Line data Source code
1 : //! A DeltaLayer represents a collection of WAL records or page images in a range of
2 : //! LSNs, and in a range of Keys. It is stored on a file on disk.
3 : //!
4 : //! Usually a delta layer only contains differences, in the form of WAL records
5 : //! against a base LSN. However, if a relation extended or a whole new relation
6 : //! is created, there would be no base for the new pages. The entries for them
7 : //! must be page images or WAL records with the 'will_init' flag set, so that
8 : //! they can be replayed without referring to an older page version.
9 : //!
10 : //! The delta files are stored in `timelines/<timeline_id>` directory. Currently,
11 : //! there are no subdirectories, and each delta file is named like this:
12 : //!
13 : //! ```text
14 : //! <key start>-<key end>__<start LSN>-<end LSN>
15 : //! ```
16 : //!
17 : //! For example:
18 : //!
19 : //! ```text
20 : //! 000000067F000032BE0000400000000020B6-000000067F000032BE0000400000000030B6__000000578C6B29-0000000057A50051
21 : //! ```
22 : //!
23 : //! Every delta file consists of three parts: "summary", "values", and
24 : //! "index". The summary is a fixed size header at the beginning of the file,
25 : //! and it contains basic information about the layer, and offsets to the other
26 : //! parts. The "index" is a B-tree, mapping from Key and LSN to an offset in the
27 : //! "values" part. The actual page images and WAL records are stored in the
28 : //! "values" part.
29 : //!
30 : use crate::config::PageServerConf;
31 : use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
32 : use crate::page_cache::{self, FileId, PAGE_SZ};
33 : use crate::repository::{Key, Value, KEY_SIZE};
34 : use crate::tenant::blob_io::BlobWriter;
35 : use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockLease, BlockReader, FileBlockReader};
36 : use crate::tenant::disk_btree::{
37 : DiskBtreeBuilder, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
38 : };
39 : use crate::tenant::storage_layer::layer::S3_UPLOAD_LIMIT;
40 : use crate::tenant::timeline::GetVectoredError;
41 : use crate::tenant::vectored_blob_io::{
42 : BlobFlag, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
43 : VectoredReadCoalesceMode, VectoredReadPlanner,
44 : };
45 : use crate::tenant::PageReconstructError;
46 : use crate::virtual_file::owned_buffers_io::io_buf_ext::{FullSlice, IoBufExt};
47 : use crate::virtual_file::{self, VirtualFile};
48 : use crate::{walrecord, TEMP_FILE_SUFFIX};
49 : use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
50 : use anyhow::{anyhow, bail, ensure, Context, Result};
51 : use bytes::BytesMut;
52 : use camino::{Utf8Path, Utf8PathBuf};
53 : use futures::StreamExt;
54 : use itertools::Itertools;
55 : use pageserver_api::config::MaxVectoredReadBytes;
56 : use pageserver_api::keyspace::KeySpace;
57 : use pageserver_api::models::ImageCompressionAlgorithm;
58 : use pageserver_api::shard::TenantShardId;
59 : use rand::{distributions::Alphanumeric, Rng};
60 : use serde::{Deserialize, Serialize};
61 : use std::collections::VecDeque;
62 : use std::fs::File;
63 : use std::io::SeekFrom;
64 : use std::ops::Range;
65 : use std::os::unix::fs::FileExt;
66 : use std::str::FromStr;
67 : use std::sync::Arc;
68 : use tokio::sync::OnceCell;
69 : use tokio_epoll_uring::IoBuf;
70 : use tracing::*;
71 :
72 : use utils::{
73 : bin_ser::BeSer,
74 : id::{TenantId, TimelineId},
75 : lsn::Lsn,
76 : };
77 :
78 : use super::{AsLayerDesc, LayerName, PersistentLayerDesc, ValuesReconstructState};
79 :
80 : ///
81 : /// Header stored in the beginning of the file
82 : ///
83 : /// After this comes the 'values' part, starting on block 1. After that,
84 : /// the 'index' starts at the block indicated by 'index_start_blk'
85 : ///
86 3132 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
87 : pub struct Summary {
88 : /// Magic value to identify this as a neon delta file. Always DELTA_FILE_MAGIC.
89 : pub magic: u16,
90 : pub format_version: u16,
91 :
92 : pub tenant_id: TenantId,
93 : pub timeline_id: TimelineId,
94 : pub key_range: Range<Key>,
95 : pub lsn_range: Range<Lsn>,
96 :
97 : /// Block number where the 'index' part of the file begins.
98 : pub index_start_blk: u32,
99 : /// Block within the 'index', where the B-tree root page is stored
100 : pub index_root_blk: u32,
101 : }
102 :
103 : impl From<&DeltaLayer> for Summary {
104 0 : fn from(layer: &DeltaLayer) -> Self {
105 0 : Self::expected(
106 0 : layer.desc.tenant_shard_id.tenant_id,
107 0 : layer.desc.timeline_id,
108 0 : layer.desc.key_range.clone(),
109 0 : layer.desc.lsn_range.clone(),
110 0 : )
111 0 : }
112 : }
113 :
114 : impl Summary {
115 3132 : pub(super) fn expected(
116 3132 : tenant_id: TenantId,
117 3132 : timeline_id: TimelineId,
118 3132 : keys: Range<Key>,
119 3132 : lsns: Range<Lsn>,
120 3132 : ) -> Self {
121 3132 : Self {
122 3132 : magic: DELTA_FILE_MAGIC,
123 3132 : format_version: STORAGE_FORMAT_VERSION,
124 3132 :
125 3132 : tenant_id,
126 3132 : timeline_id,
127 3132 : key_range: keys,
128 3132 : lsn_range: lsns,
129 3132 :
130 3132 : index_start_blk: 0,
131 3132 : index_root_blk: 0,
132 3132 : }
133 3132 : }
134 : }
135 :
136 : // Flag indicating that this version initialize the page
137 : const WILL_INIT: u64 = 1;
138 :
139 : /// Struct representing reference to BLOB in layers.
140 : ///
141 : /// Reference contains BLOB offset, and for WAL records it also contains
142 : /// `will_init` flag. The flag helps to determine the range of records
143 : /// that needs to be applied, without reading/deserializing records themselves.
144 0 : #[derive(Debug, Serialize, Deserialize, Copy, Clone)]
145 : pub struct BlobRef(pub u64);
146 :
147 : impl BlobRef {
148 1529861 : pub fn will_init(&self) -> bool {
149 1529861 : (self.0 & WILL_INIT) != 0
150 1529861 : }
151 :
152 14237519 : pub fn pos(&self) -> u64 {
153 14237519 : self.0 >> 1
154 14237519 : }
155 :
156 19413366 : pub fn new(pos: u64, will_init: bool) -> BlobRef {
157 19413366 : let mut blob_ref = pos << 1;
158 19413366 : if will_init {
159 19409010 : blob_ref |= WILL_INIT;
160 19409010 : }
161 19413366 : BlobRef(blob_ref)
162 19413366 : }
163 : }
164 :
165 : pub const DELTA_KEY_SIZE: usize = KEY_SIZE + 8;
166 : struct DeltaKey([u8; DELTA_KEY_SIZE]);
167 :
168 : /// This is the key of the B-tree index stored in the delta layer. It consists
169 : /// of the serialized representation of a Key and LSN.
170 : impl DeltaKey {
171 6192594 : fn from_slice(buf: &[u8]) -> Self {
172 6192594 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
173 6192594 : bytes.copy_from_slice(buf);
174 6192594 : DeltaKey(bytes)
175 6192594 : }
176 :
177 20153027 : fn from_key_lsn(key: &Key, lsn: Lsn) -> Self {
178 20153027 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
179 20153027 : key.write_to_byte_slice(&mut bytes[0..KEY_SIZE]);
180 20153027 : bytes[KEY_SIZE..].copy_from_slice(&u64::to_be_bytes(lsn.0));
181 20153027 : DeltaKey(bytes)
182 20153027 : }
183 :
184 6192594 : fn key(&self) -> Key {
185 6192594 : Key::from_slice(&self.0)
186 6192594 : }
187 :
188 6192594 : fn lsn(&self) -> Lsn {
189 6192594 : Lsn(u64::from_be_bytes(self.0[KEY_SIZE..].try_into().unwrap()))
190 6192594 : }
191 :
192 8044745 : fn extract_lsn_from_buf(buf: &[u8]) -> Lsn {
193 8044745 : let mut lsn_buf = [0u8; 8];
194 8044745 : lsn_buf.copy_from_slice(&buf[KEY_SIZE..]);
195 8044745 : Lsn(u64::from_be_bytes(lsn_buf))
196 8044745 : }
197 : }
198 :
199 : /// This is used only from `pagectl`. Within pageserver, all layers are
200 : /// [`crate::tenant::storage_layer::Layer`], which can hold a [`DeltaLayerInner`].
201 : pub struct DeltaLayer {
202 : path: Utf8PathBuf,
203 : pub desc: PersistentLayerDesc,
204 : inner: OnceCell<Arc<DeltaLayerInner>>,
205 : }
206 :
207 : impl std::fmt::Debug for DeltaLayer {
208 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
209 : use super::RangeDisplayDebug;
210 :
211 0 : f.debug_struct("DeltaLayer")
212 0 : .field("key_range", &RangeDisplayDebug(&self.desc.key_range))
213 0 : .field("lsn_range", &self.desc.lsn_range)
214 0 : .field("file_size", &self.desc.file_size)
215 0 : .field("inner", &self.inner)
216 0 : .finish()
217 0 : }
218 : }
219 :
220 : /// `DeltaLayerInner` is the in-memory data structure associated with an on-disk delta
221 : /// file.
222 : pub struct DeltaLayerInner {
223 : // values copied from summary
224 : index_start_blk: u32,
225 : index_root_blk: u32,
226 :
227 : file: VirtualFile,
228 : file_id: FileId,
229 :
230 : layer_key_range: Range<Key>,
231 : layer_lsn_range: Range<Lsn>,
232 :
233 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
234 : }
235 :
236 : impl DeltaLayerInner {
237 0 : pub(crate) fn layer_dbg_info(&self) -> String {
238 0 : format!(
239 0 : "delta {}..{} {}..{}",
240 0 : self.key_range().start,
241 0 : self.key_range().end,
242 0 : self.lsn_range().start,
243 0 : self.lsn_range().end
244 0 : )
245 0 : }
246 : }
247 :
248 : impl std::fmt::Debug for DeltaLayerInner {
249 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
250 0 : f.debug_struct("DeltaLayerInner")
251 0 : .field("index_start_blk", &self.index_start_blk)
252 0 : .field("index_root_blk", &self.index_root_blk)
253 0 : .finish()
254 0 : }
255 : }
256 :
257 : /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
258 : impl std::fmt::Display for DeltaLayer {
259 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
260 0 : write!(f, "{}", self.layer_desc().short_id())
261 0 : }
262 : }
263 :
264 : impl AsLayerDesc for DeltaLayer {
265 0 : fn layer_desc(&self) -> &PersistentLayerDesc {
266 0 : &self.desc
267 0 : }
268 : }
269 :
270 : impl DeltaLayer {
271 0 : pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
272 0 : self.desc.dump();
273 0 :
274 0 : if !verbose {
275 0 : return Ok(());
276 0 : }
277 :
278 0 : let inner = self.load(ctx).await?;
279 :
280 0 : inner.dump(ctx).await
281 0 : }
282 :
283 4242 : fn temp_path_for(
284 4242 : conf: &PageServerConf,
285 4242 : tenant_shard_id: &TenantShardId,
286 4242 : timeline_id: &TimelineId,
287 4242 : key_start: Key,
288 4242 : lsn_range: &Range<Lsn>,
289 4242 : ) -> Utf8PathBuf {
290 4242 : let rand_string: String = rand::thread_rng()
291 4242 : .sample_iter(&Alphanumeric)
292 4242 : .take(8)
293 4242 : .map(char::from)
294 4242 : .collect();
295 4242 :
296 4242 : conf.timeline_path(tenant_shard_id, timeline_id)
297 4242 : .join(format!(
298 4242 : "{}-XXX__{:016X}-{:016X}.{}.{}",
299 4242 : key_start,
300 4242 : u64::from(lsn_range.start),
301 4242 : u64::from(lsn_range.end),
302 4242 : rand_string,
303 4242 : TEMP_FILE_SUFFIX,
304 4242 : ))
305 4242 : }
306 :
307 : ///
308 : /// Open the underlying file and read the metadata into memory, if it's
309 : /// not loaded already.
310 : ///
311 0 : async fn load(&self, ctx: &RequestContext) -> Result<&Arc<DeltaLayerInner>> {
312 0 : // Quick exit if already loaded
313 0 : self.inner
314 0 : .get_or_try_init(|| self.load_inner(ctx))
315 0 : .await
316 0 : .with_context(|| format!("Failed to load delta layer {}", self.path()))
317 0 : }
318 :
319 0 : async fn load_inner(&self, ctx: &RequestContext) -> anyhow::Result<Arc<DeltaLayerInner>> {
320 0 : let path = self.path();
321 :
322 0 : let loaded = DeltaLayerInner::load(&path, None, None, ctx).await?;
323 :
324 : // not production code
325 0 : let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
326 0 : let expected_layer_name = self.layer_desc().layer_name();
327 0 :
328 0 : if actual_layer_name != expected_layer_name {
329 0 : println!("warning: filename does not match what is expected from in-file summary");
330 0 : println!("actual: {:?}", actual_layer_name.to_string());
331 0 : println!("expected: {:?}", expected_layer_name.to_string());
332 0 : }
333 :
334 0 : Ok(Arc::new(loaded))
335 0 : }
336 :
337 : /// Create a DeltaLayer struct representing an existing file on disk.
338 : ///
339 : /// This variant is only used for debugging purposes, by the 'pagectl' binary.
340 0 : pub fn new_for_path(path: &Utf8Path, file: File) -> Result<Self> {
341 0 : let mut summary_buf = vec![0; PAGE_SZ];
342 0 : file.read_exact_at(&mut summary_buf, 0)?;
343 0 : let summary = Summary::des_prefix(&summary_buf)?;
344 :
345 0 : let metadata = file
346 0 : .metadata()
347 0 : .context("get file metadata to determine size")?;
348 :
349 : // This function is never used for constructing layers in a running pageserver,
350 : // so it does not need an accurate TenantShardId.
351 0 : let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
352 0 :
353 0 : Ok(DeltaLayer {
354 0 : path: path.to_path_buf(),
355 0 : desc: PersistentLayerDesc::new_delta(
356 0 : tenant_shard_id,
357 0 : summary.timeline_id,
358 0 : summary.key_range,
359 0 : summary.lsn_range,
360 0 : metadata.len(),
361 0 : ),
362 0 : inner: OnceCell::new(),
363 0 : })
364 0 : }
365 :
366 : /// Path to the layer file in pageserver workdir.
367 0 : fn path(&self) -> Utf8PathBuf {
368 0 : self.path.clone()
369 0 : }
370 : }
371 :
372 : /// A builder object for constructing a new delta layer.
373 : ///
374 : /// Usage:
375 : ///
376 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
377 : ///
378 : /// 2. Write the contents by calling `put_value` for every page
379 : /// version to store in the layer.
380 : ///
381 : /// 3. Call `finish`.
382 : ///
383 : struct DeltaLayerWriterInner {
384 : pub path: Utf8PathBuf,
385 : timeline_id: TimelineId,
386 : tenant_shard_id: TenantShardId,
387 :
388 : key_start: Key,
389 : lsn_range: Range<Lsn>,
390 :
391 : tree: DiskBtreeBuilder<BlockBuf, DELTA_KEY_SIZE>,
392 :
393 : blob_writer: BlobWriter<true>,
394 :
395 : // Number of key-lsns in the layer.
396 : num_keys: usize,
397 : }
398 :
399 : impl DeltaLayerWriterInner {
400 : ///
401 : /// Start building a new delta layer.
402 : ///
403 4242 : async fn new(
404 4242 : conf: &'static PageServerConf,
405 4242 : timeline_id: TimelineId,
406 4242 : tenant_shard_id: TenantShardId,
407 4242 : key_start: Key,
408 4242 : lsn_range: Range<Lsn>,
409 4242 : ctx: &RequestContext,
410 4242 : ) -> anyhow::Result<Self> {
411 4242 : // Create the file initially with a temporary filename. We don't know
412 4242 : // the end key yet, so we cannot form the final filename yet. We will
413 4242 : // rename it when we're done.
414 4242 : //
415 4242 : // Note: This overwrites any existing file. There shouldn't be any.
416 4242 : // FIXME: throw an error instead?
417 4242 : let path =
418 4242 : DeltaLayer::temp_path_for(conf, &tenant_shard_id, &timeline_id, key_start, &lsn_range);
419 :
420 4242 : let mut file = VirtualFile::create(&path, ctx).await?;
421 : // make room for the header block
422 4242 : file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
423 4242 : let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
424 4242 :
425 4242 : // Initialize the b-tree index builder
426 4242 : let block_buf = BlockBuf::new();
427 4242 : let tree_builder = DiskBtreeBuilder::new(block_buf);
428 4242 :
429 4242 : Ok(Self {
430 4242 : path,
431 4242 : timeline_id,
432 4242 : tenant_shard_id,
433 4242 : key_start,
434 4242 : lsn_range,
435 4242 : tree: tree_builder,
436 4242 : blob_writer,
437 4242 : num_keys: 0,
438 4242 : })
439 4242 : }
440 :
441 : ///
442 : /// Append a key-value pair to the file.
443 : ///
444 : /// The values must be appended in key, lsn order.
445 : ///
446 6253074 : async fn put_value(
447 6253074 : &mut self,
448 6253074 : key: Key,
449 6253074 : lsn: Lsn,
450 6253074 : val: Value,
451 6253074 : ctx: &RequestContext,
452 6253074 : ) -> anyhow::Result<()> {
453 6253074 : let (_, res) = self
454 6253074 : .put_value_bytes(
455 6253074 : key,
456 6253074 : lsn,
457 6253074 : Value::ser(&val)?.slice_len(),
458 6253074 : val.will_init(),
459 6253074 : ctx,
460 : )
461 5946 : .await;
462 6253074 : res
463 6253074 : }
464 :
465 19413054 : async fn put_value_bytes<Buf>(
466 19413054 : &mut self,
467 19413054 : key: Key,
468 19413054 : lsn: Lsn,
469 19413054 : val: FullSlice<Buf>,
470 19413054 : will_init: bool,
471 19413054 : ctx: &RequestContext,
472 19413054 : ) -> (FullSlice<Buf>, anyhow::Result<()>)
473 19413054 : where
474 19413054 : Buf: IoBuf + Send,
475 19413054 : {
476 19413054 : assert!(
477 19413054 : self.lsn_range.start <= lsn,
478 0 : "lsn_start={}, lsn={}",
479 : self.lsn_range.start,
480 : lsn
481 : );
482 : // We don't want to use compression in delta layer creation
483 19413054 : let compression = ImageCompressionAlgorithm::Disabled;
484 19413054 : let (val, res) = self
485 19413054 : .blob_writer
486 19413054 : .write_blob_maybe_compressed(val, ctx, compression)
487 14788 : .await;
488 19413054 : let off = match res {
489 19413054 : Ok((off, _)) => off,
490 0 : Err(e) => return (val, Err(anyhow::anyhow!(e))),
491 : };
492 :
493 19413054 : let blob_ref = BlobRef::new(off, will_init);
494 19413054 :
495 19413054 : let delta_key = DeltaKey::from_key_lsn(&key, lsn);
496 19413054 : let res = self.tree.append(&delta_key.0, blob_ref.0);
497 19413054 :
498 19413054 : self.num_keys += 1;
499 19413054 :
500 19413054 : (val, res.map_err(|e| anyhow::anyhow!(e)))
501 19413054 : }
502 :
503 6071916 : fn size(&self) -> u64 {
504 6071916 : self.blob_writer.size() + self.tree.borrow_writer().size()
505 6071916 : }
506 :
507 : ///
508 : /// Finish writing the delta layer.
509 : ///
510 4206 : async fn finish(
511 4206 : self,
512 4206 : key_end: Key,
513 4206 : ctx: &RequestContext,
514 4206 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
515 4206 : let temp_path = self.path.clone();
516 29003 : let result = self.finish0(key_end, ctx).await;
517 4206 : if result.is_err() {
518 0 : tracing::info!(%temp_path, "cleaning up temporary file after error during writing");
519 0 : if let Err(e) = std::fs::remove_file(&temp_path) {
520 0 : tracing::warn!(error=%e, %temp_path, "error cleaning up temporary layer file after error during writing");
521 0 : }
522 4206 : }
523 4206 : result
524 4206 : }
525 :
526 4206 : async fn finish0(
527 4206 : self,
528 4206 : key_end: Key,
529 4206 : ctx: &RequestContext,
530 4206 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
531 4206 : let index_start_blk =
532 4206 : ((self.blob_writer.size() + PAGE_SZ as u64 - 1) / PAGE_SZ as u64) as u32;
533 :
534 4206 : let mut file = self.blob_writer.into_inner(ctx).await?;
535 :
536 : // Write out the index
537 4206 : let (index_root_blk, block_buf) = self.tree.finish()?;
538 4206 : file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
539 0 : .await?;
540 45071 : for buf in block_buf.blocks {
541 40865 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
542 40865 : res?;
543 : }
544 4206 : assert!(self.lsn_range.start < self.lsn_range.end);
545 : // Fill in the summary on blk 0
546 4206 : let summary = Summary {
547 4206 : magic: DELTA_FILE_MAGIC,
548 4206 : format_version: STORAGE_FORMAT_VERSION,
549 4206 : tenant_id: self.tenant_shard_id.tenant_id,
550 4206 : timeline_id: self.timeline_id,
551 4206 : key_range: self.key_start..key_end,
552 4206 : lsn_range: self.lsn_range.clone(),
553 4206 : index_start_blk,
554 4206 : index_root_blk,
555 4206 : };
556 4206 :
557 4206 : let mut buf = Vec::with_capacity(PAGE_SZ);
558 4206 : // TODO: could use smallvec here but it's a pain with Slice<T>
559 4206 : Summary::ser_into(&summary, &mut buf)?;
560 4206 : file.seek(SeekFrom::Start(0)).await?;
561 4206 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
562 4206 : res?;
563 :
564 4206 : let metadata = file
565 4206 : .metadata()
566 2109 : .await
567 4206 : .context("get file metadata to determine size")?;
568 :
569 : // 5GB limit for objects without multipart upload (which we don't want to use)
570 : // Make it a little bit below to account for differing GB units
571 : // https://docs.aws.amazon.com/AmazonS3/latest/userguide/upload-objects.html
572 4206 : ensure!(
573 4206 : metadata.len() <= S3_UPLOAD_LIMIT,
574 0 : "Created delta layer file at {} of size {} above limit {S3_UPLOAD_LIMIT}!",
575 0 : file.path,
576 0 : metadata.len()
577 : );
578 :
579 : // Note: Because we opened the file in write-only mode, we cannot
580 : // reuse the same VirtualFile for reading later. That's why we don't
581 : // set inner.file here. The first read will have to re-open it.
582 :
583 4206 : let desc = PersistentLayerDesc::new_delta(
584 4206 : self.tenant_shard_id,
585 4206 : self.timeline_id,
586 4206 : self.key_start..key_end,
587 4206 : self.lsn_range.clone(),
588 4206 : metadata.len(),
589 4206 : );
590 4206 :
591 4206 : // fsync the file
592 4206 : file.sync_all().await?;
593 :
594 4206 : trace!("created delta layer {}", self.path);
595 :
596 4206 : Ok((desc, self.path))
597 4206 : }
598 : }
599 :
600 : /// A builder object for constructing a new delta layer.
601 : ///
602 : /// Usage:
603 : ///
604 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
605 : ///
606 : /// 2. Write the contents by calling `put_value` for every page
607 : /// version to store in the layer.
608 : ///
609 : /// 3. Call `finish`.
610 : ///
611 : /// # Note
612 : ///
613 : /// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
614 : /// possible for the writer to drop before `finish` is actually called. So this
615 : /// could lead to odd temporary files in the directory, exhausting file system.
616 : /// This structure wraps `DeltaLayerWriterInner` and also contains `Drop`
617 : /// implementation that cleans up the temporary file in failure. It's not
618 : /// possible to do this directly in `DeltaLayerWriterInner` since `finish` moves
619 : /// out some fields, making it impossible to implement `Drop`.
620 : ///
621 : #[must_use]
622 : pub struct DeltaLayerWriter {
623 : inner: Option<DeltaLayerWriterInner>,
624 : }
625 :
626 : impl DeltaLayerWriter {
627 : ///
628 : /// Start building a new delta layer.
629 : ///
630 4242 : pub async fn new(
631 4242 : conf: &'static PageServerConf,
632 4242 : timeline_id: TimelineId,
633 4242 : tenant_shard_id: TenantShardId,
634 4242 : key_start: Key,
635 4242 : lsn_range: Range<Lsn>,
636 4242 : ctx: &RequestContext,
637 4242 : ) -> anyhow::Result<Self> {
638 4242 : Ok(Self {
639 4242 : inner: Some(
640 4242 : DeltaLayerWriterInner::new(
641 4242 : conf,
642 4242 : timeline_id,
643 4242 : tenant_shard_id,
644 4242 : key_start,
645 4242 : lsn_range,
646 4242 : ctx,
647 4242 : )
648 2152 : .await?,
649 : ),
650 : })
651 4242 : }
652 :
653 : ///
654 : /// Append a key-value pair to the file.
655 : ///
656 : /// The values must be appended in key, lsn order.
657 : ///
658 6253074 : pub async fn put_value(
659 6253074 : &mut self,
660 6253074 : key: Key,
661 6253074 : lsn: Lsn,
662 6253074 : val: Value,
663 6253074 : ctx: &RequestContext,
664 6253074 : ) -> anyhow::Result<()> {
665 6253074 : self.inner
666 6253074 : .as_mut()
667 6253074 : .unwrap()
668 6253074 : .put_value(key, lsn, val, ctx)
669 5946 : .await
670 6253074 : }
671 :
672 13159980 : pub async fn put_value_bytes<Buf>(
673 13159980 : &mut self,
674 13159980 : key: Key,
675 13159980 : lsn: Lsn,
676 13159980 : val: FullSlice<Buf>,
677 13159980 : will_init: bool,
678 13159980 : ctx: &RequestContext,
679 13159980 : ) -> (FullSlice<Buf>, anyhow::Result<()>)
680 13159980 : where
681 13159980 : Buf: IoBuf + Send,
682 13159980 : {
683 13159980 : self.inner
684 13159980 : .as_mut()
685 13159980 : .unwrap()
686 13159980 : .put_value_bytes(key, lsn, val, will_init, ctx)
687 8842 : .await
688 13159980 : }
689 :
690 6071916 : pub fn size(&self) -> u64 {
691 6071916 : self.inner.as_ref().unwrap().size()
692 6071916 : }
693 :
694 : ///
695 : /// Finish writing the delta layer.
696 : ///
697 4206 : pub(crate) async fn finish(
698 4206 : mut self,
699 4206 : key_end: Key,
700 4206 : ctx: &RequestContext,
701 4206 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
702 29003 : self.inner.take().unwrap().finish(key_end, ctx).await
703 4206 : }
704 :
705 36486 : pub(crate) fn num_keys(&self) -> usize {
706 36486 : self.inner.as_ref().unwrap().num_keys
707 36486 : }
708 :
709 45252 : pub(crate) fn estimated_size(&self) -> u64 {
710 45252 : let inner = self.inner.as_ref().unwrap();
711 45252 : inner.blob_writer.size() + inner.tree.borrow_writer().size() + PAGE_SZ as u64
712 45252 : }
713 : }
714 :
715 : impl Drop for DeltaLayerWriter {
716 4242 : fn drop(&mut self) {
717 4242 : if let Some(inner) = self.inner.take() {
718 36 : // We want to remove the virtual file here, so it's fine to not
719 36 : // having completely flushed unwritten data.
720 36 : let vfile = inner.blob_writer.into_inner_no_flush();
721 36 : vfile.remove();
722 4206 : }
723 4242 : }
724 : }
725 :
726 0 : #[derive(thiserror::Error, Debug)]
727 : pub enum RewriteSummaryError {
728 : #[error("magic mismatch")]
729 : MagicMismatch,
730 : #[error(transparent)]
731 : Other(#[from] anyhow::Error),
732 : }
733 :
734 : impl From<std::io::Error> for RewriteSummaryError {
735 0 : fn from(e: std::io::Error) -> Self {
736 0 : Self::Other(anyhow::anyhow!(e))
737 0 : }
738 : }
739 :
740 : impl DeltaLayer {
741 0 : pub async fn rewrite_summary<F>(
742 0 : path: &Utf8Path,
743 0 : rewrite: F,
744 0 : ctx: &RequestContext,
745 0 : ) -> Result<(), RewriteSummaryError>
746 0 : where
747 0 : F: Fn(Summary) -> Summary,
748 0 : {
749 0 : let mut file = VirtualFile::open_with_options(
750 0 : path,
751 0 : virtual_file::OpenOptions::new().read(true).write(true),
752 0 : ctx,
753 0 : )
754 0 : .await
755 0 : .with_context(|| format!("Failed to open file '{}'", path))?;
756 0 : let file_id = page_cache::next_file_id();
757 0 : let block_reader = FileBlockReader::new(&file, file_id);
758 0 : let summary_blk = block_reader.read_blk(0, ctx).await?;
759 0 : let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
760 0 : if actual_summary.magic != DELTA_FILE_MAGIC {
761 0 : return Err(RewriteSummaryError::MagicMismatch);
762 0 : }
763 0 :
764 0 : let new_summary = rewrite(actual_summary);
765 0 :
766 0 : let mut buf = Vec::with_capacity(PAGE_SZ);
767 0 : // TODO: could use smallvec here, but it's a pain with Slice<T>
768 0 : Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
769 0 : file.seek(SeekFrom::Start(0)).await?;
770 0 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
771 0 : res?;
772 0 : Ok(())
773 0 : }
774 : }
775 :
776 : impl DeltaLayerInner {
777 1422 : pub(crate) fn key_range(&self) -> &Range<Key> {
778 1422 : &self.layer_key_range
779 1422 : }
780 :
781 1422 : pub(crate) fn lsn_range(&self) -> &Range<Lsn> {
782 1422 : &self.layer_lsn_range
783 1422 : }
784 :
785 3132 : pub(super) async fn load(
786 3132 : path: &Utf8Path,
787 3132 : summary: Option<Summary>,
788 3132 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
789 3132 : ctx: &RequestContext,
790 3132 : ) -> anyhow::Result<Self> {
791 3132 : let file = VirtualFile::open(path, ctx)
792 1566 : .await
793 3132 : .context("open layer file")?;
794 :
795 3132 : let file_id = page_cache::next_file_id();
796 3132 :
797 3132 : let block_reader = FileBlockReader::new(&file, file_id);
798 :
799 3132 : let summary_blk = block_reader
800 3132 : .read_blk(0, ctx)
801 1579 : .await
802 3132 : .context("read first block")?;
803 :
804 : // TODO: this should be an assertion instead; see ImageLayerInner::load
805 3132 : let actual_summary =
806 3132 : Summary::des_prefix(summary_blk.as_ref()).context("deserialize first block")?;
807 :
808 3132 : if let Some(mut expected_summary) = summary {
809 : // production code path
810 3132 : expected_summary.index_start_blk = actual_summary.index_start_blk;
811 3132 : expected_summary.index_root_blk = actual_summary.index_root_blk;
812 3132 : // mask out the timeline_id, but still require the layers to be from the same tenant
813 3132 : expected_summary.timeline_id = actual_summary.timeline_id;
814 3132 :
815 3132 : if actual_summary != expected_summary {
816 0 : bail!(
817 0 : "in-file summary does not match expected summary. actual = {:?} expected = {:?}",
818 0 : actual_summary,
819 0 : expected_summary
820 0 : );
821 3132 : }
822 0 : }
823 :
824 3132 : Ok(DeltaLayerInner {
825 3132 : file,
826 3132 : file_id,
827 3132 : index_start_blk: actual_summary.index_start_blk,
828 3132 : index_root_blk: actual_summary.index_root_blk,
829 3132 : max_vectored_read_bytes,
830 3132 : layer_key_range: actual_summary.key_range,
831 3132 : layer_lsn_range: actual_summary.lsn_range,
832 3132 : })
833 3132 : }
834 :
835 : // Look up the keys in the provided keyspace and update
836 : // the reconstruct state with whatever is found.
837 : //
838 : // If the key is cached, go no further than the cached Lsn.
839 : //
840 : // Currently, the index is visited for each range, but this
841 : // can be further optimised to visit the index only once.
842 612459 : pub(super) async fn get_values_reconstruct_data(
843 612459 : &self,
844 612459 : keyspace: KeySpace,
845 612459 : lsn_range: Range<Lsn>,
846 612459 : reconstruct_state: &mut ValuesReconstructState,
847 612459 : ctx: &RequestContext,
848 612459 : ) -> Result<(), GetVectoredError> {
849 612459 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
850 612459 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
851 612459 : self.index_start_blk,
852 612459 : self.index_root_blk,
853 612459 : block_reader,
854 612459 : );
855 612459 :
856 612459 : let planner = VectoredReadPlanner::new(
857 612459 : self.max_vectored_read_bytes
858 612459 : .expect("Layer is loaded with max vectored bytes config")
859 612459 : .0
860 612459 : .into(),
861 612459 : );
862 612459 :
863 612459 : let data_end_offset = self.index_start_offset();
864 :
865 612459 : let reads = Self::plan_reads(
866 612459 : &keyspace,
867 612459 : lsn_range.clone(),
868 612459 : data_end_offset,
869 612459 : index_reader,
870 612459 : planner,
871 612459 : reconstruct_state,
872 612459 : ctx,
873 612459 : )
874 61955 : .await
875 612459 : .map_err(GetVectoredError::Other)?;
876 :
877 612459 : self.do_reads_and_update_state(reads, reconstruct_state, ctx)
878 205079 : .await;
879 :
880 612459 : reconstruct_state.on_lsn_advanced(&keyspace, lsn_range.start);
881 612459 :
882 612459 : Ok(())
883 612459 : }
884 :
885 613065 : async fn plan_reads<Reader>(
886 613065 : keyspace: &KeySpace,
887 613065 : lsn_range: Range<Lsn>,
888 613065 : data_end_offset: u64,
889 613065 : index_reader: DiskBtreeReader<Reader, DELTA_KEY_SIZE>,
890 613065 : mut planner: VectoredReadPlanner,
891 613065 : reconstruct_state: &mut ValuesReconstructState,
892 613065 : ctx: &RequestContext,
893 613065 : ) -> anyhow::Result<Vec<VectoredRead>>
894 613065 : where
895 613065 : Reader: BlockReader + Clone,
896 613065 : {
897 613065 : let ctx = RequestContextBuilder::extend(ctx)
898 613065 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
899 613065 : .build();
900 :
901 739871 : for range in keyspace.ranges.iter() {
902 739871 : let mut range_end_handled = false;
903 739871 :
904 739871 : let start_key = DeltaKey::from_key_lsn(&range.start, lsn_range.start);
905 739871 : let index_stream = index_reader.clone().into_stream(&start_key.0, &ctx);
906 739871 : let mut index_stream = std::pin::pin!(index_stream);
907 :
908 1775593 : while let Some(index_entry) = index_stream.next().await {
909 1749067 : let (raw_key, value) = index_entry?;
910 1749067 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
911 1749067 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
912 1749067 : let blob_ref = BlobRef(value);
913 1749067 :
914 1749067 : // Lsns are not monotonically increasing across keys, so we don't assert on them.
915 1749067 : assert!(key >= range.start);
916 :
917 1749067 : let outside_lsn_range = !lsn_range.contains(&lsn);
918 1749067 : let below_cached_lsn = reconstruct_state.get_cached_lsn(&key) >= Some(lsn);
919 :
920 1749067 : let flag = {
921 1749067 : if outside_lsn_range || below_cached_lsn {
922 219206 : BlobFlag::Ignore
923 1529861 : } else if blob_ref.will_init() {
924 1354991 : BlobFlag::ReplaceAll
925 : } else {
926 : // Usual path: add blob to the read
927 174870 : BlobFlag::None
928 : }
929 : };
930 :
931 1749067 : if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
932 713345 : planner.handle_range_end(blob_ref.pos());
933 713345 : range_end_handled = true;
934 713345 : break;
935 1035722 : } else {
936 1035722 : planner.handle(key, lsn, blob_ref.pos(), flag);
937 1035722 : }
938 : }
939 :
940 739871 : if !range_end_handled {
941 26526 : tracing::debug!("Handling range end fallback at {}", data_end_offset);
942 26526 : planner.handle_range_end(data_end_offset);
943 713345 : }
944 : }
945 :
946 613065 : Ok(planner.finish())
947 613065 : }
948 :
949 613059 : fn get_min_read_buffer_size(
950 613059 : planned_reads: &[VectoredRead],
951 613059 : read_size_soft_max: usize,
952 613059 : ) -> usize {
953 613059 : let Some(largest_read) = planned_reads.iter().max_by_key(|read| read.size()) else {
954 247148 : return read_size_soft_max;
955 : };
956 :
957 365911 : let largest_read_size = largest_read.size();
958 365911 : if largest_read_size > read_size_soft_max {
959 : // If the read is oversized, it should only contain one key.
960 600 : let offenders = largest_read
961 600 : .blobs_at
962 600 : .as_slice()
963 600 : .iter()
964 600 : .map(|(_, blob_meta)| format!("{}@{}", blob_meta.key, blob_meta.lsn))
965 600 : .join(", ");
966 600 : tracing::warn!(
967 0 : "Oversized vectored read ({} > {}) for keys {}",
968 : largest_read_size,
969 : read_size_soft_max,
970 : offenders
971 : );
972 365311 : }
973 :
974 365911 : largest_read_size
975 613059 : }
976 :
977 612459 : async fn do_reads_and_update_state(
978 612459 : &self,
979 612459 : reads: Vec<VectoredRead>,
980 612459 : reconstruct_state: &mut ValuesReconstructState,
981 612459 : ctx: &RequestContext,
982 612459 : ) {
983 612459 : let vectored_blob_reader = VectoredBlobReader::new(&self.file);
984 612459 : let mut ignore_key_with_err = None;
985 612459 :
986 612459 : let max_vectored_read_bytes = self
987 612459 : .max_vectored_read_bytes
988 612459 : .expect("Layer is loaded with max vectored bytes config")
989 612459 : .0
990 612459 : .into();
991 612459 : let buf_size = Self::get_min_read_buffer_size(&reads, max_vectored_read_bytes);
992 612459 : let mut buf = Some(BytesMut::with_capacity(buf_size));
993 :
994 : // Note that reads are processed in reverse order (from highest key+lsn).
995 : // This is the order that `ReconstructState` requires such that it can
996 : // track when a key is done.
997 612459 : for read in reads.into_iter().rev() {
998 403163 : let res = vectored_blob_reader
999 403163 : .read_blobs(&read, buf.take().expect("Should have a buffer"), ctx)
1000 205079 : .await;
1001 :
1002 403163 : let blobs_buf = match res {
1003 403163 : Ok(blobs_buf) => blobs_buf,
1004 0 : Err(err) => {
1005 0 : let kind = err.kind();
1006 0 : for (_, blob_meta) in read.blobs_at.as_slice() {
1007 0 : reconstruct_state.on_key_error(
1008 0 : blob_meta.key,
1009 0 : PageReconstructError::Other(anyhow!(
1010 0 : "Failed to read blobs from virtual file {}: {}",
1011 0 : self.file.path,
1012 0 : kind
1013 0 : )),
1014 0 : );
1015 0 : }
1016 :
1017 : // We have "lost" the buffer since the lower level IO api
1018 : // doesn't return the buffer on error. Allocate a new one.
1019 0 : buf = Some(BytesMut::with_capacity(buf_size));
1020 0 :
1021 0 : continue;
1022 : }
1023 : };
1024 :
1025 449421 : for meta in blobs_buf.blobs.iter().rev() {
1026 449421 : if Some(meta.meta.key) == ignore_key_with_err {
1027 0 : continue;
1028 449421 : }
1029 449421 :
1030 449421 : let value = Value::des(&blobs_buf.buf[meta.start..meta.end]);
1031 449421 : let value = match value {
1032 449421 : Ok(v) => v,
1033 0 : Err(e) => {
1034 0 : reconstruct_state.on_key_error(
1035 0 : meta.meta.key,
1036 0 : PageReconstructError::Other(anyhow!(e).context(format!(
1037 0 : "Failed to deserialize blob from virtual file {}",
1038 0 : self.file.path,
1039 0 : ))),
1040 0 : );
1041 0 :
1042 0 : ignore_key_with_err = Some(meta.meta.key);
1043 0 : continue;
1044 : }
1045 : };
1046 :
1047 : // Invariant: once a key reaches [`ValueReconstructSituation::Complete`]
1048 : // state, no further updates shall be made to it. The call below will
1049 : // panic if the invariant is violated.
1050 449421 : reconstruct_state.update_key(&meta.meta.key, meta.meta.lsn, value);
1051 : }
1052 :
1053 403163 : buf = Some(blobs_buf.buf);
1054 : }
1055 612459 : }
1056 :
1057 1218 : pub(super) async fn load_keys<'a>(
1058 1218 : &'a self,
1059 1218 : ctx: &RequestContext,
1060 1218 : ) -> Result<Vec<DeltaEntry<'a>>> {
1061 1218 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1062 1218 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1063 1218 : self.index_start_blk,
1064 1218 : self.index_root_blk,
1065 1218 : block_reader,
1066 1218 : );
1067 1218 :
1068 1218 : let mut all_keys: Vec<DeltaEntry<'_>> = Vec::new();
1069 1218 :
1070 1218 : tree_reader
1071 1218 : .visit(
1072 1218 : &[0u8; DELTA_KEY_SIZE],
1073 1218 : VisitDirection::Forwards,
1074 6192138 : |key, value| {
1075 6192138 : let delta_key = DeltaKey::from_slice(key);
1076 6192138 : let val_ref = ValueRef {
1077 6192138 : blob_ref: BlobRef(value),
1078 6192138 : layer: self,
1079 6192138 : };
1080 6192138 : let pos = BlobRef(value).pos();
1081 6192138 : if let Some(last) = all_keys.last_mut() {
1082 6190920 : // subtract offset of the current and last entries to get the size
1083 6190920 : // of the value associated with this (key, lsn) tuple
1084 6190920 : let first_pos = last.size;
1085 6190920 : last.size = pos - first_pos;
1086 6190920 : }
1087 6192138 : let entry = DeltaEntry {
1088 6192138 : key: delta_key.key(),
1089 6192138 : lsn: delta_key.lsn(),
1090 6192138 : size: pos,
1091 6192138 : val: val_ref,
1092 6192138 : };
1093 6192138 : all_keys.push(entry);
1094 6192138 : true
1095 6192138 : },
1096 1218 : &RequestContextBuilder::extend(ctx)
1097 1218 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
1098 1218 : .build(),
1099 1218 : )
1100 6357 : .await?;
1101 1218 : if let Some(last) = all_keys.last_mut() {
1102 1218 : // Last key occupies all space till end of value storage,
1103 1218 : // which corresponds to beginning of the index
1104 1218 : last.size = self.index_start_offset() - last.size;
1105 1218 : }
1106 1218 : Ok(all_keys)
1107 1218 : }
1108 :
1109 : /// Using the given writer, write out a version which has the earlier Lsns than `until`.
1110 : ///
1111 : /// Return the amount of key value records pushed to the writer.
1112 30 : pub(super) async fn copy_prefix(
1113 30 : &self,
1114 30 : writer: &mut DeltaLayerWriter,
1115 30 : until: Lsn,
1116 30 : ctx: &RequestContext,
1117 30 : ) -> anyhow::Result<usize> {
1118 : use crate::tenant::vectored_blob_io::{
1119 : BlobMeta, VectoredReadBuilder, VectoredReadExtended,
1120 : };
1121 : use futures::stream::TryStreamExt;
1122 :
1123 : #[derive(Debug)]
1124 : enum Item {
1125 : Actual(Key, Lsn, BlobRef),
1126 : Sentinel,
1127 : }
1128 :
1129 : impl From<Item> for Option<(Key, Lsn, BlobRef)> {
1130 210 : fn from(value: Item) -> Self {
1131 210 : match value {
1132 180 : Item::Actual(key, lsn, blob) => Some((key, lsn, blob)),
1133 30 : Item::Sentinel => None,
1134 : }
1135 210 : }
1136 : }
1137 :
1138 : impl Item {
1139 210 : fn offset(&self) -> Option<BlobRef> {
1140 210 : match self {
1141 180 : Item::Actual(_, _, blob) => Some(*blob),
1142 30 : Item::Sentinel => None,
1143 : }
1144 210 : }
1145 :
1146 210 : fn is_last(&self) -> bool {
1147 210 : matches!(self, Item::Sentinel)
1148 210 : }
1149 : }
1150 :
1151 30 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1152 30 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1153 30 : self.index_start_blk,
1154 30 : self.index_root_blk,
1155 30 : block_reader,
1156 30 : );
1157 30 :
1158 30 : let stream = self.stream_index_forwards(tree_reader, &[0u8; DELTA_KEY_SIZE], ctx);
1159 180 : let stream = stream.map_ok(|(key, lsn, pos)| Item::Actual(key, lsn, pos));
1160 30 : // put in a sentinel value for getting the end offset for last item, and not having to
1161 30 : // repeat the whole read part
1162 30 : let stream = stream.chain(futures::stream::once(futures::future::ready(Ok(
1163 30 : Item::Sentinel,
1164 30 : ))));
1165 30 : let mut stream = std::pin::pin!(stream);
1166 30 :
1167 30 : let mut prev: Option<(Key, Lsn, BlobRef)> = None;
1168 30 :
1169 30 : let mut read_builder: Option<VectoredReadBuilder> = None;
1170 30 : let read_mode = VectoredReadCoalesceMode::get();
1171 30 :
1172 30 : let max_read_size = self
1173 30 : .max_vectored_read_bytes
1174 30 : .map(|x| x.0.get())
1175 30 : .unwrap_or(8192);
1176 30 :
1177 30 : let mut buffer = Some(BytesMut::with_capacity(max_read_size));
1178 30 :
1179 30 : // FIXME: buffering of DeltaLayerWriter
1180 30 : let mut per_blob_copy = Vec::new();
1181 30 :
1182 30 : let mut records = 0;
1183 :
1184 240 : while let Some(item) = stream.try_next().await? {
1185 210 : tracing::debug!(?item, "popped");
1186 210 : let offset = item
1187 210 : .offset()
1188 210 : .unwrap_or(BlobRef::new(self.index_start_offset(), false));
1189 :
1190 210 : let actionable = if let Some((key, lsn, start_offset)) = prev.take() {
1191 180 : let end_offset = offset;
1192 180 :
1193 180 : Some((BlobMeta { key, lsn }, start_offset..end_offset))
1194 : } else {
1195 30 : None
1196 : };
1197 :
1198 210 : let is_last = item.is_last();
1199 210 :
1200 210 : prev = Option::from(item);
1201 210 :
1202 210 : let actionable = actionable.filter(|x| x.0.lsn < until);
1203 :
1204 210 : let builder = if let Some((meta, offsets)) = actionable {
1205 : // extend or create a new builder
1206 96 : if read_builder
1207 96 : .as_mut()
1208 96 : .map(|x| x.extend(offsets.start.pos(), offsets.end.pos(), meta))
1209 96 : .unwrap_or(VectoredReadExtended::No)
1210 96 : == VectoredReadExtended::Yes
1211 : {
1212 48 : None
1213 : } else {
1214 48 : read_builder.replace(VectoredReadBuilder::new(
1215 48 : offsets.start.pos(),
1216 48 : offsets.end.pos(),
1217 48 : meta,
1218 48 : max_read_size,
1219 48 : read_mode,
1220 48 : ))
1221 : }
1222 : } else {
1223 : // nothing to do, except perhaps flush any existing for the last element
1224 114 : None
1225 : };
1226 :
1227 : // flush the possible older builder and also the new one if the item was the last one
1228 210 : let builders = builder.into_iter();
1229 210 : let builders = if is_last {
1230 30 : builders.chain(read_builder.take())
1231 : } else {
1232 180 : builders.chain(None)
1233 : };
1234 :
1235 258 : for builder in builders {
1236 48 : let read = builder.build();
1237 48 :
1238 48 : let reader = VectoredBlobReader::new(&self.file);
1239 48 :
1240 48 : let mut buf = buffer.take().unwrap();
1241 48 :
1242 48 : buf.clear();
1243 48 : buf.reserve(read.size());
1244 48 : let res = reader.read_blobs(&read, buf, ctx).await?;
1245 :
1246 144 : for blob in res.blobs {
1247 96 : let key = blob.meta.key;
1248 96 : let lsn = blob.meta.lsn;
1249 96 : let data = &res.buf[blob.start..blob.end];
1250 96 :
1251 96 : #[cfg(debug_assertions)]
1252 96 : Value::des(data)
1253 96 : .with_context(|| {
1254 0 : format!(
1255 0 : "blob failed to deserialize for {}@{}, {}..{}: {:?}",
1256 0 : blob.meta.key,
1257 0 : blob.meta.lsn,
1258 0 : blob.start,
1259 0 : blob.end,
1260 0 : utils::Hex(data)
1261 0 : )
1262 96 : })
1263 96 : .unwrap();
1264 96 :
1265 96 : // is it an image or will_init walrecord?
1266 96 : // FIXME: this could be handled by threading the BlobRef to the
1267 96 : // VectoredReadBuilder
1268 96 : let will_init = crate::repository::ValueBytes::will_init(data)
1269 96 : .inspect_err(|_e| {
1270 0 : #[cfg(feature = "testing")]
1271 0 : tracing::error!(data=?utils::Hex(data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
1272 96 : })
1273 96 : .unwrap_or(false);
1274 96 :
1275 96 : per_blob_copy.clear();
1276 96 : per_blob_copy.extend_from_slice(data);
1277 :
1278 96 : let (tmp, res) = writer
1279 96 : .put_value_bytes(
1280 96 : key,
1281 96 : lsn,
1282 96 : std::mem::take(&mut per_blob_copy).slice_len(),
1283 96 : will_init,
1284 96 : ctx,
1285 96 : )
1286 12 : .await;
1287 96 : per_blob_copy = tmp.into_raw_slice().into_inner();
1288 96 :
1289 96 : res?;
1290 :
1291 96 : records += 1;
1292 : }
1293 :
1294 48 : buffer = Some(res.buf);
1295 : }
1296 : }
1297 :
1298 30 : assert!(
1299 30 : read_builder.is_none(),
1300 0 : "with the sentinel above loop should had handled all"
1301 : );
1302 :
1303 30 : Ok(records)
1304 30 : }
1305 :
1306 12 : pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
1307 12 : println!(
1308 12 : "index_start_blk: {}, root {}",
1309 12 : self.index_start_blk, self.index_root_blk
1310 12 : );
1311 12 :
1312 12 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1313 12 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1314 12 : self.index_start_blk,
1315 12 : self.index_root_blk,
1316 12 : block_reader,
1317 12 : );
1318 12 :
1319 12 : tree_reader.dump().await?;
1320 :
1321 12 : let keys = self.load_keys(ctx).await?;
1322 :
1323 24 : async fn dump_blob(val: &ValueRef<'_>, ctx: &RequestContext) -> anyhow::Result<String> {
1324 24 : let buf = val.load_raw(ctx).await?;
1325 24 : let val = Value::des(&buf)?;
1326 24 : let desc = match val {
1327 24 : Value::Image(img) => {
1328 24 : format!(" img {} bytes", img.len())
1329 : }
1330 0 : Value::WalRecord(rec) => {
1331 0 : let wal_desc = walrecord::describe_wal_record(&rec)?;
1332 0 : format!(
1333 0 : " rec {} bytes will_init: {} {}",
1334 0 : buf.len(),
1335 0 : rec.will_init(),
1336 0 : wal_desc
1337 0 : )
1338 : }
1339 : };
1340 24 : Ok(desc)
1341 24 : }
1342 :
1343 36 : for entry in keys {
1344 24 : let DeltaEntry { key, lsn, val, .. } = entry;
1345 24 : let desc = match dump_blob(&val, ctx).await {
1346 24 : Ok(desc) => desc,
1347 0 : Err(err) => {
1348 0 : format!("ERROR: {err}")
1349 : }
1350 : };
1351 24 : println!(" key {key} at {lsn}: {desc}");
1352 :
1353 : // Print more details about CHECKPOINT records. Would be nice to print details
1354 : // of many other record types too, but these are particularly interesting, as
1355 : // have a lot of special processing for them in walingest.rs.
1356 12 : use pageserver_api::key::CHECKPOINT_KEY;
1357 12 : use postgres_ffi::CheckPoint;
1358 24 : if key == CHECKPOINT_KEY {
1359 0 : let val = val.load(ctx).await?;
1360 0 : match val {
1361 0 : Value::Image(img) => {
1362 0 : let checkpoint = CheckPoint::decode(&img)?;
1363 0 : println!(" CHECKPOINT: {:?}", checkpoint);
1364 : }
1365 0 : Value::WalRecord(_rec) => {
1366 0 : println!(" unexpected walrecord value for checkpoint key");
1367 0 : }
1368 : }
1369 24 : }
1370 : }
1371 :
1372 12 : Ok(())
1373 12 : }
1374 :
1375 90 : fn stream_index_forwards<'a, R>(
1376 90 : &'a self,
1377 90 : reader: DiskBtreeReader<R, DELTA_KEY_SIZE>,
1378 90 : start: &'a [u8; DELTA_KEY_SIZE],
1379 90 : ctx: &'a RequestContext,
1380 90 : ) -> impl futures::stream::Stream<
1381 90 : Item = Result<(Key, Lsn, BlobRef), crate::tenant::disk_btree::DiskBtreeError>,
1382 90 : > + 'a
1383 90 : where
1384 90 : R: BlockReader + 'a,
1385 90 : {
1386 : use futures::stream::TryStreamExt;
1387 90 : let stream = reader.into_stream(start, ctx);
1388 456 : stream.map_ok(|(key, value)| {
1389 456 : let key = DeltaKey::from_slice(&key);
1390 456 : let (key, lsn) = (key.key(), key.lsn());
1391 456 : let offset = BlobRef(value);
1392 456 :
1393 456 : (key, lsn, offset)
1394 456 : })
1395 90 : }
1396 :
1397 : /// The file offset to the first block of index.
1398 : ///
1399 : /// The file structure is summary, values, and index. We often need this for the size of last blob.
1400 615393 : fn index_start_offset(&self) -> u64 {
1401 615393 : let offset = self.index_start_blk as u64 * PAGE_SZ as u64;
1402 615393 : let bref = BlobRef(offset);
1403 615393 : tracing::debug!(
1404 : index_start_blk = self.index_start_blk,
1405 : offset,
1406 0 : pos = bref.pos(),
1407 0 : "index_start_offset"
1408 : );
1409 615393 : offset
1410 615393 : }
1411 :
1412 1590 : pub(crate) fn iter<'a>(&'a self, ctx: &'a RequestContext) -> DeltaLayerIterator<'a> {
1413 1590 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1414 1590 : let tree_reader =
1415 1590 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
1416 1590 : DeltaLayerIterator {
1417 1590 : delta_layer: self,
1418 1590 : ctx,
1419 1590 : index_iter: tree_reader.iter(&[0; DELTA_KEY_SIZE], ctx),
1420 1590 : key_values_batch: std::collections::VecDeque::new(),
1421 1590 : is_end: false,
1422 1590 : planner: StreamingVectoredReadPlanner::new(
1423 1590 : 1024 * 8192, // The default value. Unit tests might use a different value. 1024 * 8K = 8MB buffer.
1424 1590 : 1024, // The default value. Unit tests might use a different value
1425 1590 : ),
1426 1590 : }
1427 1590 : }
1428 : }
1429 :
1430 : /// A set of data associated with a delta layer key and its value
1431 : pub struct DeltaEntry<'a> {
1432 : pub key: Key,
1433 : pub lsn: Lsn,
1434 : /// Size of the stored value
1435 : pub size: u64,
1436 : /// Reference to the on-disk value
1437 : pub val: ValueRef<'a>,
1438 : }
1439 :
1440 : /// Reference to an on-disk value
1441 : pub struct ValueRef<'a> {
1442 : blob_ref: BlobRef,
1443 : layer: &'a DeltaLayerInner,
1444 : }
1445 :
1446 : impl<'a> ValueRef<'a> {
1447 : /// Loads the value from disk
1448 0 : pub async fn load(&self, ctx: &RequestContext) -> Result<Value> {
1449 0 : let buf = self.load_raw(ctx).await?;
1450 0 : let val = Value::des(&buf)?;
1451 0 : Ok(val)
1452 0 : }
1453 :
1454 24 : async fn load_raw(&self, ctx: &RequestContext) -> Result<Vec<u8>> {
1455 24 : let reader = BlockCursor::new(crate::tenant::block_io::BlockReaderRef::Adapter(Adapter(
1456 24 : self.layer,
1457 24 : )));
1458 24 : let buf = reader.read_blob(self.blob_ref.pos(), ctx).await?;
1459 24 : Ok(buf)
1460 24 : }
1461 : }
1462 :
1463 : pub(crate) struct Adapter<T>(T);
1464 :
1465 : impl<T: AsRef<DeltaLayerInner>> Adapter<T> {
1466 24 : pub(crate) async fn read_blk(
1467 24 : &self,
1468 24 : blknum: u32,
1469 24 : ctx: &RequestContext,
1470 24 : ) -> Result<BlockLease, std::io::Error> {
1471 24 : let block_reader = FileBlockReader::new(&self.0.as_ref().file, self.0.as_ref().file_id);
1472 24 : block_reader.read_blk(blknum, ctx).await
1473 24 : }
1474 : }
1475 :
1476 : impl AsRef<DeltaLayerInner> for DeltaLayerInner {
1477 48 : fn as_ref(&self) -> &DeltaLayerInner {
1478 48 : self
1479 48 : }
1480 : }
1481 :
1482 : impl<'a> pageserver_compaction::interface::CompactionDeltaEntry<'a, Key> for DeltaEntry<'a> {
1483 0 : fn key(&self) -> Key {
1484 0 : self.key
1485 0 : }
1486 0 : fn lsn(&self) -> Lsn {
1487 0 : self.lsn
1488 0 : }
1489 0 : fn size(&self) -> u64 {
1490 0 : self.size
1491 0 : }
1492 : }
1493 :
1494 : pub struct DeltaLayerIterator<'a> {
1495 : delta_layer: &'a DeltaLayerInner,
1496 : ctx: &'a RequestContext,
1497 : planner: StreamingVectoredReadPlanner,
1498 : index_iter: DiskBtreeIterator<'a>,
1499 : key_values_batch: VecDeque<(Key, Lsn, Value)>,
1500 : is_end: bool,
1501 : }
1502 :
1503 : impl<'a> DeltaLayerIterator<'a> {
1504 0 : pub(crate) fn layer_dbg_info(&self) -> String {
1505 0 : self.delta_layer.layer_dbg_info()
1506 0 : }
1507 :
1508 : /// Retrieve a batch of key-value pairs into the iterator buffer.
1509 63800 : async fn next_batch(&mut self) -> anyhow::Result<()> {
1510 63800 : assert!(self.key_values_batch.is_empty());
1511 63800 : assert!(!self.is_end);
1512 :
1513 63800 : let plan = loop {
1514 6297184 : if let Some(res) = self.index_iter.next().await {
1515 6295678 : let (raw_key, value) = res?;
1516 6295678 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
1517 6295678 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
1518 6295678 : let blob_ref = BlobRef(value);
1519 6295678 : let offset = blob_ref.pos();
1520 6295678 : if let Some(batch_plan) = self.planner.handle(key, lsn, offset) {
1521 62294 : break batch_plan;
1522 6233384 : }
1523 : } else {
1524 1506 : self.is_end = true;
1525 1506 : let data_end_offset = self.delta_layer.index_start_offset();
1526 1506 : if let Some(item) = self.planner.handle_range_end(data_end_offset) {
1527 1506 : break item;
1528 : } else {
1529 0 : return Ok(()); // TODO: test empty iterator
1530 : }
1531 : }
1532 : };
1533 63800 : let vectored_blob_reader = VectoredBlobReader::new(&self.delta_layer.file);
1534 63800 : let mut next_batch = std::collections::VecDeque::new();
1535 63800 : let buf_size = plan.size();
1536 63800 : let buf = BytesMut::with_capacity(buf_size);
1537 63800 : let blobs_buf = vectored_blob_reader
1538 63800 : .read_blobs(&plan, buf, self.ctx)
1539 32616 : .await?;
1540 63800 : let frozen_buf = blobs_buf.buf.freeze();
1541 6295594 : for meta in blobs_buf.blobs.iter() {
1542 6295594 : let value = Value::des(&frozen_buf[meta.start..meta.end])?;
1543 6295594 : next_batch.push_back((meta.meta.key, meta.meta.lsn, value));
1544 : }
1545 63800 : self.key_values_batch = next_batch;
1546 63800 : Ok(())
1547 63800 : }
1548 :
1549 6297714 : pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
1550 6297714 : if self.key_values_batch.is_empty() {
1551 66476 : if self.is_end {
1552 2928 : return Ok(None);
1553 63548 : }
1554 63548 : self.next_batch().await?;
1555 6231238 : }
1556 6294786 : Ok(Some(
1557 6294786 : self.key_values_batch
1558 6294786 : .pop_front()
1559 6294786 : .expect("should not be empty"),
1560 6294786 : ))
1561 6297714 : }
1562 : }
1563 :
1564 : #[cfg(test)]
1565 : pub(crate) mod test {
1566 : use std::collections::BTreeMap;
1567 :
1568 : use itertools::MinMaxResult;
1569 : use rand::prelude::{SeedableRng, SliceRandom, StdRng};
1570 : use rand::RngCore;
1571 :
1572 : use super::*;
1573 : use crate::repository::Value;
1574 : use crate::tenant::harness::TIMELINE_ID;
1575 : use crate::tenant::storage_layer::{Layer, ResidentLayer};
1576 : use crate::tenant::vectored_blob_io::StreamingVectoredReadPlanner;
1577 : use crate::tenant::{Tenant, Timeline};
1578 : use crate::{
1579 : context::DownloadBehavior,
1580 : task_mgr::TaskKind,
1581 : tenant::{disk_btree::tests::TestDisk, harness::TenantHarness},
1582 : DEFAULT_PG_VERSION,
1583 : };
1584 : use bytes::Bytes;
1585 :
1586 : /// Construct an index for a fictional delta layer and and then
1587 : /// traverse in order to plan vectored reads for a query. Finally,
1588 : /// verify that the traversal fed the right index key and value
1589 : /// pairs into the planner.
1590 : #[tokio::test]
1591 6 : async fn test_delta_layer_index_traversal() {
1592 6 : let base_key = Key {
1593 6 : field1: 0,
1594 6 : field2: 1663,
1595 6 : field3: 12972,
1596 6 : field4: 16396,
1597 6 : field5: 0,
1598 6 : field6: 246080,
1599 6 : };
1600 6 :
1601 6 : // Populate the index with some entries
1602 6 : let entries: BTreeMap<Key, Vec<Lsn>> = BTreeMap::from([
1603 6 : (base_key, vec![Lsn(1), Lsn(5), Lsn(25), Lsn(26), Lsn(28)]),
1604 6 : (base_key.add(1), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1605 6 : (base_key.add(2), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1606 6 : (base_key.add(5), vec![Lsn(10), Lsn(15), Lsn(16), Lsn(20)]),
1607 6 : ]);
1608 6 :
1609 6 : let mut disk = TestDisk::default();
1610 6 : let mut writer = DiskBtreeBuilder::<_, DELTA_KEY_SIZE>::new(&mut disk);
1611 6 :
1612 6 : let mut disk_offset = 0;
1613 30 : for (key, lsns) in &entries {
1614 126 : for lsn in lsns {
1615 102 : let index_key = DeltaKey::from_key_lsn(key, *lsn);
1616 102 : let blob_ref = BlobRef::new(disk_offset, false);
1617 102 : writer
1618 102 : .append(&index_key.0, blob_ref.0)
1619 102 : .expect("In memory disk append should never fail");
1620 102 :
1621 102 : disk_offset += 1;
1622 102 : }
1623 6 : }
1624 6 :
1625 6 : // Prepare all the arguments for the call into `plan_reads` below
1626 6 : let (root_offset, _writer) = writer
1627 6 : .finish()
1628 6 : .expect("In memory disk finish should never fail");
1629 6 : let reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(0, root_offset, disk);
1630 6 : let planner = VectoredReadPlanner::new(100);
1631 6 : let mut reconstruct_state = ValuesReconstructState::new();
1632 6 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1633 6 :
1634 6 : let keyspace = KeySpace {
1635 6 : ranges: vec![
1636 6 : base_key..base_key.add(3),
1637 6 : base_key.add(3)..base_key.add(100),
1638 6 : ],
1639 6 : };
1640 6 : let lsn_range = Lsn(2)..Lsn(40);
1641 6 :
1642 6 : // Plan and validate
1643 6 : let vectored_reads = DeltaLayerInner::plan_reads(
1644 6 : &keyspace,
1645 6 : lsn_range.clone(),
1646 6 : disk_offset,
1647 6 : reader,
1648 6 : planner,
1649 6 : &mut reconstruct_state,
1650 6 : &ctx,
1651 6 : )
1652 6 : .await
1653 6 : .expect("Read planning should not fail");
1654 6 :
1655 6 : validate(keyspace, lsn_range, vectored_reads, entries);
1656 6 : }
1657 :
1658 6 : fn validate(
1659 6 : keyspace: KeySpace,
1660 6 : lsn_range: Range<Lsn>,
1661 6 : vectored_reads: Vec<VectoredRead>,
1662 6 : index_entries: BTreeMap<Key, Vec<Lsn>>,
1663 6 : ) {
1664 : #[derive(Debug, PartialEq, Eq)]
1665 : struct BlobSpec {
1666 : key: Key,
1667 : lsn: Lsn,
1668 : at: u64,
1669 : }
1670 :
1671 6 : let mut planned_blobs = Vec::new();
1672 46 : for read in vectored_reads {
1673 84 : for (at, meta) in read.blobs_at.as_slice() {
1674 84 : planned_blobs.push(BlobSpec {
1675 84 : key: meta.key,
1676 84 : lsn: meta.lsn,
1677 84 : at: *at,
1678 84 : });
1679 84 : }
1680 : }
1681 :
1682 6 : let mut expected_blobs = Vec::new();
1683 6 : let mut disk_offset = 0;
1684 30 : for (key, lsns) in index_entries {
1685 126 : for lsn in lsns {
1686 126 : let key_included = keyspace.ranges.iter().any(|range| range.contains(&key));
1687 102 : let lsn_included = lsn_range.contains(&lsn);
1688 102 :
1689 102 : if key_included && lsn_included {
1690 84 : expected_blobs.push(BlobSpec {
1691 84 : key,
1692 84 : lsn,
1693 84 : at: disk_offset,
1694 84 : });
1695 84 : }
1696 :
1697 102 : disk_offset += 1;
1698 : }
1699 : }
1700 :
1701 6 : assert_eq!(planned_blobs, expected_blobs);
1702 6 : }
1703 :
1704 : mod constants {
1705 : use utils::lsn::Lsn;
1706 :
1707 : /// Offset used by all lsns in this test
1708 : pub(super) const LSN_OFFSET: Lsn = Lsn(0x08);
1709 : /// Number of unique keys including in the test data
1710 : pub(super) const KEY_COUNT: u8 = 60;
1711 : /// Max number of different lsns for each key
1712 : pub(super) const MAX_ENTRIES_PER_KEY: u8 = 20;
1713 : /// Possible value sizes for each key along with a probability weight
1714 : pub(super) const VALUE_SIZES: [(usize, u8); 3] = [(100, 2), (1024, 2), (1024 * 1024, 1)];
1715 : /// Probability that there will be a gap between the current key and the next one (33.3%)
1716 : pub(super) const KEY_GAP_CHANGES: [(bool, u8); 2] = [(true, 1), (false, 2)];
1717 : /// The minimum size of a key range in all the generated reads
1718 : pub(super) const MIN_RANGE_SIZE: i128 = 10;
1719 : /// The number of ranges included in each vectored read
1720 : pub(super) const RANGES_COUNT: u8 = 2;
1721 : /// The number of vectored reads performed
1722 : pub(super) const READS_COUNT: u8 = 100;
1723 : /// Soft max size of a vectored read. Will be violated if we have to read keys
1724 : /// with values larger than the limit
1725 : pub(super) const MAX_VECTORED_READ_BYTES: usize = 64 * 1024;
1726 : }
1727 :
1728 : struct Entry {
1729 : key: Key,
1730 : lsn: Lsn,
1731 : value: Vec<u8>,
1732 : }
1733 :
1734 6 : fn generate_entries(rng: &mut StdRng) -> Vec<Entry> {
1735 6 : let mut current_key = Key::MIN;
1736 6 :
1737 6 : let mut entries = Vec::new();
1738 366 : for _ in 0..constants::KEY_COUNT {
1739 360 : let count = rng.gen_range(1..constants::MAX_ENTRIES_PER_KEY);
1740 360 : let mut lsns_iter =
1741 6780 : std::iter::successors(Some(Lsn(constants::LSN_OFFSET.0 + 0x08)), |lsn| {
1742 6780 : Some(Lsn(lsn.0 + 0x08))
1743 6780 : });
1744 360 : let mut lsns = Vec::new();
1745 7140 : while lsns.len() < count as usize {
1746 6780 : let take = rng.gen_bool(0.5);
1747 6780 : let lsn = lsns_iter.next().unwrap();
1748 6780 : if take {
1749 3336 : lsns.push(lsn);
1750 3444 : }
1751 : }
1752 :
1753 3696 : for lsn in lsns {
1754 3336 : let size = constants::VALUE_SIZES
1755 10008 : .choose_weighted(rng, |item| item.1)
1756 3336 : .unwrap()
1757 3336 : .0;
1758 3336 : let mut buf = vec![0; size];
1759 3336 : rng.fill_bytes(&mut buf);
1760 3336 :
1761 3336 : entries.push(Entry {
1762 3336 : key: current_key,
1763 3336 : lsn,
1764 3336 : value: buf,
1765 3336 : })
1766 : }
1767 :
1768 360 : let gap = constants::KEY_GAP_CHANGES
1769 720 : .choose_weighted(rng, |item| item.1)
1770 360 : .unwrap()
1771 360 : .0;
1772 360 : if gap {
1773 114 : current_key = current_key.add(2);
1774 246 : } else {
1775 246 : current_key = current_key.add(1);
1776 246 : }
1777 : }
1778 :
1779 6 : entries
1780 6 : }
1781 :
1782 : struct EntriesMeta {
1783 : key_range: Range<Key>,
1784 : lsn_range: Range<Lsn>,
1785 : index: BTreeMap<(Key, Lsn), Vec<u8>>,
1786 : }
1787 :
1788 6 : fn get_entries_meta(entries: &[Entry]) -> EntriesMeta {
1789 3336 : let key_range = match entries.iter().minmax_by_key(|e| e.key) {
1790 6 : MinMaxResult::MinMax(min, max) => min.key..max.key.next(),
1791 0 : _ => panic!("More than one entry is always expected"),
1792 : };
1793 :
1794 3336 : let lsn_range = match entries.iter().minmax_by_key(|e| e.lsn) {
1795 6 : MinMaxResult::MinMax(min, max) => min.lsn..Lsn(max.lsn.0 + 1),
1796 0 : _ => panic!("More than one entry is always expected"),
1797 : };
1798 :
1799 6 : let mut index = BTreeMap::new();
1800 3336 : for entry in entries.iter() {
1801 3336 : index.insert((entry.key, entry.lsn), entry.value.clone());
1802 3336 : }
1803 :
1804 6 : EntriesMeta {
1805 6 : key_range,
1806 6 : lsn_range,
1807 6 : index,
1808 6 : }
1809 6 : }
1810 :
1811 600 : fn pick_random_keyspace(rng: &mut StdRng, key_range: &Range<Key>) -> KeySpace {
1812 600 : let start = key_range.start.to_i128();
1813 600 : let end = key_range.end.to_i128();
1814 600 :
1815 600 : let mut keyspace = KeySpace::default();
1816 :
1817 1800 : for _ in 0..constants::RANGES_COUNT {
1818 1200 : let mut range: Option<Range<Key>> = Option::default();
1819 3732 : while range.is_none() || keyspace.overlaps(range.as_ref().unwrap()) {
1820 2532 : let range_start = rng.gen_range(start..end);
1821 2532 : let range_end_offset = range_start + constants::MIN_RANGE_SIZE;
1822 2532 : if range_end_offset >= end {
1823 300 : range = Some(Key::from_i128(range_start)..Key::from_i128(end));
1824 2232 : } else {
1825 2232 : let range_end = rng.gen_range((range_start + constants::MIN_RANGE_SIZE)..end);
1826 2232 : range = Some(Key::from_i128(range_start)..Key::from_i128(range_end));
1827 2232 : }
1828 : }
1829 1200 : keyspace.ranges.push(range.unwrap());
1830 : }
1831 :
1832 600 : keyspace
1833 600 : }
1834 :
1835 : #[tokio::test]
1836 6 : async fn test_delta_layer_vectored_read_end_to_end() -> anyhow::Result<()> {
1837 6 : let harness = TenantHarness::create("test_delta_layer_oversized_vectored_read").await?;
1838 24 : let (tenant, ctx) = harness.load().await;
1839 6 :
1840 6 : let timeline_id = TimelineId::generate();
1841 6 : let timeline = tenant
1842 6 : .create_test_timeline(timeline_id, constants::LSN_OFFSET, DEFAULT_PG_VERSION, &ctx)
1843 12 : .await?;
1844 6 :
1845 6 : tracing::info!("Generating test data ...");
1846 6 :
1847 6 : let rng = &mut StdRng::seed_from_u64(0);
1848 6 : let entries = generate_entries(rng);
1849 6 : let entries_meta = get_entries_meta(&entries);
1850 6 :
1851 6 : tracing::info!("Done generating {} entries", entries.len());
1852 6 :
1853 6 : tracing::info!("Writing test data to delta layer ...");
1854 6 : let mut writer = DeltaLayerWriter::new(
1855 6 : harness.conf,
1856 6 : timeline_id,
1857 6 : harness.tenant_shard_id,
1858 6 : entries_meta.key_range.start,
1859 6 : entries_meta.lsn_range.clone(),
1860 6 : &ctx,
1861 6 : )
1862 6 : .await?;
1863 6 :
1864 3342 : for entry in entries {
1865 3336 : let (_, res) = writer
1866 3336 : .put_value_bytes(entry.key, entry.lsn, entry.value.slice_len(), false, &ctx)
1867 645 : .await;
1868 3336 : res?;
1869 6 : }
1870 6 :
1871 15 : let (desc, path) = writer.finish(entries_meta.key_range.end, &ctx).await?;
1872 6 : let resident = Layer::finish_creating(harness.conf, &timeline, desc, &path)?;
1873 6 :
1874 6 : let inner = resident.get_as_delta(&ctx).await?;
1875 6 :
1876 6 : let file_size = inner.file.metadata().await?.len();
1877 6 : tracing::info!(
1878 6 : "Done writing test data to delta layer. Resulting file size is: {}",
1879 6 : file_size
1880 6 : );
1881 6 :
1882 606 : for i in 0..constants::READS_COUNT {
1883 600 : tracing::info!("Doing vectored read {}/{}", i + 1, constants::READS_COUNT);
1884 6 :
1885 600 : let block_reader = FileBlockReader::new(&inner.file, inner.file_id);
1886 600 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1887 600 : inner.index_start_blk,
1888 600 : inner.index_root_blk,
1889 600 : block_reader,
1890 600 : );
1891 600 :
1892 600 : let planner = VectoredReadPlanner::new(constants::MAX_VECTORED_READ_BYTES);
1893 600 : let mut reconstruct_state = ValuesReconstructState::new();
1894 600 : let keyspace = pick_random_keyspace(rng, &entries_meta.key_range);
1895 600 : let data_end_offset = inner.index_start_blk as u64 * PAGE_SZ as u64;
1896 6 :
1897 600 : let vectored_reads = DeltaLayerInner::plan_reads(
1898 600 : &keyspace,
1899 600 : entries_meta.lsn_range.clone(),
1900 600 : data_end_offset,
1901 600 : index_reader,
1902 600 : planner,
1903 600 : &mut reconstruct_state,
1904 600 : &ctx,
1905 600 : )
1906 12 : .await?;
1907 6 :
1908 600 : let vectored_blob_reader = VectoredBlobReader::new(&inner.file);
1909 600 : let buf_size = DeltaLayerInner::get_min_read_buffer_size(
1910 600 : &vectored_reads,
1911 600 : constants::MAX_VECTORED_READ_BYTES,
1912 600 : );
1913 600 : let mut buf = Some(BytesMut::with_capacity(buf_size));
1914 6 :
1915 59772 : for read in vectored_reads {
1916 59172 : let blobs_buf = vectored_blob_reader
1917 59172 : .read_blobs(&read, buf.take().expect("Should have a buffer"), &ctx)
1918 30048 : .await?;
1919 171912 : for meta in blobs_buf.blobs.iter() {
1920 171912 : let value = &blobs_buf.buf[meta.start..meta.end];
1921 171912 : assert_eq!(value, entries_meta.index[&(meta.meta.key, meta.meta.lsn)]);
1922 6 : }
1923 6 :
1924 59172 : buf = Some(blobs_buf.buf);
1925 6 : }
1926 6 : }
1927 6 :
1928 6 : Ok(())
1929 6 : }
1930 :
1931 : #[tokio::test]
1932 6 : async fn copy_delta_prefix_smoke() {
1933 6 : use crate::walrecord::NeonWalRecord;
1934 6 : use bytes::Bytes;
1935 6 :
1936 6 : let h = crate::tenant::harness::TenantHarness::create("truncate_delta_smoke")
1937 6 : .await
1938 6 : .unwrap();
1939 24 : let (tenant, ctx) = h.load().await;
1940 6 : let ctx = &ctx;
1941 6 : let timeline = tenant
1942 6 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, ctx)
1943 12 : .await
1944 6 : .unwrap();
1945 6 :
1946 6 : let initdb_layer = timeline
1947 6 : .layers
1948 6 : .read()
1949 6 : .await
1950 6 : .likely_resident_layers()
1951 6 : .next()
1952 6 : .cloned()
1953 6 : .unwrap();
1954 6 :
1955 6 : {
1956 6 : let mut writer = timeline.writer().await;
1957 6 :
1958 6 : let data = [
1959 6 : (0x20, 12, Value::Image(Bytes::from_static(b"foobar"))),
1960 6 : (
1961 6 : 0x30,
1962 6 : 12,
1963 6 : Value::WalRecord(NeonWalRecord::Postgres {
1964 6 : will_init: false,
1965 6 : rec: Bytes::from_static(b"1"),
1966 6 : }),
1967 6 : ),
1968 6 : (
1969 6 : 0x40,
1970 6 : 12,
1971 6 : Value::WalRecord(NeonWalRecord::Postgres {
1972 6 : will_init: true,
1973 6 : rec: Bytes::from_static(b"2"),
1974 6 : }),
1975 6 : ),
1976 6 : // build an oversized value so we cannot extend and existing read over
1977 6 : // this
1978 6 : (
1979 6 : 0x50,
1980 6 : 12,
1981 6 : Value::WalRecord(NeonWalRecord::Postgres {
1982 6 : will_init: true,
1983 6 : rec: {
1984 6 : let mut buf =
1985 6 : vec![0u8; tenant.conf.max_vectored_read_bytes.0.get() + 1024];
1986 6 : buf.iter_mut()
1987 6 : .enumerate()
1988 792576 : .for_each(|(i, slot)| *slot = (i % 256) as u8);
1989 6 : Bytes::from(buf)
1990 6 : },
1991 6 : }),
1992 6 : ),
1993 6 : // because the oversized read cannot be extended further, we are sure to exercise the
1994 6 : // builder created on the last round with this:
1995 6 : (
1996 6 : 0x60,
1997 6 : 12,
1998 6 : Value::WalRecord(NeonWalRecord::Postgres {
1999 6 : will_init: true,
2000 6 : rec: Bytes::from_static(b"3"),
2001 6 : }),
2002 6 : ),
2003 6 : (
2004 6 : 0x60,
2005 6 : 9,
2006 6 : Value::Image(Bytes::from_static(b"something for a different key")),
2007 6 : ),
2008 6 : ];
2009 6 :
2010 6 : let mut last_lsn = None;
2011 6 :
2012 42 : for (lsn, key, value) in data {
2013 36 : let key = Key::from_i128(key);
2014 36 : writer.put(key, Lsn(lsn), &value, ctx).await.unwrap();
2015 36 : last_lsn = Some(lsn);
2016 6 : }
2017 6 :
2018 6 : writer.finish_write(Lsn(last_lsn.unwrap()));
2019 6 : }
2020 6 : timeline.freeze_and_flush().await.unwrap();
2021 6 :
2022 6 : let new_layer = timeline
2023 6 : .layers
2024 6 : .read()
2025 6 : .await
2026 6 : .likely_resident_layers()
2027 8 : .find(|&x| x != &initdb_layer)
2028 6 : .cloned()
2029 6 : .unwrap();
2030 6 :
2031 6 : // create a copy for the timeline, so we don't overwrite the file
2032 6 : let branch = tenant
2033 6 : .branch_timeline_test(&timeline, TimelineId::generate(), None, ctx)
2034 6 : .await
2035 6 : .unwrap();
2036 6 :
2037 6 : assert_eq!(branch.get_ancestor_lsn(), Lsn(0x60));
2038 6 :
2039 6 : // truncating at 0x61 gives us a full copy, otherwise just go backwards until there's just
2040 6 : // a single key
2041 6 :
2042 36 : for truncate_at in [0x61, 0x51, 0x41, 0x31, 0x21] {
2043 30 : let truncate_at = Lsn(truncate_at);
2044 6 :
2045 30 : let mut writer = DeltaLayerWriter::new(
2046 30 : tenant.conf,
2047 30 : branch.timeline_id,
2048 30 : tenant.tenant_shard_id,
2049 30 : Key::MIN,
2050 30 : Lsn(0x11)..truncate_at,
2051 30 : ctx,
2052 30 : )
2053 15 : .await
2054 30 : .unwrap();
2055 6 :
2056 30 : let new_layer = new_layer.download_and_keep_resident().await.unwrap();
2057 30 :
2058 30 : new_layer
2059 30 : .copy_delta_prefix(&mut writer, truncate_at, ctx)
2060 45 : .await
2061 30 : .unwrap();
2062 6 :
2063 72 : let (desc, path) = writer.finish(Key::MAX, ctx).await.unwrap();
2064 30 : let copied_layer = Layer::finish_creating(tenant.conf, &branch, desc, &path).unwrap();
2065 30 :
2066 30 : copied_layer.get_as_delta(ctx).await.unwrap();
2067 30 :
2068 30 : assert_keys_and_values_eq(
2069 30 : new_layer.get_as_delta(ctx).await.unwrap(),
2070 30 : copied_layer.get_as_delta(ctx).await.unwrap(),
2071 30 : truncate_at,
2072 30 : ctx,
2073 6 : )
2074 155 : .await;
2075 6 : }
2076 6 : }
2077 :
2078 30 : async fn assert_keys_and_values_eq(
2079 30 : source: &DeltaLayerInner,
2080 30 : truncated: &DeltaLayerInner,
2081 30 : truncated_at: Lsn,
2082 30 : ctx: &RequestContext,
2083 30 : ) {
2084 : use futures::future::ready;
2085 : use futures::stream::TryStreamExt;
2086 :
2087 30 : let start_key = [0u8; DELTA_KEY_SIZE];
2088 30 :
2089 30 : let source_reader = FileBlockReader::new(&source.file, source.file_id);
2090 30 : let source_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2091 30 : source.index_start_blk,
2092 30 : source.index_root_blk,
2093 30 : &source_reader,
2094 30 : );
2095 30 : let source_stream = source.stream_index_forwards(source_tree, &start_key, ctx);
2096 180 : let source_stream = source_stream.filter(|res| match res {
2097 180 : Ok((_, lsn, _)) => ready(lsn < &truncated_at),
2098 0 : _ => ready(true),
2099 180 : });
2100 30 : let mut source_stream = std::pin::pin!(source_stream);
2101 30 :
2102 30 : let truncated_reader = FileBlockReader::new(&truncated.file, truncated.file_id);
2103 30 : let truncated_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2104 30 : truncated.index_start_blk,
2105 30 : truncated.index_root_blk,
2106 30 : &truncated_reader,
2107 30 : );
2108 30 : let truncated_stream = truncated.stream_index_forwards(truncated_tree, &start_key, ctx);
2109 30 : let mut truncated_stream = std::pin::pin!(truncated_stream);
2110 30 :
2111 30 : let mut scratch_left = Vec::new();
2112 30 : let mut scratch_right = Vec::new();
2113 :
2114 : loop {
2115 126 : let (src, truncated) = (source_stream.try_next(), truncated_stream.try_next());
2116 126 : let (src, truncated) = tokio::try_join!(src, truncated).unwrap();
2117 126 :
2118 126 : if src.is_none() {
2119 30 : assert!(truncated.is_none());
2120 30 : break;
2121 96 : }
2122 96 :
2123 96 : let (src, truncated) = (src.unwrap(), truncated.unwrap());
2124 96 :
2125 96 : // because we've filtered the source with Lsn, we should always have the same keys from both.
2126 96 : assert_eq!(src.0, truncated.0);
2127 96 : assert_eq!(src.1, truncated.1);
2128 :
2129 : // if this is needed for something else, just drop this assert.
2130 96 : assert!(
2131 96 : src.2.pos() >= truncated.2.pos(),
2132 0 : "value position should not go backwards {} vs. {}",
2133 0 : src.2.pos(),
2134 0 : truncated.2.pos()
2135 : );
2136 :
2137 96 : scratch_left.clear();
2138 96 : let src_cursor = source_reader.block_cursor();
2139 96 : let left = src_cursor.read_blob_into_buf(src.2.pos(), &mut scratch_left, ctx);
2140 96 : scratch_right.clear();
2141 96 : let trunc_cursor = truncated_reader.block_cursor();
2142 96 : let right = trunc_cursor.read_blob_into_buf(truncated.2.pos(), &mut scratch_right, ctx);
2143 96 :
2144 96 : tokio::try_join!(left, right).unwrap();
2145 96 :
2146 96 : assert_eq!(utils::Hex(&scratch_left), utils::Hex(&scratch_right));
2147 : }
2148 30 : }
2149 :
2150 54078 : pub(crate) fn sort_delta(
2151 54078 : (k1, l1, _): &(Key, Lsn, Value),
2152 54078 : (k2, l2, _): &(Key, Lsn, Value),
2153 54078 : ) -> std::cmp::Ordering {
2154 54078 : (k1, l1).cmp(&(k2, l2))
2155 54078 : }
2156 :
2157 282 : pub(crate) fn sort_delta_value(
2158 282 : (k1, l1, v1): &(Key, Lsn, Value),
2159 282 : (k2, l2, v2): &(Key, Lsn, Value),
2160 282 : ) -> std::cmp::Ordering {
2161 282 : let order_1 = if v1.is_image() { 0 } else { 1 };
2162 282 : let order_2 = if v2.is_image() { 0 } else { 1 };
2163 282 : (k1, l1, order_1).cmp(&(k2, l2, order_2))
2164 282 : }
2165 :
2166 60 : pub(crate) async fn produce_delta_layer(
2167 60 : tenant: &Tenant,
2168 60 : tline: &Arc<Timeline>,
2169 60 : mut deltas: Vec<(Key, Lsn, Value)>,
2170 60 : ctx: &RequestContext,
2171 60 : ) -> anyhow::Result<ResidentLayer> {
2172 60 : deltas.sort_by(sort_delta);
2173 60 : let (key_start, _, _) = deltas.first().unwrap();
2174 60 : let (key_max, _, _) = deltas.last().unwrap();
2175 24120 : let lsn_min = deltas.iter().map(|(_, lsn, _)| lsn).min().unwrap();
2176 24120 : let lsn_max = deltas.iter().map(|(_, lsn, _)| lsn).max().unwrap();
2177 60 : let lsn_end = Lsn(lsn_max.0 + 1);
2178 60 : let mut writer = DeltaLayerWriter::new(
2179 60 : tenant.conf,
2180 60 : tline.timeline_id,
2181 60 : tenant.tenant_shard_id,
2182 60 : *key_start,
2183 60 : (*lsn_min)..lsn_end,
2184 60 : ctx,
2185 60 : )
2186 30 : .await?;
2187 60 : let key_end = key_max.next();
2188 :
2189 24180 : for (key, lsn, value) in deltas {
2190 24120 : writer.put_value(key, lsn, value, ctx).await?;
2191 : }
2192 :
2193 174 : let (desc, path) = writer.finish(key_end, ctx).await?;
2194 60 : let delta_layer = Layer::finish_creating(tenant.conf, tline, desc, &path)?;
2195 :
2196 60 : Ok::<_, anyhow::Error>(delta_layer)
2197 60 : }
2198 :
2199 84 : async fn assert_delta_iter_equal(
2200 84 : delta_iter: &mut DeltaLayerIterator<'_>,
2201 84 : expect: &[(Key, Lsn, Value)],
2202 84 : ) {
2203 84 : let mut expect_iter = expect.iter();
2204 : loop {
2205 84084 : let o1 = delta_iter.next().await.unwrap();
2206 84084 : let o2 = expect_iter.next();
2207 84084 : assert_eq!(o1.is_some(), o2.is_some());
2208 84084 : if o1.is_none() && o2.is_none() {
2209 84 : break;
2210 84000 : }
2211 84000 : let (k1, l1, v1) = o1.unwrap();
2212 84000 : let (k2, l2, v2) = o2.unwrap();
2213 84000 : assert_eq!(&k1, k2);
2214 84000 : assert_eq!(l1, *l2);
2215 84000 : assert_eq!(&v1, v2);
2216 : }
2217 84 : }
2218 :
2219 : #[tokio::test]
2220 6 : async fn delta_layer_iterator() {
2221 6 : let harness = TenantHarness::create("delta_layer_iterator").await.unwrap();
2222 24 : let (tenant, ctx) = harness.load().await;
2223 6 :
2224 6 : let tline = tenant
2225 6 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
2226 12 : .await
2227 6 : .unwrap();
2228 6 :
2229 6000 : fn get_key(id: u32) -> Key {
2230 6000 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
2231 6000 : key.field6 = id;
2232 6000 : key
2233 6000 : }
2234 6 : const N: usize = 1000;
2235 6 : let test_deltas = (0..N)
2236 6000 : .map(|idx| {
2237 6000 : (
2238 6000 : get_key(idx as u32 / 10),
2239 6000 : Lsn(0x10 * ((idx as u64) % 10 + 1)),
2240 6000 : Value::Image(Bytes::from(format!("img{idx:05}"))),
2241 6000 : )
2242 6000 : })
2243 6 : .collect_vec();
2244 6 : let resident_layer = produce_delta_layer(&tenant, &tline, test_deltas.clone(), &ctx)
2245 24 : .await
2246 6 : .unwrap();
2247 6 : let delta_layer = resident_layer.get_as_delta(&ctx).await.unwrap();
2248 18 : for max_read_size in [1, 1024] {
2249 96 : for batch_size in [1, 2, 4, 8, 3, 7, 13] {
2250 84 : println!("running with batch_size={batch_size} max_read_size={max_read_size}");
2251 84 : // Test if the batch size is correctly determined
2252 84 : let mut iter = delta_layer.iter(&ctx);
2253 84 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2254 84 : let mut num_items = 0;
2255 336 : for _ in 0..3 {
2256 252 : iter.next_batch().await.unwrap();
2257 252 : num_items += iter.key_values_batch.len();
2258 252 : if max_read_size == 1 {
2259 6 : // every key should be a batch b/c the value is larger than max_read_size
2260 126 : assert_eq!(iter.key_values_batch.len(), 1);
2261 6 : } else {
2262 126 : assert!(iter.key_values_batch.len() <= batch_size);
2263 6 : }
2264 252 : if num_items >= N {
2265 6 : break;
2266 252 : }
2267 252 : iter.key_values_batch.clear();
2268 6 : }
2269 6 : // Test if the result is correct
2270 84 : let mut iter = delta_layer.iter(&ctx);
2271 84 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2272 28803 : assert_delta_iter_equal(&mut iter, &test_deltas).await;
2273 6 : }
2274 6 : }
2275 6 : }
2276 : }
|
