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::ops::Range;
33 : use std::os::unix::fs::FileExt;
34 : use std::str::FromStr;
35 : use std::sync::Arc;
36 : use std::sync::atomic::AtomicU64;
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 serde::{Deserialize, Serialize};
49 : use tokio::sync::OnceCell;
50 : use tokio_epoll_uring::IoBuf;
51 : use tokio_util::sync::CancellationToken;
52 : use tracing::*;
53 : use utils::bin_ser::BeSer;
54 : use utils::bin_ser::SerializeError;
55 : use utils::id::{TenantId, TimelineId};
56 : use utils::lsn::Lsn;
57 :
58 : use super::{
59 : AsLayerDesc, LayerName, OnDiskValue, OnDiskValueIo, PersistentLayerDesc, ResidentLayer,
60 : ValuesReconstructState,
61 : };
62 : use crate::config::PageServerConf;
63 : use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
64 : use crate::page_cache::{self, FileId, PAGE_SZ};
65 : use crate::tenant::blob_io::BlobWriter;
66 : use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockLease, BlockReader, FileBlockReader};
67 : use crate::tenant::disk_btree::{
68 : DiskBtreeBuilder, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
69 : };
70 : use crate::tenant::storage_layer::layer::S3_UPLOAD_LIMIT;
71 : use crate::tenant::timeline::GetVectoredError;
72 : use crate::tenant::vectored_blob_io::{
73 : BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
74 : VectoredReadPlanner,
75 : };
76 : use crate::virtual_file::TempVirtualFile;
77 : use crate::virtual_file::owned_buffers_io::io_buf_ext::{FullSlice, IoBufExt};
78 : use crate::virtual_file::owned_buffers_io::write::{Buffer, BufferedWriterShutdownMode};
79 : use crate::virtual_file::{self, IoBuffer, IoBufferMut, MaybeFatalIo, VirtualFile};
80 : use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION, TEMP_FILE_SUFFIX};
81 :
82 : ///
83 : /// Header stored in the beginning of the file
84 : ///
85 : /// After this comes the 'values' part, starting on block 1. After that,
86 : /// the 'index' starts at the block indicated by 'index_start_blk'
87 : ///
88 0 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
89 : pub struct Summary {
90 : /// Magic value to identify this as a neon delta file. Always DELTA_FILE_MAGIC.
91 : pub magic: u16,
92 : pub format_version: u16,
93 :
94 : pub tenant_id: TenantId,
95 : pub timeline_id: TimelineId,
96 : pub key_range: Range<Key>,
97 : pub lsn_range: Range<Lsn>,
98 :
99 : /// Block number where the 'index' part of the file begins.
100 : pub index_start_blk: u32,
101 : /// Block within the 'index', where the B-tree root page is stored
102 : pub index_root_blk: u32,
103 : }
104 :
105 : impl From<&DeltaLayer> for Summary {
106 0 : fn from(layer: &DeltaLayer) -> Self {
107 0 : Self::expected(
108 0 : layer.desc.tenant_shard_id.tenant_id,
109 0 : layer.desc.timeline_id,
110 0 : layer.desc.key_range.clone(),
111 0 : layer.desc.lsn_range.clone(),
112 0 : )
113 0 : }
114 : }
115 :
116 : impl Summary {
117 : /// Serializes the summary header into an aligned buffer of lenth `PAGE_SZ`.
118 8700 : pub fn ser_into_page(&self) -> Result<IoBuffer, SerializeError> {
119 8700 : let mut buf = IoBufferMut::with_capacity(PAGE_SZ);
120 8700 : Self::ser_into(self, &mut buf)?;
121 : // Pad zeroes to the buffer so the length is a multiple of the alignment.
122 8700 : buf.extend_with(0, buf.capacity() - buf.len());
123 8700 : Ok(buf.freeze())
124 8700 : }
125 :
126 6612 : pub(super) fn expected(
127 6612 : tenant_id: TenantId,
128 6612 : timeline_id: TimelineId,
129 6612 : keys: Range<Key>,
130 6612 : lsns: Range<Lsn>,
131 6612 : ) -> Self {
132 6612 : Self {
133 6612 : magic: DELTA_FILE_MAGIC,
134 6612 : format_version: STORAGE_FORMAT_VERSION,
135 6612 :
136 6612 : tenant_id,
137 6612 : timeline_id,
138 6612 : key_range: keys,
139 6612 : lsn_range: lsns,
140 6612 :
141 6612 : index_start_blk: 0,
142 6612 : index_root_blk: 0,
143 6612 : }
144 6612 : }
145 : }
146 :
147 : // Flag indicating that this version initialize the page
148 : const WILL_INIT: u64 = 1;
149 :
150 : /// Struct representing reference to BLOB in layers.
151 : ///
152 : /// Reference contains BLOB offset, and for WAL records it also contains
153 : /// `will_init` flag. The flag helps to determine the range of records
154 : /// that needs to be applied, without reading/deserializing records themselves.
155 0 : #[derive(Debug, Serialize, Deserialize, Copy, Clone)]
156 : pub struct BlobRef(pub u64);
157 :
158 : impl BlobRef {
159 28830911 : pub fn will_init(&self) -> bool {
160 28830911 : (self.0 & WILL_INIT) != 0
161 28830911 : }
162 :
163 46437219 : pub fn pos(&self) -> u64 {
164 46437219 : self.0 >> 1
165 46437219 : }
166 :
167 38952492 : pub fn new(pos: u64, will_init: bool) -> BlobRef {
168 38952492 : let mut blob_ref = pos << 1;
169 38952492 : if will_init {
170 38820996 : blob_ref |= WILL_INIT;
171 38820996 : }
172 38952492 : BlobRef(blob_ref)
173 38952492 : }
174 : }
175 :
176 : pub const DELTA_KEY_SIZE: usize = KEY_SIZE + 8;
177 : struct DeltaKey([u8; DELTA_KEY_SIZE]);
178 :
179 : /// This is the key of the B-tree index stored in the delta layer. It consists
180 : /// of the serialized representation of a Key and LSN.
181 : impl DeltaKey {
182 12385188 : fn from_slice(buf: &[u8]) -> Self {
183 12385188 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
184 12385188 : bytes.copy_from_slice(buf);
185 12385188 : DeltaKey(bytes)
186 12385188 : }
187 :
188 40401979 : fn from_key_lsn(key: &Key, lsn: Lsn) -> Self {
189 40401979 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
190 40401979 : key.write_to_byte_slice(&mut bytes[0..KEY_SIZE]);
191 40401979 : bytes[KEY_SIZE..].copy_from_slice(&u64::to_be_bytes(lsn.0));
192 40401979 : DeltaKey(bytes)
193 40401979 : }
194 :
195 12385188 : fn key(&self) -> Key {
196 12385188 : Key::from_slice(&self.0)
197 12385188 : }
198 :
199 12385188 : fn lsn(&self) -> Lsn {
200 12385188 : Lsn(u64::from_be_bytes(self.0[KEY_SIZE..].try_into().unwrap()))
201 12385188 : }
202 :
203 34051671 : fn extract_lsn_from_buf(buf: &[u8]) -> Lsn {
204 34051671 : let mut lsn_buf = [0u8; 8];
205 34051671 : lsn_buf.copy_from_slice(&buf[KEY_SIZE..]);
206 34051671 : Lsn(u64::from_be_bytes(lsn_buf))
207 34051671 : }
208 : }
209 :
210 : /// This is used only from `pagectl`. Within pageserver, all layers are
211 : /// [`crate::tenant::storage_layer::Layer`], which can hold a [`DeltaLayerInner`].
212 : pub struct DeltaLayer {
213 : path: Utf8PathBuf,
214 : pub desc: PersistentLayerDesc,
215 : inner: OnceCell<Arc<DeltaLayerInner>>,
216 : }
217 :
218 : impl std::fmt::Debug for DeltaLayer {
219 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
220 : use super::RangeDisplayDebug;
221 :
222 0 : f.debug_struct("DeltaLayer")
223 0 : .field("key_range", &RangeDisplayDebug(&self.desc.key_range))
224 0 : .field("lsn_range", &self.desc.lsn_range)
225 0 : .field("file_size", &self.desc.file_size)
226 0 : .field("inner", &self.inner)
227 0 : .finish()
228 0 : }
229 : }
230 :
231 : /// `DeltaLayerInner` is the in-memory data structure associated with an on-disk delta
232 : /// file.
233 : pub struct DeltaLayerInner {
234 : // values copied from summary
235 : index_start_blk: u32,
236 : index_root_blk: u32,
237 :
238 : file: Arc<VirtualFile>,
239 : file_id: FileId,
240 :
241 : layer_key_range: Range<Key>,
242 : layer_lsn_range: Range<Lsn>,
243 :
244 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
245 : }
246 :
247 : impl DeltaLayerInner {
248 0 : pub(crate) fn layer_dbg_info(&self) -> String {
249 0 : format!(
250 0 : "delta {}..{} {}..{}",
251 0 : self.key_range().start,
252 0 : self.key_range().end,
253 0 : self.lsn_range().start,
254 0 : self.lsn_range().end
255 0 : )
256 0 : }
257 : }
258 :
259 : impl std::fmt::Debug for DeltaLayerInner {
260 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
261 0 : f.debug_struct("DeltaLayerInner")
262 0 : .field("index_start_blk", &self.index_start_blk)
263 0 : .field("index_root_blk", &self.index_root_blk)
264 0 : .finish()
265 0 : }
266 : }
267 :
268 : /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
269 : impl std::fmt::Display for DeltaLayer {
270 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
271 0 : write!(f, "{}", self.layer_desc().short_id())
272 0 : }
273 : }
274 :
275 : impl AsLayerDesc for DeltaLayer {
276 0 : fn layer_desc(&self) -> &PersistentLayerDesc {
277 0 : &self.desc
278 0 : }
279 : }
280 :
281 : impl DeltaLayer {
282 0 : pub async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
283 0 : self.desc.dump();
284 0 :
285 0 : if !verbose {
286 0 : return Ok(());
287 0 : }
288 :
289 0 : let inner = self.load(ctx).await?;
290 :
291 0 : inner.dump(ctx).await
292 0 : }
293 :
294 8856 : fn temp_path_for(
295 8856 : conf: &PageServerConf,
296 8856 : tenant_shard_id: &TenantShardId,
297 8856 : timeline_id: &TimelineId,
298 8856 : key_start: Key,
299 8856 : lsn_range: &Range<Lsn>,
300 8856 : ) -> Utf8PathBuf {
301 : // TempVirtualFile requires us to never reuse a filename while an old
302 : // instance of TempVirtualFile created with that filename is not done dropping yet.
303 : // So, we use a monotonic counter to disambiguate the filenames.
304 : static NEXT_TEMP_DISAMBIGUATOR: AtomicU64 = AtomicU64::new(1);
305 8856 : let filename_disambiguator =
306 8856 : NEXT_TEMP_DISAMBIGUATOR.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
307 8856 :
308 8856 : conf.timeline_path(tenant_shard_id, timeline_id)
309 8856 : .join(format!(
310 8856 : "{}-XXX__{:016X}-{:016X}.{:x}.{}",
311 8856 : key_start,
312 8856 : u64::from(lsn_range.start),
313 8856 : u64::from(lsn_range.end),
314 8856 : filename_disambiguator,
315 8856 : TEMP_FILE_SUFFIX,
316 8856 : ))
317 8856 : }
318 :
319 : ///
320 : /// Open the underlying file and read the metadata into memory, if it's
321 : /// not loaded already.
322 : ///
323 0 : async fn load(&self, ctx: &RequestContext) -> Result<&Arc<DeltaLayerInner>> {
324 0 : // Quick exit if already loaded
325 0 : self.inner
326 0 : .get_or_try_init(|| self.load_inner(ctx))
327 0 : .await
328 0 : .with_context(|| format!("Failed to load delta layer {}", self.path()))
329 0 : }
330 :
331 0 : async fn load_inner(&self, ctx: &RequestContext) -> anyhow::Result<Arc<DeltaLayerInner>> {
332 0 : let path = self.path();
333 :
334 0 : let loaded = DeltaLayerInner::load(&path, None, None, ctx).await?;
335 :
336 : // not production code
337 0 : let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
338 0 : let expected_layer_name = self.layer_desc().layer_name();
339 0 :
340 0 : if actual_layer_name != expected_layer_name {
341 0 : println!("warning: filename does not match what is expected from in-file summary");
342 0 : println!("actual: {:?}", actual_layer_name.to_string());
343 0 : println!("expected: {:?}", expected_layer_name.to_string());
344 0 : }
345 :
346 0 : Ok(Arc::new(loaded))
347 0 : }
348 :
349 : /// Create a DeltaLayer struct representing an existing file on disk.
350 : ///
351 : /// This variant is only used for debugging purposes, by the 'pagectl' binary.
352 0 : pub fn new_for_path(path: &Utf8Path, file: File) -> Result<Self> {
353 0 : let mut summary_buf = vec![0; PAGE_SZ];
354 0 : file.read_exact_at(&mut summary_buf, 0)?;
355 0 : let summary = Summary::des_prefix(&summary_buf)?;
356 :
357 0 : let metadata = file
358 0 : .metadata()
359 0 : .context("get file metadata to determine size")?;
360 :
361 : // This function is never used for constructing layers in a running pageserver,
362 : // so it does not need an accurate TenantShardId.
363 0 : let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
364 0 :
365 0 : Ok(DeltaLayer {
366 0 : path: path.to_path_buf(),
367 0 : desc: PersistentLayerDesc::new_delta(
368 0 : tenant_shard_id,
369 0 : summary.timeline_id,
370 0 : summary.key_range,
371 0 : summary.lsn_range,
372 0 : metadata.len(),
373 0 : ),
374 0 : inner: OnceCell::new(),
375 0 : })
376 0 : }
377 :
378 : /// Path to the layer file in pageserver workdir.
379 0 : fn path(&self) -> Utf8PathBuf {
380 0 : self.path.clone()
381 0 : }
382 : }
383 :
384 : /// A builder object for constructing a new delta layer.
385 : ///
386 : /// Usage:
387 : ///
388 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
389 : ///
390 : /// 2. Write the contents by calling `put_value` for every page
391 : /// version to store in the layer.
392 : ///
393 : /// 3. Call `finish`.
394 : ///
395 : struct DeltaLayerWriterInner {
396 : pub path: Utf8PathBuf,
397 : timeline_id: TimelineId,
398 : tenant_shard_id: TenantShardId,
399 :
400 : key_start: Key,
401 : lsn_range: Range<Lsn>,
402 :
403 : tree: DiskBtreeBuilder<BlockBuf, DELTA_KEY_SIZE>,
404 :
405 : blob_writer: BlobWriter<TempVirtualFile>,
406 :
407 : // Number of key-lsns in the layer.
408 : num_keys: usize,
409 : }
410 :
411 : impl DeltaLayerWriterInner {
412 : ///
413 : /// Start building a new delta layer.
414 : ///
415 : #[allow(clippy::too_many_arguments)]
416 8856 : async fn new(
417 8856 : conf: &'static PageServerConf,
418 8856 : timeline_id: TimelineId,
419 8856 : tenant_shard_id: TenantShardId,
420 8856 : key_start: Key,
421 8856 : lsn_range: Range<Lsn>,
422 8856 : gate: &utils::sync::gate::Gate,
423 8856 : cancel: CancellationToken,
424 8856 : ctx: &RequestContext,
425 8856 : ) -> anyhow::Result<Self> {
426 8856 : // Create the file initially with a temporary filename. We don't know
427 8856 : // the end key yet, so we cannot form the final filename yet. We will
428 8856 : // rename it when we're done.
429 8856 : let path =
430 8856 : DeltaLayer::temp_path_for(conf, &tenant_shard_id, &timeline_id, key_start, &lsn_range);
431 8856 : let file = TempVirtualFile::new(
432 8856 : VirtualFile::open_with_options_v2(
433 8856 : &path,
434 8856 : virtual_file::OpenOptions::new()
435 8856 : .create_new(true)
436 8856 : .write(true),
437 8856 : ctx,
438 8856 : )
439 8856 : .await?,
440 8856 : gate.enter()?,
441 : );
442 :
443 : // Start at PAGE_SZ, make room for the header block
444 8856 : let blob_writer = BlobWriter::new(
445 8856 : file,
446 8856 : PAGE_SZ as u64,
447 8856 : gate,
448 8856 : cancel,
449 8856 : ctx,
450 8856 : info_span!(parent: None, "delta_layer_writer_flush_task", tenant_id=%tenant_shard_id.tenant_id, shard_id=%tenant_shard_id.shard_slug(), timeline_id=%timeline_id, path = %path),
451 0 : )?;
452 :
453 : // Initialize the b-tree index builder
454 8856 : let block_buf = BlockBuf::new();
455 8856 : let tree_builder = DiskBtreeBuilder::new(block_buf);
456 8856 :
457 8856 : Ok(Self {
458 8856 : path,
459 8856 : timeline_id,
460 8856 : tenant_shard_id,
461 8856 : key_start,
462 8856 : lsn_range,
463 8856 : tree: tree_builder,
464 8856 : blob_writer,
465 8856 : num_keys: 0,
466 8856 : })
467 8856 : }
468 :
469 : ///
470 : /// Append a key-value pair to the file.
471 : ///
472 : /// The values must be appended in key, lsn order.
473 : ///
474 12631908 : async fn put_value(
475 12631908 : &mut self,
476 12631908 : key: Key,
477 12631908 : lsn: Lsn,
478 12631908 : val: Value,
479 12631908 : ctx: &RequestContext,
480 12631908 : ) -> anyhow::Result<()> {
481 12631908 : let (_, res) = self
482 12631908 : .put_value_bytes(
483 12631908 : key,
484 12631908 : lsn,
485 12631908 : Value::ser(&val)?.slice_len(),
486 12631908 : val.will_init(),
487 12631908 : ctx,
488 12631908 : )
489 12631908 : .await;
490 12631908 : res
491 12631908 : }
492 :
493 38951868 : async fn put_value_bytes<Buf>(
494 38951868 : &mut self,
495 38951868 : key: Key,
496 38951868 : lsn: Lsn,
497 38951868 : val: FullSlice<Buf>,
498 38951868 : will_init: bool,
499 38951868 : ctx: &RequestContext,
500 38951868 : ) -> (FullSlice<Buf>, anyhow::Result<()>)
501 38951868 : where
502 38951868 : Buf: IoBuf + Send,
503 38951868 : {
504 38951868 : assert!(
505 38951868 : self.lsn_range.start <= lsn,
506 0 : "lsn_start={}, lsn={}",
507 : self.lsn_range.start,
508 : lsn
509 : );
510 : // We don't want to use compression in delta layer creation
511 38951868 : let compression = ImageCompressionAlgorithm::Disabled;
512 38951868 : let (val, res) = self
513 38951868 : .blob_writer
514 38951868 : .write_blob_maybe_compressed(val, ctx, compression)
515 38951868 : .await;
516 38951868 : let off = match res {
517 38951868 : Ok((off, _)) => off,
518 0 : Err(e) => return (val, Err(anyhow::anyhow!(e))),
519 : };
520 :
521 38951868 : let blob_ref = BlobRef::new(off, will_init);
522 38951868 :
523 38951868 : let delta_key = DeltaKey::from_key_lsn(&key, lsn);
524 38951868 : let res = self.tree.append(&delta_key.0, blob_ref.0);
525 38951868 :
526 38951868 : self.num_keys += 1;
527 38951868 :
528 38951868 : (val, res.map_err(|e| anyhow::anyhow!(e)))
529 38951868 : }
530 :
531 12143832 : fn size(&self) -> u64 {
532 12143832 : self.blob_writer.size() + self.tree.borrow_writer().size()
533 12143832 : }
534 :
535 : ///
536 : /// Finish writing the delta layer.
537 : ///
538 8700 : async fn finish(
539 8700 : self,
540 8700 : key_end: Key,
541 8700 : ctx: &RequestContext,
542 8700 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
543 8700 : let index_start_blk = self.blob_writer.size().div_ceil(PAGE_SZ as u64) as u32;
544 :
545 8700 : let file = self
546 8700 : .blob_writer
547 8700 : .shutdown(
548 8700 : BufferedWriterShutdownMode::ZeroPadToNextMultiple(PAGE_SZ),
549 8700 : ctx,
550 8700 : )
551 8700 : .await?;
552 :
553 : // Write out the index
554 8700 : let (index_root_blk, block_buf) = self.tree.finish()?;
555 8700 : let mut offset = index_start_blk as u64 * PAGE_SZ as u64;
556 :
557 : // TODO(yuchen): https://github.com/neondatabase/neon/issues/10092
558 : // Should we just replace BlockBuf::blocks with one big buffer
559 90911 : for buf in block_buf.blocks {
560 82211 : let (_buf, res) = file.write_all_at(buf.slice_len(), offset, ctx).await;
561 82211 : res?;
562 82211 : offset += PAGE_SZ as u64;
563 : }
564 8700 : assert!(self.lsn_range.start < self.lsn_range.end);
565 : // Fill in the summary on blk 0
566 8700 : let summary = Summary {
567 8700 : magic: DELTA_FILE_MAGIC,
568 8700 : format_version: STORAGE_FORMAT_VERSION,
569 8700 : tenant_id: self.tenant_shard_id.tenant_id,
570 8700 : timeline_id: self.timeline_id,
571 8700 : key_range: self.key_start..key_end,
572 8700 : lsn_range: self.lsn_range.clone(),
573 8700 : index_start_blk,
574 8700 : index_root_blk,
575 8700 : };
576 :
577 : // Writes summary at the first block (offset 0).
578 8700 : let buf = summary.ser_into_page()?;
579 8700 : let (_buf, res) = file.write_all_at(buf.slice_len(), 0, ctx).await;
580 8700 : res?;
581 :
582 8700 : let metadata = file
583 8700 : .metadata()
584 8700 : .await
585 8700 : .context("get file metadata to determine size")?;
586 :
587 : // 5GB limit for objects without multipart upload (which we don't want to use)
588 : // Make it a little bit below to account for differing GB units
589 : // https://docs.aws.amazon.com/AmazonS3/latest/userguide/upload-objects.html
590 8700 : ensure!(
591 8700 : metadata.len() <= S3_UPLOAD_LIMIT,
592 0 : "Created delta layer file at {} of size {} above limit {S3_UPLOAD_LIMIT}!",
593 0 : file.path(),
594 0 : metadata.len()
595 : );
596 :
597 : // Note: Because we opened the file in write-only mode, we cannot
598 : // reuse the same VirtualFile for reading later. That's why we don't
599 : // set inner.file here. The first read will have to re-open it.
600 :
601 8700 : let desc = PersistentLayerDesc::new_delta(
602 8700 : self.tenant_shard_id,
603 8700 : self.timeline_id,
604 8700 : self.key_start..key_end,
605 8700 : self.lsn_range.clone(),
606 8700 : metadata.len(),
607 8700 : );
608 8700 :
609 8700 : // fsync the file
610 8700 : file.sync_all()
611 8700 : .await
612 8700 : .maybe_fatal_err("delta_layer sync_all")?;
613 :
614 8700 : trace!("created delta layer {}", self.path);
615 :
616 : // The gate guard stored in `destination_file` is dropped. Callers (e.g.. flush loop or compaction)
617 : // keep the gate open also, so that it's safe for them to rename the file to its final destination.
618 8700 : file.disarm_into_inner();
619 8700 :
620 8700 : Ok((desc, self.path))
621 8700 : }
622 : }
623 :
624 : /// A builder object for constructing a new delta layer.
625 : ///
626 : /// Usage:
627 : ///
628 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
629 : ///
630 : /// 2. Write the contents by calling `put_value` for every page
631 : /// version to store in the layer.
632 : ///
633 : /// 3. Call `finish`.
634 : ///
635 : /// # Note
636 : ///
637 : /// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
638 : /// possible for the writer to drop before `finish` is actually called. So this
639 : /// could lead to odd temporary files in the directory, exhausting file system.
640 : /// This structure wraps `DeltaLayerWriterInner` and also contains `Drop`
641 : /// implementation that cleans up the temporary file in failure. It's not
642 : /// possible to do this directly in `DeltaLayerWriterInner` since `finish` moves
643 : /// out some fields, making it impossible to implement `Drop`.
644 : ///
645 : #[must_use]
646 : pub struct DeltaLayerWriter {
647 : inner: Option<DeltaLayerWriterInner>,
648 : }
649 :
650 : impl DeltaLayerWriter {
651 : ///
652 : /// Start building a new delta layer.
653 : ///
654 : #[allow(clippy::too_many_arguments)]
655 8856 : pub async fn new(
656 8856 : conf: &'static PageServerConf,
657 8856 : timeline_id: TimelineId,
658 8856 : tenant_shard_id: TenantShardId,
659 8856 : key_start: Key,
660 8856 : lsn_range: Range<Lsn>,
661 8856 : gate: &utils::sync::gate::Gate,
662 8856 : cancel: CancellationToken,
663 8856 : ctx: &RequestContext,
664 8856 : ) -> anyhow::Result<Self> {
665 8856 : Ok(Self {
666 8856 : inner: Some(
667 8856 : DeltaLayerWriterInner::new(
668 8856 : conf,
669 8856 : timeline_id,
670 8856 : tenant_shard_id,
671 8856 : key_start,
672 8856 : lsn_range,
673 8856 : gate,
674 8856 : cancel,
675 8856 : ctx,
676 8856 : )
677 8856 : .await?,
678 : ),
679 : })
680 8856 : }
681 :
682 0 : pub fn is_empty(&self) -> bool {
683 0 : self.inner.as_ref().unwrap().num_keys == 0
684 0 : }
685 :
686 : ///
687 : /// Append a key-value pair to the file.
688 : ///
689 : /// The values must be appended in key, lsn order.
690 : ///
691 12631908 : pub async fn put_value(
692 12631908 : &mut self,
693 12631908 : key: Key,
694 12631908 : lsn: Lsn,
695 12631908 : val: Value,
696 12631908 : ctx: &RequestContext,
697 12631908 : ) -> anyhow::Result<()> {
698 12631908 : self.inner
699 12631908 : .as_mut()
700 12631908 : .unwrap()
701 12631908 : .put_value(key, lsn, val, ctx)
702 12631908 : .await
703 12631908 : }
704 :
705 26319960 : pub async fn put_value_bytes<Buf>(
706 26319960 : &mut self,
707 26319960 : key: Key,
708 26319960 : lsn: Lsn,
709 26319960 : val: FullSlice<Buf>,
710 26319960 : will_init: bool,
711 26319960 : ctx: &RequestContext,
712 26319960 : ) -> (FullSlice<Buf>, anyhow::Result<()>)
713 26319960 : where
714 26319960 : Buf: IoBuf + Send,
715 26319960 : {
716 26319960 : self.inner
717 26319960 : .as_mut()
718 26319960 : .unwrap()
719 26319960 : .put_value_bytes(key, lsn, val, will_init, ctx)
720 26319960 : .await
721 26319960 : }
722 :
723 12143832 : pub fn size(&self) -> u64 {
724 12143832 : self.inner.as_ref().unwrap().size()
725 12143832 : }
726 :
727 : ///
728 : /// Finish writing the delta layer.
729 : ///
730 8700 : pub(crate) async fn finish(
731 8700 : mut self,
732 8700 : key_end: Key,
733 8700 : ctx: &RequestContext,
734 8700 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
735 8700 : self.inner.take().unwrap().finish(key_end, ctx).await
736 8700 : }
737 :
738 73524 : pub(crate) fn num_keys(&self) -> usize {
739 73524 : self.inner.as_ref().unwrap().num_keys
740 73524 : }
741 :
742 90840 : pub(crate) fn estimated_size(&self) -> u64 {
743 90840 : let inner = self.inner.as_ref().unwrap();
744 90840 : inner.blob_writer.size() + inner.tree.borrow_writer().size() + PAGE_SZ as u64
745 90840 : }
746 : }
747 :
748 : #[derive(thiserror::Error, Debug)]
749 : pub enum RewriteSummaryError {
750 : #[error("magic mismatch")]
751 : MagicMismatch,
752 : #[error(transparent)]
753 : Other(#[from] anyhow::Error),
754 : }
755 :
756 : impl From<std::io::Error> for RewriteSummaryError {
757 0 : fn from(e: std::io::Error) -> Self {
758 0 : Self::Other(anyhow::anyhow!(e))
759 0 : }
760 : }
761 :
762 : impl DeltaLayer {
763 0 : pub async fn rewrite_summary<F>(
764 0 : path: &Utf8Path,
765 0 : rewrite: F,
766 0 : ctx: &RequestContext,
767 0 : ) -> Result<(), RewriteSummaryError>
768 0 : where
769 0 : F: Fn(Summary) -> Summary,
770 0 : {
771 0 : let file = VirtualFile::open_with_options_v2(
772 0 : path,
773 0 : virtual_file::OpenOptions::new().read(true).write(true),
774 0 : ctx,
775 0 : )
776 0 : .await
777 0 : .with_context(|| format!("Failed to open file '{}'", path))?;
778 0 : let file_id = page_cache::next_file_id();
779 0 : let block_reader = FileBlockReader::new(&file, file_id);
780 0 : let summary_blk = block_reader.read_blk(0, ctx).await?;
781 0 : let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
782 0 : if actual_summary.magic != DELTA_FILE_MAGIC {
783 0 : return Err(RewriteSummaryError::MagicMismatch);
784 0 : }
785 0 :
786 0 : let new_summary = rewrite(actual_summary);
787 :
788 0 : let buf = new_summary.ser_into_page().context("serialize")?;
789 0 : let (_buf, res) = file.write_all_at(buf.slice_len(), 0, ctx).await;
790 0 : res?;
791 0 : Ok(())
792 0 : }
793 : }
794 :
795 : impl DeltaLayerInner {
796 6240 : pub(crate) fn key_range(&self) -> &Range<Key> {
797 6240 : &self.layer_key_range
798 6240 : }
799 :
800 6240 : pub(crate) fn lsn_range(&self) -> &Range<Lsn> {
801 6240 : &self.layer_lsn_range
802 6240 : }
803 :
804 6612 : pub(super) async fn load(
805 6612 : path: &Utf8Path,
806 6612 : summary: Option<Summary>,
807 6612 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
808 6612 : ctx: &RequestContext,
809 6612 : ) -> anyhow::Result<Self> {
810 6612 : let file = Arc::new(
811 6612 : VirtualFile::open_v2(path, ctx)
812 6612 : .await
813 6612 : .context("open layer file")?,
814 : );
815 :
816 6612 : let file_id = page_cache::next_file_id();
817 6612 :
818 6612 : let block_reader = FileBlockReader::new(&file, file_id);
819 :
820 6612 : let summary_blk = block_reader
821 6612 : .read_blk(0, ctx)
822 6612 : .await
823 6612 : .context("read first block")?;
824 :
825 : // TODO: this should be an assertion instead; see ImageLayerInner::load
826 6612 : let actual_summary =
827 6612 : Summary::des_prefix(summary_blk.as_ref()).context("deserialize first block")?;
828 :
829 6612 : if let Some(mut expected_summary) = summary {
830 : // production code path
831 6612 : expected_summary.index_start_blk = actual_summary.index_start_blk;
832 6612 : expected_summary.index_root_blk = actual_summary.index_root_blk;
833 6612 : // mask out the timeline_id, but still require the layers to be from the same tenant
834 6612 : expected_summary.timeline_id = actual_summary.timeline_id;
835 6612 :
836 6612 : if actual_summary != expected_summary {
837 0 : bail!(
838 0 : "in-file summary does not match expected summary. actual = {:?} expected = {:?}",
839 0 : actual_summary,
840 0 : expected_summary
841 0 : );
842 6612 : }
843 0 : }
844 :
845 6612 : Ok(DeltaLayerInner {
846 6612 : file,
847 6612 : file_id,
848 6612 : index_start_blk: actual_summary.index_start_blk,
849 6612 : index_root_blk: actual_summary.index_root_blk,
850 6612 : max_vectored_read_bytes,
851 6612 : layer_key_range: actual_summary.key_range,
852 6612 : layer_lsn_range: actual_summary.lsn_range,
853 6612 : })
854 6612 : }
855 :
856 : // Look up the keys in the provided keyspace and update
857 : // the reconstruct state with whatever is found.
858 : //
859 : // Currently, the index is visited for each range, but this
860 : // can be further optimised to visit the index only once.
861 1417975 : pub(super) async fn get_values_reconstruct_data(
862 1417975 : &self,
863 1417975 : this: ResidentLayer,
864 1417975 : keyspace: KeySpace,
865 1417975 : lsn_range: Range<Lsn>,
866 1417975 : reconstruct_state: &mut ValuesReconstructState,
867 1417975 : ctx: &RequestContext,
868 1417975 : ) -> Result<(), GetVectoredError> {
869 1417975 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
870 1417975 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
871 1417975 : self.index_start_blk,
872 1417975 : self.index_root_blk,
873 1417975 : block_reader,
874 1417975 : );
875 1417975 :
876 1417975 : let planner = VectoredReadPlanner::new(
877 1417975 : self.max_vectored_read_bytes
878 1417975 : .expect("Layer is loaded with max vectored bytes config")
879 1417975 : .0
880 1417975 : .into(),
881 1417975 : );
882 1417975 :
883 1417975 : let data_end_offset = self.index_start_offset();
884 :
885 1417975 : let reads = Self::plan_reads(
886 1417975 : &keyspace,
887 1417975 : lsn_range.clone(),
888 1417975 : data_end_offset,
889 1417975 : index_reader,
890 1417975 : planner,
891 1417975 : ctx,
892 1417975 : )
893 1417975 : .await
894 1417975 : .map_err(GetVectoredError::Other)?;
895 :
896 1417975 : self.do_reads_and_update_state(this, reads, reconstruct_state, ctx)
897 1417975 : .await;
898 :
899 1417975 : Ok(())
900 1417975 : }
901 :
902 1419187 : async fn plan_reads<Reader>(
903 1419187 : keyspace: &KeySpace,
904 1419187 : lsn_range: Range<Lsn>,
905 1419187 : data_end_offset: u64,
906 1419187 : index_reader: DiskBtreeReader<Reader, DELTA_KEY_SIZE>,
907 1419187 : mut planner: VectoredReadPlanner,
908 1419187 : ctx: &RequestContext,
909 1419187 : ) -> anyhow::Result<Vec<VectoredRead>>
910 1419187 : where
911 1419187 : Reader: BlockReader + Clone,
912 1419187 : {
913 1419187 : let ctx = RequestContextBuilder::from(ctx)
914 1419187 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
915 1419187 : .attached_child();
916 :
917 1449907 : for range in keyspace.ranges.iter() {
918 1449907 : let mut range_end_handled = false;
919 1449907 :
920 1449907 : let start_key = DeltaKey::from_key_lsn(&range.start, lsn_range.start);
921 1449907 : let index_stream = index_reader.clone().into_stream(&start_key.0, &ctx);
922 1449907 : let mut index_stream = std::pin::pin!(index_stream);
923 :
924 21509173 : while let Some(index_entry) = index_stream.next().await {
925 21457179 : let (raw_key, value) = index_entry?;
926 21457179 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
927 21457179 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
928 21457179 : let blob_ref = BlobRef(value);
929 21457179 :
930 21457179 : // Lsns are not monotonically increasing across keys, so we don't assert on them.
931 21457179 : assert!(key >= range.start);
932 :
933 21457179 : let outside_lsn_range = !lsn_range.contains(&lsn);
934 :
935 21457179 : let flag = {
936 21457179 : if outside_lsn_range {
937 5220760 : BlobFlag::Ignore
938 16236419 : } else if blob_ref.will_init() {
939 3292703 : BlobFlag::ReplaceAll
940 : } else {
941 : // Usual path: add blob to the read
942 12943716 : BlobFlag::None
943 : }
944 : };
945 :
946 21457179 : if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
947 1397913 : planner.handle_range_end(blob_ref.pos());
948 1397913 : range_end_handled = true;
949 1397913 : break;
950 20059266 : } else {
951 20059266 : planner.handle(key, lsn, blob_ref.pos(), flag);
952 20059266 : }
953 : }
954 :
955 1449907 : if !range_end_handled {
956 51994 : tracing::debug!("Handling range end fallback at {}", data_end_offset);
957 51994 : planner.handle_range_end(data_end_offset);
958 1397913 : }
959 : }
960 :
961 1419187 : Ok(planner.finish())
962 1419187 : }
963 :
964 1419175 : fn get_min_read_buffer_size(
965 1419175 : planned_reads: &[VectoredRead],
966 1419175 : read_size_soft_max: usize,
967 1419175 : ) -> usize {
968 1419175 : let Some(largest_read) = planned_reads.iter().max_by_key(|read| read.size()) else {
969 521640 : return read_size_soft_max;
970 : };
971 :
972 897535 : let largest_read_size = largest_read.size();
973 897535 : if largest_read_size > read_size_soft_max {
974 : // If the read is oversized, it should only contain one key.
975 1200 : let offenders = largest_read
976 1200 : .blobs_at
977 1200 : .as_slice()
978 1200 : .iter()
979 1200 : .filter_map(|(_, blob_meta)| {
980 1200 : if blob_meta.key.is_rel_dir_key()
981 1200 : || blob_meta.key == DBDIR_KEY
982 1200 : || blob_meta.key.is_aux_file_key()
983 : {
984 : // The size of values for these keys is unbounded and can
985 : // grow very large in pathological cases.
986 0 : None
987 : } else {
988 1200 : Some(format!("{}@{}", blob_meta.key, blob_meta.lsn))
989 : }
990 1200 : })
991 1200 : .join(", ");
992 1200 :
993 1200 : if !offenders.is_empty() {
994 1200 : tracing::warn!(
995 0 : "Oversized vectored read ({} > {}) for keys {}",
996 : largest_read_size,
997 : read_size_soft_max,
998 : offenders
999 : );
1000 0 : }
1001 896335 : }
1002 :
1003 897535 : largest_read_size
1004 1419175 : }
1005 :
1006 1417975 : async fn do_reads_and_update_state(
1007 1417975 : &self,
1008 1417975 : this: ResidentLayer,
1009 1417975 : reads: Vec<VectoredRead>,
1010 1417975 : reconstruct_state: &mut ValuesReconstructState,
1011 1417975 : ctx: &RequestContext,
1012 1417975 : ) {
1013 1417975 : let max_vectored_read_bytes = self
1014 1417975 : .max_vectored_read_bytes
1015 1417975 : .expect("Layer is loaded with max vectored bytes config")
1016 1417975 : .0
1017 1417975 : .into();
1018 1417975 : let buf_size = Self::get_min_read_buffer_size(&reads, max_vectored_read_bytes);
1019 :
1020 : // Note that reads are processed in reverse order (from highest key+lsn).
1021 : // This is the order that `ReconstructState` requires such that it can
1022 : // track when a key is done.
1023 1417975 : for read in reads.into_iter().rev() {
1024 1021751 : let mut ios: HashMap<(Key, Lsn), OnDiskValueIo> = Default::default();
1025 9442286 : for (_, blob_meta) in read.blobs_at.as_slice().iter().rev() {
1026 9442286 : let io = reconstruct_state.update_key(
1027 9442286 : &blob_meta.key,
1028 9442286 : blob_meta.lsn,
1029 9442286 : blob_meta.will_init,
1030 9442286 : );
1031 9442286 : ios.insert((blob_meta.key, blob_meta.lsn), io);
1032 9442286 : }
1033 :
1034 1021751 : let read_extend_residency = this.clone();
1035 1021751 : let read_from = self.file.clone();
1036 1021751 : let read_ctx = ctx.attached_child();
1037 1021751 : reconstruct_state
1038 1021751 : .spawn_io(async move {
1039 1021751 : let vectored_blob_reader = VectoredBlobReader::new(&read_from);
1040 1021751 : let buf = IoBufferMut::with_capacity(buf_size);
1041 :
1042 1021751 : let res = vectored_blob_reader.read_blobs(&read, buf, &read_ctx).await;
1043 1021751 : match res {
1044 1021751 : Ok(blobs_buf) => {
1045 1021751 : let view = BufView::new_slice(&blobs_buf.buf);
1046 9442286 : for meta in blobs_buf.blobs.iter().rev() {
1047 9442286 : let io = ios.remove(&(meta.meta.key, meta.meta.lsn)).unwrap();
1048 :
1049 9442286 : let blob_read = meta.read(&view).await;
1050 9442286 : let blob_read = match blob_read {
1051 9442286 : Ok(buf) => buf,
1052 0 : Err(e) => {
1053 0 : io.complete(Err(e));
1054 0 : continue;
1055 : }
1056 : };
1057 :
1058 9442286 : io.complete(Ok(OnDiskValue::WalRecordOrImage(
1059 9442286 : blob_read.into_bytes(),
1060 9442286 : )));
1061 : }
1062 :
1063 1021751 : assert!(ios.is_empty());
1064 : }
1065 0 : Err(err) => {
1066 0 : for (_, sender) in ios {
1067 0 : sender.complete(Err(std::io::Error::new(
1068 0 : err.kind(),
1069 0 : "vec read failed",
1070 0 : )));
1071 0 : }
1072 : }
1073 : }
1074 :
1075 : // keep layer resident until this IO is done; this spawned IO future generally outlives the
1076 : // call to `self` / the `Arc<DownloadedLayer>` / the `ResidentLayer` that guarantees residency
1077 1021751 : drop(read_extend_residency);
1078 1021751 : })
1079 1021751 : .await;
1080 : }
1081 1417975 : }
1082 :
1083 2436 : pub(crate) async fn index_entries<'a>(
1084 2436 : &'a self,
1085 2436 : ctx: &RequestContext,
1086 2436 : ) -> Result<Vec<DeltaEntry<'a>>> {
1087 2436 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1088 2436 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1089 2436 : self.index_start_blk,
1090 2436 : self.index_root_blk,
1091 2436 : block_reader,
1092 2436 : );
1093 2436 :
1094 2436 : let mut all_keys: Vec<DeltaEntry<'_>> = Vec::new();
1095 2436 :
1096 2436 : tree_reader
1097 2436 : .visit(
1098 2436 : &[0u8; DELTA_KEY_SIZE],
1099 2436 : VisitDirection::Forwards,
1100 12384276 : |key, value| {
1101 12384276 : let delta_key = DeltaKey::from_slice(key);
1102 12384276 : let val_ref = ValueRef {
1103 12384276 : blob_ref: BlobRef(value),
1104 12384276 : layer: self,
1105 12384276 : };
1106 12384276 : let pos = BlobRef(value).pos();
1107 12384276 : if let Some(last) = all_keys.last_mut() {
1108 12381840 : // subtract offset of the current and last entries to get the size
1109 12381840 : // of the value associated with this (key, lsn) tuple
1110 12381840 : let first_pos = last.size;
1111 12381840 : last.size = pos - first_pos;
1112 12381840 : }
1113 12384276 : let entry = DeltaEntry {
1114 12384276 : key: delta_key.key(),
1115 12384276 : lsn: delta_key.lsn(),
1116 12384276 : size: pos,
1117 12384276 : val: val_ref,
1118 12384276 : };
1119 12384276 : all_keys.push(entry);
1120 12384276 : true
1121 12384276 : },
1122 2436 : &RequestContextBuilder::from(ctx)
1123 2436 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
1124 2436 : .attached_child(),
1125 2436 : )
1126 2436 : .await?;
1127 2436 : if let Some(last) = all_keys.last_mut() {
1128 2436 : // Last key occupies all space till end of value storage,
1129 2436 : // which corresponds to beginning of the index
1130 2436 : last.size = self.index_start_offset() - last.size;
1131 2436 : }
1132 2436 : Ok(all_keys)
1133 2436 : }
1134 :
1135 : /// Using the given writer, write out a version which has the earlier Lsns than `until`.
1136 : ///
1137 : /// Return the amount of key value records pushed to the writer.
1138 60 : pub(super) async fn copy_prefix(
1139 60 : &self,
1140 60 : writer: &mut DeltaLayerWriter,
1141 60 : until: Lsn,
1142 60 : ctx: &RequestContext,
1143 60 : ) -> anyhow::Result<usize> {
1144 : use futures::stream::TryStreamExt;
1145 :
1146 : use crate::tenant::vectored_blob_io::{
1147 : BlobMeta, ChunkedVectoredReadBuilder, VectoredReadExtended,
1148 : };
1149 :
1150 : #[derive(Debug)]
1151 : enum Item {
1152 : Actual(Key, Lsn, BlobRef),
1153 : Sentinel,
1154 : }
1155 :
1156 : impl From<Item> for Option<(Key, Lsn, BlobRef)> {
1157 420 : fn from(value: Item) -> Self {
1158 420 : match value {
1159 360 : Item::Actual(key, lsn, blob) => Some((key, lsn, blob)),
1160 60 : Item::Sentinel => None,
1161 : }
1162 420 : }
1163 : }
1164 :
1165 : impl Item {
1166 420 : fn offset(&self) -> Option<BlobRef> {
1167 420 : match self {
1168 360 : Item::Actual(_, _, blob) => Some(*blob),
1169 60 : Item::Sentinel => None,
1170 : }
1171 420 : }
1172 :
1173 420 : fn is_last(&self) -> bool {
1174 420 : matches!(self, Item::Sentinel)
1175 420 : }
1176 : }
1177 :
1178 60 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1179 60 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1180 60 : self.index_start_blk,
1181 60 : self.index_root_blk,
1182 60 : block_reader,
1183 60 : );
1184 60 :
1185 60 : let stream = self.stream_index_forwards(tree_reader, &[0u8; DELTA_KEY_SIZE], ctx);
1186 360 : let stream = stream.map_ok(|(key, lsn, pos)| Item::Actual(key, lsn, pos));
1187 60 : // put in a sentinel value for getting the end offset for last item, and not having to
1188 60 : // repeat the whole read part
1189 60 : let stream = stream.chain(futures::stream::once(futures::future::ready(Ok(
1190 60 : Item::Sentinel,
1191 60 : ))));
1192 60 : let mut stream = std::pin::pin!(stream);
1193 60 :
1194 60 : let mut prev: Option<(Key, Lsn, BlobRef)> = None;
1195 60 :
1196 60 : let mut read_builder: Option<ChunkedVectoredReadBuilder> = None;
1197 60 :
1198 60 : let max_read_size = self
1199 60 : .max_vectored_read_bytes
1200 60 : .map(|x| x.0.get())
1201 60 : .unwrap_or(8192);
1202 60 :
1203 60 : let mut buffer = Some(IoBufferMut::with_capacity(max_read_size));
1204 60 :
1205 60 : // FIXME: buffering of DeltaLayerWriter
1206 60 : let mut per_blob_copy = Vec::new();
1207 60 :
1208 60 : let mut records = 0;
1209 :
1210 480 : while let Some(item) = stream.try_next().await? {
1211 420 : tracing::debug!(?item, "popped");
1212 420 : let offset = item
1213 420 : .offset()
1214 420 : .unwrap_or(BlobRef::new(self.index_start_offset(), false));
1215 :
1216 420 : let actionable = if let Some((key, lsn, start_offset)) = prev.take() {
1217 360 : let end_offset = offset;
1218 360 :
1219 360 : Some((
1220 360 : BlobMeta {
1221 360 : key,
1222 360 : lsn,
1223 360 : will_init: false,
1224 360 : },
1225 360 : start_offset..end_offset,
1226 360 : ))
1227 : } else {
1228 60 : None
1229 : };
1230 :
1231 420 : let is_last = item.is_last();
1232 420 :
1233 420 : prev = Option::from(item);
1234 420 :
1235 420 : let actionable = actionable.filter(|x| x.0.lsn < until);
1236 :
1237 420 : let builder = if let Some((meta, offsets)) = actionable {
1238 : // extend or create a new builder
1239 192 : if read_builder
1240 192 : .as_mut()
1241 192 : .map(|x| x.extend(offsets.start.pos(), offsets.end.pos(), meta))
1242 192 : .unwrap_or(VectoredReadExtended::No)
1243 192 : == VectoredReadExtended::Yes
1244 : {
1245 96 : None
1246 : } else {
1247 96 : read_builder.replace(ChunkedVectoredReadBuilder::new(
1248 96 : offsets.start.pos(),
1249 96 : offsets.end.pos(),
1250 96 : meta,
1251 96 : max_read_size,
1252 96 : ))
1253 : }
1254 : } else {
1255 : // nothing to do, except perhaps flush any existing for the last element
1256 228 : None
1257 : };
1258 :
1259 : // flush the possible older builder and also the new one if the item was the last one
1260 420 : let builders = builder.into_iter();
1261 420 : let builders = if is_last {
1262 60 : builders.chain(read_builder.take())
1263 : } else {
1264 360 : builders.chain(None)
1265 : };
1266 :
1267 516 : for builder in builders {
1268 96 : let read = builder.build();
1269 96 :
1270 96 : let reader = VectoredBlobReader::new(&self.file);
1271 96 :
1272 96 : let mut buf = buffer.take().unwrap();
1273 96 :
1274 96 : buf.clear();
1275 96 : buf.reserve(read.size());
1276 96 : let res = reader.read_blobs(&read, buf, ctx).await?;
1277 :
1278 96 : let view = BufView::new_slice(&res.buf);
1279 :
1280 288 : for blob in res.blobs {
1281 192 : let key = blob.meta.key;
1282 192 : let lsn = blob.meta.lsn;
1283 :
1284 192 : let data = blob.read(&view).await?;
1285 :
1286 : #[cfg(debug_assertions)]
1287 192 : Value::des(&data)
1288 192 : .with_context(|| {
1289 0 : format!(
1290 0 : "blob failed to deserialize for {}: {:?}",
1291 0 : blob,
1292 0 : utils::Hex(&data)
1293 0 : )
1294 192 : })
1295 192 : .unwrap();
1296 192 :
1297 192 : // is it an image or will_init walrecord?
1298 192 : // FIXME: this could be handled by threading the BlobRef to the
1299 192 : // VectoredReadBuilder
1300 192 : let will_init = pageserver_api::value::ValueBytes::will_init(&data)
1301 192 : .inspect_err(|_e| {
1302 0 : #[cfg(feature = "testing")]
1303 0 : tracing::error!(data=?utils::Hex(&data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
1304 192 : })
1305 192 : .unwrap_or(false);
1306 192 :
1307 192 : per_blob_copy.clear();
1308 192 : per_blob_copy.extend_from_slice(&data);
1309 :
1310 192 : let (tmp, res) = writer
1311 192 : .put_value_bytes(
1312 192 : key,
1313 192 : lsn,
1314 192 : std::mem::take(&mut per_blob_copy).slice_len(),
1315 192 : will_init,
1316 192 : ctx,
1317 192 : )
1318 192 : .await;
1319 192 : per_blob_copy = tmp.into_raw_slice().into_inner();
1320 192 :
1321 192 : res?;
1322 :
1323 192 : records += 1;
1324 : }
1325 :
1326 96 : buffer = Some(res.buf);
1327 : }
1328 : }
1329 :
1330 60 : assert!(
1331 60 : read_builder.is_none(),
1332 0 : "with the sentinel above loop should had handled all"
1333 : );
1334 :
1335 60 : Ok(records)
1336 60 : }
1337 :
1338 24 : pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
1339 24 : println!(
1340 24 : "index_start_blk: {}, root {}",
1341 24 : self.index_start_blk, self.index_root_blk
1342 24 : );
1343 24 :
1344 24 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1345 24 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1346 24 : self.index_start_blk,
1347 24 : self.index_root_blk,
1348 24 : block_reader,
1349 24 : );
1350 24 :
1351 24 : tree_reader.dump(ctx).await?;
1352 :
1353 24 : let keys = self.index_entries(ctx).await?;
1354 :
1355 48 : async fn dump_blob(val: &ValueRef<'_>, ctx: &RequestContext) -> anyhow::Result<String> {
1356 48 : let buf = val.load_raw(ctx).await?;
1357 48 : let val = Value::des(&buf)?;
1358 48 : let desc = match val {
1359 48 : Value::Image(img) => {
1360 48 : format!(" img {} bytes", img.len())
1361 : }
1362 0 : Value::WalRecord(rec) => {
1363 0 : let wal_desc = pageserver_api::record::describe_wal_record(&rec)?;
1364 0 : format!(
1365 0 : " rec {} bytes will_init: {} {}",
1366 0 : buf.len(),
1367 0 : rec.will_init(),
1368 0 : wal_desc
1369 0 : )
1370 : }
1371 : };
1372 48 : Ok(desc)
1373 48 : }
1374 :
1375 72 : for entry in keys {
1376 48 : let DeltaEntry { key, lsn, val, .. } = entry;
1377 48 : let desc = match dump_blob(&val, ctx).await {
1378 48 : Ok(desc) => desc,
1379 0 : Err(err) => {
1380 0 : format!("ERROR: {err}")
1381 : }
1382 : };
1383 48 : println!(" key {key} at {lsn}: {desc}");
1384 :
1385 : // Print more details about CHECKPOINT records. Would be nice to print details
1386 : // of many other record types too, but these are particularly interesting, as
1387 : // have a lot of special processing for them in walingest.rs.
1388 24 : use pageserver_api::key::CHECKPOINT_KEY;
1389 24 : use postgres_ffi::CheckPoint;
1390 48 : if key == CHECKPOINT_KEY {
1391 0 : let val = val.load(ctx).await?;
1392 0 : match val {
1393 0 : Value::Image(img) => {
1394 0 : let checkpoint = CheckPoint::decode(&img)?;
1395 0 : println!(" CHECKPOINT: {:?}", checkpoint);
1396 : }
1397 0 : Value::WalRecord(_rec) => {
1398 0 : println!(" unexpected walrecord value for checkpoint key");
1399 0 : }
1400 : }
1401 48 : }
1402 : }
1403 :
1404 24 : Ok(())
1405 24 : }
1406 :
1407 180 : fn stream_index_forwards<'a, R>(
1408 180 : &'a self,
1409 180 : reader: DiskBtreeReader<R, DELTA_KEY_SIZE>,
1410 180 : start: &'a [u8; DELTA_KEY_SIZE],
1411 180 : ctx: &'a RequestContext,
1412 180 : ) -> impl futures::stream::Stream<
1413 180 : Item = Result<(Key, Lsn, BlobRef), crate::tenant::disk_btree::DiskBtreeError>,
1414 180 : > + 'a
1415 180 : where
1416 180 : R: BlockReader + 'a,
1417 180 : {
1418 : use futures::stream::TryStreamExt;
1419 180 : let stream = reader.into_stream(start, ctx);
1420 912 : stream.map_ok(|(key, value)| {
1421 912 : let key = DeltaKey::from_slice(&key);
1422 912 : let (key, lsn) = (key.key(), key.lsn());
1423 912 : let offset = BlobRef(value);
1424 912 :
1425 912 : (key, lsn, offset)
1426 912 : })
1427 180 : }
1428 :
1429 : /// The file offset to the first block of index.
1430 : ///
1431 : /// The file structure is summary, values, and index. We often need this for the size of last blob.
1432 1424119 : fn index_start_offset(&self) -> u64 {
1433 1424119 : let offset = self.index_start_blk as u64 * PAGE_SZ as u64;
1434 1424119 : let bref = BlobRef(offset);
1435 1424119 : tracing::debug!(
1436 : index_start_blk = self.index_start_blk,
1437 : offset,
1438 0 : pos = bref.pos(),
1439 0 : "index_start_offset"
1440 : );
1441 1424119 : offset
1442 1424119 : }
1443 :
1444 3456 : pub fn iter<'a>(&'a self, ctx: &'a RequestContext) -> DeltaLayerIterator<'a> {
1445 3456 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1446 3456 : let tree_reader =
1447 3456 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
1448 3456 : DeltaLayerIterator {
1449 3456 : delta_layer: self,
1450 3456 : ctx,
1451 3456 : index_iter: tree_reader.iter(&[0; DELTA_KEY_SIZE], ctx),
1452 3456 : key_values_batch: std::collections::VecDeque::new(),
1453 3456 : is_end: false,
1454 3456 : planner: StreamingVectoredReadPlanner::new(
1455 3456 : 1024 * 8192, // The default value. Unit tests might use a different value. 1024 * 8K = 8MB buffer.
1456 3456 : 1024, // The default value. Unit tests might use a different value
1457 3456 : ),
1458 3456 : }
1459 3456 : }
1460 :
1461 : /// NB: not super efficient, but not terrible either. Should prob be an iterator.
1462 : //
1463 : // We're reusing the index traversal logical in plan_reads; would be nice to
1464 : // factor that out.
1465 0 : pub(crate) async fn load_keys(&self, ctx: &RequestContext) -> anyhow::Result<Vec<Key>> {
1466 0 : self.index_entries(ctx)
1467 0 : .await
1468 0 : .map(|entries| entries.into_iter().map(|entry| entry.key).collect())
1469 0 : }
1470 : }
1471 :
1472 : /// A set of data associated with a delta layer key and its value
1473 : pub struct DeltaEntry<'a> {
1474 : pub key: Key,
1475 : pub lsn: Lsn,
1476 : /// Size of the stored value
1477 : pub size: u64,
1478 : /// Reference to the on-disk value
1479 : pub val: ValueRef<'a>,
1480 : }
1481 :
1482 : /// Reference to an on-disk value
1483 : pub struct ValueRef<'a> {
1484 : blob_ref: BlobRef,
1485 : layer: &'a DeltaLayerInner,
1486 : }
1487 :
1488 : impl ValueRef<'_> {
1489 : /// Loads the value from disk
1490 0 : pub async fn load(&self, ctx: &RequestContext) -> Result<Value> {
1491 0 : let buf = self.load_raw(ctx).await?;
1492 0 : let val = Value::des(&buf)?;
1493 0 : Ok(val)
1494 0 : }
1495 :
1496 48 : async fn load_raw(&self, ctx: &RequestContext) -> Result<Vec<u8>> {
1497 48 : let reader = BlockCursor::new(crate::tenant::block_io::BlockReaderRef::Adapter(Adapter(
1498 48 : self.layer,
1499 48 : )));
1500 48 : let buf = reader.read_blob(self.blob_ref.pos(), ctx).await?;
1501 48 : Ok(buf)
1502 48 : }
1503 : }
1504 :
1505 : pub(crate) struct Adapter<T>(T);
1506 :
1507 : impl<T: AsRef<DeltaLayerInner>> Adapter<T> {
1508 48 : pub(crate) async fn read_blk(
1509 48 : &self,
1510 48 : blknum: u32,
1511 48 : ctx: &RequestContext,
1512 48 : ) -> Result<BlockLease, std::io::Error> {
1513 48 : let block_reader = FileBlockReader::new(&self.0.as_ref().file, self.0.as_ref().file_id);
1514 48 : block_reader.read_blk(blknum, ctx).await
1515 48 : }
1516 : }
1517 :
1518 : impl AsRef<DeltaLayerInner> for DeltaLayerInner {
1519 96 : fn as_ref(&self) -> &DeltaLayerInner {
1520 96 : self
1521 96 : }
1522 : }
1523 :
1524 : impl<'a> pageserver_compaction::interface::CompactionDeltaEntry<'a, Key> for DeltaEntry<'a> {
1525 0 : fn key(&self) -> Key {
1526 0 : self.key
1527 0 : }
1528 0 : fn lsn(&self) -> Lsn {
1529 0 : self.lsn
1530 0 : }
1531 0 : fn size(&self) -> u64 {
1532 0 : self.size
1533 0 : }
1534 : }
1535 :
1536 : pub struct DeltaLayerIterator<'a> {
1537 : delta_layer: &'a DeltaLayerInner,
1538 : ctx: &'a RequestContext,
1539 : planner: StreamingVectoredReadPlanner,
1540 : index_iter: DiskBtreeIterator<'a>,
1541 : key_values_batch: VecDeque<(Key, Lsn, Value)>,
1542 : is_end: bool,
1543 : }
1544 :
1545 : impl DeltaLayerIterator<'_> {
1546 0 : pub(crate) fn layer_dbg_info(&self) -> String {
1547 0 : self.delta_layer.layer_dbg_info()
1548 0 : }
1549 :
1550 : /// Retrieve a batch of key-value pairs into the iterator buffer.
1551 128436 : async fn next_batch(&mut self) -> anyhow::Result<()> {
1552 128436 : assert!(self.key_values_batch.is_empty());
1553 128436 : assert!(!self.is_end);
1554 :
1555 128436 : let plan = loop {
1556 12597780 : if let Some(res) = self.index_iter.next().await {
1557 12594492 : let (raw_key, value) = res?;
1558 12594492 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
1559 12594492 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
1560 12594492 : let blob_ref = BlobRef(value);
1561 12594492 : let offset = blob_ref.pos();
1562 125148 : if let Some(batch_plan) =
1563 12594492 : self.planner.handle(key, lsn, offset, blob_ref.will_init())
1564 : {
1565 125148 : break batch_plan;
1566 12469344 : }
1567 : } else {
1568 3288 : self.is_end = true;
1569 3288 : let data_end_offset = self.delta_layer.index_start_offset();
1570 3288 : if let Some(item) = self.planner.handle_range_end(data_end_offset) {
1571 3288 : break item;
1572 : } else {
1573 0 : return Ok(()); // TODO: test empty iterator
1574 : }
1575 : }
1576 : };
1577 128436 : let vectored_blob_reader = VectoredBlobReader::new(&self.delta_layer.file);
1578 128436 : let mut next_batch = std::collections::VecDeque::new();
1579 128436 : let buf_size = plan.size();
1580 128436 : let buf = IoBufferMut::with_capacity(buf_size);
1581 128436 : let blobs_buf = vectored_blob_reader
1582 128436 : .read_blobs(&plan, buf, self.ctx)
1583 128436 : .await?;
1584 128436 : let view = BufView::new_slice(&blobs_buf.buf);
1585 12594324 : for meta in blobs_buf.blobs.iter() {
1586 12594324 : let blob_read = meta.read(&view).await?;
1587 12594324 : let value = Value::des(&blob_read)?;
1588 :
1589 12594324 : next_batch.push_back((meta.meta.key, meta.meta.lsn, value));
1590 : }
1591 128436 : self.key_values_batch = next_batch;
1592 128436 : Ok(())
1593 128436 : }
1594 :
1595 12599052 : pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
1596 12599052 : if self.key_values_batch.is_empty() {
1597 134304 : if self.is_end {
1598 6372 : return Ok(None);
1599 127932 : }
1600 127932 : self.next_batch().await?;
1601 12464748 : }
1602 12592680 : Ok(Some(
1603 12592680 : self.key_values_batch
1604 12592680 : .pop_front()
1605 12592680 : .expect("should not be empty"),
1606 12592680 : ))
1607 12599052 : }
1608 : }
1609 :
1610 : #[cfg(test)]
1611 : pub(crate) mod test {
1612 : use std::collections::BTreeMap;
1613 :
1614 : use bytes::Bytes;
1615 : use itertools::MinMaxResult;
1616 : use pageserver_api::value::Value;
1617 : use rand::prelude::{SeedableRng, SliceRandom, StdRng};
1618 : use rand::{Rng, RngCore};
1619 :
1620 : use super::*;
1621 : use crate::DEFAULT_PG_VERSION;
1622 : use crate::context::DownloadBehavior;
1623 : use crate::task_mgr::TaskKind;
1624 : use crate::tenant::disk_btree::tests::TestDisk;
1625 : use crate::tenant::harness::{TIMELINE_ID, TenantHarness};
1626 : use crate::tenant::storage_layer::{Layer, ResidentLayer};
1627 : use crate::tenant::vectored_blob_io::StreamingVectoredReadPlanner;
1628 : use crate::tenant::{TenantShard, Timeline};
1629 :
1630 : /// Construct an index for a fictional delta layer and and then
1631 : /// traverse in order to plan vectored reads for a query. Finally,
1632 : /// verify that the traversal fed the right index key and value
1633 : /// pairs into the planner.
1634 : #[tokio::test]
1635 12 : async fn test_delta_layer_index_traversal() {
1636 12 : let base_key = Key {
1637 12 : field1: 0,
1638 12 : field2: 1663,
1639 12 : field3: 12972,
1640 12 : field4: 16396,
1641 12 : field5: 0,
1642 12 : field6: 246080,
1643 12 : };
1644 12 :
1645 12 : // Populate the index with some entries
1646 12 : let entries: BTreeMap<Key, Vec<Lsn>> = BTreeMap::from([
1647 12 : (base_key, vec![Lsn(1), Lsn(5), Lsn(25), Lsn(26), Lsn(28)]),
1648 12 : (base_key.add(1), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1649 12 : (base_key.add(2), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1650 12 : (base_key.add(5), vec![Lsn(10), Lsn(15), Lsn(16), Lsn(20)]),
1651 12 : ]);
1652 12 :
1653 12 : let mut disk = TestDisk::default();
1654 12 : let mut writer = DiskBtreeBuilder::<_, DELTA_KEY_SIZE>::new(&mut disk);
1655 12 :
1656 12 : let mut disk_offset = 0;
1657 60 : for (key, lsns) in &entries {
1658 252 : for lsn in lsns {
1659 204 : let index_key = DeltaKey::from_key_lsn(key, *lsn);
1660 204 : let blob_ref = BlobRef::new(disk_offset, false);
1661 204 : writer
1662 204 : .append(&index_key.0, blob_ref.0)
1663 204 : .expect("In memory disk append should never fail");
1664 204 :
1665 204 : disk_offset += 1;
1666 204 : }
1667 12 : }
1668 12 :
1669 12 : // Prepare all the arguments for the call into `plan_reads` below
1670 12 : let (root_offset, _writer) = writer
1671 12 : .finish()
1672 12 : .expect("In memory disk finish should never fail");
1673 12 : let reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(0, root_offset, disk);
1674 12 : let planner = VectoredReadPlanner::new(100);
1675 12 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1676 12 :
1677 12 : let keyspace = KeySpace {
1678 12 : ranges: vec![
1679 12 : base_key..base_key.add(3),
1680 12 : base_key.add(3)..base_key.add(100),
1681 12 : ],
1682 12 : };
1683 12 : let lsn_range = Lsn(2)..Lsn(40);
1684 12 :
1685 12 : // Plan and validate
1686 12 : let vectored_reads = DeltaLayerInner::plan_reads(
1687 12 : &keyspace,
1688 12 : lsn_range.clone(),
1689 12 : disk_offset,
1690 12 : reader,
1691 12 : planner,
1692 12 : &ctx,
1693 12 : )
1694 12 : .await
1695 12 : .expect("Read planning should not fail");
1696 12 :
1697 12 : validate(keyspace, lsn_range, vectored_reads, entries);
1698 12 : }
1699 :
1700 12 : fn validate(
1701 12 : keyspace: KeySpace,
1702 12 : lsn_range: Range<Lsn>,
1703 12 : vectored_reads: Vec<VectoredRead>,
1704 12 : index_entries: BTreeMap<Key, Vec<Lsn>>,
1705 12 : ) {
1706 : #[derive(Debug, PartialEq, Eq)]
1707 : struct BlobSpec {
1708 : key: Key,
1709 : lsn: Lsn,
1710 : at: u64,
1711 : }
1712 :
1713 12 : let mut planned_blobs = Vec::new();
1714 180 : for read in vectored_reads {
1715 168 : for (at, meta) in read.blobs_at.as_slice() {
1716 168 : planned_blobs.push(BlobSpec {
1717 168 : key: meta.key,
1718 168 : lsn: meta.lsn,
1719 168 : at: *at,
1720 168 : });
1721 168 : }
1722 : }
1723 :
1724 12 : let mut expected_blobs = Vec::new();
1725 12 : let mut disk_offset = 0;
1726 60 : for (key, lsns) in index_entries {
1727 252 : for lsn in lsns {
1728 252 : let key_included = keyspace.ranges.iter().any(|range| range.contains(&key));
1729 204 : let lsn_included = lsn_range.contains(&lsn);
1730 204 :
1731 204 : if key_included && lsn_included {
1732 168 : expected_blobs.push(BlobSpec {
1733 168 : key,
1734 168 : lsn,
1735 168 : at: disk_offset,
1736 168 : });
1737 168 : }
1738 :
1739 204 : disk_offset += 1;
1740 : }
1741 : }
1742 :
1743 12 : assert_eq!(planned_blobs, expected_blobs);
1744 12 : }
1745 :
1746 : mod constants {
1747 : use utils::lsn::Lsn;
1748 :
1749 : /// Offset used by all lsns in this test
1750 : pub(super) const LSN_OFFSET: Lsn = Lsn(0x08);
1751 : /// Number of unique keys including in the test data
1752 : pub(super) const KEY_COUNT: u8 = 60;
1753 : /// Max number of different lsns for each key
1754 : pub(super) const MAX_ENTRIES_PER_KEY: u8 = 20;
1755 : /// Possible value sizes for each key along with a probability weight
1756 : pub(super) const VALUE_SIZES: [(usize, u8); 3] = [(100, 2), (1024, 2), (1024 * 1024, 1)];
1757 : /// Probability that there will be a gap between the current key and the next one (33.3%)
1758 : pub(super) const KEY_GAP_CHANGES: [(bool, u8); 2] = [(true, 1), (false, 2)];
1759 : /// The minimum size of a key range in all the generated reads
1760 : pub(super) const MIN_RANGE_SIZE: i128 = 10;
1761 : /// The number of ranges included in each vectored read
1762 : pub(super) const RANGES_COUNT: u8 = 2;
1763 : /// The number of vectored reads performed
1764 : pub(super) const READS_COUNT: u8 = 100;
1765 : /// Soft max size of a vectored read. Will be violated if we have to read keys
1766 : /// with values larger than the limit
1767 : pub(super) const MAX_VECTORED_READ_BYTES: usize = 64 * 1024;
1768 : }
1769 :
1770 : struct Entry {
1771 : key: Key,
1772 : lsn: Lsn,
1773 : value: Vec<u8>,
1774 : }
1775 :
1776 12 : fn generate_entries(rng: &mut StdRng) -> Vec<Entry> {
1777 12 : let mut current_key = Key::MIN;
1778 12 :
1779 12 : let mut entries = Vec::new();
1780 732 : for _ in 0..constants::KEY_COUNT {
1781 720 : let count = rng.gen_range(1..constants::MAX_ENTRIES_PER_KEY);
1782 720 : let mut lsns_iter =
1783 13560 : std::iter::successors(Some(Lsn(constants::LSN_OFFSET.0 + 0x08)), |lsn| {
1784 13560 : Some(Lsn(lsn.0 + 0x08))
1785 13560 : });
1786 720 : let mut lsns = Vec::new();
1787 14280 : while lsns.len() < count as usize {
1788 13560 : let take = rng.gen_bool(0.5);
1789 13560 : let lsn = lsns_iter.next().unwrap();
1790 13560 : if take {
1791 6672 : lsns.push(lsn);
1792 6888 : }
1793 : }
1794 :
1795 7392 : for lsn in lsns {
1796 6672 : let size = constants::VALUE_SIZES
1797 20016 : .choose_weighted(rng, |item| item.1)
1798 6672 : .unwrap()
1799 6672 : .0;
1800 6672 : let mut buf = vec![0; size];
1801 6672 : rng.fill_bytes(&mut buf);
1802 6672 :
1803 6672 : entries.push(Entry {
1804 6672 : key: current_key,
1805 6672 : lsn,
1806 6672 : value: buf,
1807 6672 : })
1808 : }
1809 :
1810 720 : let gap = constants::KEY_GAP_CHANGES
1811 1440 : .choose_weighted(rng, |item| item.1)
1812 720 : .unwrap()
1813 720 : .0;
1814 720 : if gap {
1815 228 : current_key = current_key.add(2);
1816 492 : } else {
1817 492 : current_key = current_key.add(1);
1818 492 : }
1819 : }
1820 :
1821 12 : entries
1822 12 : }
1823 :
1824 : struct EntriesMeta {
1825 : key_range: Range<Key>,
1826 : lsn_range: Range<Lsn>,
1827 : index: BTreeMap<(Key, Lsn), Vec<u8>>,
1828 : }
1829 :
1830 12 : fn get_entries_meta(entries: &[Entry]) -> EntriesMeta {
1831 6672 : let key_range = match entries.iter().minmax_by_key(|e| e.key) {
1832 12 : MinMaxResult::MinMax(min, max) => min.key..max.key.next(),
1833 0 : _ => panic!("More than one entry is always expected"),
1834 : };
1835 :
1836 6672 : let lsn_range = match entries.iter().minmax_by_key(|e| e.lsn) {
1837 12 : MinMaxResult::MinMax(min, max) => min.lsn..Lsn(max.lsn.0 + 1),
1838 0 : _ => panic!("More than one entry is always expected"),
1839 : };
1840 :
1841 12 : let mut index = BTreeMap::new();
1842 6672 : for entry in entries.iter() {
1843 6672 : index.insert((entry.key, entry.lsn), entry.value.clone());
1844 6672 : }
1845 :
1846 12 : EntriesMeta {
1847 12 : key_range,
1848 12 : lsn_range,
1849 12 : index,
1850 12 : }
1851 12 : }
1852 :
1853 1200 : fn pick_random_keyspace(rng: &mut StdRng, key_range: &Range<Key>) -> KeySpace {
1854 1200 : let start = key_range.start.to_i128();
1855 1200 : let end = key_range.end.to_i128();
1856 1200 :
1857 1200 : let mut keyspace = KeySpace::default();
1858 :
1859 3600 : for _ in 0..constants::RANGES_COUNT {
1860 2400 : let mut range: Option<Range<Key>> = Option::default();
1861 7464 : while range.is_none() || keyspace.overlaps(range.as_ref().unwrap()) {
1862 5064 : let range_start = rng.gen_range(start..end);
1863 5064 : let range_end_offset = range_start + constants::MIN_RANGE_SIZE;
1864 5064 : if range_end_offset >= end {
1865 600 : range = Some(Key::from_i128(range_start)..Key::from_i128(end));
1866 4464 : } else {
1867 4464 : let range_end = rng.gen_range((range_start + constants::MIN_RANGE_SIZE)..end);
1868 4464 : range = Some(Key::from_i128(range_start)..Key::from_i128(range_end));
1869 4464 : }
1870 : }
1871 2400 : keyspace.ranges.push(range.unwrap());
1872 : }
1873 :
1874 1200 : keyspace
1875 1200 : }
1876 :
1877 : #[tokio::test]
1878 12 : async fn test_delta_layer_vectored_read_end_to_end() -> anyhow::Result<()> {
1879 12 : let harness = TenantHarness::create("test_delta_layer_oversized_vectored_read").await?;
1880 12 : let (tenant, ctx) = harness.load().await;
1881 12 :
1882 12 : let timeline_id = TimelineId::generate();
1883 12 : let timeline = tenant
1884 12 : .create_test_timeline(timeline_id, constants::LSN_OFFSET, DEFAULT_PG_VERSION, &ctx)
1885 12 : .await?;
1886 12 :
1887 12 : tracing::info!("Generating test data ...");
1888 12 :
1889 12 : let rng = &mut StdRng::seed_from_u64(0);
1890 12 : let entries = generate_entries(rng);
1891 12 : let entries_meta = get_entries_meta(&entries);
1892 12 :
1893 12 : tracing::info!("Done generating {} entries", entries.len());
1894 12 :
1895 12 : tracing::info!("Writing test data to delta layer ...");
1896 12 : let mut writer = DeltaLayerWriter::new(
1897 12 : harness.conf,
1898 12 : timeline_id,
1899 12 : harness.tenant_shard_id,
1900 12 : entries_meta.key_range.start,
1901 12 : entries_meta.lsn_range.clone(),
1902 12 : &timeline.gate,
1903 12 : timeline.cancel.clone(),
1904 12 : &ctx,
1905 12 : )
1906 12 : .await?;
1907 12 :
1908 6684 : for entry in entries {
1909 6672 : let (_, res) = writer
1910 6672 : .put_value_bytes(entry.key, entry.lsn, entry.value.slice_len(), false, &ctx)
1911 6672 : .await;
1912 6672 : res?;
1913 12 : }
1914 12 :
1915 12 : let (desc, path) = writer.finish(entries_meta.key_range.end, &ctx).await?;
1916 12 : let resident = Layer::finish_creating(harness.conf, &timeline, desc, &path)?;
1917 12 :
1918 12 : let inner = resident.get_as_delta(&ctx).await?;
1919 12 :
1920 12 : let file_size = inner.file.metadata().await?.len();
1921 12 : tracing::info!(
1922 12 : "Done writing test data to delta layer. Resulting file size is: {}",
1923 12 : file_size
1924 12 : );
1925 12 :
1926 1212 : for i in 0..constants::READS_COUNT {
1927 1200 : tracing::info!("Doing vectored read {}/{}", i + 1, constants::READS_COUNT);
1928 12 :
1929 1200 : let block_reader = FileBlockReader::new(&inner.file, inner.file_id);
1930 1200 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1931 1200 : inner.index_start_blk,
1932 1200 : inner.index_root_blk,
1933 1200 : block_reader,
1934 1200 : );
1935 1200 :
1936 1200 : let planner = VectoredReadPlanner::new(constants::MAX_VECTORED_READ_BYTES);
1937 1200 : let keyspace = pick_random_keyspace(rng, &entries_meta.key_range);
1938 1200 : let data_end_offset = inner.index_start_blk as u64 * PAGE_SZ as u64;
1939 12 :
1940 1200 : let vectored_reads = DeltaLayerInner::plan_reads(
1941 1200 : &keyspace,
1942 1200 : entries_meta.lsn_range.clone(),
1943 1200 : data_end_offset,
1944 1200 : index_reader,
1945 1200 : planner,
1946 1200 : &ctx,
1947 1200 : )
1948 1200 : .await?;
1949 12 :
1950 1200 : let vectored_blob_reader = VectoredBlobReader::new(&inner.file);
1951 1200 : let buf_size = DeltaLayerInner::get_min_read_buffer_size(
1952 1200 : &vectored_reads,
1953 1200 : constants::MAX_VECTORED_READ_BYTES,
1954 1200 : );
1955 1200 : let mut buf = Some(IoBufferMut::with_capacity(buf_size));
1956 12 :
1957 119544 : for read in vectored_reads {
1958 118344 : let blobs_buf = vectored_blob_reader
1959 118344 : .read_blobs(&read, buf.take().expect("Should have a buffer"), &ctx)
1960 118344 : .await?;
1961 118344 : let view = BufView::new_slice(&blobs_buf.buf);
1962 343824 : for meta in blobs_buf.blobs.iter() {
1963 343824 : let value = meta.read(&view).await?;
1964 343824 : assert_eq!(
1965 343824 : &value[..],
1966 343824 : &entries_meta.index[&(meta.meta.key, meta.meta.lsn)]
1967 343824 : );
1968 12 : }
1969 12 :
1970 118344 : buf = Some(blobs_buf.buf);
1971 12 : }
1972 12 : }
1973 12 :
1974 12 : Ok(())
1975 12 : }
1976 :
1977 : #[tokio::test]
1978 12 : async fn copy_delta_prefix_smoke() {
1979 12 : use bytes::Bytes;
1980 12 : use pageserver_api::record::NeonWalRecord;
1981 12 :
1982 12 : let h = crate::tenant::harness::TenantHarness::create("truncate_delta_smoke")
1983 12 : .await
1984 12 : .unwrap();
1985 12 : let (tenant, ctx) = h.load().await;
1986 12 : let ctx = &ctx;
1987 12 : let timeline = tenant
1988 12 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, ctx)
1989 12 : .await
1990 12 : .unwrap();
1991 12 : let ctx = &ctx.with_scope_timeline(&timeline);
1992 12 :
1993 12 : let initdb_layer = timeline
1994 12 : .layers
1995 12 : .read()
1996 12 : .await
1997 12 : .likely_resident_layers()
1998 12 : .next()
1999 12 : .cloned()
2000 12 : .unwrap();
2001 12 :
2002 12 : {
2003 12 : let mut writer = timeline.writer().await;
2004 12 :
2005 12 : let data = [
2006 12 : (0x20, 12, Value::Image(Bytes::from_static(b"foobar"))),
2007 12 : (
2008 12 : 0x30,
2009 12 : 12,
2010 12 : Value::WalRecord(NeonWalRecord::Postgres {
2011 12 : will_init: false,
2012 12 : rec: Bytes::from_static(b"1"),
2013 12 : }),
2014 12 : ),
2015 12 : (
2016 12 : 0x40,
2017 12 : 12,
2018 12 : Value::WalRecord(NeonWalRecord::Postgres {
2019 12 : will_init: true,
2020 12 : rec: Bytes::from_static(b"2"),
2021 12 : }),
2022 12 : ),
2023 12 : // build an oversized value so we cannot extend and existing read over
2024 12 : // this
2025 12 : (
2026 12 : 0x50,
2027 12 : 12,
2028 12 : Value::WalRecord(NeonWalRecord::Postgres {
2029 12 : will_init: true,
2030 12 : rec: {
2031 12 : let mut buf =
2032 12 : vec![0u8; tenant.conf.max_vectored_read_bytes.0.get() + 1024];
2033 12 : buf.iter_mut()
2034 12 : .enumerate()
2035 1609728 : .for_each(|(i, slot)| *slot = (i % 256) as u8);
2036 12 : Bytes::from(buf)
2037 12 : },
2038 12 : }),
2039 12 : ),
2040 12 : // because the oversized read cannot be extended further, we are sure to exercise the
2041 12 : // builder created on the last round with this:
2042 12 : (
2043 12 : 0x60,
2044 12 : 12,
2045 12 : Value::WalRecord(NeonWalRecord::Postgres {
2046 12 : will_init: true,
2047 12 : rec: Bytes::from_static(b"3"),
2048 12 : }),
2049 12 : ),
2050 12 : (
2051 12 : 0x60,
2052 12 : 9,
2053 12 : Value::Image(Bytes::from_static(b"something for a different key")),
2054 12 : ),
2055 12 : ];
2056 12 :
2057 12 : let mut last_lsn = None;
2058 12 :
2059 84 : for (lsn, key, value) in data {
2060 72 : let key = Key::from_i128(key);
2061 72 : writer.put(key, Lsn(lsn), &value, ctx).await.unwrap();
2062 72 : last_lsn = Some(lsn);
2063 12 : }
2064 12 :
2065 12 : writer.finish_write(Lsn(last_lsn.unwrap()));
2066 12 : }
2067 12 : timeline.freeze_and_flush().await.unwrap();
2068 12 :
2069 12 : let new_layer = timeline
2070 12 : .layers
2071 12 : .read()
2072 12 : .await
2073 12 : .likely_resident_layers()
2074 19 : .find(|&x| x != &initdb_layer)
2075 12 : .cloned()
2076 12 : .unwrap();
2077 12 :
2078 12 : // create a copy for the timeline, so we don't overwrite the file
2079 12 : let branch = tenant
2080 12 : .branch_timeline_test(&timeline, TimelineId::generate(), None, ctx)
2081 12 : .await
2082 12 : .unwrap();
2083 12 :
2084 12 : assert_eq!(branch.get_ancestor_lsn(), Lsn(0x60));
2085 12 :
2086 12 : // truncating at 0x61 gives us a full copy, otherwise just go backwards until there's just
2087 12 : // a single key
2088 12 :
2089 72 : for truncate_at in [0x61, 0x51, 0x41, 0x31, 0x21] {
2090 60 : let truncate_at = Lsn(truncate_at);
2091 12 :
2092 60 : let mut writer = DeltaLayerWriter::new(
2093 60 : tenant.conf,
2094 60 : branch.timeline_id,
2095 60 : tenant.tenant_shard_id,
2096 60 : Key::MIN,
2097 60 : Lsn(0x11)..truncate_at,
2098 60 : &branch.gate,
2099 60 : branch.cancel.clone(),
2100 60 : ctx,
2101 60 : )
2102 60 : .await
2103 60 : .unwrap();
2104 12 :
2105 60 : let new_layer = new_layer.download_and_keep_resident(ctx).await.unwrap();
2106 60 :
2107 60 : new_layer
2108 60 : .copy_delta_prefix(&mut writer, truncate_at, ctx)
2109 60 : .await
2110 60 : .unwrap();
2111 12 :
2112 60 : let (desc, path) = writer.finish(Key::MAX, ctx).await.unwrap();
2113 60 : let copied_layer = Layer::finish_creating(tenant.conf, &branch, desc, &path).unwrap();
2114 60 :
2115 60 : copied_layer.get_as_delta(ctx).await.unwrap();
2116 60 :
2117 60 : assert_keys_and_values_eq(
2118 60 : new_layer.get_as_delta(ctx).await.unwrap(),
2119 60 : copied_layer.get_as_delta(ctx).await.unwrap(),
2120 60 : truncate_at,
2121 60 : ctx,
2122 60 : )
2123 60 : .await;
2124 12 : }
2125 12 : }
2126 :
2127 60 : async fn assert_keys_and_values_eq(
2128 60 : source: &DeltaLayerInner,
2129 60 : truncated: &DeltaLayerInner,
2130 60 : truncated_at: Lsn,
2131 60 : ctx: &RequestContext,
2132 60 : ) {
2133 : use futures::future::ready;
2134 : use futures::stream::TryStreamExt;
2135 :
2136 60 : let start_key = [0u8; DELTA_KEY_SIZE];
2137 60 :
2138 60 : let source_reader = FileBlockReader::new(&source.file, source.file_id);
2139 60 : let source_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2140 60 : source.index_start_blk,
2141 60 : source.index_root_blk,
2142 60 : &source_reader,
2143 60 : );
2144 60 : let source_stream = source.stream_index_forwards(source_tree, &start_key, ctx);
2145 360 : let source_stream = source_stream.filter(|res| match res {
2146 360 : Ok((_, lsn, _)) => ready(lsn < &truncated_at),
2147 0 : _ => ready(true),
2148 360 : });
2149 60 : let mut source_stream = std::pin::pin!(source_stream);
2150 60 :
2151 60 : let truncated_reader = FileBlockReader::new(&truncated.file, truncated.file_id);
2152 60 : let truncated_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2153 60 : truncated.index_start_blk,
2154 60 : truncated.index_root_blk,
2155 60 : &truncated_reader,
2156 60 : );
2157 60 : let truncated_stream = truncated.stream_index_forwards(truncated_tree, &start_key, ctx);
2158 60 : let mut truncated_stream = std::pin::pin!(truncated_stream);
2159 60 :
2160 60 : let mut scratch_left = Vec::new();
2161 60 : let mut scratch_right = Vec::new();
2162 :
2163 : loop {
2164 252 : let (src, truncated) = (source_stream.try_next(), truncated_stream.try_next());
2165 252 : let (src, truncated) = tokio::try_join!(src, truncated).unwrap();
2166 252 :
2167 252 : if src.is_none() {
2168 60 : assert!(truncated.is_none());
2169 60 : break;
2170 192 : }
2171 192 :
2172 192 : let (src, truncated) = (src.unwrap(), truncated.unwrap());
2173 192 :
2174 192 : // because we've filtered the source with Lsn, we should always have the same keys from both.
2175 192 : assert_eq!(src.0, truncated.0);
2176 192 : assert_eq!(src.1, truncated.1);
2177 :
2178 : // if this is needed for something else, just drop this assert.
2179 192 : assert!(
2180 192 : src.2.pos() >= truncated.2.pos(),
2181 0 : "value position should not go backwards {} vs. {}",
2182 0 : src.2.pos(),
2183 0 : truncated.2.pos()
2184 : );
2185 :
2186 192 : scratch_left.clear();
2187 192 : let src_cursor = source_reader.block_cursor();
2188 192 : let left = src_cursor.read_blob_into_buf(src.2.pos(), &mut scratch_left, ctx);
2189 192 : scratch_right.clear();
2190 192 : let trunc_cursor = truncated_reader.block_cursor();
2191 192 : let right = trunc_cursor.read_blob_into_buf(truncated.2.pos(), &mut scratch_right, ctx);
2192 192 :
2193 192 : tokio::try_join!(left, right).unwrap();
2194 192 :
2195 192 : assert_eq!(utils::Hex(&scratch_left), utils::Hex(&scratch_right));
2196 : }
2197 60 : }
2198 :
2199 109344 : pub(crate) fn sort_delta(
2200 109344 : (k1, l1, _): &(Key, Lsn, Value),
2201 109344 : (k2, l2, _): &(Key, Lsn, Value),
2202 109344 : ) -> std::cmp::Ordering {
2203 109344 : (k1, l1).cmp(&(k2, l2))
2204 109344 : }
2205 :
2206 : #[cfg(feature = "testing")]
2207 564 : pub(crate) fn sort_delta_value(
2208 564 : (k1, l1, v1): &(Key, Lsn, Value),
2209 564 : (k2, l2, v2): &(Key, Lsn, Value),
2210 564 : ) -> std::cmp::Ordering {
2211 564 : let order_1 = if v1.is_image() { 0 } else { 1 };
2212 564 : let order_2 = if v2.is_image() { 0 } else { 1 };
2213 564 : (k1, l1, order_1).cmp(&(k2, l2, order_2))
2214 564 : }
2215 :
2216 132 : pub(crate) async fn produce_delta_layer(
2217 132 : tenant: &TenantShard,
2218 132 : tline: &Arc<Timeline>,
2219 132 : mut deltas: Vec<(Key, Lsn, Value)>,
2220 132 : ctx: &RequestContext,
2221 132 : ) -> anyhow::Result<ResidentLayer> {
2222 132 : deltas.sort_by(sort_delta);
2223 132 : let (key_start, _, _) = deltas.first().unwrap();
2224 132 : let (key_max, _, _) = deltas.last().unwrap();
2225 49440 : let lsn_min = deltas.iter().map(|(_, lsn, _)| lsn).min().unwrap();
2226 49440 : let lsn_max = deltas.iter().map(|(_, lsn, _)| lsn).max().unwrap();
2227 132 : let lsn_end = Lsn(lsn_max.0 + 1);
2228 132 : let mut writer = DeltaLayerWriter::new(
2229 132 : tenant.conf,
2230 132 : tline.timeline_id,
2231 132 : tenant.tenant_shard_id,
2232 132 : *key_start,
2233 132 : (*lsn_min)..lsn_end,
2234 132 : &tline.gate,
2235 132 : tline.cancel.clone(),
2236 132 : ctx,
2237 132 : )
2238 132 : .await?;
2239 132 : let key_end = key_max.next();
2240 :
2241 49572 : for (key, lsn, value) in deltas {
2242 49440 : writer.put_value(key, lsn, value, ctx).await?;
2243 : }
2244 :
2245 132 : let (desc, path) = writer.finish(key_end, ctx).await?;
2246 132 : let delta_layer = Layer::finish_creating(tenant.conf, tline, desc, &path)?;
2247 :
2248 132 : Ok::<_, anyhow::Error>(delta_layer)
2249 132 : }
2250 :
2251 168 : async fn assert_delta_iter_equal(
2252 168 : delta_iter: &mut DeltaLayerIterator<'_>,
2253 168 : expect: &[(Key, Lsn, Value)],
2254 168 : ) {
2255 168 : let mut expect_iter = expect.iter();
2256 : loop {
2257 168168 : let o1 = delta_iter.next().await.unwrap();
2258 168168 : let o2 = expect_iter.next();
2259 168168 : assert_eq!(o1.is_some(), o2.is_some());
2260 168168 : if o1.is_none() && o2.is_none() {
2261 168 : break;
2262 168000 : }
2263 168000 : let (k1, l1, v1) = o1.unwrap();
2264 168000 : let (k2, l2, v2) = o2.unwrap();
2265 168000 : assert_eq!(&k1, k2);
2266 168000 : assert_eq!(l1, *l2);
2267 168000 : assert_eq!(&v1, v2);
2268 : }
2269 168 : }
2270 :
2271 : #[tokio::test]
2272 12 : async fn delta_layer_iterator() {
2273 12 : let harness = TenantHarness::create("delta_layer_iterator").await.unwrap();
2274 12 : let (tenant, ctx) = harness.load().await;
2275 12 :
2276 12 : let tline = tenant
2277 12 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
2278 12 : .await
2279 12 : .unwrap();
2280 12 :
2281 12000 : fn get_key(id: u32) -> Key {
2282 12000 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
2283 12000 : key.field6 = id;
2284 12000 : key
2285 12000 : }
2286 12 : const N: usize = 1000;
2287 12 : let test_deltas = (0..N)
2288 12000 : .map(|idx| {
2289 12000 : (
2290 12000 : get_key(idx as u32 / 10),
2291 12000 : Lsn(0x10 * ((idx as u64) % 10 + 1)),
2292 12000 : Value::Image(Bytes::from(format!("img{idx:05}"))),
2293 12000 : )
2294 12000 : })
2295 12 : .collect_vec();
2296 12 : let resident_layer = produce_delta_layer(&tenant, &tline, test_deltas.clone(), &ctx)
2297 12 : .await
2298 12 : .unwrap();
2299 12 : let delta_layer = resident_layer.get_as_delta(&ctx).await.unwrap();
2300 36 : for max_read_size in [1, 1024] {
2301 192 : for batch_size in [1, 2, 4, 8, 3, 7, 13] {
2302 168 : println!("running with batch_size={batch_size} max_read_size={max_read_size}");
2303 168 : // Test if the batch size is correctly determined
2304 168 : let mut iter = delta_layer.iter(&ctx);
2305 168 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2306 168 : let mut num_items = 0;
2307 672 : for _ in 0..3 {
2308 504 : iter.next_batch().await.unwrap();
2309 504 : num_items += iter.key_values_batch.len();
2310 504 : if max_read_size == 1 {
2311 12 : // every key should be a batch b/c the value is larger than max_read_size
2312 252 : assert_eq!(iter.key_values_batch.len(), 1);
2313 12 : } else {
2314 252 : assert!(iter.key_values_batch.len() <= batch_size);
2315 12 : }
2316 504 : if num_items >= N {
2317 12 : break;
2318 504 : }
2319 504 : iter.key_values_batch.clear();
2320 12 : }
2321 12 : // Test if the result is correct
2322 168 : let mut iter = delta_layer.iter(&ctx);
2323 168 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2324 168 : assert_delta_iter_equal(&mut iter, &test_deltas).await;
2325 12 : }
2326 12 : }
2327 12 : }
2328 : }
|
