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 std::collections::{HashMap, VecDeque};
31 : use std::fs::File;
32 : use std::io::SeekFrom;
33 : use std::ops::Range;
34 : use std::os::unix::fs::FileExt;
35 : use std::str::FromStr;
36 : use std::sync::Arc;
37 :
38 : use anyhow::{Context, Result, bail, ensure};
39 : use camino::{Utf8Path, Utf8PathBuf};
40 : use futures::StreamExt;
41 : use itertools::Itertools;
42 : use pageserver_api::config::MaxVectoredReadBytes;
43 : use pageserver_api::key::{DBDIR_KEY, KEY_SIZE, Key};
44 : use pageserver_api::keyspace::KeySpace;
45 : use pageserver_api::models::ImageCompressionAlgorithm;
46 : use pageserver_api::shard::TenantShardId;
47 : use pageserver_api::value::Value;
48 : use rand::Rng;
49 : use rand::distributions::Alphanumeric;
50 : use serde::{Deserialize, Serialize};
51 : use tokio::sync::OnceCell;
52 : use tokio_epoll_uring::IoBuf;
53 : use tokio_util::sync::CancellationToken;
54 : use tracing::*;
55 : use utils::bin_ser::BeSer;
56 : use utils::id::{TenantId, TimelineId};
57 : use utils::lsn::Lsn;
58 :
59 : use super::{
60 : AsLayerDesc, LayerName, OnDiskValue, OnDiskValueIo, PersistentLayerDesc, ResidentLayer,
61 : ValuesReconstructState,
62 : };
63 : use crate::config::PageServerConf;
64 : use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
65 : use crate::page_cache::{self, FileId, PAGE_SZ};
66 : use crate::tenant::blob_io::BlobWriter;
67 : use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockLease, BlockReader, FileBlockReader};
68 : use crate::tenant::disk_btree::{
69 : DiskBtreeBuilder, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
70 : };
71 : use crate::tenant::storage_layer::layer::S3_UPLOAD_LIMIT;
72 : use crate::tenant::timeline::GetVectoredError;
73 : use crate::tenant::vectored_blob_io::{
74 : BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
75 : VectoredReadPlanner,
76 : };
77 : use crate::virtual_file::owned_buffers_io::io_buf_ext::{FullSlice, IoBufExt};
78 : use crate::virtual_file::{self, IoBufferMut, MaybeFatalIo, VirtualFile};
79 : use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
80 :
81 : ///
82 : /// Header stored in the beginning of the file
83 : ///
84 : /// After this comes the 'values' part, starting on block 1. After that,
85 : /// the 'index' starts at the block indicated by 'index_start_blk'
86 : ///
87 0 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
88 : pub struct Summary {
89 : /// Magic value to identify this as a neon delta file. Always DELTA_FILE_MAGIC.
90 : pub magic: u16,
91 : pub format_version: u16,
92 :
93 : pub tenant_id: TenantId,
94 : pub timeline_id: TimelineId,
95 : pub key_range: Range<Key>,
96 : pub lsn_range: Range<Lsn>,
97 :
98 : /// Block number where the 'index' part of the file begins.
99 : pub index_start_blk: u32,
100 : /// Block within the 'index', where the B-tree root page is stored
101 : pub index_root_blk: u32,
102 : }
103 :
104 : impl From<&DeltaLayer> for Summary {
105 0 : fn from(layer: &DeltaLayer) -> Self {
106 0 : Self::expected(
107 0 : layer.desc.tenant_shard_id.tenant_id,
108 0 : layer.desc.timeline_id,
109 0 : layer.desc.key_range.clone(),
110 0 : layer.desc.lsn_range.clone(),
111 0 : )
112 0 : }
113 : }
114 :
115 : impl Summary {
116 6612 : pub(super) fn expected(
117 6612 : tenant_id: TenantId,
118 6612 : timeline_id: TimelineId,
119 6612 : keys: Range<Key>,
120 6612 : lsns: Range<Lsn>,
121 6612 : ) -> Self {
122 6612 : Self {
123 6612 : magic: DELTA_FILE_MAGIC,
124 6612 : format_version: STORAGE_FORMAT_VERSION,
125 6612 :
126 6612 : tenant_id,
127 6612 : timeline_id,
128 6612 : key_range: keys,
129 6612 : lsn_range: lsns,
130 6612 :
131 6612 : index_start_blk: 0,
132 6612 : index_root_blk: 0,
133 6612 : }
134 6612 : }
135 : }
136 :
137 : // Flag indicating that this version initialize the page
138 : const WILL_INIT: u64 = 1;
139 :
140 : /// Struct representing reference to BLOB in layers.
141 : ///
142 : /// Reference contains BLOB offset, and for WAL records it also contains
143 : /// `will_init` flag. The flag helps to determine the range of records
144 : /// that needs to be applied, without reading/deserializing records themselves.
145 0 : #[derive(Debug, Serialize, Deserialize, Copy, Clone)]
146 : pub struct BlobRef(pub u64);
147 :
148 : impl BlobRef {
149 28828666 : pub fn will_init(&self) -> bool {
150 28828666 : (self.0 & WILL_INIT) != 0
151 28828666 : }
152 :
153 46434963 : pub fn pos(&self) -> u64 {
154 46434963 : self.0 >> 1
155 46434963 : }
156 :
157 38952492 : pub fn new(pos: u64, will_init: bool) -> BlobRef {
158 38952492 : let mut blob_ref = pos << 1;
159 38952492 : if will_init {
160 38820996 : blob_ref |= WILL_INIT;
161 38820996 : }
162 38952492 : BlobRef(blob_ref)
163 38952492 : }
164 : }
165 :
166 : pub const DELTA_KEY_SIZE: usize = KEY_SIZE + 8;
167 : struct DeltaKey([u8; DELTA_KEY_SIZE]);
168 :
169 : /// This is the key of the B-tree index stored in the delta layer. It consists
170 : /// of the serialized representation of a Key and LSN.
171 : impl DeltaKey {
172 12385188 : fn from_slice(buf: &[u8]) -> Self {
173 12385188 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
174 12385188 : bytes.copy_from_slice(buf);
175 12385188 : DeltaKey(bytes)
176 12385188 : }
177 :
178 40400324 : fn from_key_lsn(key: &Key, lsn: Lsn) -> Self {
179 40400324 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
180 40400324 : key.write_to_byte_slice(&mut bytes[0..KEY_SIZE]);
181 40400324 : bytes[KEY_SIZE..].copy_from_slice(&u64::to_be_bytes(lsn.0));
182 40400324 : DeltaKey(bytes)
183 40400324 : }
184 :
185 12385188 : fn key(&self) -> Key {
186 12385188 : Key::from_slice(&self.0)
187 12385188 : }
188 :
189 12385188 : fn lsn(&self) -> Lsn {
190 12385188 : Lsn(u64::from_be_bytes(self.0[KEY_SIZE..].try_into().unwrap()))
191 12385188 : }
192 :
193 34049415 : fn extract_lsn_from_buf(buf: &[u8]) -> Lsn {
194 34049415 : let mut lsn_buf = [0u8; 8];
195 34049415 : lsn_buf.copy_from_slice(&buf[KEY_SIZE..]);
196 34049415 : Lsn(u64::from_be_bytes(lsn_buf))
197 34049415 : }
198 : }
199 :
200 : /// This is used only from `pagectl`. Within pageserver, all layers are
201 : /// [`crate::tenant::storage_layer::Layer`], which can hold a [`DeltaLayerInner`].
202 : pub struct DeltaLayer {
203 : path: Utf8PathBuf,
204 : pub desc: PersistentLayerDesc,
205 : inner: OnceCell<Arc<DeltaLayerInner>>,
206 : }
207 :
208 : impl std::fmt::Debug for DeltaLayer {
209 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
210 : use super::RangeDisplayDebug;
211 :
212 0 : f.debug_struct("DeltaLayer")
213 0 : .field("key_range", &RangeDisplayDebug(&self.desc.key_range))
214 0 : .field("lsn_range", &self.desc.lsn_range)
215 0 : .field("file_size", &self.desc.file_size)
216 0 : .field("inner", &self.inner)
217 0 : .finish()
218 0 : }
219 : }
220 :
221 : /// `DeltaLayerInner` is the in-memory data structure associated with an on-disk delta
222 : /// file.
223 : pub struct DeltaLayerInner {
224 : // values copied from summary
225 : index_start_blk: u32,
226 : index_root_blk: u32,
227 :
228 : file: Arc<VirtualFile>,
229 : file_id: FileId,
230 :
231 : layer_key_range: Range<Key>,
232 : layer_lsn_range: Range<Lsn>,
233 :
234 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
235 : }
236 :
237 : impl DeltaLayerInner {
238 0 : pub(crate) fn layer_dbg_info(&self) -> String {
239 0 : format!(
240 0 : "delta {}..{} {}..{}",
241 0 : self.key_range().start,
242 0 : self.key_range().end,
243 0 : self.lsn_range().start,
244 0 : self.lsn_range().end
245 0 : )
246 0 : }
247 : }
248 :
249 : impl std::fmt::Debug for DeltaLayerInner {
250 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
251 0 : f.debug_struct("DeltaLayerInner")
252 0 : .field("index_start_blk", &self.index_start_blk)
253 0 : .field("index_root_blk", &self.index_root_blk)
254 0 : .finish()
255 0 : }
256 : }
257 :
258 : /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
259 : impl std::fmt::Display for DeltaLayer {
260 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
261 0 : write!(f, "{}", self.layer_desc().short_id())
262 0 : }
263 : }
264 :
265 : impl AsLayerDesc for DeltaLayer {
266 0 : fn layer_desc(&self) -> &PersistentLayerDesc {
267 0 : &self.desc
268 0 : }
269 : }
270 :
271 : impl DeltaLayer {
272 0 : pub async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
273 0 : self.desc.dump();
274 0 :
275 0 : if !verbose {
276 0 : return Ok(());
277 0 : }
278 :
279 0 : let inner = self.load(ctx).await?;
280 :
281 0 : inner.dump(ctx).await
282 0 : }
283 :
284 8856 : fn temp_path_for(
285 8856 : conf: &PageServerConf,
286 8856 : tenant_shard_id: &TenantShardId,
287 8856 : timeline_id: &TimelineId,
288 8856 : key_start: Key,
289 8856 : lsn_range: &Range<Lsn>,
290 8856 : ) -> Utf8PathBuf {
291 8856 : let rand_string: String = rand::thread_rng()
292 8856 : .sample_iter(&Alphanumeric)
293 8856 : .take(8)
294 8856 : .map(char::from)
295 8856 : .collect();
296 8856 :
297 8856 : conf.timeline_path(tenant_shard_id, timeline_id)
298 8856 : .join(format!(
299 8856 : "{}-XXX__{:016X}-{:016X}.{}.{}",
300 8856 : key_start,
301 8856 : u64::from(lsn_range.start),
302 8856 : u64::from(lsn_range.end),
303 8856 : rand_string,
304 8856 : TEMP_FILE_SUFFIX,
305 8856 : ))
306 8856 : }
307 :
308 : ///
309 : /// Open the underlying file and read the metadata into memory, if it's
310 : /// not loaded already.
311 : ///
312 0 : async fn load(&self, ctx: &RequestContext) -> Result<&Arc<DeltaLayerInner>> {
313 0 : // Quick exit if already loaded
314 0 : self.inner
315 0 : .get_or_try_init(|| self.load_inner(ctx))
316 0 : .await
317 0 : .with_context(|| format!("Failed to load delta layer {}", self.path()))
318 0 : }
319 :
320 0 : async fn load_inner(&self, ctx: &RequestContext) -> anyhow::Result<Arc<DeltaLayerInner>> {
321 0 : let path = self.path();
322 :
323 0 : let loaded = DeltaLayerInner::load(&path, None, None, ctx).await?;
324 :
325 : // not production code
326 0 : let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
327 0 : let expected_layer_name = self.layer_desc().layer_name();
328 0 :
329 0 : if actual_layer_name != expected_layer_name {
330 0 : println!("warning: filename does not match what is expected from in-file summary");
331 0 : println!("actual: {:?}", actual_layer_name.to_string());
332 0 : println!("expected: {:?}", expected_layer_name.to_string());
333 0 : }
334 :
335 0 : Ok(Arc::new(loaded))
336 0 : }
337 :
338 : /// Create a DeltaLayer struct representing an existing file on disk.
339 : ///
340 : /// This variant is only used for debugging purposes, by the 'pagectl' binary.
341 0 : pub fn new_for_path(path: &Utf8Path, file: File) -> Result<Self> {
342 0 : let mut summary_buf = vec![0; PAGE_SZ];
343 0 : file.read_exact_at(&mut summary_buf, 0)?;
344 0 : let summary = Summary::des_prefix(&summary_buf)?;
345 :
346 0 : let metadata = file
347 0 : .metadata()
348 0 : .context("get file metadata to determine size")?;
349 :
350 : // This function is never used for constructing layers in a running pageserver,
351 : // so it does not need an accurate TenantShardId.
352 0 : let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
353 0 :
354 0 : Ok(DeltaLayer {
355 0 : path: path.to_path_buf(),
356 0 : desc: PersistentLayerDesc::new_delta(
357 0 : tenant_shard_id,
358 0 : summary.timeline_id,
359 0 : summary.key_range,
360 0 : summary.lsn_range,
361 0 : metadata.len(),
362 0 : ),
363 0 : inner: OnceCell::new(),
364 0 : })
365 0 : }
366 :
367 : /// Path to the layer file in pageserver workdir.
368 0 : fn path(&self) -> Utf8PathBuf {
369 0 : self.path.clone()
370 0 : }
371 : }
372 :
373 : /// A builder object for constructing a new delta layer.
374 : ///
375 : /// Usage:
376 : ///
377 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
378 : ///
379 : /// 2. Write the contents by calling `put_value` for every page
380 : /// version to store in the layer.
381 : ///
382 : /// 3. Call `finish`.
383 : ///
384 : struct DeltaLayerWriterInner {
385 : pub path: Utf8PathBuf,
386 : timeline_id: TimelineId,
387 : tenant_shard_id: TenantShardId,
388 :
389 : key_start: Key,
390 : lsn_range: Range<Lsn>,
391 :
392 : tree: DiskBtreeBuilder<BlockBuf, DELTA_KEY_SIZE>,
393 :
394 : blob_writer: BlobWriter<true>,
395 :
396 : // Number of key-lsns in the layer.
397 : num_keys: usize,
398 : }
399 :
400 : impl DeltaLayerWriterInner {
401 : ///
402 : /// Start building a new delta layer.
403 : ///
404 : #[allow(clippy::too_many_arguments)]
405 8856 : async fn new(
406 8856 : conf: &'static PageServerConf,
407 8856 : timeline_id: TimelineId,
408 8856 : tenant_shard_id: TenantShardId,
409 8856 : key_start: Key,
410 8856 : lsn_range: Range<Lsn>,
411 8856 : gate: &utils::sync::gate::Gate,
412 8856 : cancel: CancellationToken,
413 8856 : ctx: &RequestContext,
414 8856 : ) -> anyhow::Result<Self> {
415 8856 : // Create the file initially with a temporary filename. We don't know
416 8856 : // the end key yet, so we cannot form the final filename yet. We will
417 8856 : // rename it when we're done.
418 8856 : //
419 8856 : // Note: This overwrites any existing file. There shouldn't be any.
420 8856 : // FIXME: throw an error instead?
421 8856 : let path =
422 8856 : DeltaLayer::temp_path_for(conf, &tenant_shard_id, &timeline_id, key_start, &lsn_range);
423 :
424 8856 : let mut file = VirtualFile::create(&path, ctx).await?;
425 : // make room for the header block
426 8856 : file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
427 8856 : let blob_writer = BlobWriter::new(file, PAGE_SZ as u64, gate, cancel, ctx);
428 8856 :
429 8856 : // Initialize the b-tree index builder
430 8856 : let block_buf = BlockBuf::new();
431 8856 : let tree_builder = DiskBtreeBuilder::new(block_buf);
432 8856 :
433 8856 : Ok(Self {
434 8856 : path,
435 8856 : timeline_id,
436 8856 : tenant_shard_id,
437 8856 : key_start,
438 8856 : lsn_range,
439 8856 : tree: tree_builder,
440 8856 : blob_writer,
441 8856 : num_keys: 0,
442 8856 : })
443 8856 : }
444 :
445 : ///
446 : /// Append a key-value pair to the file.
447 : ///
448 : /// The values must be appended in key, lsn order.
449 : ///
450 12631908 : async fn put_value(
451 12631908 : &mut self,
452 12631908 : key: Key,
453 12631908 : lsn: Lsn,
454 12631908 : val: Value,
455 12631908 : ctx: &RequestContext,
456 12631908 : ) -> anyhow::Result<()> {
457 12631908 : let (_, res) = self
458 12631908 : .put_value_bytes(
459 12631908 : key,
460 12631908 : lsn,
461 12631908 : Value::ser(&val)?.slice_len(),
462 12631908 : val.will_init(),
463 12631908 : ctx,
464 12631908 : )
465 12631908 : .await;
466 12631908 : res
467 12631908 : }
468 :
469 38951868 : async fn put_value_bytes<Buf>(
470 38951868 : &mut self,
471 38951868 : key: Key,
472 38951868 : lsn: Lsn,
473 38951868 : val: FullSlice<Buf>,
474 38951868 : will_init: bool,
475 38951868 : ctx: &RequestContext,
476 38951868 : ) -> (FullSlice<Buf>, anyhow::Result<()>)
477 38951868 : where
478 38951868 : Buf: IoBuf + Send,
479 38951868 : {
480 38951868 : assert!(
481 38951868 : self.lsn_range.start <= lsn,
482 0 : "lsn_start={}, lsn={}",
483 : self.lsn_range.start,
484 : lsn
485 : );
486 : // We don't want to use compression in delta layer creation
487 38951868 : let compression = ImageCompressionAlgorithm::Disabled;
488 38951868 : let (val, res) = self
489 38951868 : .blob_writer
490 38951868 : .write_blob_maybe_compressed(val, ctx, compression)
491 38951868 : .await;
492 38951868 : let off = match res {
493 38951868 : Ok((off, _)) => off,
494 0 : Err(e) => return (val, Err(anyhow::anyhow!(e))),
495 : };
496 :
497 38951868 : let blob_ref = BlobRef::new(off, will_init);
498 38951868 :
499 38951868 : let delta_key = DeltaKey::from_key_lsn(&key, lsn);
500 38951868 : let res = self.tree.append(&delta_key.0, blob_ref.0);
501 38951868 :
502 38951868 : self.num_keys += 1;
503 38951868 :
504 38951868 : (val, res.map_err(|e| anyhow::anyhow!(e)))
505 38951868 : }
506 :
507 12143832 : fn size(&self) -> u64 {
508 12143832 : self.blob_writer.size() + self.tree.borrow_writer().size()
509 12143832 : }
510 :
511 : ///
512 : /// Finish writing the delta layer.
513 : ///
514 8700 : async fn finish(
515 8700 : self,
516 8700 : key_end: Key,
517 8700 : ctx: &RequestContext,
518 8700 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
519 8700 : let temp_path = self.path.clone();
520 8700 : let result = self.finish0(key_end, ctx).await;
521 8700 : if let Err(ref e) = result {
522 0 : tracing::info!(%temp_path, "cleaning up temporary file after error during writing: {e}");
523 0 : if let Err(e) = std::fs::remove_file(&temp_path) {
524 0 : tracing::warn!(error=%e, %temp_path, "error cleaning up temporary layer file after error during writing");
525 0 : }
526 8700 : }
527 8700 : result
528 8700 : }
529 :
530 8700 : async fn finish0(
531 8700 : self,
532 8700 : key_end: Key,
533 8700 : ctx: &RequestContext,
534 8700 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
535 8700 : let index_start_blk = self.blob_writer.size().div_ceil(PAGE_SZ as u64) as u32;
536 :
537 8700 : let mut file = self.blob_writer.into_inner(ctx).await?;
538 :
539 : // Write out the index
540 8700 : let (index_root_blk, block_buf) = self.tree.finish()?;
541 8700 : file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
542 8700 : .await?;
543 90971 : for buf in block_buf.blocks {
544 82271 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
545 82271 : res?;
546 : }
547 8700 : assert!(self.lsn_range.start < self.lsn_range.end);
548 : // Fill in the summary on blk 0
549 8700 : let summary = Summary {
550 8700 : magic: DELTA_FILE_MAGIC,
551 8700 : format_version: STORAGE_FORMAT_VERSION,
552 8700 : tenant_id: self.tenant_shard_id.tenant_id,
553 8700 : timeline_id: self.timeline_id,
554 8700 : key_range: self.key_start..key_end,
555 8700 : lsn_range: self.lsn_range.clone(),
556 8700 : index_start_blk,
557 8700 : index_root_blk,
558 8700 : };
559 8700 :
560 8700 : let mut buf = Vec::with_capacity(PAGE_SZ);
561 8700 : // TODO: could use smallvec here but it's a pain with Slice<T>
562 8700 : Summary::ser_into(&summary, &mut buf)?;
563 8700 : file.seek(SeekFrom::Start(0)).await?;
564 8700 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
565 8700 : res?;
566 :
567 8700 : let metadata = file
568 8700 : .metadata()
569 8700 : .await
570 8700 : .context("get file metadata to determine size")?;
571 :
572 : // 5GB limit for objects without multipart upload (which we don't want to use)
573 : // Make it a little bit below to account for differing GB units
574 : // https://docs.aws.amazon.com/AmazonS3/latest/userguide/upload-objects.html
575 8700 : ensure!(
576 8700 : metadata.len() <= S3_UPLOAD_LIMIT,
577 0 : "Created delta layer file at {} of size {} above limit {S3_UPLOAD_LIMIT}!",
578 0 : file.path(),
579 0 : metadata.len()
580 : );
581 :
582 : // Note: Because we opened the file in write-only mode, we cannot
583 : // reuse the same VirtualFile for reading later. That's why we don't
584 : // set inner.file here. The first read will have to re-open it.
585 :
586 8700 : let desc = PersistentLayerDesc::new_delta(
587 8700 : self.tenant_shard_id,
588 8700 : self.timeline_id,
589 8700 : self.key_start..key_end,
590 8700 : self.lsn_range.clone(),
591 8700 : metadata.len(),
592 8700 : );
593 8700 :
594 8700 : // fsync the file
595 8700 : file.sync_all()
596 8700 : .await
597 8700 : .maybe_fatal_err("delta_layer sync_all")?;
598 :
599 8700 : trace!("created delta layer {}", self.path);
600 :
601 8700 : Ok((desc, self.path))
602 8700 : }
603 : }
604 :
605 : /// A builder object for constructing a new delta layer.
606 : ///
607 : /// Usage:
608 : ///
609 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
610 : ///
611 : /// 2. Write the contents by calling `put_value` for every page
612 : /// version to store in the layer.
613 : ///
614 : /// 3. Call `finish`.
615 : ///
616 : /// # Note
617 : ///
618 : /// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
619 : /// possible for the writer to drop before `finish` is actually called. So this
620 : /// could lead to odd temporary files in the directory, exhausting file system.
621 : /// This structure wraps `DeltaLayerWriterInner` and also contains `Drop`
622 : /// implementation that cleans up the temporary file in failure. It's not
623 : /// possible to do this directly in `DeltaLayerWriterInner` since `finish` moves
624 : /// out some fields, making it impossible to implement `Drop`.
625 : ///
626 : #[must_use]
627 : pub struct DeltaLayerWriter {
628 : inner: Option<DeltaLayerWriterInner>,
629 : }
630 :
631 : impl DeltaLayerWriter {
632 : ///
633 : /// Start building a new delta layer.
634 : ///
635 : #[allow(clippy::too_many_arguments)]
636 8856 : pub async fn new(
637 8856 : conf: &'static PageServerConf,
638 8856 : timeline_id: TimelineId,
639 8856 : tenant_shard_id: TenantShardId,
640 8856 : key_start: Key,
641 8856 : lsn_range: Range<Lsn>,
642 8856 : gate: &utils::sync::gate::Gate,
643 8856 : cancel: CancellationToken,
644 8856 : ctx: &RequestContext,
645 8856 : ) -> anyhow::Result<Self> {
646 8856 : Ok(Self {
647 8856 : inner: Some(
648 8856 : DeltaLayerWriterInner::new(
649 8856 : conf,
650 8856 : timeline_id,
651 8856 : tenant_shard_id,
652 8856 : key_start,
653 8856 : lsn_range,
654 8856 : gate,
655 8856 : cancel,
656 8856 : ctx,
657 8856 : )
658 8856 : .await?,
659 : ),
660 : })
661 8856 : }
662 :
663 0 : pub fn is_empty(&self) -> bool {
664 0 : self.inner.as_ref().unwrap().num_keys == 0
665 0 : }
666 :
667 : ///
668 : /// Append a key-value pair to the file.
669 : ///
670 : /// The values must be appended in key, lsn order.
671 : ///
672 12631908 : pub async fn put_value(
673 12631908 : &mut self,
674 12631908 : key: Key,
675 12631908 : lsn: Lsn,
676 12631908 : val: Value,
677 12631908 : ctx: &RequestContext,
678 12631908 : ) -> anyhow::Result<()> {
679 12631908 : self.inner
680 12631908 : .as_mut()
681 12631908 : .unwrap()
682 12631908 : .put_value(key, lsn, val, ctx)
683 12631908 : .await
684 12631908 : }
685 :
686 26319960 : pub async fn put_value_bytes<Buf>(
687 26319960 : &mut self,
688 26319960 : key: Key,
689 26319960 : lsn: Lsn,
690 26319960 : val: FullSlice<Buf>,
691 26319960 : will_init: bool,
692 26319960 : ctx: &RequestContext,
693 26319960 : ) -> (FullSlice<Buf>, anyhow::Result<()>)
694 26319960 : where
695 26319960 : Buf: IoBuf + Send,
696 26319960 : {
697 26319960 : self.inner
698 26319960 : .as_mut()
699 26319960 : .unwrap()
700 26319960 : .put_value_bytes(key, lsn, val, will_init, ctx)
701 26319960 : .await
702 26319960 : }
703 :
704 12143832 : pub fn size(&self) -> u64 {
705 12143832 : self.inner.as_ref().unwrap().size()
706 12143832 : }
707 :
708 : ///
709 : /// Finish writing the delta layer.
710 : ///
711 8700 : pub(crate) async fn finish(
712 8700 : mut self,
713 8700 : key_end: Key,
714 8700 : ctx: &RequestContext,
715 8700 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
716 8700 : self.inner.take().unwrap().finish(key_end, ctx).await
717 8700 : }
718 :
719 73524 : pub(crate) fn num_keys(&self) -> usize {
720 73524 : self.inner.as_ref().unwrap().num_keys
721 73524 : }
722 :
723 90840 : pub(crate) fn estimated_size(&self) -> u64 {
724 90840 : let inner = self.inner.as_ref().unwrap();
725 90840 : inner.blob_writer.size() + inner.tree.borrow_writer().size() + PAGE_SZ as u64
726 90840 : }
727 : }
728 :
729 : impl Drop for DeltaLayerWriter {
730 8856 : fn drop(&mut self) {
731 8856 : if let Some(inner) = self.inner.take() {
732 156 : // We want to remove the virtual file here, so it's fine to not
733 156 : // having completely flushed unwritten data.
734 156 : let vfile = inner.blob_writer.into_inner_no_flush();
735 156 : vfile.remove();
736 8700 : }
737 8856 : }
738 : }
739 :
740 : #[derive(thiserror::Error, Debug)]
741 : pub enum RewriteSummaryError {
742 : #[error("magic mismatch")]
743 : MagicMismatch,
744 : #[error(transparent)]
745 : Other(#[from] anyhow::Error),
746 : }
747 :
748 : impl From<std::io::Error> for RewriteSummaryError {
749 0 : fn from(e: std::io::Error) -> Self {
750 0 : Self::Other(anyhow::anyhow!(e))
751 0 : }
752 : }
753 :
754 : impl DeltaLayer {
755 0 : pub async fn rewrite_summary<F>(
756 0 : path: &Utf8Path,
757 0 : rewrite: F,
758 0 : ctx: &RequestContext,
759 0 : ) -> Result<(), RewriteSummaryError>
760 0 : where
761 0 : F: Fn(Summary) -> Summary,
762 0 : {
763 0 : let mut file = VirtualFile::open_with_options(
764 0 : path,
765 0 : virtual_file::OpenOptions::new().read(true).write(true),
766 0 : ctx,
767 0 : )
768 0 : .await
769 0 : .with_context(|| format!("Failed to open file '{}'", path))?;
770 0 : let file_id = page_cache::next_file_id();
771 0 : let block_reader = FileBlockReader::new(&file, file_id);
772 0 : let summary_blk = block_reader.read_blk(0, ctx).await?;
773 0 : let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
774 0 : if actual_summary.magic != DELTA_FILE_MAGIC {
775 0 : return Err(RewriteSummaryError::MagicMismatch);
776 0 : }
777 0 :
778 0 : let new_summary = rewrite(actual_summary);
779 0 :
780 0 : let mut buf = Vec::with_capacity(PAGE_SZ);
781 0 : // TODO: could use smallvec here, but it's a pain with Slice<T>
782 0 : Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
783 0 : file.seek(SeekFrom::Start(0)).await?;
784 0 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
785 0 : res?;
786 0 : Ok(())
787 0 : }
788 : }
789 :
790 : impl DeltaLayerInner {
791 6240 : pub(crate) fn key_range(&self) -> &Range<Key> {
792 6240 : &self.layer_key_range
793 6240 : }
794 :
795 6240 : pub(crate) fn lsn_range(&self) -> &Range<Lsn> {
796 6240 : &self.layer_lsn_range
797 6240 : }
798 :
799 6612 : pub(super) async fn load(
800 6612 : path: &Utf8Path,
801 6612 : summary: Option<Summary>,
802 6612 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
803 6612 : ctx: &RequestContext,
804 6612 : ) -> anyhow::Result<Self> {
805 6612 : let file = Arc::new(
806 6612 : VirtualFile::open_v2(path, ctx)
807 6612 : .await
808 6612 : .context("open layer file")?,
809 : );
810 :
811 6612 : let file_id = page_cache::next_file_id();
812 6612 :
813 6612 : let block_reader = FileBlockReader::new(&file, file_id);
814 :
815 6612 : let summary_blk = block_reader
816 6612 : .read_blk(0, ctx)
817 6612 : .await
818 6612 : .context("read first block")?;
819 :
820 : // TODO: this should be an assertion instead; see ImageLayerInner::load
821 6612 : let actual_summary =
822 6612 : Summary::des_prefix(summary_blk.as_ref()).context("deserialize first block")?;
823 :
824 6612 : if let Some(mut expected_summary) = summary {
825 : // production code path
826 6612 : expected_summary.index_start_blk = actual_summary.index_start_blk;
827 6612 : expected_summary.index_root_blk = actual_summary.index_root_blk;
828 6612 : // mask out the timeline_id, but still require the layers to be from the same tenant
829 6612 : expected_summary.timeline_id = actual_summary.timeline_id;
830 6612 :
831 6612 : if actual_summary != expected_summary {
832 0 : bail!(
833 0 : "in-file summary does not match expected summary. actual = {:?} expected = {:?}",
834 0 : actual_summary,
835 0 : expected_summary
836 0 : );
837 6612 : }
838 0 : }
839 :
840 6612 : Ok(DeltaLayerInner {
841 6612 : file,
842 6612 : file_id,
843 6612 : index_start_blk: actual_summary.index_start_blk,
844 6612 : index_root_blk: actual_summary.index_root_blk,
845 6612 : max_vectored_read_bytes,
846 6612 : layer_key_range: actual_summary.key_range,
847 6612 : layer_lsn_range: actual_summary.lsn_range,
848 6612 : })
849 6612 : }
850 :
851 : // Look up the keys in the provided keyspace and update
852 : // the reconstruct state with whatever is found.
853 : //
854 : // Currently, the index is visited for each range, but this
855 : // can be further optimised to visit the index only once.
856 1416320 : pub(super) async fn get_values_reconstruct_data(
857 1416320 : &self,
858 1416320 : this: ResidentLayer,
859 1416320 : keyspace: KeySpace,
860 1416320 : lsn_range: Range<Lsn>,
861 1416320 : reconstruct_state: &mut ValuesReconstructState,
862 1416320 : ctx: &RequestContext,
863 1416320 : ) -> Result<(), GetVectoredError> {
864 1416320 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
865 1416320 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
866 1416320 : self.index_start_blk,
867 1416320 : self.index_root_blk,
868 1416320 : block_reader,
869 1416320 : );
870 1416320 :
871 1416320 : let planner = VectoredReadPlanner::new(
872 1416320 : self.max_vectored_read_bytes
873 1416320 : .expect("Layer is loaded with max vectored bytes config")
874 1416320 : .0
875 1416320 : .into(),
876 1416320 : );
877 1416320 :
878 1416320 : let data_end_offset = self.index_start_offset();
879 :
880 1416320 : let reads = Self::plan_reads(
881 1416320 : &keyspace,
882 1416320 : lsn_range.clone(),
883 1416320 : data_end_offset,
884 1416320 : index_reader,
885 1416320 : planner,
886 1416320 : ctx,
887 1416320 : )
888 1416320 : .await
889 1416320 : .map_err(GetVectoredError::Other)?;
890 :
891 1416320 : self.do_reads_and_update_state(this, reads, reconstruct_state, ctx)
892 1416320 : .await;
893 :
894 1416320 : Ok(())
895 1416320 : }
896 :
897 1417532 : async fn plan_reads<Reader>(
898 1417532 : keyspace: &KeySpace,
899 1417532 : lsn_range: Range<Lsn>,
900 1417532 : data_end_offset: u64,
901 1417532 : index_reader: DiskBtreeReader<Reader, DELTA_KEY_SIZE>,
902 1417532 : mut planner: VectoredReadPlanner,
903 1417532 : ctx: &RequestContext,
904 1417532 : ) -> anyhow::Result<Vec<VectoredRead>>
905 1417532 : where
906 1417532 : Reader: BlockReader + Clone,
907 1417532 : {
908 1417532 : let ctx = RequestContextBuilder::from(ctx)
909 1417532 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
910 1417532 : .attached_child();
911 :
912 1448252 : for range in keyspace.ranges.iter() {
913 1448252 : let mut range_end_handled = false;
914 1448252 :
915 1448252 : let start_key = DeltaKey::from_key_lsn(&range.start, lsn_range.start);
916 1448252 : let index_stream = index_reader.clone().into_stream(&start_key.0, &ctx);
917 1448252 : let mut index_stream = std::pin::pin!(index_stream);
918 :
919 21506875 : while let Some(index_entry) = index_stream.next().await {
920 21454923 : let (raw_key, value) = index_entry?;
921 21454923 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
922 21454923 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
923 21454923 : let blob_ref = BlobRef(value);
924 21454923 :
925 21454923 : // Lsns are not monotonically increasing across keys, so we don't assert on them.
926 21454923 : assert!(key >= range.start);
927 :
928 21454923 : let outside_lsn_range = !lsn_range.contains(&lsn);
929 :
930 21454923 : let flag = {
931 21454923 : if outside_lsn_range {
932 5220749 : BlobFlag::Ignore
933 16234174 : } else if blob_ref.will_init() {
934 3290458 : BlobFlag::ReplaceAll
935 : } else {
936 : // Usual path: add blob to the read
937 12943716 : BlobFlag::None
938 : }
939 : };
940 :
941 21454923 : if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
942 1396300 : planner.handle_range_end(blob_ref.pos());
943 1396300 : range_end_handled = true;
944 1396300 : break;
945 20058623 : } else {
946 20058623 : planner.handle(key, lsn, blob_ref.pos(), flag);
947 20058623 : }
948 : }
949 :
950 1448252 : if !range_end_handled {
951 51952 : tracing::debug!("Handling range end fallback at {}", data_end_offset);
952 51952 : planner.handle_range_end(data_end_offset);
953 1396300 : }
954 : }
955 :
956 1417532 : Ok(planner.finish())
957 1417532 : }
958 :
959 1417520 : fn get_min_read_buffer_size(
960 1417520 : planned_reads: &[VectoredRead],
961 1417520 : read_size_soft_max: usize,
962 1417520 : ) -> usize {
963 1417520 : let Some(largest_read) = planned_reads.iter().max_by_key(|read| read.size()) else {
964 519091 : return read_size_soft_max;
965 : };
966 :
967 898429 : let largest_read_size = largest_read.size();
968 898429 : if largest_read_size > read_size_soft_max {
969 : // If the read is oversized, it should only contain one key.
970 1200 : let offenders = largest_read
971 1200 : .blobs_at
972 1200 : .as_slice()
973 1200 : .iter()
974 1200 : .filter_map(|(_, blob_meta)| {
975 1200 : if blob_meta.key.is_rel_dir_key()
976 1200 : || blob_meta.key == DBDIR_KEY
977 1200 : || blob_meta.key.is_aux_file_key()
978 : {
979 : // The size of values for these keys is unbounded and can
980 : // grow very large in pathological cases.
981 0 : None
982 : } else {
983 1200 : Some(format!("{}@{}", blob_meta.key, blob_meta.lsn))
984 : }
985 1200 : })
986 1200 : .join(", ");
987 1200 :
988 1200 : if !offenders.is_empty() {
989 1200 : tracing::warn!(
990 0 : "Oversized vectored read ({} > {}) for keys {}",
991 : largest_read_size,
992 : read_size_soft_max,
993 : offenders
994 : );
995 0 : }
996 897229 : }
997 :
998 898429 : largest_read_size
999 1417520 : }
1000 :
1001 1416320 : async fn do_reads_and_update_state(
1002 1416320 : &self,
1003 1416320 : this: ResidentLayer,
1004 1416320 : reads: Vec<VectoredRead>,
1005 1416320 : reconstruct_state: &mut ValuesReconstructState,
1006 1416320 : ctx: &RequestContext,
1007 1416320 : ) {
1008 1416320 : let max_vectored_read_bytes = self
1009 1416320 : .max_vectored_read_bytes
1010 1416320 : .expect("Layer is loaded with max vectored bytes config")
1011 1416320 : .0
1012 1416320 : .into();
1013 1416320 : let buf_size = Self::get_min_read_buffer_size(&reads, max_vectored_read_bytes);
1014 :
1015 : // Note that reads are processed in reverse order (from highest key+lsn).
1016 : // This is the order that `ReconstructState` requires such that it can
1017 : // track when a key is done.
1018 1416320 : for read in reads.into_iter().rev() {
1019 1022790 : let mut ios: HashMap<(Key, Lsn), OnDiskValueIo> = Default::default();
1020 9444102 : for (_, blob_meta) in read.blobs_at.as_slice().iter().rev() {
1021 9444102 : let io = reconstruct_state.update_key(
1022 9444102 : &blob_meta.key,
1023 9444102 : blob_meta.lsn,
1024 9444102 : blob_meta.will_init,
1025 9444102 : );
1026 9444102 : ios.insert((blob_meta.key, blob_meta.lsn), io);
1027 9444102 : }
1028 :
1029 1022790 : let read_extend_residency = this.clone();
1030 1022790 : let read_from = self.file.clone();
1031 1022790 : let read_ctx = ctx.attached_child();
1032 1022790 : reconstruct_state
1033 1022790 : .spawn_io(async move {
1034 1022790 : let vectored_blob_reader = VectoredBlobReader::new(&read_from);
1035 1022790 : let buf = IoBufferMut::with_capacity(buf_size);
1036 :
1037 1022790 : let res = vectored_blob_reader.read_blobs(&read, buf, &read_ctx).await;
1038 1022790 : match res {
1039 1022790 : Ok(blobs_buf) => {
1040 1022790 : let view = BufView::new_slice(&blobs_buf.buf);
1041 9444102 : for meta in blobs_buf.blobs.iter().rev() {
1042 9444102 : let io = ios.remove(&(meta.meta.key, meta.meta.lsn)).unwrap();
1043 :
1044 9444102 : let blob_read = meta.read(&view).await;
1045 9444102 : let blob_read = match blob_read {
1046 9444102 : Ok(buf) => buf,
1047 0 : Err(e) => {
1048 0 : io.complete(Err(e));
1049 0 : continue;
1050 : }
1051 : };
1052 :
1053 9444102 : io.complete(Ok(OnDiskValue::WalRecordOrImage(
1054 9444102 : blob_read.into_bytes(),
1055 9444102 : )));
1056 : }
1057 :
1058 1022790 : assert!(ios.is_empty());
1059 : }
1060 0 : Err(err) => {
1061 0 : for (_, sender) in ios {
1062 0 : sender.complete(Err(std::io::Error::new(
1063 0 : err.kind(),
1064 0 : "vec read failed",
1065 0 : )));
1066 0 : }
1067 : }
1068 : }
1069 :
1070 : // keep layer resident until this IO is done; this spawned IO future generally outlives the
1071 : // call to `self` / the `Arc<DownloadedLayer>` / the `ResidentLayer` that guarantees residency
1072 1022790 : drop(read_extend_residency);
1073 1022790 : })
1074 1022790 : .await;
1075 : }
1076 1416320 : }
1077 :
1078 2436 : pub(crate) async fn index_entries<'a>(
1079 2436 : &'a self,
1080 2436 : ctx: &RequestContext,
1081 2436 : ) -> Result<Vec<DeltaEntry<'a>>> {
1082 2436 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1083 2436 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1084 2436 : self.index_start_blk,
1085 2436 : self.index_root_blk,
1086 2436 : block_reader,
1087 2436 : );
1088 2436 :
1089 2436 : let mut all_keys: Vec<DeltaEntry<'_>> = Vec::new();
1090 2436 :
1091 2436 : tree_reader
1092 2436 : .visit(
1093 2436 : &[0u8; DELTA_KEY_SIZE],
1094 2436 : VisitDirection::Forwards,
1095 12384276 : |key, value| {
1096 12384276 : let delta_key = DeltaKey::from_slice(key);
1097 12384276 : let val_ref = ValueRef {
1098 12384276 : blob_ref: BlobRef(value),
1099 12384276 : layer: self,
1100 12384276 : };
1101 12384276 : let pos = BlobRef(value).pos();
1102 12384276 : if let Some(last) = all_keys.last_mut() {
1103 12381840 : // subtract offset of the current and last entries to get the size
1104 12381840 : // of the value associated with this (key, lsn) tuple
1105 12381840 : let first_pos = last.size;
1106 12381840 : last.size = pos - first_pos;
1107 12381840 : }
1108 12384276 : let entry = DeltaEntry {
1109 12384276 : key: delta_key.key(),
1110 12384276 : lsn: delta_key.lsn(),
1111 12384276 : size: pos,
1112 12384276 : val: val_ref,
1113 12384276 : };
1114 12384276 : all_keys.push(entry);
1115 12384276 : true
1116 12384276 : },
1117 2436 : &RequestContextBuilder::from(ctx)
1118 2436 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
1119 2436 : .attached_child(),
1120 2436 : )
1121 2436 : .await?;
1122 2436 : if let Some(last) = all_keys.last_mut() {
1123 2436 : // Last key occupies all space till end of value storage,
1124 2436 : // which corresponds to beginning of the index
1125 2436 : last.size = self.index_start_offset() - last.size;
1126 2436 : }
1127 2436 : Ok(all_keys)
1128 2436 : }
1129 :
1130 : /// Using the given writer, write out a version which has the earlier Lsns than `until`.
1131 : ///
1132 : /// Return the amount of key value records pushed to the writer.
1133 60 : pub(super) async fn copy_prefix(
1134 60 : &self,
1135 60 : writer: &mut DeltaLayerWriter,
1136 60 : until: Lsn,
1137 60 : ctx: &RequestContext,
1138 60 : ) -> anyhow::Result<usize> {
1139 : use futures::stream::TryStreamExt;
1140 :
1141 : use crate::tenant::vectored_blob_io::{
1142 : BlobMeta, ChunkedVectoredReadBuilder, VectoredReadExtended,
1143 : };
1144 :
1145 : #[derive(Debug)]
1146 : enum Item {
1147 : Actual(Key, Lsn, BlobRef),
1148 : Sentinel,
1149 : }
1150 :
1151 : impl From<Item> for Option<(Key, Lsn, BlobRef)> {
1152 420 : fn from(value: Item) -> Self {
1153 420 : match value {
1154 360 : Item::Actual(key, lsn, blob) => Some((key, lsn, blob)),
1155 60 : Item::Sentinel => None,
1156 : }
1157 420 : }
1158 : }
1159 :
1160 : impl Item {
1161 420 : fn offset(&self) -> Option<BlobRef> {
1162 420 : match self {
1163 360 : Item::Actual(_, _, blob) => Some(*blob),
1164 60 : Item::Sentinel => None,
1165 : }
1166 420 : }
1167 :
1168 420 : fn is_last(&self) -> bool {
1169 420 : matches!(self, Item::Sentinel)
1170 420 : }
1171 : }
1172 :
1173 60 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1174 60 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1175 60 : self.index_start_blk,
1176 60 : self.index_root_blk,
1177 60 : block_reader,
1178 60 : );
1179 60 :
1180 60 : let stream = self.stream_index_forwards(tree_reader, &[0u8; DELTA_KEY_SIZE], ctx);
1181 360 : let stream = stream.map_ok(|(key, lsn, pos)| Item::Actual(key, lsn, pos));
1182 60 : // put in a sentinel value for getting the end offset for last item, and not having to
1183 60 : // repeat the whole read part
1184 60 : let stream = stream.chain(futures::stream::once(futures::future::ready(Ok(
1185 60 : Item::Sentinel,
1186 60 : ))));
1187 60 : let mut stream = std::pin::pin!(stream);
1188 60 :
1189 60 : let mut prev: Option<(Key, Lsn, BlobRef)> = None;
1190 60 :
1191 60 : let mut read_builder: Option<ChunkedVectoredReadBuilder> = None;
1192 60 :
1193 60 : let max_read_size = self
1194 60 : .max_vectored_read_bytes
1195 60 : .map(|x| x.0.get())
1196 60 : .unwrap_or(8192);
1197 60 :
1198 60 : let mut buffer = Some(IoBufferMut::with_capacity(max_read_size));
1199 60 :
1200 60 : // FIXME: buffering of DeltaLayerWriter
1201 60 : let mut per_blob_copy = Vec::new();
1202 60 :
1203 60 : let mut records = 0;
1204 :
1205 480 : while let Some(item) = stream.try_next().await? {
1206 420 : tracing::debug!(?item, "popped");
1207 420 : let offset = item
1208 420 : .offset()
1209 420 : .unwrap_or(BlobRef::new(self.index_start_offset(), false));
1210 :
1211 420 : let actionable = if let Some((key, lsn, start_offset)) = prev.take() {
1212 360 : let end_offset = offset;
1213 360 :
1214 360 : Some((
1215 360 : BlobMeta {
1216 360 : key,
1217 360 : lsn,
1218 360 : will_init: false,
1219 360 : },
1220 360 : start_offset..end_offset,
1221 360 : ))
1222 : } else {
1223 60 : None
1224 : };
1225 :
1226 420 : let is_last = item.is_last();
1227 420 :
1228 420 : prev = Option::from(item);
1229 420 :
1230 420 : let actionable = actionable.filter(|x| x.0.lsn < until);
1231 :
1232 420 : let builder = if let Some((meta, offsets)) = actionable {
1233 : // extend or create a new builder
1234 192 : if read_builder
1235 192 : .as_mut()
1236 192 : .map(|x| x.extend(offsets.start.pos(), offsets.end.pos(), meta))
1237 192 : .unwrap_or(VectoredReadExtended::No)
1238 192 : == VectoredReadExtended::Yes
1239 : {
1240 96 : None
1241 : } else {
1242 96 : read_builder.replace(ChunkedVectoredReadBuilder::new(
1243 96 : offsets.start.pos(),
1244 96 : offsets.end.pos(),
1245 96 : meta,
1246 96 : max_read_size,
1247 96 : ))
1248 : }
1249 : } else {
1250 : // nothing to do, except perhaps flush any existing for the last element
1251 228 : None
1252 : };
1253 :
1254 : // flush the possible older builder and also the new one if the item was the last one
1255 420 : let builders = builder.into_iter();
1256 420 : let builders = if is_last {
1257 60 : builders.chain(read_builder.take())
1258 : } else {
1259 360 : builders.chain(None)
1260 : };
1261 :
1262 516 : for builder in builders {
1263 96 : let read = builder.build();
1264 96 :
1265 96 : let reader = VectoredBlobReader::new(&self.file);
1266 96 :
1267 96 : let mut buf = buffer.take().unwrap();
1268 96 :
1269 96 : buf.clear();
1270 96 : buf.reserve(read.size());
1271 96 : let res = reader.read_blobs(&read, buf, ctx).await?;
1272 :
1273 96 : let view = BufView::new_slice(&res.buf);
1274 :
1275 288 : for blob in res.blobs {
1276 192 : let key = blob.meta.key;
1277 192 : let lsn = blob.meta.lsn;
1278 :
1279 192 : let data = blob.read(&view).await?;
1280 :
1281 : #[cfg(debug_assertions)]
1282 192 : Value::des(&data)
1283 192 : .with_context(|| {
1284 0 : format!(
1285 0 : "blob failed to deserialize for {}: {:?}",
1286 0 : blob,
1287 0 : utils::Hex(&data)
1288 0 : )
1289 192 : })
1290 192 : .unwrap();
1291 192 :
1292 192 : // is it an image or will_init walrecord?
1293 192 : // FIXME: this could be handled by threading the BlobRef to the
1294 192 : // VectoredReadBuilder
1295 192 : let will_init = pageserver_api::value::ValueBytes::will_init(&data)
1296 192 : .inspect_err(|_e| {
1297 0 : #[cfg(feature = "testing")]
1298 0 : tracing::error!(data=?utils::Hex(&data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
1299 192 : })
1300 192 : .unwrap_or(false);
1301 192 :
1302 192 : per_blob_copy.clear();
1303 192 : per_blob_copy.extend_from_slice(&data);
1304 :
1305 192 : let (tmp, res) = writer
1306 192 : .put_value_bytes(
1307 192 : key,
1308 192 : lsn,
1309 192 : std::mem::take(&mut per_blob_copy).slice_len(),
1310 192 : will_init,
1311 192 : ctx,
1312 192 : )
1313 192 : .await;
1314 192 : per_blob_copy = tmp.into_raw_slice().into_inner();
1315 192 :
1316 192 : res?;
1317 :
1318 192 : records += 1;
1319 : }
1320 :
1321 96 : buffer = Some(res.buf);
1322 : }
1323 : }
1324 :
1325 60 : assert!(
1326 60 : read_builder.is_none(),
1327 0 : "with the sentinel above loop should had handled all"
1328 : );
1329 :
1330 60 : Ok(records)
1331 60 : }
1332 :
1333 24 : pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
1334 24 : println!(
1335 24 : "index_start_blk: {}, root {}",
1336 24 : self.index_start_blk, self.index_root_blk
1337 24 : );
1338 24 :
1339 24 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1340 24 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1341 24 : self.index_start_blk,
1342 24 : self.index_root_blk,
1343 24 : block_reader,
1344 24 : );
1345 24 :
1346 24 : tree_reader.dump(ctx).await?;
1347 :
1348 24 : let keys = self.index_entries(ctx).await?;
1349 :
1350 48 : async fn dump_blob(val: &ValueRef<'_>, ctx: &RequestContext) -> anyhow::Result<String> {
1351 48 : let buf = val.load_raw(ctx).await?;
1352 48 : let val = Value::des(&buf)?;
1353 48 : let desc = match val {
1354 48 : Value::Image(img) => {
1355 48 : format!(" img {} bytes", img.len())
1356 : }
1357 0 : Value::WalRecord(rec) => {
1358 0 : let wal_desc = pageserver_api::record::describe_wal_record(&rec)?;
1359 0 : format!(
1360 0 : " rec {} bytes will_init: {} {}",
1361 0 : buf.len(),
1362 0 : rec.will_init(),
1363 0 : wal_desc
1364 0 : )
1365 : }
1366 : };
1367 48 : Ok(desc)
1368 48 : }
1369 :
1370 72 : for entry in keys {
1371 48 : let DeltaEntry { key, lsn, val, .. } = entry;
1372 48 : let desc = match dump_blob(&val, ctx).await {
1373 48 : Ok(desc) => desc,
1374 0 : Err(err) => {
1375 0 : format!("ERROR: {err}")
1376 : }
1377 : };
1378 48 : println!(" key {key} at {lsn}: {desc}");
1379 :
1380 : // Print more details about CHECKPOINT records. Would be nice to print details
1381 : // of many other record types too, but these are particularly interesting, as
1382 : // have a lot of special processing for them in walingest.rs.
1383 24 : use pageserver_api::key::CHECKPOINT_KEY;
1384 24 : use postgres_ffi::CheckPoint;
1385 48 : if key == CHECKPOINT_KEY {
1386 0 : let val = val.load(ctx).await?;
1387 0 : match val {
1388 0 : Value::Image(img) => {
1389 0 : let checkpoint = CheckPoint::decode(&img)?;
1390 0 : println!(" CHECKPOINT: {:?}", checkpoint);
1391 : }
1392 0 : Value::WalRecord(_rec) => {
1393 0 : println!(" unexpected walrecord value for checkpoint key");
1394 0 : }
1395 : }
1396 48 : }
1397 : }
1398 :
1399 24 : Ok(())
1400 24 : }
1401 :
1402 180 : fn stream_index_forwards<'a, R>(
1403 180 : &'a self,
1404 180 : reader: DiskBtreeReader<R, DELTA_KEY_SIZE>,
1405 180 : start: &'a [u8; DELTA_KEY_SIZE],
1406 180 : ctx: &'a RequestContext,
1407 180 : ) -> impl futures::stream::Stream<
1408 180 : Item = Result<(Key, Lsn, BlobRef), crate::tenant::disk_btree::DiskBtreeError>,
1409 180 : > + 'a
1410 180 : where
1411 180 : R: BlockReader + 'a,
1412 180 : {
1413 : use futures::stream::TryStreamExt;
1414 180 : let stream = reader.into_stream(start, ctx);
1415 912 : stream.map_ok(|(key, value)| {
1416 912 : let key = DeltaKey::from_slice(&key);
1417 912 : let (key, lsn) = (key.key(), key.lsn());
1418 912 : let offset = BlobRef(value);
1419 912 :
1420 912 : (key, lsn, offset)
1421 912 : })
1422 180 : }
1423 :
1424 : /// The file offset to the first block of index.
1425 : ///
1426 : /// The file structure is summary, values, and index. We often need this for the size of last blob.
1427 1422464 : fn index_start_offset(&self) -> u64 {
1428 1422464 : let offset = self.index_start_blk as u64 * PAGE_SZ as u64;
1429 1422464 : let bref = BlobRef(offset);
1430 1422464 : tracing::debug!(
1431 : index_start_blk = self.index_start_blk,
1432 : offset,
1433 0 : pos = bref.pos(),
1434 0 : "index_start_offset"
1435 : );
1436 1422464 : offset
1437 1422464 : }
1438 :
1439 3456 : pub fn iter<'a>(&'a self, ctx: &'a RequestContext) -> DeltaLayerIterator<'a> {
1440 3456 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1441 3456 : let tree_reader =
1442 3456 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
1443 3456 : DeltaLayerIterator {
1444 3456 : delta_layer: self,
1445 3456 : ctx,
1446 3456 : index_iter: tree_reader.iter(&[0; DELTA_KEY_SIZE], ctx),
1447 3456 : key_values_batch: std::collections::VecDeque::new(),
1448 3456 : is_end: false,
1449 3456 : planner: StreamingVectoredReadPlanner::new(
1450 3456 : 1024 * 8192, // The default value. Unit tests might use a different value. 1024 * 8K = 8MB buffer.
1451 3456 : 1024, // The default value. Unit tests might use a different value
1452 3456 : ),
1453 3456 : }
1454 3456 : }
1455 :
1456 : /// NB: not super efficient, but not terrible either. Should prob be an iterator.
1457 : //
1458 : // We're reusing the index traversal logical in plan_reads; would be nice to
1459 : // factor that out.
1460 0 : pub(crate) async fn load_keys(&self, ctx: &RequestContext) -> anyhow::Result<Vec<Key>> {
1461 0 : self.index_entries(ctx)
1462 0 : .await
1463 0 : .map(|entries| entries.into_iter().map(|entry| entry.key).collect())
1464 0 : }
1465 : }
1466 :
1467 : /// A set of data associated with a delta layer key and its value
1468 : pub struct DeltaEntry<'a> {
1469 : pub key: Key,
1470 : pub lsn: Lsn,
1471 : /// Size of the stored value
1472 : pub size: u64,
1473 : /// Reference to the on-disk value
1474 : pub val: ValueRef<'a>,
1475 : }
1476 :
1477 : /// Reference to an on-disk value
1478 : pub struct ValueRef<'a> {
1479 : blob_ref: BlobRef,
1480 : layer: &'a DeltaLayerInner,
1481 : }
1482 :
1483 : impl ValueRef<'_> {
1484 : /// Loads the value from disk
1485 0 : pub async fn load(&self, ctx: &RequestContext) -> Result<Value> {
1486 0 : let buf = self.load_raw(ctx).await?;
1487 0 : let val = Value::des(&buf)?;
1488 0 : Ok(val)
1489 0 : }
1490 :
1491 48 : async fn load_raw(&self, ctx: &RequestContext) -> Result<Vec<u8>> {
1492 48 : let reader = BlockCursor::new(crate::tenant::block_io::BlockReaderRef::Adapter(Adapter(
1493 48 : self.layer,
1494 48 : )));
1495 48 : let buf = reader.read_blob(self.blob_ref.pos(), ctx).await?;
1496 48 : Ok(buf)
1497 48 : }
1498 : }
1499 :
1500 : pub(crate) struct Adapter<T>(T);
1501 :
1502 : impl<T: AsRef<DeltaLayerInner>> Adapter<T> {
1503 48 : pub(crate) async fn read_blk(
1504 48 : &self,
1505 48 : blknum: u32,
1506 48 : ctx: &RequestContext,
1507 48 : ) -> Result<BlockLease, std::io::Error> {
1508 48 : let block_reader = FileBlockReader::new(&self.0.as_ref().file, self.0.as_ref().file_id);
1509 48 : block_reader.read_blk(blknum, ctx).await
1510 48 : }
1511 : }
1512 :
1513 : impl AsRef<DeltaLayerInner> for DeltaLayerInner {
1514 96 : fn as_ref(&self) -> &DeltaLayerInner {
1515 96 : self
1516 96 : }
1517 : }
1518 :
1519 : impl<'a> pageserver_compaction::interface::CompactionDeltaEntry<'a, Key> for DeltaEntry<'a> {
1520 0 : fn key(&self) -> Key {
1521 0 : self.key
1522 0 : }
1523 0 : fn lsn(&self) -> Lsn {
1524 0 : self.lsn
1525 0 : }
1526 0 : fn size(&self) -> u64 {
1527 0 : self.size
1528 0 : }
1529 : }
1530 :
1531 : pub struct DeltaLayerIterator<'a> {
1532 : delta_layer: &'a DeltaLayerInner,
1533 : ctx: &'a RequestContext,
1534 : planner: StreamingVectoredReadPlanner,
1535 : index_iter: DiskBtreeIterator<'a>,
1536 : key_values_batch: VecDeque<(Key, Lsn, Value)>,
1537 : is_end: bool,
1538 : }
1539 :
1540 : impl DeltaLayerIterator<'_> {
1541 0 : pub(crate) fn layer_dbg_info(&self) -> String {
1542 0 : self.delta_layer.layer_dbg_info()
1543 0 : }
1544 :
1545 : /// Retrieve a batch of key-value pairs into the iterator buffer.
1546 128436 : async fn next_batch(&mut self) -> anyhow::Result<()> {
1547 128436 : assert!(self.key_values_batch.is_empty());
1548 128436 : assert!(!self.is_end);
1549 :
1550 128436 : let plan = loop {
1551 12597780 : if let Some(res) = self.index_iter.next().await {
1552 12594492 : let (raw_key, value) = res?;
1553 12594492 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
1554 12594492 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
1555 12594492 : let blob_ref = BlobRef(value);
1556 12594492 : let offset = blob_ref.pos();
1557 125148 : if let Some(batch_plan) =
1558 12594492 : self.planner.handle(key, lsn, offset, blob_ref.will_init())
1559 : {
1560 125148 : break batch_plan;
1561 12469344 : }
1562 : } else {
1563 3288 : self.is_end = true;
1564 3288 : let data_end_offset = self.delta_layer.index_start_offset();
1565 3288 : if let Some(item) = self.planner.handle_range_end(data_end_offset) {
1566 3288 : break item;
1567 : } else {
1568 0 : return Ok(()); // TODO: test empty iterator
1569 : }
1570 : }
1571 : };
1572 128436 : let vectored_blob_reader = VectoredBlobReader::new(&self.delta_layer.file);
1573 128436 : let mut next_batch = std::collections::VecDeque::new();
1574 128436 : let buf_size = plan.size();
1575 128436 : let buf = IoBufferMut::with_capacity(buf_size);
1576 128436 : let blobs_buf = vectored_blob_reader
1577 128436 : .read_blobs(&plan, buf, self.ctx)
1578 128436 : .await?;
1579 128436 : let view = BufView::new_slice(&blobs_buf.buf);
1580 12594324 : for meta in blobs_buf.blobs.iter() {
1581 12594324 : let blob_read = meta.read(&view).await?;
1582 12594324 : let value = Value::des(&blob_read)?;
1583 :
1584 12594324 : next_batch.push_back((meta.meta.key, meta.meta.lsn, value));
1585 : }
1586 128436 : self.key_values_batch = next_batch;
1587 128436 : Ok(())
1588 128436 : }
1589 :
1590 12599052 : pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
1591 12599052 : if self.key_values_batch.is_empty() {
1592 134304 : if self.is_end {
1593 6372 : return Ok(None);
1594 127932 : }
1595 127932 : self.next_batch().await?;
1596 12464748 : }
1597 12592680 : Ok(Some(
1598 12592680 : self.key_values_batch
1599 12592680 : .pop_front()
1600 12592680 : .expect("should not be empty"),
1601 12592680 : ))
1602 12599052 : }
1603 : }
1604 :
1605 : #[cfg(test)]
1606 : pub(crate) mod test {
1607 : use std::collections::BTreeMap;
1608 :
1609 : use bytes::Bytes;
1610 : use itertools::MinMaxResult;
1611 : use pageserver_api::value::Value;
1612 : use rand::RngCore;
1613 : use rand::prelude::{SeedableRng, SliceRandom, StdRng};
1614 :
1615 : use super::*;
1616 : use crate::DEFAULT_PG_VERSION;
1617 : use crate::context::DownloadBehavior;
1618 : use crate::task_mgr::TaskKind;
1619 : use crate::tenant::disk_btree::tests::TestDisk;
1620 : use crate::tenant::harness::{TIMELINE_ID, TenantHarness};
1621 : use crate::tenant::storage_layer::{Layer, ResidentLayer};
1622 : use crate::tenant::vectored_blob_io::StreamingVectoredReadPlanner;
1623 : use crate::tenant::{TenantShard, Timeline};
1624 :
1625 : /// Construct an index for a fictional delta layer and and then
1626 : /// traverse in order to plan vectored reads for a query. Finally,
1627 : /// verify that the traversal fed the right index key and value
1628 : /// pairs into the planner.
1629 : #[tokio::test]
1630 12 : async fn test_delta_layer_index_traversal() {
1631 12 : let base_key = Key {
1632 12 : field1: 0,
1633 12 : field2: 1663,
1634 12 : field3: 12972,
1635 12 : field4: 16396,
1636 12 : field5: 0,
1637 12 : field6: 246080,
1638 12 : };
1639 12 :
1640 12 : // Populate the index with some entries
1641 12 : let entries: BTreeMap<Key, Vec<Lsn>> = BTreeMap::from([
1642 12 : (base_key, vec![Lsn(1), Lsn(5), Lsn(25), Lsn(26), Lsn(28)]),
1643 12 : (base_key.add(1), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1644 12 : (base_key.add(2), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1645 12 : (base_key.add(5), vec![Lsn(10), Lsn(15), Lsn(16), Lsn(20)]),
1646 12 : ]);
1647 12 :
1648 12 : let mut disk = TestDisk::default();
1649 12 : let mut writer = DiskBtreeBuilder::<_, DELTA_KEY_SIZE>::new(&mut disk);
1650 12 :
1651 12 : let mut disk_offset = 0;
1652 60 : for (key, lsns) in &entries {
1653 252 : for lsn in lsns {
1654 204 : let index_key = DeltaKey::from_key_lsn(key, *lsn);
1655 204 : let blob_ref = BlobRef::new(disk_offset, false);
1656 204 : writer
1657 204 : .append(&index_key.0, blob_ref.0)
1658 204 : .expect("In memory disk append should never fail");
1659 204 :
1660 204 : disk_offset += 1;
1661 204 : }
1662 12 : }
1663 12 :
1664 12 : // Prepare all the arguments for the call into `plan_reads` below
1665 12 : let (root_offset, _writer) = writer
1666 12 : .finish()
1667 12 : .expect("In memory disk finish should never fail");
1668 12 : let reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(0, root_offset, disk);
1669 12 : let planner = VectoredReadPlanner::new(100);
1670 12 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1671 12 :
1672 12 : let keyspace = KeySpace {
1673 12 : ranges: vec![
1674 12 : base_key..base_key.add(3),
1675 12 : base_key.add(3)..base_key.add(100),
1676 12 : ],
1677 12 : };
1678 12 : let lsn_range = Lsn(2)..Lsn(40);
1679 12 :
1680 12 : // Plan and validate
1681 12 : let vectored_reads = DeltaLayerInner::plan_reads(
1682 12 : &keyspace,
1683 12 : lsn_range.clone(),
1684 12 : disk_offset,
1685 12 : reader,
1686 12 : planner,
1687 12 : &ctx,
1688 12 : )
1689 12 : .await
1690 12 : .expect("Read planning should not fail");
1691 12 :
1692 12 : validate(keyspace, lsn_range, vectored_reads, entries);
1693 12 : }
1694 :
1695 12 : fn validate(
1696 12 : keyspace: KeySpace,
1697 12 : lsn_range: Range<Lsn>,
1698 12 : vectored_reads: Vec<VectoredRead>,
1699 12 : index_entries: BTreeMap<Key, Vec<Lsn>>,
1700 12 : ) {
1701 : #[derive(Debug, PartialEq, Eq)]
1702 : struct BlobSpec {
1703 : key: Key,
1704 : lsn: Lsn,
1705 : at: u64,
1706 : }
1707 :
1708 12 : let mut planned_blobs = Vec::new();
1709 180 : for read in vectored_reads {
1710 168 : for (at, meta) in read.blobs_at.as_slice() {
1711 168 : planned_blobs.push(BlobSpec {
1712 168 : key: meta.key,
1713 168 : lsn: meta.lsn,
1714 168 : at: *at,
1715 168 : });
1716 168 : }
1717 : }
1718 :
1719 12 : let mut expected_blobs = Vec::new();
1720 12 : let mut disk_offset = 0;
1721 60 : for (key, lsns) in index_entries {
1722 252 : for lsn in lsns {
1723 252 : let key_included = keyspace.ranges.iter().any(|range| range.contains(&key));
1724 204 : let lsn_included = lsn_range.contains(&lsn);
1725 204 :
1726 204 : if key_included && lsn_included {
1727 168 : expected_blobs.push(BlobSpec {
1728 168 : key,
1729 168 : lsn,
1730 168 : at: disk_offset,
1731 168 : });
1732 168 : }
1733 :
1734 204 : disk_offset += 1;
1735 : }
1736 : }
1737 :
1738 12 : assert_eq!(planned_blobs, expected_blobs);
1739 12 : }
1740 :
1741 : mod constants {
1742 : use utils::lsn::Lsn;
1743 :
1744 : /// Offset used by all lsns in this test
1745 : pub(super) const LSN_OFFSET: Lsn = Lsn(0x08);
1746 : /// Number of unique keys including in the test data
1747 : pub(super) const KEY_COUNT: u8 = 60;
1748 : /// Max number of different lsns for each key
1749 : pub(super) const MAX_ENTRIES_PER_KEY: u8 = 20;
1750 : /// Possible value sizes for each key along with a probability weight
1751 : pub(super) const VALUE_SIZES: [(usize, u8); 3] = [(100, 2), (1024, 2), (1024 * 1024, 1)];
1752 : /// Probability that there will be a gap between the current key and the next one (33.3%)
1753 : pub(super) const KEY_GAP_CHANGES: [(bool, u8); 2] = [(true, 1), (false, 2)];
1754 : /// The minimum size of a key range in all the generated reads
1755 : pub(super) const MIN_RANGE_SIZE: i128 = 10;
1756 : /// The number of ranges included in each vectored read
1757 : pub(super) const RANGES_COUNT: u8 = 2;
1758 : /// The number of vectored reads performed
1759 : pub(super) const READS_COUNT: u8 = 100;
1760 : /// Soft max size of a vectored read. Will be violated if we have to read keys
1761 : /// with values larger than the limit
1762 : pub(super) const MAX_VECTORED_READ_BYTES: usize = 64 * 1024;
1763 : }
1764 :
1765 : struct Entry {
1766 : key: Key,
1767 : lsn: Lsn,
1768 : value: Vec<u8>,
1769 : }
1770 :
1771 12 : fn generate_entries(rng: &mut StdRng) -> Vec<Entry> {
1772 12 : let mut current_key = Key::MIN;
1773 12 :
1774 12 : let mut entries = Vec::new();
1775 732 : for _ in 0..constants::KEY_COUNT {
1776 720 : let count = rng.gen_range(1..constants::MAX_ENTRIES_PER_KEY);
1777 720 : let mut lsns_iter =
1778 13560 : std::iter::successors(Some(Lsn(constants::LSN_OFFSET.0 + 0x08)), |lsn| {
1779 13560 : Some(Lsn(lsn.0 + 0x08))
1780 13560 : });
1781 720 : let mut lsns = Vec::new();
1782 14280 : while lsns.len() < count as usize {
1783 13560 : let take = rng.gen_bool(0.5);
1784 13560 : let lsn = lsns_iter.next().unwrap();
1785 13560 : if take {
1786 6672 : lsns.push(lsn);
1787 6888 : }
1788 : }
1789 :
1790 7392 : for lsn in lsns {
1791 6672 : let size = constants::VALUE_SIZES
1792 20016 : .choose_weighted(rng, |item| item.1)
1793 6672 : .unwrap()
1794 6672 : .0;
1795 6672 : let mut buf = vec![0; size];
1796 6672 : rng.fill_bytes(&mut buf);
1797 6672 :
1798 6672 : entries.push(Entry {
1799 6672 : key: current_key,
1800 6672 : lsn,
1801 6672 : value: buf,
1802 6672 : })
1803 : }
1804 :
1805 720 : let gap = constants::KEY_GAP_CHANGES
1806 1440 : .choose_weighted(rng, |item| item.1)
1807 720 : .unwrap()
1808 720 : .0;
1809 720 : if gap {
1810 228 : current_key = current_key.add(2);
1811 492 : } else {
1812 492 : current_key = current_key.add(1);
1813 492 : }
1814 : }
1815 :
1816 12 : entries
1817 12 : }
1818 :
1819 : struct EntriesMeta {
1820 : key_range: Range<Key>,
1821 : lsn_range: Range<Lsn>,
1822 : index: BTreeMap<(Key, Lsn), Vec<u8>>,
1823 : }
1824 :
1825 12 : fn get_entries_meta(entries: &[Entry]) -> EntriesMeta {
1826 6672 : let key_range = match entries.iter().minmax_by_key(|e| e.key) {
1827 12 : MinMaxResult::MinMax(min, max) => min.key..max.key.next(),
1828 0 : _ => panic!("More than one entry is always expected"),
1829 : };
1830 :
1831 6672 : let lsn_range = match entries.iter().minmax_by_key(|e| e.lsn) {
1832 12 : MinMaxResult::MinMax(min, max) => min.lsn..Lsn(max.lsn.0 + 1),
1833 0 : _ => panic!("More than one entry is always expected"),
1834 : };
1835 :
1836 12 : let mut index = BTreeMap::new();
1837 6672 : for entry in entries.iter() {
1838 6672 : index.insert((entry.key, entry.lsn), entry.value.clone());
1839 6672 : }
1840 :
1841 12 : EntriesMeta {
1842 12 : key_range,
1843 12 : lsn_range,
1844 12 : index,
1845 12 : }
1846 12 : }
1847 :
1848 1200 : fn pick_random_keyspace(rng: &mut StdRng, key_range: &Range<Key>) -> KeySpace {
1849 1200 : let start = key_range.start.to_i128();
1850 1200 : let end = key_range.end.to_i128();
1851 1200 :
1852 1200 : let mut keyspace = KeySpace::default();
1853 :
1854 3600 : for _ in 0..constants::RANGES_COUNT {
1855 2400 : let mut range: Option<Range<Key>> = Option::default();
1856 7464 : while range.is_none() || keyspace.overlaps(range.as_ref().unwrap()) {
1857 5064 : let range_start = rng.gen_range(start..end);
1858 5064 : let range_end_offset = range_start + constants::MIN_RANGE_SIZE;
1859 5064 : if range_end_offset >= end {
1860 600 : range = Some(Key::from_i128(range_start)..Key::from_i128(end));
1861 4464 : } else {
1862 4464 : let range_end = rng.gen_range((range_start + constants::MIN_RANGE_SIZE)..end);
1863 4464 : range = Some(Key::from_i128(range_start)..Key::from_i128(range_end));
1864 4464 : }
1865 : }
1866 2400 : keyspace.ranges.push(range.unwrap());
1867 : }
1868 :
1869 1200 : keyspace
1870 1200 : }
1871 :
1872 : #[tokio::test]
1873 12 : async fn test_delta_layer_vectored_read_end_to_end() -> anyhow::Result<()> {
1874 12 : let harness = TenantHarness::create("test_delta_layer_oversized_vectored_read").await?;
1875 12 : let (tenant, ctx) = harness.load().await;
1876 12 :
1877 12 : let timeline_id = TimelineId::generate();
1878 12 : let timeline = tenant
1879 12 : .create_test_timeline(timeline_id, constants::LSN_OFFSET, DEFAULT_PG_VERSION, &ctx)
1880 12 : .await?;
1881 12 :
1882 12 : tracing::info!("Generating test data ...");
1883 12 :
1884 12 : let rng = &mut StdRng::seed_from_u64(0);
1885 12 : let entries = generate_entries(rng);
1886 12 : let entries_meta = get_entries_meta(&entries);
1887 12 :
1888 12 : tracing::info!("Done generating {} entries", entries.len());
1889 12 :
1890 12 : tracing::info!("Writing test data to delta layer ...");
1891 12 : let mut writer = DeltaLayerWriter::new(
1892 12 : harness.conf,
1893 12 : timeline_id,
1894 12 : harness.tenant_shard_id,
1895 12 : entries_meta.key_range.start,
1896 12 : entries_meta.lsn_range.clone(),
1897 12 : &timeline.gate,
1898 12 : timeline.cancel.clone(),
1899 12 : &ctx,
1900 12 : )
1901 12 : .await?;
1902 12 :
1903 6684 : for entry in entries {
1904 6672 : let (_, res) = writer
1905 6672 : .put_value_bytes(entry.key, entry.lsn, entry.value.slice_len(), false, &ctx)
1906 6672 : .await;
1907 6672 : res?;
1908 12 : }
1909 12 :
1910 12 : let (desc, path) = writer.finish(entries_meta.key_range.end, &ctx).await?;
1911 12 : let resident = Layer::finish_creating(harness.conf, &timeline, desc, &path)?;
1912 12 :
1913 12 : let inner = resident.get_as_delta(&ctx).await?;
1914 12 :
1915 12 : let file_size = inner.file.metadata().await?.len();
1916 12 : tracing::info!(
1917 12 : "Done writing test data to delta layer. Resulting file size is: {}",
1918 12 : file_size
1919 12 : );
1920 12 :
1921 1212 : for i in 0..constants::READS_COUNT {
1922 1200 : tracing::info!("Doing vectored read {}/{}", i + 1, constants::READS_COUNT);
1923 12 :
1924 1200 : let block_reader = FileBlockReader::new(&inner.file, inner.file_id);
1925 1200 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1926 1200 : inner.index_start_blk,
1927 1200 : inner.index_root_blk,
1928 1200 : block_reader,
1929 1200 : );
1930 1200 :
1931 1200 : let planner = VectoredReadPlanner::new(constants::MAX_VECTORED_READ_BYTES);
1932 1200 : let keyspace = pick_random_keyspace(rng, &entries_meta.key_range);
1933 1200 : let data_end_offset = inner.index_start_blk as u64 * PAGE_SZ as u64;
1934 12 :
1935 1200 : let vectored_reads = DeltaLayerInner::plan_reads(
1936 1200 : &keyspace,
1937 1200 : entries_meta.lsn_range.clone(),
1938 1200 : data_end_offset,
1939 1200 : index_reader,
1940 1200 : planner,
1941 1200 : &ctx,
1942 1200 : )
1943 1200 : .await?;
1944 12 :
1945 1200 : let vectored_blob_reader = VectoredBlobReader::new(&inner.file);
1946 1200 : let buf_size = DeltaLayerInner::get_min_read_buffer_size(
1947 1200 : &vectored_reads,
1948 1200 : constants::MAX_VECTORED_READ_BYTES,
1949 1200 : );
1950 1200 : let mut buf = Some(IoBufferMut::with_capacity(buf_size));
1951 12 :
1952 119544 : for read in vectored_reads {
1953 118344 : let blobs_buf = vectored_blob_reader
1954 118344 : .read_blobs(&read, buf.take().expect("Should have a buffer"), &ctx)
1955 118344 : .await?;
1956 118344 : let view = BufView::new_slice(&blobs_buf.buf);
1957 343824 : for meta in blobs_buf.blobs.iter() {
1958 343824 : let value = meta.read(&view).await?;
1959 343824 : assert_eq!(
1960 343824 : &value[..],
1961 343824 : &entries_meta.index[&(meta.meta.key, meta.meta.lsn)]
1962 343824 : );
1963 12 : }
1964 12 :
1965 118344 : buf = Some(blobs_buf.buf);
1966 12 : }
1967 12 : }
1968 12 :
1969 12 : Ok(())
1970 12 : }
1971 :
1972 : #[tokio::test]
1973 12 : async fn copy_delta_prefix_smoke() {
1974 12 : use bytes::Bytes;
1975 12 : use pageserver_api::record::NeonWalRecord;
1976 12 :
1977 12 : let h = crate::tenant::harness::TenantHarness::create("truncate_delta_smoke")
1978 12 : .await
1979 12 : .unwrap();
1980 12 : let (tenant, ctx) = h.load().await;
1981 12 : let ctx = &ctx;
1982 12 : let timeline = tenant
1983 12 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, ctx)
1984 12 : .await
1985 12 : .unwrap();
1986 12 : let ctx = &ctx.with_scope_timeline(&timeline);
1987 12 :
1988 12 : let initdb_layer = timeline
1989 12 : .layers
1990 12 : .read()
1991 12 : .await
1992 12 : .likely_resident_layers()
1993 12 : .next()
1994 12 : .cloned()
1995 12 : .unwrap();
1996 12 :
1997 12 : {
1998 12 : let mut writer = timeline.writer().await;
1999 12 :
2000 12 : let data = [
2001 12 : (0x20, 12, Value::Image(Bytes::from_static(b"foobar"))),
2002 12 : (
2003 12 : 0x30,
2004 12 : 12,
2005 12 : Value::WalRecord(NeonWalRecord::Postgres {
2006 12 : will_init: false,
2007 12 : rec: Bytes::from_static(b"1"),
2008 12 : }),
2009 12 : ),
2010 12 : (
2011 12 : 0x40,
2012 12 : 12,
2013 12 : Value::WalRecord(NeonWalRecord::Postgres {
2014 12 : will_init: true,
2015 12 : rec: Bytes::from_static(b"2"),
2016 12 : }),
2017 12 : ),
2018 12 : // build an oversized value so we cannot extend and existing read over
2019 12 : // this
2020 12 : (
2021 12 : 0x50,
2022 12 : 12,
2023 12 : Value::WalRecord(NeonWalRecord::Postgres {
2024 12 : will_init: true,
2025 12 : rec: {
2026 12 : let mut buf =
2027 12 : vec![0u8; tenant.conf.max_vectored_read_bytes.0.get() + 1024];
2028 12 : buf.iter_mut()
2029 12 : .enumerate()
2030 1609728 : .for_each(|(i, slot)| *slot = (i % 256) as u8);
2031 12 : Bytes::from(buf)
2032 12 : },
2033 12 : }),
2034 12 : ),
2035 12 : // because the oversized read cannot be extended further, we are sure to exercise the
2036 12 : // builder created on the last round with this:
2037 12 : (
2038 12 : 0x60,
2039 12 : 12,
2040 12 : Value::WalRecord(NeonWalRecord::Postgres {
2041 12 : will_init: true,
2042 12 : rec: Bytes::from_static(b"3"),
2043 12 : }),
2044 12 : ),
2045 12 : (
2046 12 : 0x60,
2047 12 : 9,
2048 12 : Value::Image(Bytes::from_static(b"something for a different key")),
2049 12 : ),
2050 12 : ];
2051 12 :
2052 12 : let mut last_lsn = None;
2053 12 :
2054 84 : for (lsn, key, value) in data {
2055 72 : let key = Key::from_i128(key);
2056 72 : writer.put(key, Lsn(lsn), &value, ctx).await.unwrap();
2057 72 : last_lsn = Some(lsn);
2058 12 : }
2059 12 :
2060 12 : writer.finish_write(Lsn(last_lsn.unwrap()));
2061 12 : }
2062 12 : timeline.freeze_and_flush().await.unwrap();
2063 12 :
2064 12 : let new_layer = timeline
2065 12 : .layers
2066 12 : .read()
2067 12 : .await
2068 12 : .likely_resident_layers()
2069 17 : .find(|&x| x != &initdb_layer)
2070 12 : .cloned()
2071 12 : .unwrap();
2072 12 :
2073 12 : // create a copy for the timeline, so we don't overwrite the file
2074 12 : let branch = tenant
2075 12 : .branch_timeline_test(&timeline, TimelineId::generate(), None, ctx)
2076 12 : .await
2077 12 : .unwrap();
2078 12 :
2079 12 : assert_eq!(branch.get_ancestor_lsn(), Lsn(0x60));
2080 12 :
2081 12 : // truncating at 0x61 gives us a full copy, otherwise just go backwards until there's just
2082 12 : // a single key
2083 12 :
2084 72 : for truncate_at in [0x61, 0x51, 0x41, 0x31, 0x21] {
2085 60 : let truncate_at = Lsn(truncate_at);
2086 12 :
2087 60 : let mut writer = DeltaLayerWriter::new(
2088 60 : tenant.conf,
2089 60 : branch.timeline_id,
2090 60 : tenant.tenant_shard_id,
2091 60 : Key::MIN,
2092 60 : Lsn(0x11)..truncate_at,
2093 60 : &branch.gate,
2094 60 : branch.cancel.clone(),
2095 60 : ctx,
2096 60 : )
2097 60 : .await
2098 60 : .unwrap();
2099 12 :
2100 60 : let new_layer = new_layer.download_and_keep_resident(ctx).await.unwrap();
2101 60 :
2102 60 : new_layer
2103 60 : .copy_delta_prefix(&mut writer, truncate_at, ctx)
2104 60 : .await
2105 60 : .unwrap();
2106 12 :
2107 60 : let (desc, path) = writer.finish(Key::MAX, ctx).await.unwrap();
2108 60 : let copied_layer = Layer::finish_creating(tenant.conf, &branch, desc, &path).unwrap();
2109 60 :
2110 60 : copied_layer.get_as_delta(ctx).await.unwrap();
2111 60 :
2112 60 : assert_keys_and_values_eq(
2113 60 : new_layer.get_as_delta(ctx).await.unwrap(),
2114 60 : copied_layer.get_as_delta(ctx).await.unwrap(),
2115 60 : truncate_at,
2116 60 : ctx,
2117 60 : )
2118 60 : .await;
2119 12 : }
2120 12 : }
2121 :
2122 60 : async fn assert_keys_and_values_eq(
2123 60 : source: &DeltaLayerInner,
2124 60 : truncated: &DeltaLayerInner,
2125 60 : truncated_at: Lsn,
2126 60 : ctx: &RequestContext,
2127 60 : ) {
2128 : use futures::future::ready;
2129 : use futures::stream::TryStreamExt;
2130 :
2131 60 : let start_key = [0u8; DELTA_KEY_SIZE];
2132 60 :
2133 60 : let source_reader = FileBlockReader::new(&source.file, source.file_id);
2134 60 : let source_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2135 60 : source.index_start_blk,
2136 60 : source.index_root_blk,
2137 60 : &source_reader,
2138 60 : );
2139 60 : let source_stream = source.stream_index_forwards(source_tree, &start_key, ctx);
2140 360 : let source_stream = source_stream.filter(|res| match res {
2141 360 : Ok((_, lsn, _)) => ready(lsn < &truncated_at),
2142 0 : _ => ready(true),
2143 360 : });
2144 60 : let mut source_stream = std::pin::pin!(source_stream);
2145 60 :
2146 60 : let truncated_reader = FileBlockReader::new(&truncated.file, truncated.file_id);
2147 60 : let truncated_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2148 60 : truncated.index_start_blk,
2149 60 : truncated.index_root_blk,
2150 60 : &truncated_reader,
2151 60 : );
2152 60 : let truncated_stream = truncated.stream_index_forwards(truncated_tree, &start_key, ctx);
2153 60 : let mut truncated_stream = std::pin::pin!(truncated_stream);
2154 60 :
2155 60 : let mut scratch_left = Vec::new();
2156 60 : let mut scratch_right = Vec::new();
2157 :
2158 : loop {
2159 252 : let (src, truncated) = (source_stream.try_next(), truncated_stream.try_next());
2160 252 : let (src, truncated) = tokio::try_join!(src, truncated).unwrap();
2161 252 :
2162 252 : if src.is_none() {
2163 60 : assert!(truncated.is_none());
2164 60 : break;
2165 192 : }
2166 192 :
2167 192 : let (src, truncated) = (src.unwrap(), truncated.unwrap());
2168 192 :
2169 192 : // because we've filtered the source with Lsn, we should always have the same keys from both.
2170 192 : assert_eq!(src.0, truncated.0);
2171 192 : assert_eq!(src.1, truncated.1);
2172 :
2173 : // if this is needed for something else, just drop this assert.
2174 192 : assert!(
2175 192 : src.2.pos() >= truncated.2.pos(),
2176 0 : "value position should not go backwards {} vs. {}",
2177 0 : src.2.pos(),
2178 0 : truncated.2.pos()
2179 : );
2180 :
2181 192 : scratch_left.clear();
2182 192 : let src_cursor = source_reader.block_cursor();
2183 192 : let left = src_cursor.read_blob_into_buf(src.2.pos(), &mut scratch_left, ctx);
2184 192 : scratch_right.clear();
2185 192 : let trunc_cursor = truncated_reader.block_cursor();
2186 192 : let right = trunc_cursor.read_blob_into_buf(truncated.2.pos(), &mut scratch_right, ctx);
2187 192 :
2188 192 : tokio::try_join!(left, right).unwrap();
2189 192 :
2190 192 : assert_eq!(utils::Hex(&scratch_left), utils::Hex(&scratch_right));
2191 : }
2192 60 : }
2193 :
2194 109344 : pub(crate) fn sort_delta(
2195 109344 : (k1, l1, _): &(Key, Lsn, Value),
2196 109344 : (k2, l2, _): &(Key, Lsn, Value),
2197 109344 : ) -> std::cmp::Ordering {
2198 109344 : (k1, l1).cmp(&(k2, l2))
2199 109344 : }
2200 :
2201 : #[cfg(feature = "testing")]
2202 564 : pub(crate) fn sort_delta_value(
2203 564 : (k1, l1, v1): &(Key, Lsn, Value),
2204 564 : (k2, l2, v2): &(Key, Lsn, Value),
2205 564 : ) -> std::cmp::Ordering {
2206 564 : let order_1 = if v1.is_image() { 0 } else { 1 };
2207 564 : let order_2 = if v2.is_image() { 0 } else { 1 };
2208 564 : (k1, l1, order_1).cmp(&(k2, l2, order_2))
2209 564 : }
2210 :
2211 132 : pub(crate) async fn produce_delta_layer(
2212 132 : tenant: &TenantShard,
2213 132 : tline: &Arc<Timeline>,
2214 132 : mut deltas: Vec<(Key, Lsn, Value)>,
2215 132 : ctx: &RequestContext,
2216 132 : ) -> anyhow::Result<ResidentLayer> {
2217 132 : deltas.sort_by(sort_delta);
2218 132 : let (key_start, _, _) = deltas.first().unwrap();
2219 132 : let (key_max, _, _) = deltas.last().unwrap();
2220 49440 : let lsn_min = deltas.iter().map(|(_, lsn, _)| lsn).min().unwrap();
2221 49440 : let lsn_max = deltas.iter().map(|(_, lsn, _)| lsn).max().unwrap();
2222 132 : let lsn_end = Lsn(lsn_max.0 + 1);
2223 132 : let mut writer = DeltaLayerWriter::new(
2224 132 : tenant.conf,
2225 132 : tline.timeline_id,
2226 132 : tenant.tenant_shard_id,
2227 132 : *key_start,
2228 132 : (*lsn_min)..lsn_end,
2229 132 : &tline.gate,
2230 132 : tline.cancel.clone(),
2231 132 : ctx,
2232 132 : )
2233 132 : .await?;
2234 132 : let key_end = key_max.next();
2235 :
2236 49572 : for (key, lsn, value) in deltas {
2237 49440 : writer.put_value(key, lsn, value, ctx).await?;
2238 : }
2239 :
2240 132 : let (desc, path) = writer.finish(key_end, ctx).await?;
2241 132 : let delta_layer = Layer::finish_creating(tenant.conf, tline, desc, &path)?;
2242 :
2243 132 : Ok::<_, anyhow::Error>(delta_layer)
2244 132 : }
2245 :
2246 168 : async fn assert_delta_iter_equal(
2247 168 : delta_iter: &mut DeltaLayerIterator<'_>,
2248 168 : expect: &[(Key, Lsn, Value)],
2249 168 : ) {
2250 168 : let mut expect_iter = expect.iter();
2251 : loop {
2252 168168 : let o1 = delta_iter.next().await.unwrap();
2253 168168 : let o2 = expect_iter.next();
2254 168168 : assert_eq!(o1.is_some(), o2.is_some());
2255 168168 : if o1.is_none() && o2.is_none() {
2256 168 : break;
2257 168000 : }
2258 168000 : let (k1, l1, v1) = o1.unwrap();
2259 168000 : let (k2, l2, v2) = o2.unwrap();
2260 168000 : assert_eq!(&k1, k2);
2261 168000 : assert_eq!(l1, *l2);
2262 168000 : assert_eq!(&v1, v2);
2263 : }
2264 168 : }
2265 :
2266 : #[tokio::test]
2267 12 : async fn delta_layer_iterator() {
2268 12 : let harness = TenantHarness::create("delta_layer_iterator").await.unwrap();
2269 12 : let (tenant, ctx) = harness.load().await;
2270 12 :
2271 12 : let tline = tenant
2272 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
2273 12 : .await
2274 12 : .unwrap();
2275 12 :
2276 12000 : fn get_key(id: u32) -> Key {
2277 12000 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
2278 12000 : key.field6 = id;
2279 12000 : key
2280 12000 : }
2281 12 : const N: usize = 1000;
2282 12 : let test_deltas = (0..N)
2283 12000 : .map(|idx| {
2284 12000 : (
2285 12000 : get_key(idx as u32 / 10),
2286 12000 : Lsn(0x10 * ((idx as u64) % 10 + 1)),
2287 12000 : Value::Image(Bytes::from(format!("img{idx:05}"))),
2288 12000 : )
2289 12000 : })
2290 12 : .collect_vec();
2291 12 : let resident_layer = produce_delta_layer(&tenant, &tline, test_deltas.clone(), &ctx)
2292 12 : .await
2293 12 : .unwrap();
2294 12 : let delta_layer = resident_layer.get_as_delta(&ctx).await.unwrap();
2295 36 : for max_read_size in [1, 1024] {
2296 192 : for batch_size in [1, 2, 4, 8, 3, 7, 13] {
2297 168 : println!("running with batch_size={batch_size} max_read_size={max_read_size}");
2298 168 : // Test if the batch size is correctly determined
2299 168 : let mut iter = delta_layer.iter(&ctx);
2300 168 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2301 168 : let mut num_items = 0;
2302 672 : for _ in 0..3 {
2303 504 : iter.next_batch().await.unwrap();
2304 504 : num_items += iter.key_values_batch.len();
2305 504 : if max_read_size == 1 {
2306 12 : // every key should be a batch b/c the value is larger than max_read_size
2307 252 : assert_eq!(iter.key_values_batch.len(), 1);
2308 12 : } else {
2309 252 : assert!(iter.key_values_batch.len() <= batch_size);
2310 12 : }
2311 504 : if num_items >= N {
2312 12 : break;
2313 504 : }
2314 504 : iter.key_values_batch.clear();
2315 12 : }
2316 12 : // Test if the result is correct
2317 168 : let mut iter = delta_layer.iter(&ctx);
2318 168 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2319 168 : assert_delta_iter_equal(&mut iter, &test_deltas).await;
2320 12 : }
2321 12 : }
2322 12 : }
2323 : }
|
