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::tenant::blob_io::BlobWriter;
34 : use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockLease, BlockReader, FileBlockReader};
35 : use crate::tenant::disk_btree::{
36 : DiskBtreeBuilder, DiskBtreeIterator, DiskBtreeReader, VisitDirection,
37 : };
38 : use crate::tenant::storage_layer::layer::S3_UPLOAD_LIMIT;
39 : use crate::tenant::timeline::GetVectoredError;
40 : use crate::tenant::vectored_blob_io::{
41 : BlobFlag, BufView, StreamingVectoredReadPlanner, VectoredBlobReader, VectoredRead,
42 : VectoredReadPlanner,
43 : };
44 : use crate::virtual_file::owned_buffers_io::io_buf_ext::{FullSlice, IoBufExt};
45 : use crate::virtual_file::IoBufferMut;
46 : use crate::virtual_file::{self, MaybeFatalIo, VirtualFile};
47 : use crate::TEMP_FILE_SUFFIX;
48 : use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
49 : use anyhow::{bail, ensure, Context, Result};
50 : use camino::{Utf8Path, Utf8PathBuf};
51 : use futures::StreamExt;
52 : use itertools::Itertools;
53 : use pageserver_api::config::MaxVectoredReadBytes;
54 : use pageserver_api::key::DBDIR_KEY;
55 : use pageserver_api::key::{Key, KEY_SIZE};
56 : use pageserver_api::keyspace::KeySpace;
57 : use pageserver_api::models::ImageCompressionAlgorithm;
58 : use pageserver_api::shard::TenantShardId;
59 : use pageserver_api::value::Value;
60 : use rand::{distributions::Alphanumeric, Rng};
61 : use serde::{Deserialize, Serialize};
62 : use std::collections::{HashMap, VecDeque};
63 : use std::fs::File;
64 : use std::io::SeekFrom;
65 : use std::ops::Range;
66 : use std::os::unix::fs::FileExt;
67 : use std::str::FromStr;
68 : use std::sync::Arc;
69 : use tokio::sync::OnceCell;
70 : use tokio_epoll_uring::IoBuf;
71 : use tracing::*;
72 :
73 : use utils::{
74 : bin_ser::BeSer,
75 : id::{TenantId, TimelineId},
76 : lsn::Lsn,
77 : };
78 :
79 : use super::{
80 : AsLayerDesc, LayerName, OnDiskValue, OnDiskValueIo, PersistentLayerDesc, ResidentLayer,
81 : ValuesReconstructState,
82 : };
83 :
84 : ///
85 : /// Header stored in the beginning of the file
86 : ///
87 : /// After this comes the 'values' part, starting on block 1. After that,
88 : /// the 'index' starts at the block indicated by 'index_start_blk'
89 : ///
90 0 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
91 : pub struct Summary {
92 : /// Magic value to identify this as a neon delta file. Always DELTA_FILE_MAGIC.
93 : pub magic: u16,
94 : pub format_version: u16,
95 :
96 : pub tenant_id: TenantId,
97 : pub timeline_id: TimelineId,
98 : pub key_range: Range<Key>,
99 : pub lsn_range: Range<Lsn>,
100 :
101 : /// Block number where the 'index' part of the file begins.
102 : pub index_start_blk: u32,
103 : /// Block within the 'index', where the B-tree root page is stored
104 : pub index_root_blk: u32,
105 : }
106 :
107 : impl From<&DeltaLayer> for Summary {
108 0 : fn from(layer: &DeltaLayer) -> Self {
109 0 : Self::expected(
110 0 : layer.desc.tenant_shard_id.tenant_id,
111 0 : layer.desc.timeline_id,
112 0 : layer.desc.key_range.clone(),
113 0 : layer.desc.lsn_range.clone(),
114 0 : )
115 0 : }
116 : }
117 :
118 : impl Summary {
119 2188 : pub(super) fn expected(
120 2188 : tenant_id: TenantId,
121 2188 : timeline_id: TimelineId,
122 2188 : keys: Range<Key>,
123 2188 : lsns: Range<Lsn>,
124 2188 : ) -> Self {
125 2188 : Self {
126 2188 : magic: DELTA_FILE_MAGIC,
127 2188 : format_version: STORAGE_FORMAT_VERSION,
128 2188 :
129 2188 : tenant_id,
130 2188 : timeline_id,
131 2188 : key_range: keys,
132 2188 : lsn_range: lsns,
133 2188 :
134 2188 : index_start_blk: 0,
135 2188 : index_root_blk: 0,
136 2188 : }
137 2188 : }
138 : }
139 :
140 : // Flag indicating that this version initialize the page
141 : const WILL_INIT: u64 = 1;
142 :
143 : /// Struct representing reference to BLOB in layers.
144 : ///
145 : /// Reference contains BLOB offset, and for WAL records it also contains
146 : /// `will_init` flag. The flag helps to determine the range of records
147 : /// that needs to be applied, without reading/deserializing records themselves.
148 0 : #[derive(Debug, Serialize, Deserialize, Copy, Clone)]
149 : pub struct BlobRef(pub u64);
150 :
151 : impl BlobRef {
152 5367868 : pub fn will_init(&self) -> bool {
153 5367868 : (self.0 & WILL_INIT) != 0
154 5367868 : }
155 :
156 9581012 : pub fn pos(&self) -> u64 {
157 9581012 : self.0 >> 1
158 9581012 : }
159 :
160 12942952 : pub fn new(pos: u64, will_init: bool) -> BlobRef {
161 12942952 : let mut blob_ref = pos << 1;
162 12942952 : if will_init {
163 12939832 : blob_ref |= WILL_INIT;
164 12939832 : }
165 12942952 : BlobRef(blob_ref)
166 12942952 : }
167 : }
168 :
169 : pub const DELTA_KEY_SIZE: usize = KEY_SIZE + 8;
170 : struct DeltaKey([u8; DELTA_KEY_SIZE]);
171 :
172 : /// This is the key of the B-tree index stored in the delta layer. It consists
173 : /// of the serialized representation of a Key and LSN.
174 : impl DeltaKey {
175 4128396 : fn from_slice(buf: &[u8]) -> Self {
176 4128396 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
177 4128396 : bytes.copy_from_slice(buf);
178 4128396 : DeltaKey(bytes)
179 4128396 : }
180 :
181 13378036 : fn from_key_lsn(key: &Key, lsn: Lsn) -> Self {
182 13378036 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
183 13378036 : key.write_to_byte_slice(&mut bytes[0..KEY_SIZE]);
184 13378036 : bytes[KEY_SIZE..].copy_from_slice(&u64::to_be_bytes(lsn.0));
185 13378036 : DeltaKey(bytes)
186 13378036 : }
187 :
188 4128396 : fn key(&self) -> Key {
189 4128396 : Key::from_slice(&self.0)
190 4128396 : }
191 :
192 4128396 : fn lsn(&self) -> Lsn {
193 4128396 : Lsn(u64::from_be_bytes(self.0[KEY_SIZE..].try_into().unwrap()))
194 4128396 : }
195 :
196 5452496 : fn extract_lsn_from_buf(buf: &[u8]) -> Lsn {
197 5452496 : let mut lsn_buf = [0u8; 8];
198 5452496 : lsn_buf.copy_from_slice(&buf[KEY_SIZE..]);
199 5452496 : Lsn(u64::from_be_bytes(lsn_buf))
200 5452496 : }
201 : }
202 :
203 : /// This is used only from `pagectl`. Within pageserver, all layers are
204 : /// [`crate::tenant::storage_layer::Layer`], which can hold a [`DeltaLayerInner`].
205 : pub struct DeltaLayer {
206 : path: Utf8PathBuf,
207 : pub desc: PersistentLayerDesc,
208 : inner: OnceCell<Arc<DeltaLayerInner>>,
209 : }
210 :
211 : impl std::fmt::Debug for DeltaLayer {
212 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
213 : use super::RangeDisplayDebug;
214 :
215 0 : f.debug_struct("DeltaLayer")
216 0 : .field("key_range", &RangeDisplayDebug(&self.desc.key_range))
217 0 : .field("lsn_range", &self.desc.lsn_range)
218 0 : .field("file_size", &self.desc.file_size)
219 0 : .field("inner", &self.inner)
220 0 : .finish()
221 0 : }
222 : }
223 :
224 : /// `DeltaLayerInner` is the in-memory data structure associated with an on-disk delta
225 : /// file.
226 : pub struct DeltaLayerInner {
227 : // values copied from summary
228 : index_start_blk: u32,
229 : index_root_blk: u32,
230 :
231 : file: Arc<VirtualFile>,
232 : file_id: FileId,
233 :
234 : layer_key_range: Range<Key>,
235 : layer_lsn_range: Range<Lsn>,
236 :
237 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
238 : }
239 :
240 : impl DeltaLayerInner {
241 0 : pub(crate) fn layer_dbg_info(&self) -> String {
242 0 : format!(
243 0 : "delta {}..{} {}..{}",
244 0 : self.key_range().start,
245 0 : self.key_range().end,
246 0 : self.lsn_range().start,
247 0 : self.lsn_range().end
248 0 : )
249 0 : }
250 : }
251 :
252 : impl std::fmt::Debug for DeltaLayerInner {
253 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
254 0 : f.debug_struct("DeltaLayerInner")
255 0 : .field("index_start_blk", &self.index_start_blk)
256 0 : .field("index_root_blk", &self.index_root_blk)
257 0 : .finish()
258 0 : }
259 : }
260 :
261 : /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
262 : impl std::fmt::Display for DeltaLayer {
263 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
264 0 : write!(f, "{}", self.layer_desc().short_id())
265 0 : }
266 : }
267 :
268 : impl AsLayerDesc for DeltaLayer {
269 0 : fn layer_desc(&self) -> &PersistentLayerDesc {
270 0 : &self.desc
271 0 : }
272 : }
273 :
274 : impl DeltaLayer {
275 0 : pub async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
276 0 : self.desc.dump();
277 0 :
278 0 : if !verbose {
279 0 : return Ok(());
280 0 : }
281 :
282 0 : let inner = self.load(ctx).await?;
283 :
284 0 : inner.dump(ctx).await
285 0 : }
286 :
287 2932 : fn temp_path_for(
288 2932 : conf: &PageServerConf,
289 2932 : tenant_shard_id: &TenantShardId,
290 2932 : timeline_id: &TimelineId,
291 2932 : key_start: Key,
292 2932 : lsn_range: &Range<Lsn>,
293 2932 : ) -> Utf8PathBuf {
294 2932 : let rand_string: String = rand::thread_rng()
295 2932 : .sample_iter(&Alphanumeric)
296 2932 : .take(8)
297 2932 : .map(char::from)
298 2932 : .collect();
299 2932 :
300 2932 : conf.timeline_path(tenant_shard_id, timeline_id)
301 2932 : .join(format!(
302 2932 : "{}-XXX__{:016X}-{:016X}.{}.{}",
303 2932 : key_start,
304 2932 : u64::from(lsn_range.start),
305 2932 : u64::from(lsn_range.end),
306 2932 : rand_string,
307 2932 : TEMP_FILE_SUFFIX,
308 2932 : ))
309 2932 : }
310 :
311 : ///
312 : /// Open the underlying file and read the metadata into memory, if it's
313 : /// not loaded already.
314 : ///
315 0 : async fn load(&self, ctx: &RequestContext) -> Result<&Arc<DeltaLayerInner>> {
316 0 : // Quick exit if already loaded
317 0 : self.inner
318 0 : .get_or_try_init(|| self.load_inner(ctx))
319 0 : .await
320 0 : .with_context(|| format!("Failed to load delta layer {}", self.path()))
321 0 : }
322 :
323 0 : async fn load_inner(&self, ctx: &RequestContext) -> anyhow::Result<Arc<DeltaLayerInner>> {
324 0 : let path = self.path();
325 :
326 0 : let loaded = DeltaLayerInner::load(&path, None, None, ctx).await?;
327 :
328 : // not production code
329 0 : let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
330 0 : let expected_layer_name = self.layer_desc().layer_name();
331 0 :
332 0 : if actual_layer_name != expected_layer_name {
333 0 : println!("warning: filename does not match what is expected from in-file summary");
334 0 : println!("actual: {:?}", actual_layer_name.to_string());
335 0 : println!("expected: {:?}", expected_layer_name.to_string());
336 0 : }
337 :
338 0 : Ok(Arc::new(loaded))
339 0 : }
340 :
341 : /// Create a DeltaLayer struct representing an existing file on disk.
342 : ///
343 : /// This variant is only used for debugging purposes, by the 'pagectl' binary.
344 0 : pub fn new_for_path(path: &Utf8Path, file: File) -> Result<Self> {
345 0 : let mut summary_buf = vec![0; PAGE_SZ];
346 0 : file.read_exact_at(&mut summary_buf, 0)?;
347 0 : let summary = Summary::des_prefix(&summary_buf)?;
348 :
349 0 : let metadata = file
350 0 : .metadata()
351 0 : .context("get file metadata to determine size")?;
352 :
353 : // This function is never used for constructing layers in a running pageserver,
354 : // so it does not need an accurate TenantShardId.
355 0 : let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
356 0 :
357 0 : Ok(DeltaLayer {
358 0 : path: path.to_path_buf(),
359 0 : desc: PersistentLayerDesc::new_delta(
360 0 : tenant_shard_id,
361 0 : summary.timeline_id,
362 0 : summary.key_range,
363 0 : summary.lsn_range,
364 0 : metadata.len(),
365 0 : ),
366 0 : inner: OnceCell::new(),
367 0 : })
368 0 : }
369 :
370 : /// Path to the layer file in pageserver workdir.
371 0 : fn path(&self) -> Utf8PathBuf {
372 0 : self.path.clone()
373 0 : }
374 : }
375 :
376 : /// A builder object for constructing a new delta layer.
377 : ///
378 : /// Usage:
379 : ///
380 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
381 : ///
382 : /// 2. Write the contents by calling `put_value` for every page
383 : /// version to store in the layer.
384 : ///
385 : /// 3. Call `finish`.
386 : ///
387 : struct DeltaLayerWriterInner {
388 : pub path: Utf8PathBuf,
389 : timeline_id: TimelineId,
390 : tenant_shard_id: TenantShardId,
391 :
392 : key_start: Key,
393 : lsn_range: Range<Lsn>,
394 :
395 : tree: DiskBtreeBuilder<BlockBuf, DELTA_KEY_SIZE>,
396 :
397 : blob_writer: BlobWriter<true>,
398 :
399 : // Number of key-lsns in the layer.
400 : num_keys: usize,
401 : }
402 :
403 : impl DeltaLayerWriterInner {
404 : ///
405 : /// Start building a new delta layer.
406 : ///
407 2932 : async fn new(
408 2932 : conf: &'static PageServerConf,
409 2932 : timeline_id: TimelineId,
410 2932 : tenant_shard_id: TenantShardId,
411 2932 : key_start: Key,
412 2932 : lsn_range: Range<Lsn>,
413 2932 : ctx: &RequestContext,
414 2932 : ) -> anyhow::Result<Self> {
415 2932 : // Create the file initially with a temporary filename. We don't know
416 2932 : // the end key yet, so we cannot form the final filename yet. We will
417 2932 : // rename it when we're done.
418 2932 : //
419 2932 : // Note: This overwrites any existing file. There shouldn't be any.
420 2932 : // FIXME: throw an error instead?
421 2932 : let path =
422 2932 : DeltaLayer::temp_path_for(conf, &tenant_shard_id, &timeline_id, key_start, &lsn_range);
423 :
424 2932 : let mut file = VirtualFile::create(&path, ctx).await?;
425 : // make room for the header block
426 2932 : file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
427 2932 : let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
428 2932 :
429 2932 : // Initialize the b-tree index builder
430 2932 : let block_buf = BlockBuf::new();
431 2932 : let tree_builder = DiskBtreeBuilder::new(block_buf);
432 2932 :
433 2932 : Ok(Self {
434 2932 : path,
435 2932 : timeline_id,
436 2932 : tenant_shard_id,
437 2932 : key_start,
438 2932 : lsn_range,
439 2932 : tree: tree_builder,
440 2932 : blob_writer,
441 2932 : num_keys: 0,
442 2932 : })
443 2932 : }
444 :
445 : ///
446 : /// Append a key-value pair to the file.
447 : ///
448 : /// The values must be appended in key, lsn order.
449 : ///
450 4169424 : async fn put_value(
451 4169424 : &mut self,
452 4169424 : key: Key,
453 4169424 : lsn: Lsn,
454 4169424 : val: Value,
455 4169424 : ctx: &RequestContext,
456 4169424 : ) -> anyhow::Result<()> {
457 4169424 : let (_, res) = self
458 4169424 : .put_value_bytes(
459 4169424 : key,
460 4169424 : lsn,
461 4169424 : Value::ser(&val)?.slice_len(),
462 4169424 : val.will_init(),
463 4169424 : ctx,
464 4169424 : )
465 4169424 : .await;
466 4169424 : res
467 4169424 : }
468 :
469 12942744 : async fn put_value_bytes<Buf>(
470 12942744 : &mut self,
471 12942744 : key: Key,
472 12942744 : lsn: Lsn,
473 12942744 : val: FullSlice<Buf>,
474 12942744 : will_init: bool,
475 12942744 : ctx: &RequestContext,
476 12942744 : ) -> (FullSlice<Buf>, anyhow::Result<()>)
477 12942744 : where
478 12942744 : Buf: IoBuf + Send,
479 12942744 : {
480 12942744 : assert!(
481 12942744 : self.lsn_range.start <= lsn,
482 0 : "lsn_start={}, lsn={}",
483 : self.lsn_range.start,
484 : lsn
485 : );
486 : // We don't want to use compression in delta layer creation
487 12942744 : let compression = ImageCompressionAlgorithm::Disabled;
488 12942744 : let (val, res) = self
489 12942744 : .blob_writer
490 12942744 : .write_blob_maybe_compressed(val, ctx, compression)
491 12942744 : .await;
492 12942744 : let off = match res {
493 12942744 : Ok((off, _)) => off,
494 0 : Err(e) => return (val, Err(anyhow::anyhow!(e))),
495 : };
496 :
497 12942744 : let blob_ref = BlobRef::new(off, will_init);
498 12942744 :
499 12942744 : let delta_key = DeltaKey::from_key_lsn(&key, lsn);
500 12942744 : let res = self.tree.append(&delta_key.0, blob_ref.0);
501 12942744 :
502 12942744 : self.num_keys += 1;
503 12942744 :
504 12942744 : (val, res.map_err(|e| anyhow::anyhow!(e)))
505 12942744 : }
506 :
507 4047944 : fn size(&self) -> u64 {
508 4047944 : self.blob_writer.size() + self.tree.borrow_writer().size()
509 4047944 : }
510 :
511 : ///
512 : /// Finish writing the delta layer.
513 : ///
514 2880 : async fn finish(
515 2880 : self,
516 2880 : key_end: Key,
517 2880 : ctx: &RequestContext,
518 2880 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
519 2880 : let temp_path = self.path.clone();
520 2880 : let result = self.finish0(key_end, ctx).await;
521 2880 : if let Err(ref e) = result {
522 0 : tracing::info!(%temp_path, "cleaning up temporary file after error during writing: {e}");
523 0 : if let Err(e) = std::fs::remove_file(&temp_path) {
524 0 : tracing::warn!(error=%e, %temp_path, "error cleaning up temporary layer file after error during writing");
525 0 : }
526 2880 : }
527 2880 : result
528 2880 : }
529 :
530 2880 : async fn finish0(
531 2880 : self,
532 2880 : key_end: Key,
533 2880 : ctx: &RequestContext,
534 2880 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
535 2880 : let index_start_blk = self.blob_writer.size().div_ceil(PAGE_SZ as u64) as u32;
536 :
537 2880 : let mut file = self.blob_writer.into_inner(ctx).await?;
538 :
539 : // Write out the index
540 2880 : let (index_root_blk, block_buf) = self.tree.finish()?;
541 2880 : file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
542 2880 : .await?;
543 30180 : for buf in block_buf.blocks {
544 27300 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
545 27300 : res?;
546 : }
547 2880 : assert!(self.lsn_range.start < self.lsn_range.end);
548 : // Fill in the summary on blk 0
549 2880 : let summary = Summary {
550 2880 : magic: DELTA_FILE_MAGIC,
551 2880 : format_version: STORAGE_FORMAT_VERSION,
552 2880 : tenant_id: self.tenant_shard_id.tenant_id,
553 2880 : timeline_id: self.timeline_id,
554 2880 : key_range: self.key_start..key_end,
555 2880 : lsn_range: self.lsn_range.clone(),
556 2880 : index_start_blk,
557 2880 : index_root_blk,
558 2880 : };
559 2880 :
560 2880 : let mut buf = Vec::with_capacity(PAGE_SZ);
561 2880 : // TODO: could use smallvec here but it's a pain with Slice<T>
562 2880 : Summary::ser_into(&summary, &mut buf)?;
563 2880 : file.seek(SeekFrom::Start(0)).await?;
564 2880 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
565 2880 : res?;
566 :
567 2880 : let metadata = file
568 2880 : .metadata()
569 2880 : .await
570 2880 : .context("get file metadata to determine size")?;
571 :
572 : // 5GB limit for objects without multipart upload (which we don't want to use)
573 : // Make it a little bit below to account for differing GB units
574 : // https://docs.aws.amazon.com/AmazonS3/latest/userguide/upload-objects.html
575 2880 : ensure!(
576 2880 : metadata.len() <= S3_UPLOAD_LIMIT,
577 0 : "Created delta layer file at {} of size {} above limit {S3_UPLOAD_LIMIT}!",
578 0 : file.path(),
579 0 : metadata.len()
580 : );
581 :
582 : // Note: Because we opened the file in write-only mode, we cannot
583 : // reuse the same VirtualFile for reading later. That's why we don't
584 : // set inner.file here. The first read will have to re-open it.
585 :
586 2880 : let desc = PersistentLayerDesc::new_delta(
587 2880 : self.tenant_shard_id,
588 2880 : self.timeline_id,
589 2880 : self.key_start..key_end,
590 2880 : self.lsn_range.clone(),
591 2880 : metadata.len(),
592 2880 : );
593 2880 :
594 2880 : // fsync the file
595 2880 : file.sync_all()
596 2880 : .await
597 2880 : .maybe_fatal_err("delta_layer sync_all")?;
598 :
599 2880 : trace!("created delta layer {}", self.path);
600 :
601 2880 : Ok((desc, self.path))
602 2880 : }
603 : }
604 :
605 : /// A builder object for constructing a new delta layer.
606 : ///
607 : /// Usage:
608 : ///
609 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
610 : ///
611 : /// 2. Write the contents by calling `put_value` for every page
612 : /// version to store in the layer.
613 : ///
614 : /// 3. Call `finish`.
615 : ///
616 : /// # Note
617 : ///
618 : /// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
619 : /// possible for the writer to drop before `finish` is actually called. So this
620 : /// could lead to odd temporary files in the directory, exhausting file system.
621 : /// This structure wraps `DeltaLayerWriterInner` and also contains `Drop`
622 : /// implementation that cleans up the temporary file in failure. It's not
623 : /// possible to do this directly in `DeltaLayerWriterInner` since `finish` moves
624 : /// out some fields, making it impossible to implement `Drop`.
625 : ///
626 : #[must_use]
627 : pub struct DeltaLayerWriter {
628 : inner: Option<DeltaLayerWriterInner>,
629 : }
630 :
631 : impl DeltaLayerWriter {
632 : ///
633 : /// Start building a new delta layer.
634 : ///
635 2932 : pub async fn new(
636 2932 : conf: &'static PageServerConf,
637 2932 : timeline_id: TimelineId,
638 2932 : tenant_shard_id: TenantShardId,
639 2932 : key_start: Key,
640 2932 : lsn_range: Range<Lsn>,
641 2932 : ctx: &RequestContext,
642 2932 : ) -> anyhow::Result<Self> {
643 2932 : Ok(Self {
644 2932 : inner: Some(
645 2932 : DeltaLayerWriterInner::new(
646 2932 : conf,
647 2932 : timeline_id,
648 2932 : tenant_shard_id,
649 2932 : key_start,
650 2932 : lsn_range,
651 2932 : ctx,
652 2932 : )
653 2932 : .await?,
654 : ),
655 : })
656 2932 : }
657 :
658 0 : pub fn is_empty(&self) -> bool {
659 0 : self.inner.as_ref().unwrap().num_keys == 0
660 0 : }
661 :
662 : ///
663 : /// Append a key-value pair to the file.
664 : ///
665 : /// The values must be appended in key, lsn order.
666 : ///
667 4169424 : pub async fn put_value(
668 4169424 : &mut self,
669 4169424 : key: Key,
670 4169424 : lsn: Lsn,
671 4169424 : val: Value,
672 4169424 : ctx: &RequestContext,
673 4169424 : ) -> anyhow::Result<()> {
674 4169424 : self.inner
675 4169424 : .as_mut()
676 4169424 : .unwrap()
677 4169424 : .put_value(key, lsn, val, ctx)
678 4169424 : .await
679 4169424 : }
680 :
681 8773320 : pub async fn put_value_bytes<Buf>(
682 8773320 : &mut self,
683 8773320 : key: Key,
684 8773320 : lsn: Lsn,
685 8773320 : val: FullSlice<Buf>,
686 8773320 : will_init: bool,
687 8773320 : ctx: &RequestContext,
688 8773320 : ) -> (FullSlice<Buf>, anyhow::Result<()>)
689 8773320 : where
690 8773320 : Buf: IoBuf + Send,
691 8773320 : {
692 8773320 : self.inner
693 8773320 : .as_mut()
694 8773320 : .unwrap()
695 8773320 : .put_value_bytes(key, lsn, val, will_init, ctx)
696 8773320 : .await
697 8773320 : }
698 :
699 4047944 : pub fn size(&self) -> u64 {
700 4047944 : self.inner.as_ref().unwrap().size()
701 4047944 : }
702 :
703 : ///
704 : /// Finish writing the delta layer.
705 : ///
706 2880 : pub(crate) async fn finish(
707 2880 : mut self,
708 2880 : key_end: Key,
709 2880 : ctx: &RequestContext,
710 2880 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
711 2880 : self.inner.take().unwrap().finish(key_end, ctx).await
712 2880 : }
713 :
714 24508 : pub(crate) fn num_keys(&self) -> usize {
715 24508 : self.inner.as_ref().unwrap().num_keys
716 24508 : }
717 :
718 30280 : pub(crate) fn estimated_size(&self) -> u64 {
719 30280 : let inner = self.inner.as_ref().unwrap();
720 30280 : inner.blob_writer.size() + inner.tree.borrow_writer().size() + PAGE_SZ as u64
721 30280 : }
722 : }
723 :
724 : impl Drop for DeltaLayerWriter {
725 2932 : fn drop(&mut self) {
726 2932 : if let Some(inner) = self.inner.take() {
727 52 : // We want to remove the virtual file here, so it's fine to not
728 52 : // having completely flushed unwritten data.
729 52 : let vfile = inner.blob_writer.into_inner_no_flush();
730 52 : vfile.remove();
731 2880 : }
732 2932 : }
733 : }
734 :
735 : #[derive(thiserror::Error, Debug)]
736 : pub enum RewriteSummaryError {
737 : #[error("magic mismatch")]
738 : MagicMismatch,
739 : #[error(transparent)]
740 : Other(#[from] anyhow::Error),
741 : }
742 :
743 : impl From<std::io::Error> for RewriteSummaryError {
744 0 : fn from(e: std::io::Error) -> Self {
745 0 : Self::Other(anyhow::anyhow!(e))
746 0 : }
747 : }
748 :
749 : impl DeltaLayer {
750 0 : pub async fn rewrite_summary<F>(
751 0 : path: &Utf8Path,
752 0 : rewrite: F,
753 0 : ctx: &RequestContext,
754 0 : ) -> Result<(), RewriteSummaryError>
755 0 : where
756 0 : F: Fn(Summary) -> Summary,
757 0 : {
758 0 : let mut file = VirtualFile::open_with_options(
759 0 : path,
760 0 : virtual_file::OpenOptions::new().read(true).write(true),
761 0 : ctx,
762 0 : )
763 0 : .await
764 0 : .with_context(|| format!("Failed to open file '{}'", path))?;
765 0 : let file_id = page_cache::next_file_id();
766 0 : let block_reader = FileBlockReader::new(&file, file_id);
767 0 : let summary_blk = block_reader.read_blk(0, ctx).await?;
768 0 : let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
769 0 : if actual_summary.magic != DELTA_FILE_MAGIC {
770 0 : return Err(RewriteSummaryError::MagicMismatch);
771 0 : }
772 0 :
773 0 : let new_summary = rewrite(actual_summary);
774 0 :
775 0 : let mut buf = Vec::with_capacity(PAGE_SZ);
776 0 : // TODO: could use smallvec here, but it's a pain with Slice<T>
777 0 : Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
778 0 : file.seek(SeekFrom::Start(0)).await?;
779 0 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
780 0 : res?;
781 0 : Ok(())
782 0 : }
783 : }
784 :
785 : impl DeltaLayerInner {
786 2072 : pub(crate) fn key_range(&self) -> &Range<Key> {
787 2072 : &self.layer_key_range
788 2072 : }
789 :
790 2072 : pub(crate) fn lsn_range(&self) -> &Range<Lsn> {
791 2072 : &self.layer_lsn_range
792 2072 : }
793 :
794 2188 : pub(super) async fn load(
795 2188 : path: &Utf8Path,
796 2188 : summary: Option<Summary>,
797 2188 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
798 2188 : ctx: &RequestContext,
799 2188 : ) -> anyhow::Result<Self> {
800 2188 : let file = Arc::new(
801 2188 : VirtualFile::open_v2(path, ctx)
802 2188 : .await
803 2188 : .context("open layer file")?,
804 : );
805 :
806 2188 : let file_id = page_cache::next_file_id();
807 2188 :
808 2188 : let block_reader = FileBlockReader::new(&file, file_id);
809 :
810 2188 : let summary_blk = block_reader
811 2188 : .read_blk(0, ctx)
812 2188 : .await
813 2188 : .context("read first block")?;
814 :
815 : // TODO: this should be an assertion instead; see ImageLayerInner::load
816 2188 : let actual_summary =
817 2188 : Summary::des_prefix(summary_blk.as_ref()).context("deserialize first block")?;
818 :
819 2188 : if let Some(mut expected_summary) = summary {
820 : // production code path
821 2188 : expected_summary.index_start_blk = actual_summary.index_start_blk;
822 2188 : expected_summary.index_root_blk = actual_summary.index_root_blk;
823 2188 : // mask out the timeline_id, but still require the layers to be from the same tenant
824 2188 : expected_summary.timeline_id = actual_summary.timeline_id;
825 2188 :
826 2188 : if actual_summary != expected_summary {
827 0 : bail!(
828 0 : "in-file summary does not match expected summary. actual = {:?} expected = {:?}",
829 0 : actual_summary,
830 0 : expected_summary
831 0 : );
832 2188 : }
833 0 : }
834 :
835 2188 : Ok(DeltaLayerInner {
836 2188 : file,
837 2188 : file_id,
838 2188 : index_start_blk: actual_summary.index_start_blk,
839 2188 : index_root_blk: actual_summary.index_root_blk,
840 2188 : max_vectored_read_bytes,
841 2188 : layer_key_range: actual_summary.key_range,
842 2188 : layer_lsn_range: actual_summary.lsn_range,
843 2188 : })
844 2188 : }
845 :
846 : // Look up the keys in the provided keyspace and update
847 : // the reconstruct state with whatever is found.
848 : //
849 : // Currently, the index is visited for each range, but this
850 : // can be further optimised to visit the index only once.
851 434416 : pub(super) async fn get_values_reconstruct_data(
852 434416 : &self,
853 434416 : this: ResidentLayer,
854 434416 : keyspace: KeySpace,
855 434416 : lsn_range: Range<Lsn>,
856 434416 : reconstruct_state: &mut ValuesReconstructState,
857 434416 : ctx: &RequestContext,
858 434416 : ) -> Result<(), GetVectoredError> {
859 434416 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
860 434416 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
861 434416 : self.index_start_blk,
862 434416 : self.index_root_blk,
863 434416 : block_reader,
864 434416 : );
865 434416 :
866 434416 : let planner = VectoredReadPlanner::new(
867 434416 : self.max_vectored_read_bytes
868 434416 : .expect("Layer is loaded with max vectored bytes config")
869 434416 : .0
870 434416 : .into(),
871 434416 : );
872 434416 :
873 434416 : let data_end_offset = self.index_start_offset();
874 :
875 434416 : let reads = Self::plan_reads(
876 434416 : &keyspace,
877 434416 : lsn_range.clone(),
878 434416 : data_end_offset,
879 434416 : index_reader,
880 434416 : planner,
881 434416 : ctx,
882 434416 : )
883 434416 : .await
884 434416 : .map_err(GetVectoredError::Other)?;
885 :
886 434416 : self.do_reads_and_update_state(this, reads, reconstruct_state, ctx)
887 434416 : .await;
888 :
889 434416 : Ok(())
890 434416 : }
891 :
892 434820 : async fn plan_reads<Reader>(
893 434820 : keyspace: &KeySpace,
894 434820 : lsn_range: Range<Lsn>,
895 434820 : data_end_offset: u64,
896 434820 : index_reader: DiskBtreeReader<Reader, DELTA_KEY_SIZE>,
897 434820 : mut planner: VectoredReadPlanner,
898 434820 : ctx: &RequestContext,
899 434820 : ) -> anyhow::Result<Vec<VectoredRead>>
900 434820 : where
901 434820 : Reader: BlockReader + Clone,
902 434820 : {
903 434820 : let ctx = RequestContextBuilder::extend(ctx)
904 434820 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
905 434820 : .build();
906 :
907 435224 : for range in keyspace.ranges.iter() {
908 435224 : let mut range_end_handled = false;
909 435224 :
910 435224 : let start_key = DeltaKey::from_key_lsn(&range.start, lsn_range.start);
911 435224 : let index_stream = index_reader.clone().into_stream(&start_key.0, &ctx);
912 435224 : let mut index_stream = std::pin::pin!(index_stream);
913 :
914 1271068 : while let Some(index_entry) = index_stream.next().await {
915 1254344 : let (raw_key, value) = index_entry?;
916 1254344 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
917 1254344 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
918 1254344 : let blob_ref = BlobRef(value);
919 1254344 :
920 1254344 : // Lsns are not monotonically increasing across keys, so we don't assert on them.
921 1254344 : assert!(key >= range.start);
922 :
923 1254344 : let outside_lsn_range = !lsn_range.contains(&lsn);
924 :
925 1254344 : let flag = {
926 1254344 : if outside_lsn_range {
927 84628 : BlobFlag::Ignore
928 1169716 : } else if blob_ref.will_init() {
929 1052464 : BlobFlag::ReplaceAll
930 : } else {
931 : // Usual path: add blob to the read
932 117252 : BlobFlag::None
933 : }
934 : };
935 :
936 1254344 : if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
937 418500 : planner.handle_range_end(blob_ref.pos());
938 418500 : range_end_handled = true;
939 418500 : break;
940 835844 : } else {
941 835844 : planner.handle(key, lsn, blob_ref.pos(), flag);
942 835844 : }
943 : }
944 :
945 435224 : if !range_end_handled {
946 16724 : tracing::debug!("Handling range end fallback at {}", data_end_offset);
947 16724 : planner.handle_range_end(data_end_offset);
948 418500 : }
949 : }
950 :
951 434820 : Ok(planner.finish())
952 434820 : }
953 :
954 434816 : fn get_min_read_buffer_size(
955 434816 : planned_reads: &[VectoredRead],
956 434816 : read_size_soft_max: usize,
957 434816 : ) -> usize {
958 434816 : let Some(largest_read) = planned_reads.iter().max_by_key(|read| read.size()) else {
959 173307 : return read_size_soft_max;
960 : };
961 :
962 261509 : let largest_read_size = largest_read.size();
963 261509 : if largest_read_size > read_size_soft_max {
964 : // If the read is oversized, it should only contain one key.
965 400 : let offenders = largest_read
966 400 : .blobs_at
967 400 : .as_slice()
968 400 : .iter()
969 400 : .filter_map(|(_, blob_meta)| {
970 400 : if blob_meta.key.is_rel_dir_key() || blob_meta.key == DBDIR_KEY {
971 : // The size of values for these keys is unbounded and can
972 : // grow very large in pathological cases.
973 0 : None
974 : } else {
975 400 : Some(format!("{}@{}", blob_meta.key, blob_meta.lsn))
976 : }
977 400 : })
978 400 : .join(", ");
979 400 :
980 400 : if !offenders.is_empty() {
981 400 : tracing::warn!(
982 0 : "Oversized vectored read ({} > {}) for keys {}",
983 : largest_read_size,
984 : read_size_soft_max,
985 : offenders
986 : );
987 0 : }
988 261109 : }
989 :
990 261509 : largest_read_size
991 434816 : }
992 :
993 434416 : async fn do_reads_and_update_state(
994 434416 : &self,
995 434416 : this: ResidentLayer,
996 434416 : reads: Vec<VectoredRead>,
997 434416 : reconstruct_state: &mut ValuesReconstructState,
998 434416 : ctx: &RequestContext,
999 434416 : ) {
1000 434416 : let max_vectored_read_bytes = self
1001 434416 : .max_vectored_read_bytes
1002 434416 : .expect("Layer is loaded with max vectored bytes config")
1003 434416 : .0
1004 434416 : .into();
1005 434416 : let buf_size = Self::get_min_read_buffer_size(&reads, max_vectored_read_bytes);
1006 :
1007 : // Note that reads are processed in reverse order (from highest key+lsn).
1008 : // This is the order that `ReconstructState` requires such that it can
1009 : // track when a key is done.
1010 434416 : for read in reads.into_iter().rev() {
1011 267135 : let mut ios: HashMap<(Key, Lsn), OnDiskValueIo> = Default::default();
1012 374489 : for (_, blob_meta) in read.blobs_at.as_slice().iter().rev() {
1013 374489 : let io = reconstruct_state.update_key(
1014 374489 : &blob_meta.key,
1015 374489 : blob_meta.lsn,
1016 374489 : blob_meta.will_init,
1017 374489 : );
1018 374489 : ios.insert((blob_meta.key, blob_meta.lsn), io);
1019 374489 : }
1020 :
1021 267135 : let read_extend_residency = this.clone();
1022 267135 : let read_from = self.file.clone();
1023 267135 : let read_ctx = ctx.attached_child();
1024 267135 : reconstruct_state
1025 267135 : .spawn_io(async move {
1026 267135 : let vectored_blob_reader = VectoredBlobReader::new(&read_from);
1027 267135 : let buf = IoBufferMut::with_capacity(buf_size);
1028 :
1029 267135 : let res = vectored_blob_reader.read_blobs(&read, buf, &read_ctx).await;
1030 267135 : match res {
1031 267135 : Ok(blobs_buf) => {
1032 267135 : let view = BufView::new_slice(&blobs_buf.buf);
1033 374489 : for meta in blobs_buf.blobs.iter().rev() {
1034 374489 : let io = ios.remove(&(meta.meta.key, meta.meta.lsn)).unwrap();
1035 :
1036 374489 : let blob_read = meta.read(&view).await;
1037 374489 : let blob_read = match blob_read {
1038 374489 : Ok(buf) => buf,
1039 0 : Err(e) => {
1040 0 : io.complete(Err(e));
1041 0 : continue;
1042 : }
1043 : };
1044 :
1045 374489 : io.complete(Ok(OnDiskValue::WalRecordOrImage(
1046 374489 : blob_read.into_bytes(),
1047 374489 : )));
1048 : }
1049 :
1050 267135 : assert!(ios.is_empty());
1051 : }
1052 0 : Err(err) => {
1053 0 : for (_, sender) in ios {
1054 0 : sender.complete(Err(std::io::Error::new(
1055 0 : err.kind(),
1056 0 : "vec read failed",
1057 0 : )));
1058 0 : }
1059 : }
1060 : }
1061 :
1062 : // keep layer resident until this IO is done; this spawned IO future generally outlives the
1063 : // call to `self` / the `Arc<DownloadedLayer>` / the `ResidentLayer` that guarantees residency
1064 267135 : drop(read_extend_residency);
1065 267135 : })
1066 267135 : .await;
1067 : }
1068 434416 : }
1069 :
1070 812 : pub(crate) async fn index_entries<'a>(
1071 812 : &'a self,
1072 812 : ctx: &RequestContext,
1073 812 : ) -> Result<Vec<DeltaEntry<'a>>> {
1074 812 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1075 812 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1076 812 : self.index_start_blk,
1077 812 : self.index_root_blk,
1078 812 : block_reader,
1079 812 : );
1080 812 :
1081 812 : let mut all_keys: Vec<DeltaEntry<'_>> = Vec::new();
1082 812 :
1083 812 : tree_reader
1084 812 : .visit(
1085 812 : &[0u8; DELTA_KEY_SIZE],
1086 812 : VisitDirection::Forwards,
1087 4128092 : |key, value| {
1088 4128092 : let delta_key = DeltaKey::from_slice(key);
1089 4128092 : let val_ref = ValueRef {
1090 4128092 : blob_ref: BlobRef(value),
1091 4128092 : layer: self,
1092 4128092 : };
1093 4128092 : let pos = BlobRef(value).pos();
1094 4128092 : if let Some(last) = all_keys.last_mut() {
1095 4127280 : // subtract offset of the current and last entries to get the size
1096 4127280 : // of the value associated with this (key, lsn) tuple
1097 4127280 : let first_pos = last.size;
1098 4127280 : last.size = pos - first_pos;
1099 4127280 : }
1100 4128092 : let entry = DeltaEntry {
1101 4128092 : key: delta_key.key(),
1102 4128092 : lsn: delta_key.lsn(),
1103 4128092 : size: pos,
1104 4128092 : val: val_ref,
1105 4128092 : };
1106 4128092 : all_keys.push(entry);
1107 4128092 : true
1108 4128092 : },
1109 812 : &RequestContextBuilder::extend(ctx)
1110 812 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
1111 812 : .build(),
1112 812 : )
1113 812 : .await?;
1114 812 : if let Some(last) = all_keys.last_mut() {
1115 812 : // Last key occupies all space till end of value storage,
1116 812 : // which corresponds to beginning of the index
1117 812 : last.size = self.index_start_offset() - last.size;
1118 812 : }
1119 812 : Ok(all_keys)
1120 812 : }
1121 :
1122 : /// Using the given writer, write out a version which has the earlier Lsns than `until`.
1123 : ///
1124 : /// Return the amount of key value records pushed to the writer.
1125 20 : pub(super) async fn copy_prefix(
1126 20 : &self,
1127 20 : writer: &mut DeltaLayerWriter,
1128 20 : until: Lsn,
1129 20 : ctx: &RequestContext,
1130 20 : ) -> anyhow::Result<usize> {
1131 : use crate::tenant::vectored_blob_io::{
1132 : BlobMeta, ChunkedVectoredReadBuilder, VectoredReadExtended,
1133 : };
1134 : use futures::stream::TryStreamExt;
1135 :
1136 : #[derive(Debug)]
1137 : enum Item {
1138 : Actual(Key, Lsn, BlobRef),
1139 : Sentinel,
1140 : }
1141 :
1142 : impl From<Item> for Option<(Key, Lsn, BlobRef)> {
1143 140 : fn from(value: Item) -> Self {
1144 140 : match value {
1145 120 : Item::Actual(key, lsn, blob) => Some((key, lsn, blob)),
1146 20 : Item::Sentinel => None,
1147 : }
1148 140 : }
1149 : }
1150 :
1151 : impl Item {
1152 140 : fn offset(&self) -> Option<BlobRef> {
1153 140 : match self {
1154 120 : Item::Actual(_, _, blob) => Some(*blob),
1155 20 : Item::Sentinel => None,
1156 : }
1157 140 : }
1158 :
1159 140 : fn is_last(&self) -> bool {
1160 140 : matches!(self, Item::Sentinel)
1161 140 : }
1162 : }
1163 :
1164 20 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1165 20 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1166 20 : self.index_start_blk,
1167 20 : self.index_root_blk,
1168 20 : block_reader,
1169 20 : );
1170 20 :
1171 20 : let stream = self.stream_index_forwards(tree_reader, &[0u8; DELTA_KEY_SIZE], ctx);
1172 120 : let stream = stream.map_ok(|(key, lsn, pos)| Item::Actual(key, lsn, pos));
1173 20 : // put in a sentinel value for getting the end offset for last item, and not having to
1174 20 : // repeat the whole read part
1175 20 : let stream = stream.chain(futures::stream::once(futures::future::ready(Ok(
1176 20 : Item::Sentinel,
1177 20 : ))));
1178 20 : let mut stream = std::pin::pin!(stream);
1179 20 :
1180 20 : let mut prev: Option<(Key, Lsn, BlobRef)> = None;
1181 20 :
1182 20 : let mut read_builder: Option<ChunkedVectoredReadBuilder> = None;
1183 20 :
1184 20 : let max_read_size = self
1185 20 : .max_vectored_read_bytes
1186 20 : .map(|x| x.0.get())
1187 20 : .unwrap_or(8192);
1188 20 :
1189 20 : let mut buffer = Some(IoBufferMut::with_capacity(max_read_size));
1190 20 :
1191 20 : // FIXME: buffering of DeltaLayerWriter
1192 20 : let mut per_blob_copy = Vec::new();
1193 20 :
1194 20 : let mut records = 0;
1195 :
1196 160 : while let Some(item) = stream.try_next().await? {
1197 140 : tracing::debug!(?item, "popped");
1198 140 : let offset = item
1199 140 : .offset()
1200 140 : .unwrap_or(BlobRef::new(self.index_start_offset(), false));
1201 :
1202 140 : let actionable = if let Some((key, lsn, start_offset)) = prev.take() {
1203 120 : let end_offset = offset;
1204 120 :
1205 120 : Some((
1206 120 : BlobMeta {
1207 120 : key,
1208 120 : lsn,
1209 120 : will_init: false,
1210 120 : },
1211 120 : start_offset..end_offset,
1212 120 : ))
1213 : } else {
1214 20 : None
1215 : };
1216 :
1217 140 : let is_last = item.is_last();
1218 140 :
1219 140 : prev = Option::from(item);
1220 140 :
1221 140 : let actionable = actionable.filter(|x| x.0.lsn < until);
1222 :
1223 140 : let builder = if let Some((meta, offsets)) = actionable {
1224 : // extend or create a new builder
1225 64 : if read_builder
1226 64 : .as_mut()
1227 64 : .map(|x| x.extend(offsets.start.pos(), offsets.end.pos(), meta))
1228 64 : .unwrap_or(VectoredReadExtended::No)
1229 64 : == VectoredReadExtended::Yes
1230 : {
1231 32 : None
1232 : } else {
1233 32 : read_builder.replace(ChunkedVectoredReadBuilder::new(
1234 32 : offsets.start.pos(),
1235 32 : offsets.end.pos(),
1236 32 : meta,
1237 32 : max_read_size,
1238 32 : ))
1239 : }
1240 : } else {
1241 : // nothing to do, except perhaps flush any existing for the last element
1242 76 : None
1243 : };
1244 :
1245 : // flush the possible older builder and also the new one if the item was the last one
1246 140 : let builders = builder.into_iter();
1247 140 : let builders = if is_last {
1248 20 : builders.chain(read_builder.take())
1249 : } else {
1250 120 : builders.chain(None)
1251 : };
1252 :
1253 172 : for builder in builders {
1254 32 : let read = builder.build();
1255 32 :
1256 32 : let reader = VectoredBlobReader::new(&self.file);
1257 32 :
1258 32 : let mut buf = buffer.take().unwrap();
1259 32 :
1260 32 : buf.clear();
1261 32 : buf.reserve(read.size());
1262 32 : let res = reader.read_blobs(&read, buf, ctx).await?;
1263 :
1264 32 : let view = BufView::new_slice(&res.buf);
1265 :
1266 96 : for blob in res.blobs {
1267 64 : let key = blob.meta.key;
1268 64 : let lsn = blob.meta.lsn;
1269 :
1270 64 : let data = blob.read(&view).await?;
1271 :
1272 : #[cfg(debug_assertions)]
1273 64 : Value::des(&data)
1274 64 : .with_context(|| {
1275 0 : format!(
1276 0 : "blob failed to deserialize for {}: {:?}",
1277 0 : blob,
1278 0 : utils::Hex(&data)
1279 0 : )
1280 64 : })
1281 64 : .unwrap();
1282 64 :
1283 64 : // is it an image or will_init walrecord?
1284 64 : // FIXME: this could be handled by threading the BlobRef to the
1285 64 : // VectoredReadBuilder
1286 64 : let will_init = pageserver_api::value::ValueBytes::will_init(&data)
1287 64 : .inspect_err(|_e| {
1288 0 : #[cfg(feature = "testing")]
1289 0 : tracing::error!(data=?utils::Hex(&data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
1290 64 : })
1291 64 : .unwrap_or(false);
1292 64 :
1293 64 : per_blob_copy.clear();
1294 64 : per_blob_copy.extend_from_slice(&data);
1295 :
1296 64 : let (tmp, res) = writer
1297 64 : .put_value_bytes(
1298 64 : key,
1299 64 : lsn,
1300 64 : std::mem::take(&mut per_blob_copy).slice_len(),
1301 64 : will_init,
1302 64 : ctx,
1303 64 : )
1304 64 : .await;
1305 64 : per_blob_copy = tmp.into_raw_slice().into_inner();
1306 64 :
1307 64 : res?;
1308 :
1309 64 : records += 1;
1310 : }
1311 :
1312 32 : buffer = Some(res.buf);
1313 : }
1314 : }
1315 :
1316 20 : assert!(
1317 20 : read_builder.is_none(),
1318 0 : "with the sentinel above loop should had handled all"
1319 : );
1320 :
1321 20 : Ok(records)
1322 20 : }
1323 :
1324 8 : pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
1325 8 : println!(
1326 8 : "index_start_blk: {}, root {}",
1327 8 : self.index_start_blk, self.index_root_blk
1328 8 : );
1329 8 :
1330 8 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1331 8 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1332 8 : self.index_start_blk,
1333 8 : self.index_root_blk,
1334 8 : block_reader,
1335 8 : );
1336 8 :
1337 8 : tree_reader.dump().await?;
1338 :
1339 8 : let keys = self.index_entries(ctx).await?;
1340 :
1341 16 : async fn dump_blob(val: &ValueRef<'_>, ctx: &RequestContext) -> anyhow::Result<String> {
1342 16 : let buf = val.load_raw(ctx).await?;
1343 16 : let val = Value::des(&buf)?;
1344 16 : let desc = match val {
1345 16 : Value::Image(img) => {
1346 16 : format!(" img {} bytes", img.len())
1347 : }
1348 0 : Value::WalRecord(rec) => {
1349 0 : let wal_desc = pageserver_api::record::describe_wal_record(&rec)?;
1350 0 : format!(
1351 0 : " rec {} bytes will_init: {} {}",
1352 0 : buf.len(),
1353 0 : rec.will_init(),
1354 0 : wal_desc
1355 0 : )
1356 : }
1357 : };
1358 16 : Ok(desc)
1359 16 : }
1360 :
1361 24 : for entry in keys {
1362 16 : let DeltaEntry { key, lsn, val, .. } = entry;
1363 16 : let desc = match dump_blob(&val, ctx).await {
1364 16 : Ok(desc) => desc,
1365 0 : Err(err) => {
1366 0 : format!("ERROR: {err}")
1367 : }
1368 : };
1369 16 : println!(" key {key} at {lsn}: {desc}");
1370 :
1371 : // Print more details about CHECKPOINT records. Would be nice to print details
1372 : // of many other record types too, but these are particularly interesting, as
1373 : // have a lot of special processing for them in walingest.rs.
1374 8 : use pageserver_api::key::CHECKPOINT_KEY;
1375 8 : use postgres_ffi::CheckPoint;
1376 16 : if key == CHECKPOINT_KEY {
1377 0 : let val = val.load(ctx).await?;
1378 0 : match val {
1379 0 : Value::Image(img) => {
1380 0 : let checkpoint = CheckPoint::decode(&img)?;
1381 0 : println!(" CHECKPOINT: {:?}", checkpoint);
1382 : }
1383 0 : Value::WalRecord(_rec) => {
1384 0 : println!(" unexpected walrecord value for checkpoint key");
1385 0 : }
1386 : }
1387 16 : }
1388 : }
1389 :
1390 8 : Ok(())
1391 8 : }
1392 :
1393 60 : fn stream_index_forwards<'a, R>(
1394 60 : &'a self,
1395 60 : reader: DiskBtreeReader<R, DELTA_KEY_SIZE>,
1396 60 : start: &'a [u8; DELTA_KEY_SIZE],
1397 60 : ctx: &'a RequestContext,
1398 60 : ) -> impl futures::stream::Stream<
1399 60 : Item = Result<(Key, Lsn, BlobRef), crate::tenant::disk_btree::DiskBtreeError>,
1400 60 : > + 'a
1401 60 : where
1402 60 : R: BlockReader + 'a,
1403 60 : {
1404 : use futures::stream::TryStreamExt;
1405 60 : let stream = reader.into_stream(start, ctx);
1406 304 : stream.map_ok(|(key, value)| {
1407 304 : let key = DeltaKey::from_slice(&key);
1408 304 : let (key, lsn) = (key.key(), key.lsn());
1409 304 : let offset = BlobRef(value);
1410 304 :
1411 304 : (key, lsn, offset)
1412 304 : })
1413 60 : }
1414 :
1415 : /// The file offset to the first block of index.
1416 : ///
1417 : /// The file structure is summary, values, and index. We often need this for the size of last blob.
1418 436460 : fn index_start_offset(&self) -> u64 {
1419 436460 : let offset = self.index_start_blk as u64 * PAGE_SZ as u64;
1420 436460 : let bref = BlobRef(offset);
1421 436460 : tracing::debug!(
1422 : index_start_blk = self.index_start_blk,
1423 : offset,
1424 0 : pos = bref.pos(),
1425 0 : "index_start_offset"
1426 : );
1427 436460 : offset
1428 436460 : }
1429 :
1430 1148 : pub fn iter<'a>(&'a self, ctx: &'a RequestContext) -> DeltaLayerIterator<'a> {
1431 1148 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1432 1148 : let tree_reader =
1433 1148 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
1434 1148 : DeltaLayerIterator {
1435 1148 : delta_layer: self,
1436 1148 : ctx,
1437 1148 : index_iter: tree_reader.iter(&[0; DELTA_KEY_SIZE], ctx),
1438 1148 : key_values_batch: std::collections::VecDeque::new(),
1439 1148 : is_end: false,
1440 1148 : planner: StreamingVectoredReadPlanner::new(
1441 1148 : 1024 * 8192, // The default value. Unit tests might use a different value. 1024 * 8K = 8MB buffer.
1442 1148 : 1024, // The default value. Unit tests might use a different value
1443 1148 : ),
1444 1148 : }
1445 1148 : }
1446 :
1447 : /// NB: not super efficient, but not terrible either. Should prob be an iterator.
1448 : //
1449 : // We're reusing the index traversal logical in plan_reads; would be nice to
1450 : // factor that out.
1451 0 : pub(crate) async fn load_keys(&self, ctx: &RequestContext) -> anyhow::Result<Vec<Key>> {
1452 0 : self.index_entries(ctx)
1453 0 : .await
1454 0 : .map(|entries| entries.into_iter().map(|entry| entry.key).collect())
1455 0 : }
1456 : }
1457 :
1458 : /// A set of data associated with a delta layer key and its value
1459 : pub struct DeltaEntry<'a> {
1460 : pub key: Key,
1461 : pub lsn: Lsn,
1462 : /// Size of the stored value
1463 : pub size: u64,
1464 : /// Reference to the on-disk value
1465 : pub val: ValueRef<'a>,
1466 : }
1467 :
1468 : /// Reference to an on-disk value
1469 : pub struct ValueRef<'a> {
1470 : blob_ref: BlobRef,
1471 : layer: &'a DeltaLayerInner,
1472 : }
1473 :
1474 : impl ValueRef<'_> {
1475 : /// Loads the value from disk
1476 0 : pub async fn load(&self, ctx: &RequestContext) -> Result<Value> {
1477 0 : let buf = self.load_raw(ctx).await?;
1478 0 : let val = Value::des(&buf)?;
1479 0 : Ok(val)
1480 0 : }
1481 :
1482 16 : async fn load_raw(&self, ctx: &RequestContext) -> Result<Vec<u8>> {
1483 16 : let reader = BlockCursor::new(crate::tenant::block_io::BlockReaderRef::Adapter(Adapter(
1484 16 : self.layer,
1485 16 : )));
1486 16 : let buf = reader.read_blob(self.blob_ref.pos(), ctx).await?;
1487 16 : Ok(buf)
1488 16 : }
1489 : }
1490 :
1491 : pub(crate) struct Adapter<T>(T);
1492 :
1493 : impl<T: AsRef<DeltaLayerInner>> Adapter<T> {
1494 16 : pub(crate) async fn read_blk(
1495 16 : &self,
1496 16 : blknum: u32,
1497 16 : ctx: &RequestContext,
1498 16 : ) -> Result<BlockLease, std::io::Error> {
1499 16 : let block_reader = FileBlockReader::new(&self.0.as_ref().file, self.0.as_ref().file_id);
1500 16 : block_reader.read_blk(blknum, ctx).await
1501 16 : }
1502 : }
1503 :
1504 : impl AsRef<DeltaLayerInner> for DeltaLayerInner {
1505 32 : fn as_ref(&self) -> &DeltaLayerInner {
1506 32 : self
1507 32 : }
1508 : }
1509 :
1510 : impl<'a> pageserver_compaction::interface::CompactionDeltaEntry<'a, Key> for DeltaEntry<'a> {
1511 0 : fn key(&self) -> Key {
1512 0 : self.key
1513 0 : }
1514 0 : fn lsn(&self) -> Lsn {
1515 0 : self.lsn
1516 0 : }
1517 0 : fn size(&self) -> u64 {
1518 0 : self.size
1519 0 : }
1520 : }
1521 :
1522 : pub struct DeltaLayerIterator<'a> {
1523 : delta_layer: &'a DeltaLayerInner,
1524 : ctx: &'a RequestContext,
1525 : planner: StreamingVectoredReadPlanner,
1526 : index_iter: DiskBtreeIterator<'a>,
1527 : key_values_batch: VecDeque<(Key, Lsn, Value)>,
1528 : is_end: bool,
1529 : }
1530 :
1531 : impl DeltaLayerIterator<'_> {
1532 0 : pub(crate) fn layer_dbg_info(&self) -> String {
1533 0 : self.delta_layer.layer_dbg_info()
1534 0 : }
1535 :
1536 : /// Retrieve a batch of key-value pairs into the iterator buffer.
1537 42808 : async fn next_batch(&mut self) -> anyhow::Result<()> {
1538 42808 : assert!(self.key_values_batch.is_empty());
1539 42808 : assert!(!self.is_end);
1540 :
1541 42808 : let plan = loop {
1542 4199244 : if let Some(res) = self.index_iter.next().await {
1543 4198152 : let (raw_key, value) = res?;
1544 4198152 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
1545 4198152 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
1546 4198152 : let blob_ref = BlobRef(value);
1547 4198152 : let offset = blob_ref.pos();
1548 41716 : if let Some(batch_plan) =
1549 4198152 : self.planner.handle(key, lsn, offset, blob_ref.will_init())
1550 : {
1551 41716 : break batch_plan;
1552 4156436 : }
1553 : } else {
1554 1092 : self.is_end = true;
1555 1092 : let data_end_offset = self.delta_layer.index_start_offset();
1556 1092 : if let Some(item) = self.planner.handle_range_end(data_end_offset) {
1557 1092 : break item;
1558 : } else {
1559 0 : return Ok(()); // TODO: test empty iterator
1560 : }
1561 : }
1562 : };
1563 42808 : let vectored_blob_reader = VectoredBlobReader::new(&self.delta_layer.file);
1564 42808 : let mut next_batch = std::collections::VecDeque::new();
1565 42808 : let buf_size = plan.size();
1566 42808 : let buf = IoBufferMut::with_capacity(buf_size);
1567 42808 : let blobs_buf = vectored_blob_reader
1568 42808 : .read_blobs(&plan, buf, self.ctx)
1569 42808 : .await?;
1570 42808 : let view = BufView::new_slice(&blobs_buf.buf);
1571 4198096 : for meta in blobs_buf.blobs.iter() {
1572 4198096 : let blob_read = meta.read(&view).await?;
1573 4198096 : let value = Value::des(&blob_read)?;
1574 :
1575 4198096 : next_batch.push_back((meta.meta.key, meta.meta.lsn, value));
1576 : }
1577 42808 : self.key_values_batch = next_batch;
1578 42808 : Ok(())
1579 42808 : }
1580 :
1581 4199668 : pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
1582 4199668 : if self.key_values_batch.is_empty() {
1583 44760 : if self.is_end {
1584 2120 : return Ok(None);
1585 42640 : }
1586 42640 : self.next_batch().await?;
1587 4154908 : }
1588 4197548 : Ok(Some(
1589 4197548 : self.key_values_batch
1590 4197548 : .pop_front()
1591 4197548 : .expect("should not be empty"),
1592 4197548 : ))
1593 4199668 : }
1594 : }
1595 :
1596 : #[cfg(test)]
1597 : pub(crate) mod test {
1598 : use std::collections::BTreeMap;
1599 :
1600 : use itertools::MinMaxResult;
1601 : use rand::prelude::{SeedableRng, SliceRandom, StdRng};
1602 : use rand::RngCore;
1603 :
1604 : use super::*;
1605 : use crate::tenant::harness::TIMELINE_ID;
1606 : use crate::tenant::storage_layer::{Layer, ResidentLayer};
1607 : use crate::tenant::vectored_blob_io::StreamingVectoredReadPlanner;
1608 : use crate::tenant::{Tenant, Timeline};
1609 : use crate::{
1610 : context::DownloadBehavior,
1611 : task_mgr::TaskKind,
1612 : tenant::{disk_btree::tests::TestDisk, harness::TenantHarness},
1613 : DEFAULT_PG_VERSION,
1614 : };
1615 : use bytes::Bytes;
1616 : use pageserver_api::value::Value;
1617 :
1618 : /// Construct an index for a fictional delta layer and and then
1619 : /// traverse in order to plan vectored reads for a query. Finally,
1620 : /// verify that the traversal fed the right index key and value
1621 : /// pairs into the planner.
1622 : #[tokio::test]
1623 4 : async fn test_delta_layer_index_traversal() {
1624 4 : let base_key = Key {
1625 4 : field1: 0,
1626 4 : field2: 1663,
1627 4 : field3: 12972,
1628 4 : field4: 16396,
1629 4 : field5: 0,
1630 4 : field6: 246080,
1631 4 : };
1632 4 :
1633 4 : // Populate the index with some entries
1634 4 : let entries: BTreeMap<Key, Vec<Lsn>> = BTreeMap::from([
1635 4 : (base_key, vec![Lsn(1), Lsn(5), Lsn(25), Lsn(26), Lsn(28)]),
1636 4 : (base_key.add(1), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1637 4 : (base_key.add(2), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1638 4 : (base_key.add(5), vec![Lsn(10), Lsn(15), Lsn(16), Lsn(20)]),
1639 4 : ]);
1640 4 :
1641 4 : let mut disk = TestDisk::default();
1642 4 : let mut writer = DiskBtreeBuilder::<_, DELTA_KEY_SIZE>::new(&mut disk);
1643 4 :
1644 4 : let mut disk_offset = 0;
1645 20 : for (key, lsns) in &entries {
1646 84 : for lsn in lsns {
1647 68 : let index_key = DeltaKey::from_key_lsn(key, *lsn);
1648 68 : let blob_ref = BlobRef::new(disk_offset, false);
1649 68 : writer
1650 68 : .append(&index_key.0, blob_ref.0)
1651 68 : .expect("In memory disk append should never fail");
1652 68 :
1653 68 : disk_offset += 1;
1654 68 : }
1655 4 : }
1656 4 :
1657 4 : // Prepare all the arguments for the call into `plan_reads` below
1658 4 : let (root_offset, _writer) = writer
1659 4 : .finish()
1660 4 : .expect("In memory disk finish should never fail");
1661 4 : let reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(0, root_offset, disk);
1662 4 : let planner = VectoredReadPlanner::new(100);
1663 4 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1664 4 :
1665 4 : let keyspace = KeySpace {
1666 4 : ranges: vec![
1667 4 : base_key..base_key.add(3),
1668 4 : base_key.add(3)..base_key.add(100),
1669 4 : ],
1670 4 : };
1671 4 : let lsn_range = Lsn(2)..Lsn(40);
1672 4 :
1673 4 : // Plan and validate
1674 4 : let vectored_reads = DeltaLayerInner::plan_reads(
1675 4 : &keyspace,
1676 4 : lsn_range.clone(),
1677 4 : disk_offset,
1678 4 : reader,
1679 4 : planner,
1680 4 : &ctx,
1681 4 : )
1682 4 : .await
1683 4 : .expect("Read planning should not fail");
1684 4 :
1685 4 : validate(keyspace, lsn_range, vectored_reads, entries);
1686 4 : }
1687 :
1688 4 : fn validate(
1689 4 : keyspace: KeySpace,
1690 4 : lsn_range: Range<Lsn>,
1691 4 : vectored_reads: Vec<VectoredRead>,
1692 4 : index_entries: BTreeMap<Key, Vec<Lsn>>,
1693 4 : ) {
1694 : #[derive(Debug, PartialEq, Eq)]
1695 : struct BlobSpec {
1696 : key: Key,
1697 : lsn: Lsn,
1698 : at: u64,
1699 : }
1700 :
1701 4 : let mut planned_blobs = Vec::new();
1702 60 : for read in vectored_reads {
1703 56 : for (at, meta) in read.blobs_at.as_slice() {
1704 56 : planned_blobs.push(BlobSpec {
1705 56 : key: meta.key,
1706 56 : lsn: meta.lsn,
1707 56 : at: *at,
1708 56 : });
1709 56 : }
1710 : }
1711 :
1712 4 : let mut expected_blobs = Vec::new();
1713 4 : let mut disk_offset = 0;
1714 20 : for (key, lsns) in index_entries {
1715 84 : for lsn in lsns {
1716 84 : let key_included = keyspace.ranges.iter().any(|range| range.contains(&key));
1717 68 : let lsn_included = lsn_range.contains(&lsn);
1718 68 :
1719 68 : if key_included && lsn_included {
1720 56 : expected_blobs.push(BlobSpec {
1721 56 : key,
1722 56 : lsn,
1723 56 : at: disk_offset,
1724 56 : });
1725 56 : }
1726 :
1727 68 : disk_offset += 1;
1728 : }
1729 : }
1730 :
1731 4 : assert_eq!(planned_blobs, expected_blobs);
1732 4 : }
1733 :
1734 : mod constants {
1735 : use utils::lsn::Lsn;
1736 :
1737 : /// Offset used by all lsns in this test
1738 : pub(super) const LSN_OFFSET: Lsn = Lsn(0x08);
1739 : /// Number of unique keys including in the test data
1740 : pub(super) const KEY_COUNT: u8 = 60;
1741 : /// Max number of different lsns for each key
1742 : pub(super) const MAX_ENTRIES_PER_KEY: u8 = 20;
1743 : /// Possible value sizes for each key along with a probability weight
1744 : pub(super) const VALUE_SIZES: [(usize, u8); 3] = [(100, 2), (1024, 2), (1024 * 1024, 1)];
1745 : /// Probability that there will be a gap between the current key and the next one (33.3%)
1746 : pub(super) const KEY_GAP_CHANGES: [(bool, u8); 2] = [(true, 1), (false, 2)];
1747 : /// The minimum size of a key range in all the generated reads
1748 : pub(super) const MIN_RANGE_SIZE: i128 = 10;
1749 : /// The number of ranges included in each vectored read
1750 : pub(super) const RANGES_COUNT: u8 = 2;
1751 : /// The number of vectored reads performed
1752 : pub(super) const READS_COUNT: u8 = 100;
1753 : /// Soft max size of a vectored read. Will be violated if we have to read keys
1754 : /// with values larger than the limit
1755 : pub(super) const MAX_VECTORED_READ_BYTES: usize = 64 * 1024;
1756 : }
1757 :
1758 : struct Entry {
1759 : key: Key,
1760 : lsn: Lsn,
1761 : value: Vec<u8>,
1762 : }
1763 :
1764 4 : fn generate_entries(rng: &mut StdRng) -> Vec<Entry> {
1765 4 : let mut current_key = Key::MIN;
1766 4 :
1767 4 : let mut entries = Vec::new();
1768 244 : for _ in 0..constants::KEY_COUNT {
1769 240 : let count = rng.gen_range(1..constants::MAX_ENTRIES_PER_KEY);
1770 240 : let mut lsns_iter =
1771 4520 : std::iter::successors(Some(Lsn(constants::LSN_OFFSET.0 + 0x08)), |lsn| {
1772 4520 : Some(Lsn(lsn.0 + 0x08))
1773 4520 : });
1774 240 : let mut lsns = Vec::new();
1775 4760 : while lsns.len() < count as usize {
1776 4520 : let take = rng.gen_bool(0.5);
1777 4520 : let lsn = lsns_iter.next().unwrap();
1778 4520 : if take {
1779 2224 : lsns.push(lsn);
1780 2296 : }
1781 : }
1782 :
1783 2464 : for lsn in lsns {
1784 2224 : let size = constants::VALUE_SIZES
1785 6672 : .choose_weighted(rng, |item| item.1)
1786 2224 : .unwrap()
1787 2224 : .0;
1788 2224 : let mut buf = vec![0; size];
1789 2224 : rng.fill_bytes(&mut buf);
1790 2224 :
1791 2224 : entries.push(Entry {
1792 2224 : key: current_key,
1793 2224 : lsn,
1794 2224 : value: buf,
1795 2224 : })
1796 : }
1797 :
1798 240 : let gap = constants::KEY_GAP_CHANGES
1799 480 : .choose_weighted(rng, |item| item.1)
1800 240 : .unwrap()
1801 240 : .0;
1802 240 : if gap {
1803 76 : current_key = current_key.add(2);
1804 164 : } else {
1805 164 : current_key = current_key.add(1);
1806 164 : }
1807 : }
1808 :
1809 4 : entries
1810 4 : }
1811 :
1812 : struct EntriesMeta {
1813 : key_range: Range<Key>,
1814 : lsn_range: Range<Lsn>,
1815 : index: BTreeMap<(Key, Lsn), Vec<u8>>,
1816 : }
1817 :
1818 4 : fn get_entries_meta(entries: &[Entry]) -> EntriesMeta {
1819 2224 : let key_range = match entries.iter().minmax_by_key(|e| e.key) {
1820 4 : MinMaxResult::MinMax(min, max) => min.key..max.key.next(),
1821 0 : _ => panic!("More than one entry is always expected"),
1822 : };
1823 :
1824 2224 : let lsn_range = match entries.iter().minmax_by_key(|e| e.lsn) {
1825 4 : MinMaxResult::MinMax(min, max) => min.lsn..Lsn(max.lsn.0 + 1),
1826 0 : _ => panic!("More than one entry is always expected"),
1827 : };
1828 :
1829 4 : let mut index = BTreeMap::new();
1830 2224 : for entry in entries.iter() {
1831 2224 : index.insert((entry.key, entry.lsn), entry.value.clone());
1832 2224 : }
1833 :
1834 4 : EntriesMeta {
1835 4 : key_range,
1836 4 : lsn_range,
1837 4 : index,
1838 4 : }
1839 4 : }
1840 :
1841 400 : fn pick_random_keyspace(rng: &mut StdRng, key_range: &Range<Key>) -> KeySpace {
1842 400 : let start = key_range.start.to_i128();
1843 400 : let end = key_range.end.to_i128();
1844 400 :
1845 400 : let mut keyspace = KeySpace::default();
1846 :
1847 1200 : for _ in 0..constants::RANGES_COUNT {
1848 800 : let mut range: Option<Range<Key>> = Option::default();
1849 2488 : while range.is_none() || keyspace.overlaps(range.as_ref().unwrap()) {
1850 1688 : let range_start = rng.gen_range(start..end);
1851 1688 : let range_end_offset = range_start + constants::MIN_RANGE_SIZE;
1852 1688 : if range_end_offset >= end {
1853 200 : range = Some(Key::from_i128(range_start)..Key::from_i128(end));
1854 1488 : } else {
1855 1488 : let range_end = rng.gen_range((range_start + constants::MIN_RANGE_SIZE)..end);
1856 1488 : range = Some(Key::from_i128(range_start)..Key::from_i128(range_end));
1857 1488 : }
1858 : }
1859 800 : keyspace.ranges.push(range.unwrap());
1860 : }
1861 :
1862 400 : keyspace
1863 400 : }
1864 :
1865 : #[tokio::test]
1866 4 : async fn test_delta_layer_vectored_read_end_to_end() -> anyhow::Result<()> {
1867 4 : let harness = TenantHarness::create("test_delta_layer_oversized_vectored_read").await?;
1868 4 : let (tenant, ctx) = harness.load().await;
1869 4 :
1870 4 : let timeline_id = TimelineId::generate();
1871 4 : let timeline = tenant
1872 4 : .create_test_timeline(timeline_id, constants::LSN_OFFSET, DEFAULT_PG_VERSION, &ctx)
1873 4 : .await?;
1874 4 :
1875 4 : tracing::info!("Generating test data ...");
1876 4 :
1877 4 : let rng = &mut StdRng::seed_from_u64(0);
1878 4 : let entries = generate_entries(rng);
1879 4 : let entries_meta = get_entries_meta(&entries);
1880 4 :
1881 4 : tracing::info!("Done generating {} entries", entries.len());
1882 4 :
1883 4 : tracing::info!("Writing test data to delta layer ...");
1884 4 : let mut writer = DeltaLayerWriter::new(
1885 4 : harness.conf,
1886 4 : timeline_id,
1887 4 : harness.tenant_shard_id,
1888 4 : entries_meta.key_range.start,
1889 4 : entries_meta.lsn_range.clone(),
1890 4 : &ctx,
1891 4 : )
1892 4 : .await?;
1893 4 :
1894 2228 : for entry in entries {
1895 2224 : let (_, res) = writer
1896 2224 : .put_value_bytes(entry.key, entry.lsn, entry.value.slice_len(), false, &ctx)
1897 2224 : .await;
1898 2224 : res?;
1899 4 : }
1900 4 :
1901 4 : let (desc, path) = writer.finish(entries_meta.key_range.end, &ctx).await?;
1902 4 : let resident = Layer::finish_creating(harness.conf, &timeline, desc, &path)?;
1903 4 :
1904 4 : let inner = resident.get_as_delta(&ctx).await?;
1905 4 :
1906 4 : let file_size = inner.file.metadata().await?.len();
1907 4 : tracing::info!(
1908 4 : "Done writing test data to delta layer. Resulting file size is: {}",
1909 4 : file_size
1910 4 : );
1911 4 :
1912 404 : for i in 0..constants::READS_COUNT {
1913 400 : tracing::info!("Doing vectored read {}/{}", i + 1, constants::READS_COUNT);
1914 4 :
1915 400 : let block_reader = FileBlockReader::new(&inner.file, inner.file_id);
1916 400 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1917 400 : inner.index_start_blk,
1918 400 : inner.index_root_blk,
1919 400 : block_reader,
1920 400 : );
1921 400 :
1922 400 : let planner = VectoredReadPlanner::new(constants::MAX_VECTORED_READ_BYTES);
1923 400 : let keyspace = pick_random_keyspace(rng, &entries_meta.key_range);
1924 400 : let data_end_offset = inner.index_start_blk as u64 * PAGE_SZ as u64;
1925 4 :
1926 400 : let vectored_reads = DeltaLayerInner::plan_reads(
1927 400 : &keyspace,
1928 400 : entries_meta.lsn_range.clone(),
1929 400 : data_end_offset,
1930 400 : index_reader,
1931 400 : planner,
1932 400 : &ctx,
1933 400 : )
1934 400 : .await?;
1935 4 :
1936 400 : let vectored_blob_reader = VectoredBlobReader::new(&inner.file);
1937 400 : let buf_size = DeltaLayerInner::get_min_read_buffer_size(
1938 400 : &vectored_reads,
1939 400 : constants::MAX_VECTORED_READ_BYTES,
1940 400 : );
1941 400 : let mut buf = Some(IoBufferMut::with_capacity(buf_size));
1942 4 :
1943 39848 : for read in vectored_reads {
1944 39448 : let blobs_buf = vectored_blob_reader
1945 39448 : .read_blobs(&read, buf.take().expect("Should have a buffer"), &ctx)
1946 39448 : .await?;
1947 39448 : let view = BufView::new_slice(&blobs_buf.buf);
1948 114608 : for meta in blobs_buf.blobs.iter() {
1949 114608 : let value = meta.read(&view).await?;
1950 114608 : assert_eq!(
1951 114608 : &value[..],
1952 114608 : &entries_meta.index[&(meta.meta.key, meta.meta.lsn)]
1953 114608 : );
1954 4 : }
1955 4 :
1956 39448 : buf = Some(blobs_buf.buf);
1957 4 : }
1958 4 : }
1959 4 :
1960 4 : Ok(())
1961 4 : }
1962 :
1963 : #[tokio::test]
1964 4 : async fn copy_delta_prefix_smoke() {
1965 4 : use bytes::Bytes;
1966 4 : use pageserver_api::record::NeonWalRecord;
1967 4 :
1968 4 : let h = crate::tenant::harness::TenantHarness::create("truncate_delta_smoke")
1969 4 : .await
1970 4 : .unwrap();
1971 4 : let (tenant, ctx) = h.load().await;
1972 4 : let ctx = &ctx;
1973 4 : let timeline = tenant
1974 4 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, ctx)
1975 4 : .await
1976 4 : .unwrap();
1977 4 :
1978 4 : let initdb_layer = timeline
1979 4 : .layers
1980 4 : .read()
1981 4 : .await
1982 4 : .likely_resident_layers()
1983 4 : .next()
1984 4 : .cloned()
1985 4 : .unwrap();
1986 4 :
1987 4 : {
1988 4 : let mut writer = timeline.writer().await;
1989 4 :
1990 4 : let data = [
1991 4 : (0x20, 12, Value::Image(Bytes::from_static(b"foobar"))),
1992 4 : (
1993 4 : 0x30,
1994 4 : 12,
1995 4 : Value::WalRecord(NeonWalRecord::Postgres {
1996 4 : will_init: false,
1997 4 : rec: Bytes::from_static(b"1"),
1998 4 : }),
1999 4 : ),
2000 4 : (
2001 4 : 0x40,
2002 4 : 12,
2003 4 : Value::WalRecord(NeonWalRecord::Postgres {
2004 4 : will_init: true,
2005 4 : rec: Bytes::from_static(b"2"),
2006 4 : }),
2007 4 : ),
2008 4 : // build an oversized value so we cannot extend and existing read over
2009 4 : // this
2010 4 : (
2011 4 : 0x50,
2012 4 : 12,
2013 4 : Value::WalRecord(NeonWalRecord::Postgres {
2014 4 : will_init: true,
2015 4 : rec: {
2016 4 : let mut buf =
2017 4 : vec![0u8; tenant.conf.max_vectored_read_bytes.0.get() + 1024];
2018 4 : buf.iter_mut()
2019 4 : .enumerate()
2020 536576 : .for_each(|(i, slot)| *slot = (i % 256) as u8);
2021 4 : Bytes::from(buf)
2022 4 : },
2023 4 : }),
2024 4 : ),
2025 4 : // because the oversized read cannot be extended further, we are sure to exercise the
2026 4 : // builder created on the last round with this:
2027 4 : (
2028 4 : 0x60,
2029 4 : 12,
2030 4 : Value::WalRecord(NeonWalRecord::Postgres {
2031 4 : will_init: true,
2032 4 : rec: Bytes::from_static(b"3"),
2033 4 : }),
2034 4 : ),
2035 4 : (
2036 4 : 0x60,
2037 4 : 9,
2038 4 : Value::Image(Bytes::from_static(b"something for a different key")),
2039 4 : ),
2040 4 : ];
2041 4 :
2042 4 : let mut last_lsn = None;
2043 4 :
2044 28 : for (lsn, key, value) in data {
2045 24 : let key = Key::from_i128(key);
2046 24 : writer.put(key, Lsn(lsn), &value, ctx).await.unwrap();
2047 24 : last_lsn = Some(lsn);
2048 4 : }
2049 4 :
2050 4 : writer.finish_write(Lsn(last_lsn.unwrap()));
2051 4 : }
2052 4 : timeline.freeze_and_flush().await.unwrap();
2053 4 :
2054 4 : let new_layer = timeline
2055 4 : .layers
2056 4 : .read()
2057 4 : .await
2058 4 : .likely_resident_layers()
2059 5 : .find(|&x| x != &initdb_layer)
2060 4 : .cloned()
2061 4 : .unwrap();
2062 4 :
2063 4 : // create a copy for the timeline, so we don't overwrite the file
2064 4 : let branch = tenant
2065 4 : .branch_timeline_test(&timeline, TimelineId::generate(), None, ctx)
2066 4 : .await
2067 4 : .unwrap();
2068 4 :
2069 4 : assert_eq!(branch.get_ancestor_lsn(), Lsn(0x60));
2070 4 :
2071 4 : // truncating at 0x61 gives us a full copy, otherwise just go backwards until there's just
2072 4 : // a single key
2073 4 :
2074 24 : for truncate_at in [0x61, 0x51, 0x41, 0x31, 0x21] {
2075 20 : let truncate_at = Lsn(truncate_at);
2076 4 :
2077 20 : let mut writer = DeltaLayerWriter::new(
2078 20 : tenant.conf,
2079 20 : branch.timeline_id,
2080 20 : tenant.tenant_shard_id,
2081 20 : Key::MIN,
2082 20 : Lsn(0x11)..truncate_at,
2083 20 : ctx,
2084 20 : )
2085 20 : .await
2086 20 : .unwrap();
2087 4 :
2088 20 : let new_layer = new_layer.download_and_keep_resident().await.unwrap();
2089 20 :
2090 20 : new_layer
2091 20 : .copy_delta_prefix(&mut writer, truncate_at, ctx)
2092 20 : .await
2093 20 : .unwrap();
2094 4 :
2095 20 : let (desc, path) = writer.finish(Key::MAX, ctx).await.unwrap();
2096 20 : let copied_layer = Layer::finish_creating(tenant.conf, &branch, desc, &path).unwrap();
2097 20 :
2098 20 : copied_layer.get_as_delta(ctx).await.unwrap();
2099 20 :
2100 20 : assert_keys_and_values_eq(
2101 20 : new_layer.get_as_delta(ctx).await.unwrap(),
2102 20 : copied_layer.get_as_delta(ctx).await.unwrap(),
2103 20 : truncate_at,
2104 20 : ctx,
2105 20 : )
2106 20 : .await;
2107 4 : }
2108 4 : }
2109 :
2110 20 : async fn assert_keys_and_values_eq(
2111 20 : source: &DeltaLayerInner,
2112 20 : truncated: &DeltaLayerInner,
2113 20 : truncated_at: Lsn,
2114 20 : ctx: &RequestContext,
2115 20 : ) {
2116 : use futures::future::ready;
2117 : use futures::stream::TryStreamExt;
2118 :
2119 20 : let start_key = [0u8; DELTA_KEY_SIZE];
2120 20 :
2121 20 : let source_reader = FileBlockReader::new(&source.file, source.file_id);
2122 20 : let source_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2123 20 : source.index_start_blk,
2124 20 : source.index_root_blk,
2125 20 : &source_reader,
2126 20 : );
2127 20 : let source_stream = source.stream_index_forwards(source_tree, &start_key, ctx);
2128 120 : let source_stream = source_stream.filter(|res| match res {
2129 120 : Ok((_, lsn, _)) => ready(lsn < &truncated_at),
2130 0 : _ => ready(true),
2131 120 : });
2132 20 : let mut source_stream = std::pin::pin!(source_stream);
2133 20 :
2134 20 : let truncated_reader = FileBlockReader::new(&truncated.file, truncated.file_id);
2135 20 : let truncated_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2136 20 : truncated.index_start_blk,
2137 20 : truncated.index_root_blk,
2138 20 : &truncated_reader,
2139 20 : );
2140 20 : let truncated_stream = truncated.stream_index_forwards(truncated_tree, &start_key, ctx);
2141 20 : let mut truncated_stream = std::pin::pin!(truncated_stream);
2142 20 :
2143 20 : let mut scratch_left = Vec::new();
2144 20 : let mut scratch_right = Vec::new();
2145 :
2146 : loop {
2147 84 : let (src, truncated) = (source_stream.try_next(), truncated_stream.try_next());
2148 84 : let (src, truncated) = tokio::try_join!(src, truncated).unwrap();
2149 84 :
2150 84 : if src.is_none() {
2151 20 : assert!(truncated.is_none());
2152 20 : break;
2153 64 : }
2154 64 :
2155 64 : let (src, truncated) = (src.unwrap(), truncated.unwrap());
2156 64 :
2157 64 : // because we've filtered the source with Lsn, we should always have the same keys from both.
2158 64 : assert_eq!(src.0, truncated.0);
2159 64 : assert_eq!(src.1, truncated.1);
2160 :
2161 : // if this is needed for something else, just drop this assert.
2162 64 : assert!(
2163 64 : src.2.pos() >= truncated.2.pos(),
2164 0 : "value position should not go backwards {} vs. {}",
2165 0 : src.2.pos(),
2166 0 : truncated.2.pos()
2167 : );
2168 :
2169 64 : scratch_left.clear();
2170 64 : let src_cursor = source_reader.block_cursor();
2171 64 : let left = src_cursor.read_blob_into_buf(src.2.pos(), &mut scratch_left, ctx);
2172 64 : scratch_right.clear();
2173 64 : let trunc_cursor = truncated_reader.block_cursor();
2174 64 : let right = trunc_cursor.read_blob_into_buf(truncated.2.pos(), &mut scratch_right, ctx);
2175 64 :
2176 64 : tokio::try_join!(left, right).unwrap();
2177 64 :
2178 64 : assert_eq!(utils::Hex(&scratch_left), utils::Hex(&scratch_right));
2179 : }
2180 20 : }
2181 :
2182 36448 : pub(crate) fn sort_delta(
2183 36448 : (k1, l1, _): &(Key, Lsn, Value),
2184 36448 : (k2, l2, _): &(Key, Lsn, Value),
2185 36448 : ) -> std::cmp::Ordering {
2186 36448 : (k1, l1).cmp(&(k2, l2))
2187 36448 : }
2188 :
2189 : #[cfg(feature = "testing")]
2190 188 : pub(crate) fn sort_delta_value(
2191 188 : (k1, l1, v1): &(Key, Lsn, Value),
2192 188 : (k2, l2, v2): &(Key, Lsn, Value),
2193 188 : ) -> std::cmp::Ordering {
2194 188 : let order_1 = if v1.is_image() { 0 } else { 1 };
2195 188 : let order_2 = if v2.is_image() { 0 } else { 1 };
2196 188 : (k1, l1, order_1).cmp(&(k2, l2, order_2))
2197 188 : }
2198 :
2199 44 : pub(crate) async fn produce_delta_layer(
2200 44 : tenant: &Tenant,
2201 44 : tline: &Arc<Timeline>,
2202 44 : mut deltas: Vec<(Key, Lsn, Value)>,
2203 44 : ctx: &RequestContext,
2204 44 : ) -> anyhow::Result<ResidentLayer> {
2205 44 : deltas.sort_by(sort_delta);
2206 44 : let (key_start, _, _) = deltas.first().unwrap();
2207 44 : let (key_max, _, _) = deltas.last().unwrap();
2208 16480 : let lsn_min = deltas.iter().map(|(_, lsn, _)| lsn).min().unwrap();
2209 16480 : let lsn_max = deltas.iter().map(|(_, lsn, _)| lsn).max().unwrap();
2210 44 : let lsn_end = Lsn(lsn_max.0 + 1);
2211 44 : let mut writer = DeltaLayerWriter::new(
2212 44 : tenant.conf,
2213 44 : tline.timeline_id,
2214 44 : tenant.tenant_shard_id,
2215 44 : *key_start,
2216 44 : (*lsn_min)..lsn_end,
2217 44 : ctx,
2218 44 : )
2219 44 : .await?;
2220 44 : let key_end = key_max.next();
2221 :
2222 16524 : for (key, lsn, value) in deltas {
2223 16480 : writer.put_value(key, lsn, value, ctx).await?;
2224 : }
2225 :
2226 44 : let (desc, path) = writer.finish(key_end, ctx).await?;
2227 44 : let delta_layer = Layer::finish_creating(tenant.conf, tline, desc, &path)?;
2228 :
2229 44 : Ok::<_, anyhow::Error>(delta_layer)
2230 44 : }
2231 :
2232 56 : async fn assert_delta_iter_equal(
2233 56 : delta_iter: &mut DeltaLayerIterator<'_>,
2234 56 : expect: &[(Key, Lsn, Value)],
2235 56 : ) {
2236 56 : let mut expect_iter = expect.iter();
2237 : loop {
2238 56056 : let o1 = delta_iter.next().await.unwrap();
2239 56056 : let o2 = expect_iter.next();
2240 56056 : assert_eq!(o1.is_some(), o2.is_some());
2241 56056 : if o1.is_none() && o2.is_none() {
2242 56 : break;
2243 56000 : }
2244 56000 : let (k1, l1, v1) = o1.unwrap();
2245 56000 : let (k2, l2, v2) = o2.unwrap();
2246 56000 : assert_eq!(&k1, k2);
2247 56000 : assert_eq!(l1, *l2);
2248 56000 : assert_eq!(&v1, v2);
2249 : }
2250 56 : }
2251 :
2252 : #[tokio::test]
2253 4 : async fn delta_layer_iterator() {
2254 4 : let harness = TenantHarness::create("delta_layer_iterator").await.unwrap();
2255 4 : let (tenant, ctx) = harness.load().await;
2256 4 :
2257 4 : let tline = tenant
2258 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
2259 4 : .await
2260 4 : .unwrap();
2261 4 :
2262 4000 : fn get_key(id: u32) -> Key {
2263 4000 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
2264 4000 : key.field6 = id;
2265 4000 : key
2266 4000 : }
2267 4 : const N: usize = 1000;
2268 4 : let test_deltas = (0..N)
2269 4000 : .map(|idx| {
2270 4000 : (
2271 4000 : get_key(idx as u32 / 10),
2272 4000 : Lsn(0x10 * ((idx as u64) % 10 + 1)),
2273 4000 : Value::Image(Bytes::from(format!("img{idx:05}"))),
2274 4000 : )
2275 4000 : })
2276 4 : .collect_vec();
2277 4 : let resident_layer = produce_delta_layer(&tenant, &tline, test_deltas.clone(), &ctx)
2278 4 : .await
2279 4 : .unwrap();
2280 4 : let delta_layer = resident_layer.get_as_delta(&ctx).await.unwrap();
2281 12 : for max_read_size in [1, 1024] {
2282 64 : for batch_size in [1, 2, 4, 8, 3, 7, 13] {
2283 56 : println!("running with batch_size={batch_size} max_read_size={max_read_size}");
2284 56 : // Test if the batch size is correctly determined
2285 56 : let mut iter = delta_layer.iter(&ctx);
2286 56 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2287 56 : let mut num_items = 0;
2288 224 : for _ in 0..3 {
2289 168 : iter.next_batch().await.unwrap();
2290 168 : num_items += iter.key_values_batch.len();
2291 168 : if max_read_size == 1 {
2292 4 : // every key should be a batch b/c the value is larger than max_read_size
2293 84 : assert_eq!(iter.key_values_batch.len(), 1);
2294 4 : } else {
2295 84 : assert!(iter.key_values_batch.len() <= batch_size);
2296 4 : }
2297 168 : if num_items >= N {
2298 4 : break;
2299 168 : }
2300 168 : iter.key_values_batch.clear();
2301 4 : }
2302 4 : // Test if the result is correct
2303 56 : let mut iter = delta_layer.iter(&ctx);
2304 56 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2305 56 : assert_delta_iter_equal(&mut iter, &test_deltas).await;
2306 4 : }
2307 4 : }
2308 4 : }
2309 : }
|
