : let (raw_key, value) = res?;
1613 2098510 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
1614 2098510 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
1615 2098510 : let blob_ref = BlobRef(value);
1616 2098510 : let offset = blob_ref.pos();
1617 2098510 : if let Some(batch_plan) = self.planner.handle(key, lsn, offset) {
1618 20718 : break batch_plan;
1619 2077792 : }
1620 : } else {
1621 506 : self.is_end = true;
1622 506 : let data_end_offset = self.delta_layer.index_start_offset();
1623 506 : if let Some(item) = self.planner.handle_range_end(data_end_offset) {
1624 506 : break item;
1625 : } else {
1626 0 : return Ok(()); // TODO: test empty iterator
1627 : }
1628 : }
1629 : };
1630 21224 : let vectored_blob_reader = VectoredBlobReader::new(&self.delta_layer.file);
1631 21224 : let mut next_batch = std::collections::VecDeque::new();
1632 21224 : let buf_size = plan.size();
1633 21224 : let buf = BytesMut::with_capacity(buf_size);
1634 21224 : let blobs_buf = vectored_blob_reader
1635 21224 : .read_blobs(&plan, buf, self.ctx)
1636 10836 : .await?;
1637 21224 : let frozen_buf = blobs_buf.buf.freeze();
1638 2098482 : for meta in blobs_buf.blobs.iter() {
1639 2098482 : let value = Value::des(&frozen_buf[meta.start..meta.end])?;
1640 2098482 : next_batch.push_back((meta.meta.key, meta.meta.lsn, value));
1641 : }
1642 21224 : self.key_values_batch = next_batch;
1643 21224 : Ok(())
1644 21224 : }
1645 :
1646 2099196 : pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
1647 2099196 : if self.key_values_batch.is_empty() {
1648 22124 : if self.is_end {
1649 984 : return Ok(None);
1650 21140 : }
1651 21140 : self.next_batch().await?;
1652 2077072 : }
1653 2098212 : Ok(Some(
1654 2098212 : self.key_values_batch
1655 2098212 : .pop_front()
1656 2098212 : .expect("should not be empty"),
1657 2098212 : ))
1658 2099196 : }
1659 : }
1660 :
1661 : #[cfg(test)]
1662 : pub(crate) mod test {
1663 : use std::collections::BTreeMap;
1664 :
1665 : use itertools::MinMaxResult;
1666 : use rand::prelude::{SeedableRng, SliceRandom, StdRng};
1667 : use rand::RngCore;
1668 :
1669 : use super::*;
1670 : use crate::repository::Value;
1671 : use crate::tenant::harness::TIMELINE_ID;
1672 : use crate::tenant::vectored_blob_io::StreamingVectoredReadPlanner;
1673 : use crate::tenant::Tenant;
1674 : use crate::{
1675 : context::DownloadBehavior,
1676 : task_mgr::TaskKind,
1677 : tenant::{disk_btree::tests::TestDisk, harness::TenantHarness},
1678 : DEFAULT_PG_VERSION,
1679 : };
1680 : use bytes::Bytes;
1681 :
1682 : /// Construct an index for a fictional delta layer and and then
1683 : /// traverse in order to plan vectored reads for a query. Finally,
1684 : /// verify that the traversal fed the right index key and value
1685 : /// pairs into the planner.
1686 : #[tokio::test]
1687 2 : async fn test_delta_layer_index_traversal() {
1688 2 : let base_key = Key {
1689 2 : field1: 0,
1690 2 : field2: 1663,
1691 2 : field3: 12972,
1692 2 : field4: 16396,
1693 2 : field5: 0,
1694 2 : field6: 246080,
1695 2 : };
1696 2 :
1697 2 : // Populate the index with some entries
1698 2 : let entries: BTreeMap<Key, Vec<Lsn>> = BTreeMap::from([
1699 2 : (base_key, vec![Lsn(1), Lsn(5), Lsn(25), Lsn(26), Lsn(28)]),
1700 2 : (base_key.add(1), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1701 2 : (base_key.add(2), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1702 2 : (base_key.add(5), vec![Lsn(10), Lsn(15), Lsn(16), Lsn(20)]),
1703 2 : ]);
1704 2 :
1705 2 : let mut disk = TestDisk::default();
1706 2 : let mut writer = DiskBtreeBuilder::<_, DELTA_KEY_SIZE>::new(&mut disk);
1707 2 :
1708 2 : let mut disk_offset = 0;
1709 10 : for (key, lsns) in &entries {
1710 42 : for lsn in lsns {
1711 34 : let index_key = DeltaKey::from_key_lsn(key, *lsn);
1712 34 : let blob_ref = BlobRef::new(disk_offset, false);
1713 34 : writer
1714 34 : .append(&index_key.0, blob_ref.0)
1715 34 : .expect("In memory disk append should never fail");
1716 34 :
1717 34 : disk_offset += 1;
1718 34 : }
1719 2 : }
1720 2 :
1721 2 : // Prepare all the arguments for the call into `plan_reads` below
1722 2 : let (root_offset, _writer) = writer
1723 2 : .finish()
1724 2 : .expect("In memory disk finish should never fail");
1725 2 : let reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(0, root_offset, disk);
1726 2 : let planner = VectoredReadPlanner::new(100);
1727 2 : let mut reconstruct_state = ValuesReconstructState::new();
1728 2 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1729 2 :
1730 2 : let keyspace = KeySpace {
1731 2 : ranges: vec![
1732 2 : base_key..base_key.add(3),
1733 2 : base_key.add(3)..base_key.add(100),
1734 2 : ],
1735 2 : };
1736 2 : let lsn_range = Lsn(2)..Lsn(40);
1737 2 :
1738 2 : // Plan and validate
1739 2 : let vectored_reads = DeltaLayerInner::plan_reads(
1740 2 : &keyspace,
1741 2 : lsn_range.clone(),
1742 2 : disk_offset,
1743 2 : reader,
1744 2 : planner,
1745 2 : &mut reconstruct_state,
1746 2 : &ctx,
1747 2 : )
1748 2 : .await
1749 2 : .expect("Read planning should not fail");
1750 2 :
1751 2 : validate(keyspace, lsn_range, vectored_reads, entries);
1752 2 : }
1753 :
1754 2 : fn validate(
1755 2 : keyspace: KeySpace,
1756 2 : lsn_range: Range<Lsn>,
1757 2 : vectored_reads: Vec<VectoredRead>,
1758 2 : index_entries: BTreeMap<Key, Vec<Lsn>>,
1759 2 : ) {
1760 2 : #[derive(Debug, PartialEq, Eq)]
1761 2 : struct BlobSpec {
1762 2 : key: Key,
1763 2 : lsn: Lsn,
1764 2 : at: u64,
1765 2 : }
1766 2 :
1767 2 : let mut planned_blobs = Vec::new();
1768 8 : for read in vectored_reads {
1769 28 : for (at, meta) in read.blobs_at.as_slice() {
1770 28 : planned_blobs.push(BlobSpec {
1771 28 : key: meta.key,
1772 28 : lsn: meta.lsn,
1773 28 : at: *at,
1774 28 : });
1775 28 : }
1776 : }
1777 :
1778 2 : let mut expected_blobs = Vec::new();
1779 2 : let mut disk_offset = 0;
1780 10 : for (key, lsns) in index_entries {
1781 42 : for lsn in lsns {
1782 42 : let key_included = keyspace.ranges.iter().any(|range| range.contains(&key));
1783 34 : let lsn_included = lsn_range.contains(&lsn);
1784 34 :
1785 34 : if key_included && lsn_included {
1786 28 : expected_blobs.push(BlobSpec {
1787 28 : key,
1788 28 : lsn,
1789 28 : at: disk_offset,
1790 28 : });
1791 28 : }
1792 :
1793 34 : disk_offset += 1;
1794 : }
1795 : }
1796 :
1797 2 : assert_eq!(planned_blobs, expected_blobs);
1798 2 : }
1799 :
1800 : mod constants {
1801 : use utils::lsn::Lsn;
1802 :
1803 : /// Offset used by all lsns in this test
1804 : pub(super) const LSN_OFFSET: Lsn = Lsn(0x08);
1805 : /// Number of unique keys including in the test data
1806 : pub(super) const KEY_COUNT: u8 = 60;
1807 : /// Max number of different lsns for each key
1808 : pub(super) const MAX_ENTRIES_PER_KEY: u8 = 20;
1809 : /// Possible value sizes for each key along with a probability weight
1810 : pub(super) const VALUE_SIZES: [(usize, u8); 3] = [(100, 2), (1024, 2), (1024 * 1024, 1)];
1811 : /// Probability that there will be a gap between the current key and the next one (33.3%)
1812 : pub(super) const KEY_GAP_CHANGES: [(bool, u8); 2] = [(true, 1), (false, 2)];
1813 : /// The minimum size of a key range in all the generated reads
1814 : pub(super) const MIN_RANGE_SIZE: i128 = 10;
1815 : /// The number of ranges included in each vectored read
1816 : pub(super) const RANGES_COUNT: u8 = 2;
1817 : /// The number of vectored reads performed
1818 : pub(super) const READS_COUNT: u8 = 100;
1819 : /// Soft max size of a vectored read. Will be violated if we have to read keys
1820 : /// with values larger than the limit
1821 : pub(super) const MAX_VECTORED_READ_BYTES: usize = 64 * 1024;
1822 : }
1823 :
1824 : struct Entry {
1825 : key: Key,
1826 : lsn: Lsn,
1827 : value: Vec<u8>,
1828 : }
1829 :
1830 2 : fn generate_entries(rng: &mut StdRng) -> Vec<Entry> {
1831 2 : let mut current_key = Key::MIN;
1832 2 :
1833 2 : let mut entries = Vec::new();
1834 122 : for _ in 0..constants::KEY_COUNT {
1835 120 : let count = rng.gen_range(1..constants::MAX_ENTRIES_PER_KEY);
1836 120 : let mut lsns_iter =
1837 2260 : std::iter::successors(Some(Lsn(constants::LSN_OFFSET.0 + 0x08)), |lsn| {
1838 2260 : Some(Lsn(lsn.0 + 0x08))
1839 2260 : });
1840 120 : let mut lsns = Vec::new();
1841 2380 : while lsns.len() < count as usize {
1842 2260 : let take = rng.gen_bool(0.5);
1843 2260 : let lsn = lsns_iter.next().unwrap();
1844 2260 : if take {
1845 1112 : lsns.push(lsn);
1846 1148 : }
1847 : }
1848 :
1849 1232 : for lsn in lsns {
1850 1112 : let size = constants::VALUE_SIZES
1851 3336 : .choose_weighted(rng, |item| item.1)
1852 1112 : .unwrap()
1853 1112 : .0;
1854 1112 : let mut buf = vec![0; size];
1855 1112 : rng.fill_bytes(&mut buf);
1856 1112 :
1857 1112 : entries.push(Entry {
1858 1112 : key: current_key,
1859 1112 : lsn,
1860 1112 : value: buf,
1861 1112 : })
1862 : }
1863 :
1864 120 : let gap = constants::KEY_GAP_CHANGES
1865 240 : .choose_weighted(rng, |item| item.1)
1866 120 : .unwrap()
1867 120 : .0;
1868 120 : if gap {
1869 38 : current_key = current_key.add(2);
1870 82 : } else {
1871 82 : current_key = current_key.add(1);
1872 82 : }
1873 : }
1874 :
1875 2 : entries
1876 2 : }
1877 :
1878 : struct EntriesMeta {
1879 : key_range: Range<Key>,
1880 : lsn_range: Range<Lsn>,
1881 : index: BTreeMap<(Key, Lsn), Vec<u8>>,
1882 : }
1883 :
1884 2 : fn get_entries_meta(entries: &[Entry]) -> EntriesMeta {
1885 1112 : let key_range = match entries.iter().minmax_by_key(|e| e.key) {
1886 2 : MinMaxResult::MinMax(min, max) => min.key..max.key.next(),
1887 0 : _ => panic!("More than one entry is always expected"),
1888 : };
1889 :
1890 1112 : let lsn_range = match entries.iter().minmax_by_key(|e| e.lsn) {
1891 2 : MinMaxResult::MinMax(min, max) => min.lsn..Lsn(max.lsn.0 + 1),
1892 0 : _ => panic!("More than one entry is always expected"),
1893 : };
1894 :
1895 2 : let mut index = BTreeMap::new();
1896 1112 : for entry in entries.iter() {
1897 1112 : index.insert((entry.key, entry.lsn), entry.value.clone());
1898 1112 : }
1899 :
1900 2 : EntriesMeta {
1901 2 : key_range,
1902 2 : lsn_range,
1903 2 : index,
1904 2 : }
1905 2 : }
1906 :
1907 200 : fn pick_random_keyspace(rng: &mut StdRng, key_range: &Range<Key>) -> KeySpace {
1908 200 : let start = key_range.start.to_i128();
1909 200 : let end = key_range.end.to_i128();
1910 200 :
1911 200 : let mut keyspace = KeySpace::default();
1912 :
1913 600 : for _ in 0..constants::RANGES_COUNT {
1914 400 : let mut range: Option<Range<Key>> = Option::default();
1915 1244 : while range.is_none() || keyspace.overlaps(range.as_ref().unwrap()) {
1916 844 : let range_start = rng.gen_range(start..end);
1917 844 : let range_end_offset = range_start + constants::MIN_RANGE_SIZE;
1918 844 : if range_end_offset >= end {
1919 100 : range = Some(Key::from_i128(range_start)..Key::from_i128(end));
1920 744 : } else {
1921 744 : let range_end = rng.gen_range((range_start + constants::MIN_RANGE_SIZE)..end);
1922 744 : range = Some(Key::from_i128(range_start)..Key::from_i128(range_end));
1923 744 : }
1924 : }
1925 400 : keyspace.ranges.push(range.unwrap());
1926 : }
1927 :
1928 200 : keyspace
1929 200 : }
1930 :
1931 : #[tokio::test]
1932 2 : async fn test_delta_layer_vectored_read_end_to_end() -> anyhow::Result<()> {
1933 2 : let harness = TenantHarness::create("test_delta_layer_oversized_vectored_read").await?;
1934 8 : let (tenant, ctx) = harness.load().await;
1935 2 :
1936 2 : let timeline_id = TimelineId::generate();
1937 2 : let timeline = tenant
1938 2 : .create_test_timeline(timeline_id, constants::LSN_OFFSET, DEFAULT_PG_VERSION, &ctx)
1939 4 : .await?;
1940 2 :
1941 2 : tracing::info!("Generating test data ...");
1942 2 :
1943 2 : let rng = &mut StdRng::seed_from_u64(0);
1944 2 : let entries = generate_entries(rng);
1945 2 : let entries_meta = get_entries_meta(&entries);
1946 2 :
1947 2 : tracing::info!("Done generating {} entries", entries.len());
1948 2 :
1949 2 : tracing::info!("Writing test data to delta layer ...");
1950 2 : let mut writer = DeltaLayerWriter::new(
1951 2 : harness.conf,
1952 2 : timeline_id,
1953 2 : harness.tenant_shard_id,
1954 2 : entries_meta.key_range.start,
1955 2 : entries_meta.lsn_range.clone(),
1956 2 : &ctx,
1957 2 : )
1958 2 : .await?;
1959 2 :
1960 1114 : for entry in entries {
1961 1112 : let (_, res) = writer
1962 1112 : .put_value_bytes(entry.key, entry.lsn, entry.value, false, &ctx)
1963 215 : .await;
1964 1112 : res?;
1965 2 : }
1966 2 :
1967 2 : let resident = writer
1968 2 : .finish(entries_meta.key_range.end, &timeline, &ctx)
1969 5 : .await?;
1970 2 :
1971 2 : let inner = resident.get_as_delta(&ctx).await?;
1972 2 :
1973 2 : let file_size = inner.file.metadata().await?.len();
1974 2 : tracing::info!(
1975 2 : "Done writing test data to delta layer. Resulting file size is: {}",
1976 2 : file_size
1977 2 : );
1978 2 :
1979 202 : for i in 0..constants::READS_COUNT {
1980 200 : tracing::info!("Doing vectored read {}/{}", i + 1, constants::READS_COUNT);
1981 2 :
1982 200 : let block_reader = FileBlockReader::new(&inner.file, inner.file_id);
1983 200 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1984 200 : inner.index_start_blk,
1985 200 : inner.index_root_blk,
1986 200 : block_reader,
1987 200 : );
1988 200 :
1989 200 : let planner = VectoredReadPlanner::new(constants::MAX_VECTORED_READ_BYTES);
1990 200 : let mut reconstruct_state = ValuesReconstructState::new();
1991 200 : let keyspace = pick_random_keyspace(rng, &entries_meta.key_range);
1992 200 : let data_end_offset = inner.index_start_blk as u64 * PAGE_SZ as u64;
1993 2 :
1994 200 : let vectored_reads = DeltaLayerInner::plan_reads(
1995 200 : &keyspace,
1996 200 : entries_meta.lsn_range.clone(),
1997 200 : data_end_offset,
1998 200 : index_reader,
1999 200 : planner,
2000 200 : &mut reconstruct_state,
2001 200 : &ctx,
2002 200 : )
2003 4 : .await?;
2004 2 :
2005 200 : let vectored_blob_reader = VectoredBlobReader::new(&inner.file);
2006 200 : let buf_size = DeltaLayerInner::get_min_read_buffer_size(
2007 200 : &vectored_reads,
2008 200 : constants::MAX_VECTORED_READ_BYTES,
2009 200 : );
2010 200 : let mut buf = Some(BytesMut::with_capacity(buf_size));
2011 2 :
2012 19924 : for read in vectored_reads {
2013 19724 : let blobs_buf = vectored_blob_reader
2014 19724 : .read_blobs(&read, buf.take().expect("Should have a buffer"), &ctx)
2015 10016 : .await?;
2016 57304 : for meta in blobs_buf.blobs.iter() {
2017 57304 : let value = &blobs_buf.buf[meta.start..meta.end];
2018 57304 : assert_eq!(value, entries_meta.index[&(meta.meta.key, meta.meta.lsn)]);
2019 2 : }
2020 2 :
2021 19724 : buf = Some(blobs_buf.buf);
2022 2 : }
2023 2 : }
2024 2 :
2025 2 : Ok(())
2026 2 : }
2027 :
2028 : #[tokio::test]
2029 2 : async fn copy_delta_prefix_smoke() {
2030 2 : use crate::walrecord::NeonWalRecord;
2031 2 : use bytes::Bytes;
2032 2 :
2033 2 : let h = crate::tenant::harness::TenantHarness::create("truncate_delta_smoke")
2034 2 : .await
2035 2 : .unwrap();
2036 8 : let (tenant, ctx) = h.load().await;
2037 2 : let ctx = &ctx;
2038 2 : let timeline = tenant
2039 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, ctx)
2040 4 : .await
2041 2 : .unwrap();
2042 2 :
2043 2 : let initdb_layer = timeline
2044 2 : .layers
2045 2 : .read()
2046 2 : .await
2047 2 : .likely_resident_layers()
2048 2 : .next()
2049 2 : .unwrap();
2050 2 :
2051 2 : {
2052 2 : let mut writer = timeline.writer().await;
2053 2 :
2054 2 : let data = [
2055 2 : (0x20, 12, Value::Image(Bytes::from_static(b"foobar"))),
2056 2 : (
2057 2 : 0x30,
2058 2 : 12,
2059 2 : Value::WalRecord(NeonWalRecord::Postgres {
2060 2 : will_init: false,
2061 2 : rec: Bytes::from_static(b"1"),
2062 2 : }),
2063 2 : ),
2064 2 : (
2065 2 : 0x40,
2066 2 : 12,
2067 2 : Value::WalRecord(NeonWalRecord::Postgres {
2068 2 : will_init: true,
2069 2 : rec: Bytes::from_static(b"2"),
2070 2 : }),
2071 2 : ),
2072 2 : // build an oversized value so we cannot extend and existing read over
2073 2 : // this
2074 2 : (
2075 2 : 0x50,
2076 2 : 12,
2077 2 : Value::WalRecord(NeonWalRecord::Postgres {
2078 2 : will_init: true,
2079 2 : rec: {
2080 2 : let mut buf =
2081 2 : vec![0u8; tenant.conf.max_vectored_read_bytes.0.get() + 1024];
2082 2 : buf.iter_mut()
2083 2 : .enumerate()
2084 264192 : .for_each(|(i, slot)| *slot = (i % 256) as u8);
2085 2 : Bytes::from(buf)
2086 2 : },
2087 2 : }),
2088 2 : ),
2089 2 : // because the oversized read cannot be extended further, we are sure to exercise the
2090 2 : // builder created on the last round with this:
2091 2 : (
2092 2 : 0x60,
2093 2 : 12,
2094 2 : Value::WalRecord(NeonWalRecord::Postgres {
2095 2 : will_init: true,
2096 2 : rec: Bytes::from_static(b"3"),
2097 2 : }),
2098 2 : ),
2099 2 : (
2100 2 : 0x60,
2101 2 : 9,
2102 2 : Value::Image(Bytes::from_static(b"something for a different key")),
2103 2 : ),
2104 2 : ];
2105 2 :
2106 2 : let mut last_lsn = None;
2107 2 :
2108 14 : for (lsn, key, value) in data {
2109 12 : let key = Key::from_i128(key);
2110 12 : writer.put(key, Lsn(lsn), &value, ctx).await.unwrap();
2111 12 : last_lsn = Some(lsn);
2112 2 : }
2113 2 :
2114 2 : writer.finish_write(Lsn(last_lsn.unwrap()));
2115 2 : }
2116 2 : timeline.freeze_and_flush().await.unwrap();
2117 2 :
2118 2 : let new_layer = timeline
2119 2 : .layers
2120 2 : .read()
2121 2 : .await
2122 2 : .likely_resident_layers()
2123 4 : .find(|x| x != &initdb_layer)
2124 2 : .unwrap();
2125 2 :
2126 2 : // create a copy for the timeline, so we don't overwrite the file
2127 2 : let branch = tenant
2128 2 : .branch_timeline_test(&timeline, TimelineId::generate(), None, ctx)
2129 2 : .await
2130 2 : .unwrap();
2131 2 :
2132 2 : assert_eq!(branch.get_ancestor_lsn(), Lsn(0x60));
2133 2 :
2134 2 : // truncating at 0x61 gives us a full copy, otherwise just go backwards until there's just
2135 2 : // a single key
2136 2 :
2137 12 : for truncate_at in [0x61, 0x51, 0x41, 0x31, 0x21] {
2138 10 : let truncate_at = Lsn(truncate_at);
2139 2 :
2140 10 : let mut writer = DeltaLayerWriter::new(
2141 10 : tenant.conf,
2142 10 : branch.timeline_id,
2143 10 : tenant.tenant_shard_id,
2144 10 : Key::MIN,
2145 10 : Lsn(0x11)..truncate_at,
2146 10 : ctx,
2147 10 : )
2148 5 : .await
2149 10 : .unwrap();
2150 2 :
2151 10 : let new_layer = new_layer.download_and_keep_resident().await.unwrap();
2152 10 :
2153 10 : new_layer
2154 10 : .copy_delta_prefix(&mut writer, truncate_at, ctx)
2155 15 : .await
2156 10 : .unwrap();
2157 2 :
2158 24 : let copied_layer = writer.finish(Key::MAX, &branch, ctx).await.unwrap();
2159 10 :
2160 10 : copied_layer.get_as_delta(ctx).await.unwrap();
2161 10 :
2162 10 : assert_keys_and_values_eq(
2163 10 : new_layer.get_as_delta(ctx).await.unwrap(),
2164 10 : copied_layer.get_as_delta(ctx).await.unwrap(),
2165 10 : truncate_at,
2166 10 : ctx,
2167 2 : )
2168 47 : .await;
2169 2 : }
2170 2 : }
2171 :
2172 10 : async fn assert_keys_and_values_eq(
2173 10 : source: &DeltaLayerInner,
2174 10 : truncated: &DeltaLayerInner,
2175 10 : truncated_at: Lsn,
2176 10 : ctx: &RequestContext,
2177 10 : ) {
2178 10 : use futures::future::ready;
2179 10 : use futures::stream::TryStreamExt;
2180 10 :
2181 10 : let start_key = [0u8; DELTA_KEY_SIZE];
2182 10 :
2183 10 : let source_reader = FileBlockReader::new(&source.file, source.file_id);
2184 10 : let source_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2185 10 : source.index_start_blk,
2186 10 : source.index_root_blk,
2187 10 : &source_reader,
2188 10 : );
2189 10 : let source_stream = source.stream_index_forwards(source_tree, &start_key, ctx);
2190 60 : let source_stream = source_stream.filter(|res| match res {
2191 60 : Ok((_, lsn, _)) => ready(lsn < &truncated_at),
2192 0 : _ => ready(true),
2193 60 : });
2194 10 : let mut source_stream = std::pin::pin!(source_stream);
2195 10 :
2196 10 : let truncated_reader = FileBlockReader::new(&truncated.file, truncated.file_id);
2197 10 : let truncated_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2198 10 : truncated.index_start_blk,
2199 10 : truncated.index_root_blk,
2200 10 : &truncated_reader,
2201 10 : );
2202 10 : let truncated_stream = truncated.stream_index_forwards(truncated_tree, &start_key, ctx);
2203 10 : let mut truncated_stream = std::pin::pin!(truncated_stream);
2204 10 :
2205 10 : let mut scratch_left = Vec::new();
2206 10 : let mut scratch_right = Vec::new();
2207 :
2208 : loop {
2209 42 : let (src, truncated) = (source_stream.try_next(), truncated_stream.try_next());
2210 42 : let (src, truncated) = tokio::try_join!(src, truncated).unwrap();
2211 42 :
2212 42 : if src.is_none() {
2213 10 : assert!(truncated.is_none());
2214 10 : break;
2215 32 : }
2216 32 :
2217 32 : let (src, truncated) = (src.unwrap(), truncated.unwrap());
2218 32 :
2219 32 : // because we've filtered the source with Lsn, we should always have the same keys from both.
2220 32 : assert_eq!(src.0, truncated.0);
2221 32 : assert_eq!(src.1, truncated.1);
2222 :
2223 : // if this is needed for something else, just drop this assert.
2224 32 : assert!(
2225 32 : src.2.pos() >= truncated.2.pos(),
2226 0 : "value position should not go backwards {} vs. {}",
2227 0 : src.2.pos(),
2228 0 : truncated.2.pos()
2229 : );
2230 :
2231 32 : scratch_left.clear();
2232 32 : let src_cursor = source_reader.block_cursor();
2233 32 : let left = src_cursor.read_blob_into_buf(src.2.pos(), &mut scratch_left, ctx);
2234 32 : scratch_right.clear();
2235 32 : let trunc_cursor = truncated_reader.block_cursor();
2236 32 : let right = trunc_cursor.read_blob_into_buf(truncated.2.pos(), &mut scratch_right, ctx);
2237 :
2238 32 : tokio::try_join!(left, right).unwrap();
2239 32 :
2240 32 : assert_eq!(utils::Hex(&scratch_left), utils::Hex(&scratch_right));
2241 : }
2242 10 : }
2243 :
2244 18026 : pub(crate) fn sort_delta(
2245 18026 : (k1, l1, _): &(Key, Lsn, Value),
2246 18026 : (k2, l2, _): &(Key, Lsn, Value),
2247 18026 : ) -> std::cmp::Ordering {
2248 18026 : (k1, l1).cmp(&(k2, l2))
2249 18026 : }
2250 :
2251 94 : pub(crate) fn sort_delta_value(
2252 94 : (k1, l1, v1): &(Key, Lsn, Value),
2253 94 : (k2, l2, v2): &(Key, Lsn, Value),
2254 94 : ) -> std::cmp::Ordering {
2255 94 : let order_1 = if v1.is_image() { 0 } else { 1 };
2256 94 : let order_2 = if v2.is_image() { 0 } else { 1 };
2257 94 : (k1, l1, order_1).cmp(&(k2, l2, order_2))
2258 94 : }
2259 :
2260 20 : pub(crate) async fn produce_delta_layer(
2261 20 : tenant: &Tenant,
2262 20 : tline: &Arc<Timeline>,
2263 20 : mut deltas: Vec<(Key, Lsn, Value)>,
2264 20 : ctx: &RequestContext,
2265 20 : ) -> anyhow::Result<ResidentLayer> {
2266 20 : deltas.sort_by(sort_delta);
2267 20 : let (key_start, _, _) = deltas.first().unwrap();
2268 20 : let (key_max, _, _) = deltas.last().unwrap();
2269 8040 : let lsn_min = deltas.iter().map(|(_, lsn, _)| lsn).min().unwrap();
2270 8040 : let lsn_max = deltas.iter().map(|(_, lsn, _)| lsn).max().unwrap();
2271 20 : let lsn_end = Lsn(lsn_max.0 + 1);
2272 20 : let mut writer = DeltaLayerWriter::new(
2273 20 : tenant.conf,
2274 20 : tline.timeline_id,
2275 20 : tenant.tenant_shard_id,
2276 20 : *key_start,
2277 20 : (*lsn_min)..lsn_end,
2278 20 : ctx,
2279 20 : )
2280 10 : .await?;
2281 20 : let key_end = key_max.next();
2282 :
2283 8060 : for (key, lsn, value) in deltas {
2284 8040 : writer.put_value(key, lsn, value, ctx).await?;
2285 : }
2286 58 : let delta_layer = writer.finish(key_end, tline, ctx).await?;
2287 :
2288 20 : Ok::<_, anyhow::Error>(delta_layer)
2289 20 : }
2290 :
2291 28 : async fn assert_delta_iter_equal(
2292 28 : delta_iter: &mut DeltaLayerIterator<'_>,
2293 28 : expect: &[(Key, Lsn, Value)],
2294 28 : ) {
2295 28 : let mut expect_iter = expect.iter();
2296 : loop {
2297 28028 : let o1 = delta_iter.next().await.unwrap();
2298 28028 : let o2 = expect_iter.next();
2299 28028 : assert_eq!(o1.is_some(), o2.is_some());
2300 28028 : if o1.is_none() && o2.is_none() {
2301 28 : break;
2302 28000 : }
2303 28000 : let (k1, l1, v1) = o1.unwrap();
2304 28000 : let (k2, l2, v2) = o2.unwrap();
2305 28000 : assert_eq!(&k1, k2);
2306 28000 : assert_eq!(l1, *l2);
2307 28000 : assert_eq!(&v1, v2);
2308 : }
2309 28 : }
2310 :
2311 : #[tokio::test]
2312 2 : async fn delta_layer_iterator() {
2313 2 : let harness = TenantHarness::create("delta_layer_iterator").await.unwrap();
2314 8 : let (tenant, ctx) = harness.load().await;
2315 2 :
2316 2 : let tline = tenant
2317 2 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
2318 4 : .await
2319 2 : .unwrap();
2320 2 :
2321 2000 : fn get_key(id: u32) -> Key {
2322 2000 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
2323 2000 : key.field6 = id;
2324 2000 : key
2325 2000 : }
2326 2 : const N: usize = 1000;
2327 2 : let test_deltas = (0..N)
2328 2000 : .map(|idx| {
2329 2000 : (
2330 2000 : get_key(idx as u32 / 10),
2331 2000 : Lsn(0x10 * ((idx as u64) % 10 + 1)),
2332 2000 : Value::Image(Bytes::from(format!("img{idx:05}"))),
2333 2000 : )
2334 2000 : })
2335 2 : .collect_vec();
2336 2 : let resident_layer = produce_delta_layer(&tenant, &tline, test_deltas.clone(), &ctx)
2337 8 : .await
2338 2 : .unwrap();
2339 2 : let delta_layer = resident_layer.get_as_delta(&ctx).await.unwrap();
2340 6 : for max_read_size in [1, 1024] {
2341 32 : for batch_size in [1, 2, 4, 8, 3, 7, 13] {
2342 28 : println!("running with batch_size={batch_size} max_read_size={max_read_size}");
2343 28 : // Test if the batch size is correctly determined
2344 28 : let mut iter = delta_layer.iter(&ctx);
2345 28 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2346 28 : let mut num_items = 0;
2347 112 : for _ in 0..3 {
2348 84 : iter.next_batch().await.unwrap();
2349 84 : num_items += iter.key_values_batch.len();
2350 84 : if max_read_size == 1 {
2351 2 : // every key should be a batch b/c the value is larger than max_read_size
2352 42 : assert_eq!(iter.key_values_batch.len(), 1);
2353 2 : } else {
2354 42 : assert_eq!(iter.key_values_batch.len(), batch_size);
2355 2 : }
2356 84 : if num_items >= N {
2357 2 : break;
2358 84 : }
2359 84 : iter.key_values_batch.clear();
2360 2 : }
2361 2 : // Test if the result is correct
2362 28 : let mut iter = delta_layer.iter(&ctx);
2363 28 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2364 9577 : assert_delta_iter_equal(&mut iter, &test_deltas).await;
2365 2 : }
2366 2 : }
2367 2 : }
2368 : }
|
