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::{Key, DBDIR_KEY, KEY_SIZE};
55 : use pageserver_api::keyspace::KeySpace;
56 : use pageserver_api::models::ImageCompressionAlgorithm;
57 : use pageserver_api::shard::TenantShardId;
58 : use pageserver_api::value::Value;
59 : use rand::{distributions::Alphanumeric, Rng};
60 : use serde::{Deserialize, Serialize};
61 : use std::collections::{HashMap, 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::{
79 : AsLayerDesc, LayerName, OnDiskValue, OnDiskValueIo, PersistentLayerDesc, ResidentLayer,
80 : ValuesReconstructState,
81 : };
82 :
83 : ///
84 : /// Header stored in the beginning of the file
85 : ///
86 : /// After this comes the 'values' part, starting on block 1. After that,
87 : /// the 'index' starts at the block indicated by 'index_start_blk'
88 : ///
89 0 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
90 : pub struct Summary {
91 : /// Magic value to identify this as a neon delta file. Always DELTA_FILE_MAGIC.
92 : pub magic: u16,
93 : pub format_version: u16,
94 :
95 : pub tenant_id: TenantId,
96 : pub timeline_id: TimelineId,
97 : pub key_range: Range<Key>,
98 : pub lsn_range: Range<Lsn>,
99 :
100 : /// Block number where the 'index' part of the file begins.
101 : pub index_start_blk: u32,
102 : /// Block within the 'index', where the B-tree root page is stored
103 : pub index_root_blk: u32,
104 : }
105 :
106 : impl From<&DeltaLayer> for Summary {
107 0 : fn from(layer: &DeltaLayer) -> Self {
108 0 : Self::expected(
109 0 : layer.desc.tenant_shard_id.tenant_id,
110 0 : layer.desc.timeline_id,
111 0 : layer.desc.key_range.clone(),
112 0 : layer.desc.lsn_range.clone(),
113 0 : )
114 0 : }
115 : }
116 :
117 : impl Summary {
118 2188 : pub(super) fn expected(
119 2188 : tenant_id: TenantId,
120 2188 : timeline_id: TimelineId,
121 2188 : keys: Range<Key>,
122 2188 : lsns: Range<Lsn>,
123 2188 : ) -> Self {
124 2188 : Self {
125 2188 : magic: DELTA_FILE_MAGIC,
126 2188 : format_version: STORAGE_FORMAT_VERSION,
127 2188 :
128 2188 : tenant_id,
129 2188 : timeline_id,
130 2188 : key_range: keys,
131 2188 : lsn_range: lsns,
132 2188 :
133 2188 : index_start_blk: 0,
134 2188 : index_root_blk: 0,
135 2188 : }
136 2188 : }
137 : }
138 :
139 : // Flag indicating that this version initialize the page
140 : const WILL_INIT: u64 = 1;
141 :
142 : /// Struct representing reference to BLOB in layers.
143 : ///
144 : /// Reference contains BLOB offset, and for WAL records it also contains
145 : /// `will_init` flag. The flag helps to determine the range of records
146 : /// that needs to be applied, without reading/deserializing records themselves.
147 0 : #[derive(Debug, Serialize, Deserialize, Copy, Clone)]
148 : pub struct BlobRef(pub u64);
149 :
150 : impl BlobRef {
151 5369047 : pub fn will_init(&self) -> bool {
152 5369047 : (self.0 & WILL_INIT) != 0
153 5369047 : }
154 :
155 9582195 : pub fn pos(&self) -> u64 {
156 9582195 : self.0 >> 1
157 9582195 : }
158 :
159 12942956 : pub fn new(pos: u64, will_init: bool) -> BlobRef {
160 12942956 : let mut blob_ref = pos << 1;
161 12942956 : if will_init {
162 12939836 : blob_ref |= WILL_INIT;
163 12939836 : }
164 12942956 : BlobRef(blob_ref)
165 12942956 : }
166 : }
167 :
168 : pub const DELTA_KEY_SIZE: usize = KEY_SIZE + 8;
169 : struct DeltaKey([u8; DELTA_KEY_SIZE]);
170 :
171 : /// This is the key of the B-tree index stored in the delta layer. It consists
172 : /// of the serialized representation of a Key and LSN.
173 : impl DeltaKey {
174 4128396 : fn from_slice(buf: &[u8]) -> Self {
175 4128396 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
176 4128396 : bytes.copy_from_slice(buf);
177 4128396 : DeltaKey(bytes)
178 4128396 : }
179 :
180 13377996 : fn from_key_lsn(key: &Key, lsn: Lsn) -> Self {
181 13377996 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
182 13377996 : key.write_to_byte_slice(&mut bytes[0..KEY_SIZE]);
183 13377996 : bytes[KEY_SIZE..].copy_from_slice(&u64::to_be_bytes(lsn.0));
184 13377996 : DeltaKey(bytes)
185 13377996 : }
186 :
187 4128396 : fn key(&self) -> Key {
188 4128396 : Key::from_slice(&self.0)
189 4128396 : }
190 :
191 4128396 : fn lsn(&self) -> Lsn {
192 4128396 : Lsn(u64::from_be_bytes(self.0[KEY_SIZE..].try_into().unwrap()))
193 4128396 : }
194 :
195 5453679 : fn extract_lsn_from_buf(buf: &[u8]) -> Lsn {
196 5453679 : let mut lsn_buf = [0u8; 8];
197 5453679 : lsn_buf.copy_from_slice(&buf[KEY_SIZE..]);
198 5453679 : Lsn(u64::from_be_bytes(lsn_buf))
199 5453679 : }
200 : }
201 :
202 : /// This is used only from `pagectl`. Within pageserver, all layers are
203 : /// [`crate::tenant::storage_layer::Layer`], which can hold a [`DeltaLayerInner`].
204 : pub struct DeltaLayer {
205 : path: Utf8PathBuf,
206 : pub desc: PersistentLayerDesc,
207 : inner: OnceCell<Arc<DeltaLayerInner>>,
208 : }
209 :
210 : impl std::fmt::Debug for DeltaLayer {
211 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
212 : use super::RangeDisplayDebug;
213 :
214 0 : f.debug_struct("DeltaLayer")
215 0 : .field("key_range", &RangeDisplayDebug(&self.desc.key_range))
216 0 : .field("lsn_range", &self.desc.lsn_range)
217 0 : .field("file_size", &self.desc.file_size)
218 0 : .field("inner", &self.inner)
219 0 : .finish()
220 0 : }
221 : }
222 :
223 : /// `DeltaLayerInner` is the in-memory data structure associated with an on-disk delta
224 : /// file.
225 : pub struct DeltaLayerInner {
226 : // values copied from summary
227 : index_start_blk: u32,
228 : index_root_blk: u32,
229 :
230 : file: Arc<VirtualFile>,
231 : file_id: FileId,
232 :
233 : layer_key_range: Range<Key>,
234 : layer_lsn_range: Range<Lsn>,
235 :
236 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
237 : }
238 :
239 : impl DeltaLayerInner {
240 0 : pub(crate) fn layer_dbg_info(&self) -> String {
241 0 : format!(
242 0 : "delta {}..{} {}..{}",
243 0 : self.key_range().start,
244 0 : self.key_range().end,
245 0 : self.lsn_range().start,
246 0 : self.lsn_range().end
247 0 : )
248 0 : }
249 : }
250 :
251 : impl std::fmt::Debug for DeltaLayerInner {
252 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
253 0 : f.debug_struct("DeltaLayerInner")
254 0 : .field("index_start_blk", &self.index_start_blk)
255 0 : .field("index_root_blk", &self.index_root_blk)
256 0 : .finish()
257 0 : }
258 : }
259 :
260 : /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
261 : impl std::fmt::Display for DeltaLayer {
262 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
263 0 : write!(f, "{}", self.layer_desc().short_id())
264 0 : }
265 : }
266 :
267 : impl AsLayerDesc for DeltaLayer {
268 0 : fn layer_desc(&self) -> &PersistentLayerDesc {
269 0 : &self.desc
270 0 : }
271 : }
272 :
273 : impl DeltaLayer {
274 0 : pub async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
275 0 : self.desc.dump();
276 0 :
277 0 : if !verbose {
278 0 : return Ok(());
279 0 : }
280 :
281 0 : let inner = self.load(ctx).await?;
282 :
283 0 : inner.dump(ctx).await
284 0 : }
285 :
286 2936 : fn temp_path_for(
287 2936 : conf: &PageServerConf,
288 2936 : tenant_shard_id: &TenantShardId,
289 2936 : timeline_id: &TimelineId,
290 2936 : key_start: Key,
291 2936 : lsn_range: &Range<Lsn>,
292 2936 : ) -> Utf8PathBuf {
293 2936 : let rand_string: String = rand::thread_rng()
294 2936 : .sample_iter(&Alphanumeric)
295 2936 : .take(8)
296 2936 : .map(char::from)
297 2936 : .collect();
298 2936 :
299 2936 : conf.timeline_path(tenant_shard_id, timeline_id)
300 2936 : .join(format!(
301 2936 : "{}-XXX__{:016X}-{:016X}.{}.{}",
302 2936 : key_start,
303 2936 : u64::from(lsn_range.start),
304 2936 : u64::from(lsn_range.end),
305 2936 : rand_string,
306 2936 : TEMP_FILE_SUFFIX,
307 2936 : ))
308 2936 : }
309 :
310 : ///
311 : /// Open the underlying file and read the metadata into memory, if it's
312 : /// not loaded already.
313 : ///
314 0 : async fn load(&self, ctx: &RequestContext) -> Result<&Arc<DeltaLayerInner>> {
315 0 : // Quick exit if already loaded
316 0 : self.inner
317 0 : .get_or_try_init(|| self.load_inner(ctx))
318 0 : .await
319 0 : .with_context(|| format!("Failed to load delta layer {}", self.path()))
320 0 : }
321 :
322 0 : async fn load_inner(&self, ctx: &RequestContext) -> anyhow::Result<Arc<DeltaLayerInner>> {
323 0 : let path = self.path();
324 :
325 0 : let loaded = DeltaLayerInner::load(&path, None, None, ctx).await?;
326 :
327 : // not production code
328 0 : let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
329 0 : let expected_layer_name = self.layer_desc().layer_name();
330 0 :
331 0 : if actual_layer_name != expected_layer_name {
332 0 : println!("warning: filename does not match what is expected from in-file summary");
333 0 : println!("actual: {:?}", actual_layer_name.to_string());
334 0 : println!("expected: {:?}", expected_layer_name.to_string());
335 0 : }
336 :
337 0 : Ok(Arc::new(loaded))
338 0 : }
339 :
340 : /// Create a DeltaLayer struct representing an existing file on disk.
341 : ///
342 : /// This variant is only used for debugging purposes, by the 'pagectl' binary.
343 0 : pub fn new_for_path(path: &Utf8Path, file: File) -> Result<Self> {
344 0 : let mut summary_buf = vec![0; PAGE_SZ];
345 0 : file.read_exact_at(&mut summary_buf, 0)?;
346 0 : let summary = Summary::des_prefix(&summary_buf)?;
347 :
348 0 : let metadata = file
349 0 : .metadata()
350 0 : .context("get file metadata to determine size")?;
351 :
352 : // This function is never used for constructing layers in a running pageserver,
353 : // so it does not need an accurate TenantShardId.
354 0 : let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
355 0 :
356 0 : Ok(DeltaLayer {
357 0 : path: path.to_path_buf(),
358 0 : desc: PersistentLayerDesc::new_delta(
359 0 : tenant_shard_id,
360 0 : summary.timeline_id,
361 0 : summary.key_range,
362 0 : summary.lsn_range,
363 0 : metadata.len(),
364 0 : ),
365 0 : inner: OnceCell::new(),
366 0 : })
367 0 : }
368 :
369 : /// Path to the layer file in pageserver workdir.
370 0 : fn path(&self) -> Utf8PathBuf {
371 0 : self.path.clone()
372 0 : }
373 : }
374 :
375 : /// A builder object for constructing a new delta layer.
376 : ///
377 : /// Usage:
378 : ///
379 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
380 : ///
381 : /// 2. Write the contents by calling `put_value` for every page
382 : /// version to store in the layer.
383 : ///
384 : /// 3. Call `finish`.
385 : ///
386 : struct DeltaLayerWriterInner {
387 : pub path: Utf8PathBuf,
388 : timeline_id: TimelineId,
389 : tenant_shard_id: TenantShardId,
390 :
391 : key_start: Key,
392 : lsn_range: Range<Lsn>,
393 :
394 : tree: DiskBtreeBuilder<BlockBuf, DELTA_KEY_SIZE>,
395 :
396 : blob_writer: BlobWriter<true>,
397 :
398 : // Number of key-lsns in the layer.
399 : num_keys: usize,
400 : }
401 :
402 : impl DeltaLayerWriterInner {
403 : ///
404 : /// Start building a new delta layer.
405 : ///
406 2936 : async fn new(
407 2936 : conf: &'static PageServerConf,
408 2936 : timeline_id: TimelineId,
409 2936 : tenant_shard_id: TenantShardId,
410 2936 : key_start: Key,
411 2936 : lsn_range: Range<Lsn>,
412 2936 : ctx: &RequestContext,
413 2936 : ) -> anyhow::Result<Self> {
414 2936 : // Create the file initially with a temporary filename. We don't know
415 2936 : // the end key yet, so we cannot form the final filename yet. We will
416 2936 : // rename it when we're done.
417 2936 : //
418 2936 : // Note: This overwrites any existing file. There shouldn't be any.
419 2936 : // FIXME: throw an error instead?
420 2936 : let path =
421 2936 : DeltaLayer::temp_path_for(conf, &tenant_shard_id, &timeline_id, key_start, &lsn_range);
422 :
423 2936 : let mut file = VirtualFile::create(&path, ctx).await?;
424 : // make room for the header block
425 2936 : file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
426 2936 : let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
427 2936 :
428 2936 : // Initialize the b-tree index builder
429 2936 : let block_buf = BlockBuf::new();
430 2936 : let tree_builder = DiskBtreeBuilder::new(block_buf);
431 2936 :
432 2936 : Ok(Self {
433 2936 : path,
434 2936 : timeline_id,
435 2936 : tenant_shard_id,
436 2936 : key_start,
437 2936 : lsn_range,
438 2936 : tree: tree_builder,
439 2936 : blob_writer,
440 2936 : num_keys: 0,
441 2936 : })
442 2936 : }
443 :
444 : ///
445 : /// Append a key-value pair to the file.
446 : ///
447 : /// The values must be appended in key, lsn order.
448 : ///
449 4169428 : async fn put_value(
450 4169428 : &mut self,
451 4169428 : key: Key,
452 4169428 : lsn: Lsn,
453 4169428 : val: Value,
454 4169428 : ctx: &RequestContext,
455 4169428 : ) -> anyhow::Result<()> {
456 4169428 : let (_, res) = self
457 4169428 : .put_value_bytes(
458 4169428 : key,
459 4169428 : lsn,
460 4169428 : Value::ser(&val)?.slice_len(),
461 4169428 : val.will_init(),
462 4169428 : ctx,
463 4169428 : )
464 4169428 : .await;
465 4169428 : res
466 4169428 : }
467 :
468 12942748 : async fn put_value_bytes<Buf>(
469 12942748 : &mut self,
470 12942748 : key: Key,
471 12942748 : lsn: Lsn,
472 12942748 : val: FullSlice<Buf>,
473 12942748 : will_init: bool,
474 12942748 : ctx: &RequestContext,
475 12942748 : ) -> (FullSlice<Buf>, anyhow::Result<()>)
476 12942748 : where
477 12942748 : Buf: IoBuf + Send,
478 12942748 : {
479 12942748 : assert!(
480 12942748 : self.lsn_range.start <= lsn,
481 0 : "lsn_start={}, lsn={}",
482 : self.lsn_range.start,
483 : lsn
484 : );
485 : // We don't want to use compression in delta layer creation
486 12942748 : let compression = ImageCompressionAlgorithm::Disabled;
487 12942748 : let (val, res) = self
488 12942748 : .blob_writer
489 12942748 : .write_blob_maybe_compressed(val, ctx, compression)
490 12942748 : .await;
491 12942748 : let off = match res {
492 12942748 : Ok((off, _)) => off,
493 0 : Err(e) => return (val, Err(anyhow::anyhow!(e))),
494 : };
495 :
496 12942748 : let blob_ref = BlobRef::new(off, will_init);
497 12942748 :
498 12942748 : let delta_key = DeltaKey::from_key_lsn(&key, lsn);
499 12942748 : let res = self.tree.append(&delta_key.0, blob_ref.0);
500 12942748 :
501 12942748 : self.num_keys += 1;
502 12942748 :
503 12942748 : (val, res.map_err(|e| anyhow::anyhow!(e)))
504 12942748 : }
505 :
506 4047944 : fn size(&self) -> u64 {
507 4047944 : self.blob_writer.size() + self.tree.borrow_writer().size()
508 4047944 : }
509 :
510 : ///
511 : /// Finish writing the delta layer.
512 : ///
513 2884 : async fn finish(
514 2884 : self,
515 2884 : key_end: Key,
516 2884 : ctx: &RequestContext,
517 2884 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
518 2884 : let temp_path = self.path.clone();
519 2884 : let result = self.finish0(key_end, ctx).await;
520 2884 : if let Err(ref e) = result {
521 0 : tracing::info!(%temp_path, "cleaning up temporary file after error during writing: {e}");
522 0 : if let Err(e) = std::fs::remove_file(&temp_path) {
523 0 : tracing::warn!(error=%e, %temp_path, "error cleaning up temporary layer file after error during writing");
524 0 : }
525 2884 : }
526 2884 : result
527 2884 : }
528 :
529 2884 : async fn finish0(
530 2884 : self,
531 2884 : key_end: Key,
532 2884 : ctx: &RequestContext,
533 2884 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
534 2884 : let index_start_blk = self.blob_writer.size().div_ceil(PAGE_SZ as u64) as u32;
535 :
536 2884 : let mut file = self.blob_writer.into_inner(ctx).await?;
537 :
538 : // Write out the index
539 2884 : let (index_root_blk, block_buf) = self.tree.finish()?;
540 2884 : file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
541 2884 : .await?;
542 30194 : for buf in block_buf.blocks {
543 27310 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
544 27310 : res?;
545 : }
546 2884 : assert!(self.lsn_range.start < self.lsn_range.end);
547 : // Fill in the summary on blk 0
548 2884 : let summary = Summary {
549 2884 : magic: DELTA_FILE_MAGIC,
550 2884 : format_version: STORAGE_FORMAT_VERSION,
551 2884 : tenant_id: self.tenant_shard_id.tenant_id,
552 2884 : timeline_id: self.timeline_id,
553 2884 : key_range: self.key_start..key_end,
554 2884 : lsn_range: self.lsn_range.clone(),
555 2884 : index_start_blk,
556 2884 : index_root_blk,
557 2884 : };
558 2884 :
559 2884 : let mut buf = Vec::with_capacity(PAGE_SZ);
560 2884 : // TODO: could use smallvec here but it's a pain with Slice<T>
561 2884 : Summary::ser_into(&summary, &mut buf)?;
562 2884 : file.seek(SeekFrom::Start(0)).await?;
563 2884 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
564 2884 : res?;
565 :
566 2884 : let metadata = file
567 2884 : .metadata()
568 2884 : .await
569 2884 : .context("get file metadata to determine size")?;
570 :
571 : // 5GB limit for objects without multipart upload (which we don't want to use)
572 : // Make it a little bit below to account for differing GB units
573 : // https://docs.aws.amazon.com/AmazonS3/latest/userguide/upload-objects.html
574 2884 : ensure!(
575 2884 : metadata.len() <= S3_UPLOAD_LIMIT,
576 0 : "Created delta layer file at {} of size {} above limit {S3_UPLOAD_LIMIT}!",
577 0 : file.path(),
578 0 : metadata.len()
579 : );
580 :
581 : // Note: Because we opened the file in write-only mode, we cannot
582 : // reuse the same VirtualFile for reading later. That's why we don't
583 : // set inner.file here. The first read will have to re-open it.
584 :
585 2884 : let desc = PersistentLayerDesc::new_delta(
586 2884 : self.tenant_shard_id,
587 2884 : self.timeline_id,
588 2884 : self.key_start..key_end,
589 2884 : self.lsn_range.clone(),
590 2884 : metadata.len(),
591 2884 : );
592 2884 :
593 2884 : // fsync the file
594 2884 : file.sync_all()
595 2884 : .await
596 2884 : .maybe_fatal_err("delta_layer sync_all")?;
597 :
598 2884 : trace!("created delta layer {}", self.path);
599 :
600 2884 : Ok((desc, self.path))
601 2884 : }
602 : }
603 :
604 : /// A builder object for constructing a new delta layer.
605 : ///
606 : /// Usage:
607 : ///
608 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
609 : ///
610 : /// 2. Write the contents by calling `put_value` for every page
611 : /// version to store in the layer.
612 : ///
613 : /// 3. Call `finish`.
614 : ///
615 : /// # Note
616 : ///
617 : /// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
618 : /// possible for the writer to drop before `finish` is actually called. So this
619 : /// could lead to odd temporary files in the directory, exhausting file system.
620 : /// This structure wraps `DeltaLayerWriterInner` and also contains `Drop`
621 : /// implementation that cleans up the temporary file in failure. It's not
622 : /// possible to do this directly in `DeltaLayerWriterInner` since `finish` moves
623 : /// out some fields, making it impossible to implement `Drop`.
624 : ///
625 : #[must_use]
626 : pub struct DeltaLayerWriter {
627 : inner: Option<DeltaLayerWriterInner>,
628 : }
629 :
630 : impl DeltaLayerWriter {
631 : ///
632 : /// Start building a new delta layer.
633 : ///
634 2936 : pub async fn new(
635 2936 : conf: &'static PageServerConf,
636 2936 : timeline_id: TimelineId,
637 2936 : tenant_shard_id: TenantShardId,
638 2936 : key_start: Key,
639 2936 : lsn_range: Range<Lsn>,
640 2936 : ctx: &RequestContext,
641 2936 : ) -> anyhow::Result<Self> {
642 2936 : Ok(Self {
643 2936 : inner: Some(
644 2936 : DeltaLayerWriterInner::new(
645 2936 : conf,
646 2936 : timeline_id,
647 2936 : tenant_shard_id,
648 2936 : key_start,
649 2936 : lsn_range,
650 2936 : ctx,
651 2936 : )
652 2936 : .await?,
653 : ),
654 : })
655 2936 : }
656 :
657 0 : pub fn is_empty(&self) -> bool {
658 0 : self.inner.as_ref().unwrap().num_keys == 0
659 0 : }
660 :
661 : ///
662 : /// Append a key-value pair to the file.
663 : ///
664 : /// The values must be appended in key, lsn order.
665 : ///
666 4169428 : pub async fn put_value(
667 4169428 : &mut self,
668 4169428 : key: Key,
669 4169428 : lsn: Lsn,
670 4169428 : val: Value,
671 4169428 : ctx: &RequestContext,
672 4169428 : ) -> anyhow::Result<()> {
673 4169428 : self.inner
674 4169428 : .as_mut()
675 4169428 : .unwrap()
676 4169428 : .put_value(key, lsn, val, ctx)
677 4169428 : .await
678 4169428 : }
679 :
680 8773320 : pub async fn put_value_bytes<Buf>(
681 8773320 : &mut self,
682 8773320 : key: Key,
683 8773320 : lsn: Lsn,
684 8773320 : val: FullSlice<Buf>,
685 8773320 : will_init: bool,
686 8773320 : ctx: &RequestContext,
687 8773320 : ) -> (FullSlice<Buf>, anyhow::Result<()>)
688 8773320 : where
689 8773320 : Buf: IoBuf + Send,
690 8773320 : {
691 8773320 : self.inner
692 8773320 : .as_mut()
693 8773320 : .unwrap()
694 8773320 : .put_value_bytes(key, lsn, val, will_init, ctx)
695 8773320 : .await
696 8773320 : }
697 :
698 4047944 : pub fn size(&self) -> u64 {
699 4047944 : self.inner.as_ref().unwrap().size()
700 4047944 : }
701 :
702 : ///
703 : /// Finish writing the delta layer.
704 : ///
705 2884 : pub(crate) async fn finish(
706 2884 : mut self,
707 2884 : key_end: Key,
708 2884 : ctx: &RequestContext,
709 2884 : ) -> anyhow::Result<(PersistentLayerDesc, Utf8PathBuf)> {
710 2884 : self.inner.take().unwrap().finish(key_end, ctx).await
711 2884 : }
712 :
713 24508 : pub(crate) fn num_keys(&self) -> usize {
714 24508 : self.inner.as_ref().unwrap().num_keys
715 24508 : }
716 :
717 30280 : pub(crate) fn estimated_size(&self) -> u64 {
718 30280 : let inner = self.inner.as_ref().unwrap();
719 30280 : inner.blob_writer.size() + inner.tree.borrow_writer().size() + PAGE_SZ as u64
720 30280 : }
721 : }
722 :
723 : impl Drop for DeltaLayerWriter {
724 2936 : fn drop(&mut self) {
725 2936 : if let Some(inner) = self.inner.take() {
726 52 : // We want to remove the virtual file here, so it's fine to not
727 52 : // having completely flushed unwritten data.
728 52 : let vfile = inner.blob_writer.into_inner_no_flush();
729 52 : vfile.remove();
730 2884 : }
731 2936 : }
732 : }
733 :
734 : #[derive(thiserror::Error, Debug)]
735 : pub enum RewriteSummaryError {
736 : #[error("magic mismatch")]
737 : MagicMismatch,
738 : #[error(transparent)]
739 : Other(#[from] anyhow::Error),
740 : }
741 :
742 : impl From<std::io::Error> for RewriteSummaryError {
743 0 : fn from(e: std::io::Error) -> Self {
744 0 : Self::Other(anyhow::anyhow!(e))
745 0 : }
746 : }
747 :
748 : impl DeltaLayer {
749 0 : pub async fn rewrite_summary<F>(
750 0 : path: &Utf8Path,
751 0 : rewrite: F,
752 0 : ctx: &RequestContext,
753 0 : ) -> Result<(), RewriteSummaryError>
754 0 : where
755 0 : F: Fn(Summary) -> Summary,
756 0 : {
757 0 : let mut file = VirtualFile::open_with_options(
758 0 : path,
759 0 : virtual_file::OpenOptions::new().read(true).write(true),
760 0 : ctx,
761 0 : )
762 0 : .await
763 0 : .with_context(|| format!("Failed to open file '{}'", path))?;
764 0 : let file_id = page_cache::next_file_id();
765 0 : let block_reader = FileBlockReader::new(&file, file_id);
766 0 : let summary_blk = block_reader.read_blk(0, ctx).await?;
767 0 : let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
768 0 : if actual_summary.magic != DELTA_FILE_MAGIC {
769 0 : return Err(RewriteSummaryError::MagicMismatch);
770 0 : }
771 0 :
772 0 : let new_summary = rewrite(actual_summary);
773 0 :
774 0 : let mut buf = Vec::with_capacity(PAGE_SZ);
775 0 : // TODO: could use smallvec here, but it's a pain with Slice<T>
776 0 : Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
777 0 : file.seek(SeekFrom::Start(0)).await?;
778 0 : let (_buf, res) = file.write_all(buf.slice_len(), ctx).await;
779 0 : res?;
780 0 : Ok(())
781 0 : }
782 : }
783 :
784 : impl DeltaLayerInner {
785 2072 : pub(crate) fn key_range(&self) -> &Range<Key> {
786 2072 : &self.layer_key_range
787 2072 : }
788 :
789 2072 : pub(crate) fn lsn_range(&self) -> &Range<Lsn> {
790 2072 : &self.layer_lsn_range
791 2072 : }
792 :
793 2188 : pub(super) async fn load(
794 2188 : path: &Utf8Path,
795 2188 : summary: Option<Summary>,
796 2188 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
797 2188 : ctx: &RequestContext,
798 2188 : ) -> anyhow::Result<Self> {
799 2188 : let file = Arc::new(
800 2188 : VirtualFile::open_v2(path, ctx)
801 2188 : .await
802 2188 : .context("open layer file")?,
803 : );
804 :
805 2188 : let file_id = page_cache::next_file_id();
806 2188 :
807 2188 : let block_reader = FileBlockReader::new(&file, file_id);
808 :
809 2188 : let summary_blk = block_reader
810 2188 : .read_blk(0, ctx)
811 2188 : .await
812 2188 : .context("read first block")?;
813 :
814 : // TODO: this should be an assertion instead; see ImageLayerInner::load
815 2188 : let actual_summary =
816 2188 : Summary::des_prefix(summary_blk.as_ref()).context("deserialize first block")?;
817 :
818 2188 : if let Some(mut expected_summary) = summary {
819 : // production code path
820 2188 : expected_summary.index_start_blk = actual_summary.index_start_blk;
821 2188 : expected_summary.index_root_blk = actual_summary.index_root_blk;
822 2188 : // mask out the timeline_id, but still require the layers to be from the same tenant
823 2188 : expected_summary.timeline_id = actual_summary.timeline_id;
824 2188 :
825 2188 : if actual_summary != expected_summary {
826 0 : bail!(
827 0 : "in-file summary does not match expected summary. actual = {:?} expected = {:?}",
828 0 : actual_summary,
829 0 : expected_summary
830 0 : );
831 2188 : }
832 0 : }
833 :
834 2188 : Ok(DeltaLayerInner {
835 2188 : file,
836 2188 : file_id,
837 2188 : index_start_blk: actual_summary.index_start_blk,
838 2188 : index_root_blk: actual_summary.index_root_blk,
839 2188 : max_vectored_read_bytes,
840 2188 : layer_key_range: actual_summary.key_range,
841 2188 : layer_lsn_range: actual_summary.lsn_range,
842 2188 : })
843 2188 : }
844 :
845 : // Look up the keys in the provided keyspace and update
846 : // the reconstruct state with whatever is found.
847 : //
848 : // Currently, the index is visited for each range, but this
849 : // can be further optimised to visit the index only once.
850 434324 : pub(super) async fn get_values_reconstruct_data(
851 434324 : &self,
852 434324 : this: ResidentLayer,
853 434324 : keyspace: KeySpace,
854 434324 : lsn_range: Range<Lsn>,
855 434324 : reconstruct_state: &mut ValuesReconstructState,
856 434324 : ctx: &RequestContext,
857 434324 : ) -> Result<(), GetVectoredError> {
858 434324 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
859 434324 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
860 434324 : self.index_start_blk,
861 434324 : self.index_root_blk,
862 434324 : block_reader,
863 434324 : );
864 434324 :
865 434324 : let planner = VectoredReadPlanner::new(
866 434324 : self.max_vectored_read_bytes
867 434324 : .expect("Layer is loaded with max vectored bytes config")
868 434324 : .0
869 434324 : .into(),
870 434324 : );
871 434324 :
872 434324 : let data_end_offset = self.index_start_offset();
873 :
874 434324 : let reads = Self::plan_reads(
875 434324 : &keyspace,
876 434324 : lsn_range.clone(),
877 434324 : data_end_offset,
878 434324 : index_reader,
879 434324 : planner,
880 434324 : ctx,
881 434324 : )
882 434324 : .await
883 434324 : .map_err(GetVectoredError::Other)?;
884 :
885 434324 : self.do_reads_and_update_state(this, reads, reconstruct_state, ctx)
886 434324 : .await;
887 :
888 434324 : Ok(())
889 434324 : }
890 :
891 434728 : async fn plan_reads<Reader>(
892 434728 : keyspace: &KeySpace,
893 434728 : lsn_range: Range<Lsn>,
894 434728 : data_end_offset: u64,
895 434728 : index_reader: DiskBtreeReader<Reader, DELTA_KEY_SIZE>,
896 434728 : mut planner: VectoredReadPlanner,
897 434728 : ctx: &RequestContext,
898 434728 : ) -> anyhow::Result<Vec<VectoredRead>>
899 434728 : where
900 434728 : Reader: BlockReader + Clone,
901 434728 : {
902 434728 : let ctx = RequestContextBuilder::extend(ctx)
903 434728 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
904 434728 : .build();
905 :
906 435180 : for range in keyspace.ranges.iter() {
907 435180 : let mut range_end_handled = false;
908 435180 :
909 435180 : let start_key = DeltaKey::from_key_lsn(&range.start, lsn_range.start);
910 435180 : let index_stream = index_reader.clone().into_stream(&start_key.0, &ctx);
911 435180 : let mut index_stream = std::pin::pin!(index_stream);
912 :
913 1272171 : while let Some(index_entry) = index_stream.next().await {
914 1255527 : let (raw_key, value) = index_entry?;
915 1255527 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
916 1255527 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
917 1255527 : let blob_ref = BlobRef(value);
918 1255527 :
919 1255527 : // Lsns are not monotonically increasing across keys, so we don't assert on them.
920 1255527 : assert!(key >= range.start);
921 :
922 1255527 : let outside_lsn_range = !lsn_range.contains(&lsn);
923 :
924 1255527 : let flag = {
925 1255527 : if outside_lsn_range {
926 84632 : BlobFlag::Ignore
927 1170895 : } else if blob_ref.will_init() {
928 1053643 : BlobFlag::ReplaceAll
929 : } else {
930 : // Usual path: add blob to the read
931 117252 : BlobFlag::None
932 : }
933 : };
934 :
935 1255527 : if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
936 418536 : planner.handle_range_end(blob_ref.pos());
937 418536 : range_end_handled = true;
938 418536 : break;
939 836991 : } else {
940 836991 : planner.handle(key, lsn, blob_ref.pos(), flag);
941 836991 : }
942 : }
943 :
944 435180 : if !range_end_handled {
945 16644 : tracing::debug!("Handling range end fallback at {}", data_end_offset);
946 16644 : planner.handle_range_end(data_end_offset);
947 418536 : }
948 : }
949 :
950 434728 : Ok(planner.finish())
951 434728 : }
952 :
953 434724 : fn get_min_read_buffer_size(
954 434724 : planned_reads: &[VectoredRead],
955 434724 : read_size_soft_max: usize,
956 434724 : ) -> usize {
957 434724 : let Some(largest_read) = planned_reads.iter().max_by_key(|read| read.size()) else {
958 172929 : return read_size_soft_max;
959 : };
960 :
961 261795 : let largest_read_size = largest_read.size();
962 261795 : if largest_read_size > read_size_soft_max {
963 : // If the read is oversized, it should only contain one key.
964 400 : let offenders = largest_read
965 400 : .blobs_at
966 400 : .as_slice()
967 400 : .iter()
968 400 : .filter_map(|(_, blob_meta)| {
969 400 : if blob_meta.key.is_rel_dir_key()
970 400 : || blob_meta.key == DBDIR_KEY
971 400 : || blob_meta.key.is_aux_file_key()
972 : {
973 : // The size of values for these keys is unbounded and can
974 : // grow very large in pathological cases.
975 0 : None
976 : } else {
977 400 : Some(format!("{}@{}", blob_meta.key, blob_meta.lsn))
978 : }
979 400 : })
980 400 : .join(", ");
981 400 :
982 400 : if !offenders.is_empty() {
983 400 : tracing::warn!(
984 0 : "Oversized vectored read ({} > {}) for keys {}",
985 : largest_read_size,
986 : read_size_soft_max,
987 : offenders
988 : );
989 0 : }
990 261395 : }
991 :
992 261795 : largest_read_size
993 434724 : }
994 :
995 434324 : async fn do_reads_and_update_state(
996 434324 : &self,
997 434324 : this: ResidentLayer,
998 434324 : reads: Vec<VectoredRead>,
999 434324 : reconstruct_state: &mut ValuesReconstructState,
1000 434324 : ctx: &RequestContext,
1001 434324 : ) {
1002 434324 : let max_vectored_read_bytes = self
1003 434324 : .max_vectored_read_bytes
1004 434324 : .expect("Layer is loaded with max vectored bytes config")
1005 434324 : .0
1006 434324 : .into();
1007 434324 : let buf_size = Self::get_min_read_buffer_size(&reads, max_vectored_read_bytes);
1008 :
1009 : // Note that reads are processed in reverse order (from highest key+lsn).
1010 : // This is the order that `ReconstructState` requires such that it can
1011 : // track when a key is done.
1012 434324 : for read in reads.into_iter().rev() {
1013 267425 : let mut ios: HashMap<(Key, Lsn), OnDiskValueIo> = Default::default();
1014 374766 : for (_, blob_meta) in read.blobs_at.as_slice().iter().rev() {
1015 374766 : let io = reconstruct_state.update_key(
1016 374766 : &blob_meta.key,
1017 374766 : blob_meta.lsn,
1018 374766 : blob_meta.will_init,
1019 374766 : );
1020 374766 : ios.insert((blob_meta.key, blob_meta.lsn), io);
1021 374766 : }
1022 :
1023 267425 : let read_extend_residency = this.clone();
1024 267425 : let read_from = self.file.clone();
1025 267425 : let read_ctx = ctx.attached_child();
1026 267425 : reconstruct_state
1027 267425 : .spawn_io(async move {
1028 267425 : let vectored_blob_reader = VectoredBlobReader::new(&read_from);
1029 267425 : let buf = IoBufferMut::with_capacity(buf_size);
1030 :
1031 267425 : let res = vectored_blob_reader.read_blobs(&read, buf, &read_ctx).await;
1032 267425 : match res {
1033 267425 : Ok(blobs_buf) => {
1034 267425 : let view = BufView::new_slice(&blobs_buf.buf);
1035 374766 : for meta in blobs_buf.blobs.iter().rev() {
1036 374766 : let io = ios.remove(&(meta.meta.key, meta.meta.lsn)).unwrap();
1037 :
1038 374766 : let blob_read = meta.read(&view).await;
1039 374766 : let blob_read = match blob_read {
1040 374766 : Ok(buf) => buf,
1041 0 : Err(e) => {
1042 0 : io.complete(Err(e));
1043 0 : continue;
1044 : }
1045 : };
1046 :
1047 374766 : io.complete(Ok(OnDiskValue::WalRecordOrImage(
1048 374766 : blob_read.into_bytes(),
1049 374766 : )));
1050 : }
1051 :
1052 267425 : assert!(ios.is_empty());
1053 : }
1054 0 : Err(err) => {
1055 0 : for (_, sender) in ios {
1056 0 : sender.complete(Err(std::io::Error::new(
1057 0 : err.kind(),
1058 0 : "vec read failed",
1059 0 : )));
1060 0 : }
1061 : }
1062 : }
1063 :
1064 : // keep layer resident until this IO is done; this spawned IO future generally outlives the
1065 : // call to `self` / the `Arc<DownloadedLayer>` / the `ResidentLayer` that guarantees residency
1066 267425 : drop(read_extend_residency);
1067 267425 : })
1068 267425 : .await;
1069 : }
1070 434324 : }
1071 :
1072 812 : pub(crate) async fn index_entries<'a>(
1073 812 : &'a self,
1074 812 : ctx: &RequestContext,
1075 812 : ) -> Result<Vec<DeltaEntry<'a>>> {
1076 812 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1077 812 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1078 812 : self.index_start_blk,
1079 812 : self.index_root_blk,
1080 812 : block_reader,
1081 812 : );
1082 812 :
1083 812 : let mut all_keys: Vec<DeltaEntry<'_>> = Vec::new();
1084 812 :
1085 812 : tree_reader
1086 812 : .visit(
1087 812 : &[0u8; DELTA_KEY_SIZE],
1088 812 : VisitDirection::Forwards,
1089 4128092 : |key, value| {
1090 4128092 : let delta_key = DeltaKey::from_slice(key);
1091 4128092 : let val_ref = ValueRef {
1092 4128092 : blob_ref: BlobRef(value),
1093 4128092 : layer: self,
1094 4128092 : };
1095 4128092 : let pos = BlobRef(value).pos();
1096 4128092 : if let Some(last) = all_keys.last_mut() {
1097 4127280 : // subtract offset of the current and last entries to get the size
1098 4127280 : // of the value associated with this (key, lsn) tuple
1099 4127280 : let first_pos = last.size;
1100 4127280 : last.size = pos - first_pos;
1101 4127280 : }
1102 4128092 : let entry = DeltaEntry {
1103 4128092 : key: delta_key.key(),
1104 4128092 : lsn: delta_key.lsn(),
1105 4128092 : size: pos,
1106 4128092 : val: val_ref,
1107 4128092 : };
1108 4128092 : all_keys.push(entry);
1109 4128092 : true
1110 4128092 : },
1111 812 : &RequestContextBuilder::extend(ctx)
1112 812 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
1113 812 : .build(),
1114 812 : )
1115 812 : .await?;
1116 812 : if let Some(last) = all_keys.last_mut() {
1117 812 : // Last key occupies all space till end of value storage,
1118 812 : // which corresponds to beginning of the index
1119 812 : last.size = self.index_start_offset() - last.size;
1120 812 : }
1121 812 : Ok(all_keys)
1122 812 : }
1123 :
1124 : /// Using the given writer, write out a version which has the earlier Lsns than `until`.
1125 : ///
1126 : /// Return the amount of key value records pushed to the writer.
1127 20 : pub(super) async fn copy_prefix(
1128 20 : &self,
1129 20 : writer: &mut DeltaLayerWriter,
1130 20 : until: Lsn,
1131 20 : ctx: &RequestContext,
1132 20 : ) -> anyhow::Result<usize> {
1133 : use crate::tenant::vectored_blob_io::{
1134 : BlobMeta, ChunkedVectoredReadBuilder, VectoredReadExtended,
1135 : };
1136 : use futures::stream::TryStreamExt;
1137 :
1138 : #[derive(Debug)]
1139 : enum Item {
1140 : Actual(Key, Lsn, BlobRef),
1141 : Sentinel,
1142 : }
1143 :
1144 : impl From<Item> for Option<(Key, Lsn, BlobRef)> {
1145 140 : fn from(value: Item) -> Self {
1146 140 : match value {
1147 120 : Item::Actual(key, lsn, blob) => Some((key, lsn, blob)),
1148 20 : Item::Sentinel => None,
1149 : }
1150 140 : }
1151 : }
1152 :
1153 : impl Item {
1154 140 : fn offset(&self) -> Option<BlobRef> {
1155 140 : match self {
1156 120 : Item::Actual(_, _, blob) => Some(*blob),
1157 20 : Item::Sentinel => None,
1158 : }
1159 140 : }
1160 :
1161 140 : fn is_last(&self) -> bool {
1162 140 : matches!(self, Item::Sentinel)
1163 140 : }
1164 : }
1165 :
1166 20 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1167 20 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1168 20 : self.index_start_blk,
1169 20 : self.index_root_blk,
1170 20 : block_reader,
1171 20 : );
1172 20 :
1173 20 : let stream = self.stream_index_forwards(tree_reader, &[0u8; DELTA_KEY_SIZE], ctx);
1174 120 : let stream = stream.map_ok(|(key, lsn, pos)| Item::Actual(key, lsn, pos));
1175 20 : // put in a sentinel value for getting the end offset for last item, and not having to
1176 20 : // repeat the whole read part
1177 20 : let stream = stream.chain(futures::stream::once(futures::future::ready(Ok(
1178 20 : Item::Sentinel,
1179 20 : ))));
1180 20 : let mut stream = std::pin::pin!(stream);
1181 20 :
1182 20 : let mut prev: Option<(Key, Lsn, BlobRef)> = None;
1183 20 :
1184 20 : let mut read_builder: Option<ChunkedVectoredReadBuilder> = None;
1185 20 :
1186 20 : let max_read_size = self
1187 20 : .max_vectored_read_bytes
1188 20 : .map(|x| x.0.get())
1189 20 : .unwrap_or(8192);
1190 20 :
1191 20 : let mut buffer = Some(IoBufferMut::with_capacity(max_read_size));
1192 20 :
1193 20 : // FIXME: buffering of DeltaLayerWriter
1194 20 : let mut per_blob_copy = Vec::new();
1195 20 :
1196 20 : let mut records = 0;
1197 :
1198 160 : while let Some(item) = stream.try_next().await? {
1199 140 : tracing::debug!(?item, "popped");
1200 140 : let offset = item
1201 140 : .offset()
1202 140 : .unwrap_or(BlobRef::new(self.index_start_offset(), false));
1203 :
1204 140 : let actionable = if let Some((key, lsn, start_offset)) = prev.take() {
1205 120 : let end_offset = offset;
1206 120 :
1207 120 : Some((
1208 120 : BlobMeta {
1209 120 : key,
1210 120 : lsn,
1211 120 : will_init: false,
1212 120 : },
1213 120 : start_offset..end_offset,
1214 120 : ))
1215 : } else {
1216 20 : None
1217 : };
1218 :
1219 140 : let is_last = item.is_last();
1220 140 :
1221 140 : prev = Option::from(item);
1222 140 :
1223 140 : let actionable = actionable.filter(|x| x.0.lsn < until);
1224 :
1225 140 : let builder = if let Some((meta, offsets)) = actionable {
1226 : // extend or create a new builder
1227 64 : if read_builder
1228 64 : .as_mut()
1229 64 : .map(|x| x.extend(offsets.start.pos(), offsets.end.pos(), meta))
1230 64 : .unwrap_or(VectoredReadExtended::No)
1231 64 : == VectoredReadExtended::Yes
1232 : {
1233 32 : None
1234 : } else {
1235 32 : read_builder.replace(ChunkedVectoredReadBuilder::new(
1236 32 : offsets.start.pos(),
1237 32 : offsets.end.pos(),
1238 32 : meta,
1239 32 : max_read_size,
1240 32 : ))
1241 : }
1242 : } else {
1243 : // nothing to do, except perhaps flush any existing for the last element
1244 76 : None
1245 : };
1246 :
1247 : // flush the possible older builder and also the new one if the item was the last one
1248 140 : let builders = builder.into_iter();
1249 140 : let builders = if is_last {
1250 20 : builders.chain(read_builder.take())
1251 : } else {
1252 120 : builders.chain(None)
1253 : };
1254 :
1255 172 : for builder in builders {
1256 32 : let read = builder.build();
1257 32 :
1258 32 : let reader = VectoredBlobReader::new(&self.file);
1259 32 :
1260 32 : let mut buf = buffer.take().unwrap();
1261 32 :
1262 32 : buf.clear();
1263 32 : buf.reserve(read.size());
1264 32 : let res = reader.read_blobs(&read, buf, ctx).await?;
1265 :
1266 32 : let view = BufView::new_slice(&res.buf);
1267 :
1268 96 : for blob in res.blobs {
1269 64 : let key = blob.meta.key;
1270 64 : let lsn = blob.meta.lsn;
1271 :
1272 64 : let data = blob.read(&view).await?;
1273 :
1274 : #[cfg(debug_assertions)]
1275 64 : Value::des(&data)
1276 64 : .with_context(|| {
1277 0 : format!(
1278 0 : "blob failed to deserialize for {}: {:?}",
1279 0 : blob,
1280 0 : utils::Hex(&data)
1281 0 : )
1282 64 : })
1283 64 : .unwrap();
1284 64 :
1285 64 : // is it an image or will_init walrecord?
1286 64 : // FIXME: this could be handled by threading the BlobRef to the
1287 64 : // VectoredReadBuilder
1288 64 : let will_init = pageserver_api::value::ValueBytes::will_init(&data)
1289 64 : .inspect_err(|_e| {
1290 0 : #[cfg(feature = "testing")]
1291 0 : tracing::error!(data=?utils::Hex(&data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
1292 64 : })
1293 64 : .unwrap_or(false);
1294 64 :
1295 64 : per_blob_copy.clear();
1296 64 : per_blob_copy.extend_from_slice(&data);
1297 :
1298 64 : let (tmp, res) = writer
1299 64 : .put_value_bytes(
1300 64 : key,
1301 64 : lsn,
1302 64 : std::mem::take(&mut per_blob_copy).slice_len(),
1303 64 : will_init,
1304 64 : ctx,
1305 64 : )
1306 64 : .await;
1307 64 : per_blob_copy = tmp.into_raw_slice().into_inner();
1308 64 :
1309 64 : res?;
1310 :
1311 64 : records += 1;
1312 : }
1313 :
1314 32 : buffer = Some(res.buf);
1315 : }
1316 : }
1317 :
1318 20 : assert!(
1319 20 : read_builder.is_none(),
1320 0 : "with the sentinel above loop should had handled all"
1321 : );
1322 :
1323 20 : Ok(records)
1324 20 : }
1325 :
1326 8 : pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
1327 8 : println!(
1328 8 : "index_start_blk: {}, root {}",
1329 8 : self.index_start_blk, self.index_root_blk
1330 8 : );
1331 8 :
1332 8 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1333 8 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1334 8 : self.index_start_blk,
1335 8 : self.index_root_blk,
1336 8 : block_reader,
1337 8 : );
1338 8 :
1339 8 : tree_reader.dump().await?;
1340 :
1341 8 : let keys = self.index_entries(ctx).await?;
1342 :
1343 16 : async fn dump_blob(val: &ValueRef<'_>, ctx: &RequestContext) -> anyhow::Result<String> {
1344 16 : let buf = val.load_raw(ctx).await?;
1345 16 : let val = Value::des(&buf)?;
1346 16 : let desc = match val {
1347 16 : Value::Image(img) => {
1348 16 : format!(" img {} bytes", img.len())
1349 : }
1350 0 : Value::WalRecord(rec) => {
1351 0 : let wal_desc = pageserver_api::record::describe_wal_record(&rec)?;
1352 0 : format!(
1353 0 : " rec {} bytes will_init: {} {}",
1354 0 : buf.len(),
1355 0 : rec.will_init(),
1356 0 : wal_desc
1357 0 : )
1358 : }
1359 : };
1360 16 : Ok(desc)
1361 16 : }
1362 :
1363 24 : for entry in keys {
1364 16 : let DeltaEntry { key, lsn, val, .. } = entry;
1365 16 : let desc = match dump_blob(&val, ctx).await {
1366 16 : Ok(desc) => desc,
1367 0 : Err(err) => {
1368 0 : format!("ERROR: {err}")
1369 : }
1370 : };
1371 16 : println!(" key {key} at {lsn}: {desc}");
1372 :
1373 : // Print more details about CHECKPOINT records. Would be nice to print details
1374 : // of many other record types too, but these are particularly interesting, as
1375 : // have a lot of special processing for them in walingest.rs.
1376 8 : use pageserver_api::key::CHECKPOINT_KEY;
1377 8 : use postgres_ffi::CheckPoint;
1378 16 : if key == CHECKPOINT_KEY {
1379 0 : let val = val.load(ctx).await?;
1380 0 : match val {
1381 0 : Value::Image(img) => {
1382 0 : let checkpoint = CheckPoint::decode(&img)?;
1383 0 : println!(" CHECKPOINT: {:?}", checkpoint);
1384 : }
1385 0 : Value::WalRecord(_rec) => {
1386 0 : println!(" unexpected walrecord value for checkpoint key");
1387 0 : }
1388 : }
1389 16 : }
1390 : }
1391 :
1392 8 : Ok(())
1393 8 : }
1394 :
1395 60 : fn stream_index_forwards<'a, R>(
1396 60 : &'a self,
1397 60 : reader: DiskBtreeReader<R, DELTA_KEY_SIZE>,
1398 60 : start: &'a [u8; DELTA_KEY_SIZE],
1399 60 : ctx: &'a RequestContext,
1400 60 : ) -> impl futures::stream::Stream<
1401 60 : Item = Result<(Key, Lsn, BlobRef), crate::tenant::disk_btree::DiskBtreeError>,
1402 60 : > + 'a
1403 60 : where
1404 60 : R: BlockReader + 'a,
1405 60 : {
1406 : use futures::stream::TryStreamExt;
1407 60 : let stream = reader.into_stream(start, ctx);
1408 304 : stream.map_ok(|(key, value)| {
1409 304 : let key = DeltaKey::from_slice(&key);
1410 304 : let (key, lsn) = (key.key(), key.lsn());
1411 304 : let offset = BlobRef(value);
1412 304 :
1413 304 : (key, lsn, offset)
1414 304 : })
1415 60 : }
1416 :
1417 : /// The file offset to the first block of index.
1418 : ///
1419 : /// The file structure is summary, values, and index. We often need this for the size of last blob.
1420 436368 : fn index_start_offset(&self) -> u64 {
1421 436368 : let offset = self.index_start_blk as u64 * PAGE_SZ as u64;
1422 436368 : let bref = BlobRef(offset);
1423 436368 : tracing::debug!(
1424 : index_start_blk = self.index_start_blk,
1425 : offset,
1426 0 : pos = bref.pos(),
1427 0 : "index_start_offset"
1428 : );
1429 436368 : offset
1430 436368 : }
1431 :
1432 1148 : pub fn iter<'a>(&'a self, ctx: &'a RequestContext) -> DeltaLayerIterator<'a> {
1433 1148 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1434 1148 : let tree_reader =
1435 1148 : DiskBtreeReader::new(self.index_start_blk, self.index_root_blk, block_reader);
1436 1148 : DeltaLayerIterator {
1437 1148 : delta_layer: self,
1438 1148 : ctx,
1439 1148 : index_iter: tree_reader.iter(&[0; DELTA_KEY_SIZE], ctx),
1440 1148 : key_values_batch: std::collections::VecDeque::new(),
1441 1148 : is_end: false,
1442 1148 : planner: StreamingVectoredReadPlanner::new(
1443 1148 : 1024 * 8192, // The default value. Unit tests might use a different value. 1024 * 8K = 8MB buffer.
1444 1148 : 1024, // The default value. Unit tests might use a different value
1445 1148 : ),
1446 1148 : }
1447 1148 : }
1448 :
1449 : /// NB: not super efficient, but not terrible either. Should prob be an iterator.
1450 : //
1451 : // We're reusing the index traversal logical in plan_reads; would be nice to
1452 : // factor that out.
1453 0 : pub(crate) async fn load_keys(&self, ctx: &RequestContext) -> anyhow::Result<Vec<Key>> {
1454 0 : self.index_entries(ctx)
1455 0 : .await
1456 0 : .map(|entries| entries.into_iter().map(|entry| entry.key).collect())
1457 0 : }
1458 : }
1459 :
1460 : /// A set of data associated with a delta layer key and its value
1461 : pub struct DeltaEntry<'a> {
1462 : pub key: Key,
1463 : pub lsn: Lsn,
1464 : /// Size of the stored value
1465 : pub size: u64,
1466 : /// Reference to the on-disk value
1467 : pub val: ValueRef<'a>,
1468 : }
1469 :
1470 : /// Reference to an on-disk value
1471 : pub struct ValueRef<'a> {
1472 : blob_ref: BlobRef,
1473 : layer: &'a DeltaLayerInner,
1474 : }
1475 :
1476 : impl ValueRef<'_> {
1477 : /// Loads the value from disk
1478 0 : pub async fn load(&self, ctx: &RequestContext) -> Result<Value> {
1479 0 : let buf = self.load_raw(ctx).await?;
1480 0 : let val = Value::des(&buf)?;
1481 0 : Ok(val)
1482 0 : }
1483 :
1484 16 : async fn load_raw(&self, ctx: &RequestContext) -> Result<Vec<u8>> {
1485 16 : let reader = BlockCursor::new(crate::tenant::block_io::BlockReaderRef::Adapter(Adapter(
1486 16 : self.layer,
1487 16 : )));
1488 16 : let buf = reader.read_blob(self.blob_ref.pos(), ctx).await?;
1489 16 : Ok(buf)
1490 16 : }
1491 : }
1492 :
1493 : pub(crate) struct Adapter<T>(T);
1494 :
1495 : impl<T: AsRef<DeltaLayerInner>> Adapter<T> {
1496 16 : pub(crate) async fn read_blk(
1497 16 : &self,
1498 16 : blknum: u32,
1499 16 : ctx: &RequestContext,
1500 16 : ) -> Result<BlockLease, std::io::Error> {
1501 16 : let block_reader = FileBlockReader::new(&self.0.as_ref().file, self.0.as_ref().file_id);
1502 16 : block_reader.read_blk(blknum, ctx).await
1503 16 : }
1504 : }
1505 :
1506 : impl AsRef<DeltaLayerInner> for DeltaLayerInner {
1507 32 : fn as_ref(&self) -> &DeltaLayerInner {
1508 32 : self
1509 32 : }
1510 : }
1511 :
1512 : impl<'a> pageserver_compaction::interface::CompactionDeltaEntry<'a, Key> for DeltaEntry<'a> {
1513 0 : fn key(&self) -> Key {
1514 0 : self.key
1515 0 : }
1516 0 : fn lsn(&self) -> Lsn {
1517 0 : self.lsn
1518 0 : }
1519 0 : fn size(&self) -> u64 {
1520 0 : self.size
1521 0 : }
1522 : }
1523 :
1524 : pub struct DeltaLayerIterator<'a> {
1525 : delta_layer: &'a DeltaLayerInner,
1526 : ctx: &'a RequestContext,
1527 : planner: StreamingVectoredReadPlanner,
1528 : index_iter: DiskBtreeIterator<'a>,
1529 : key_values_batch: VecDeque<(Key, Lsn, Value)>,
1530 : is_end: bool,
1531 : }
1532 :
1533 : impl DeltaLayerIterator<'_> {
1534 0 : pub(crate) fn layer_dbg_info(&self) -> String {
1535 0 : self.delta_layer.layer_dbg_info()
1536 0 : }
1537 :
1538 : /// Retrieve a batch of key-value pairs into the iterator buffer.
1539 42808 : async fn next_batch(&mut self) -> anyhow::Result<()> {
1540 42808 : assert!(self.key_values_batch.is_empty());
1541 42808 : assert!(!self.is_end);
1542 :
1543 42808 : let plan = loop {
1544 4199244 : if let Some(res) = self.index_iter.next().await {
1545 4198152 : let (raw_key, value) = res?;
1546 4198152 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
1547 4198152 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
1548 4198152 : let blob_ref = BlobRef(value);
1549 4198152 : let offset = blob_ref.pos();
1550 41716 : if let Some(batch_plan) =
1551 4198152 : self.planner.handle(key, lsn, offset, blob_ref.will_init())
1552 : {
1553 41716 : break batch_plan;
1554 4156436 : }
1555 : } else {
1556 1092 : self.is_end = true;
1557 1092 : let data_end_offset = self.delta_layer.index_start_offset();
1558 1092 : if let Some(item) = self.planner.handle_range_end(data_end_offset) {
1559 1092 : break item;
1560 : } else {
1561 0 : return Ok(()); // TODO: test empty iterator
1562 : }
1563 : }
1564 : };
1565 42808 : let vectored_blob_reader = VectoredBlobReader::new(&self.delta_layer.file);
1566 42808 : let mut next_batch = std::collections::VecDeque::new();
1567 42808 : let buf_size = plan.size();
1568 42808 : let buf = IoBufferMut::with_capacity(buf_size);
1569 42808 : let blobs_buf = vectored_blob_reader
1570 42808 : .read_blobs(&plan, buf, self.ctx)
1571 42808 : .await?;
1572 42808 : let view = BufView::new_slice(&blobs_buf.buf);
1573 4198096 : for meta in blobs_buf.blobs.iter() {
1574 4198096 : let blob_read = meta.read(&view).await?;
1575 4198096 : let value = Value::des(&blob_read)?;
1576 :
1577 4198096 : next_batch.push_back((meta.meta.key, meta.meta.lsn, value));
1578 : }
1579 42808 : self.key_values_batch = next_batch;
1580 42808 : Ok(())
1581 42808 : }
1582 :
1583 4199668 : pub async fn next(&mut self) -> anyhow::Result<Option<(Key, Lsn, Value)>> {
1584 4199668 : if self.key_values_batch.is_empty() {
1585 44760 : if self.is_end {
1586 2120 : return Ok(None);
1587 42640 : }
1588 42640 : self.next_batch().await?;
1589 4154908 : }
1590 4197548 : Ok(Some(
1591 4197548 : self.key_values_batch
1592 4197548 : .pop_front()
1593 4197548 : .expect("should not be empty"),
1594 4197548 : ))
1595 4199668 : }
1596 : }
1597 :
1598 : #[cfg(test)]
1599 : pub(crate) mod test {
1600 : use std::collections::BTreeMap;
1601 :
1602 : use itertools::MinMaxResult;
1603 : use rand::prelude::{SeedableRng, SliceRandom, StdRng};
1604 : use rand::RngCore;
1605 :
1606 : use super::*;
1607 : use crate::tenant::harness::TIMELINE_ID;
1608 : use crate::tenant::storage_layer::{Layer, ResidentLayer};
1609 : use crate::tenant::vectored_blob_io::StreamingVectoredReadPlanner;
1610 : use crate::tenant::{Tenant, Timeline};
1611 : use crate::{
1612 : context::DownloadBehavior,
1613 : task_mgr::TaskKind,
1614 : tenant::{disk_btree::tests::TestDisk, harness::TenantHarness},
1615 : DEFAULT_PG_VERSION,
1616 : };
1617 : use bytes::Bytes;
1618 : use pageserver_api::value::Value;
1619 :
1620 : /// Construct an index for a fictional delta layer and and then
1621 : /// traverse in order to plan vectored reads for a query. Finally,
1622 : /// verify that the traversal fed the right index key and value
1623 : /// pairs into the planner.
1624 : #[tokio::test]
1625 4 : async fn test_delta_layer_index_traversal() {
1626 4 : let base_key = Key {
1627 4 : field1: 0,
1628 4 : field2: 1663,
1629 4 : field3: 12972,
1630 4 : field4: 16396,
1631 4 : field5: 0,
1632 4 : field6: 246080,
1633 4 : };
1634 4 :
1635 4 : // Populate the index with some entries
1636 4 : let entries: BTreeMap<Key, Vec<Lsn>> = BTreeMap::from([
1637 4 : (base_key, vec![Lsn(1), Lsn(5), Lsn(25), Lsn(26), Lsn(28)]),
1638 4 : (base_key.add(1), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1639 4 : (base_key.add(2), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1640 4 : (base_key.add(5), vec![Lsn(10), Lsn(15), Lsn(16), Lsn(20)]),
1641 4 : ]);
1642 4 :
1643 4 : let mut disk = TestDisk::default();
1644 4 : let mut writer = DiskBtreeBuilder::<_, DELTA_KEY_SIZE>::new(&mut disk);
1645 4 :
1646 4 : let mut disk_offset = 0;
1647 20 : for (key, lsns) in &entries {
1648 84 : for lsn in lsns {
1649 68 : let index_key = DeltaKey::from_key_lsn(key, *lsn);
1650 68 : let blob_ref = BlobRef::new(disk_offset, false);
1651 68 : writer
1652 68 : .append(&index_key.0, blob_ref.0)
1653 68 : .expect("In memory disk append should never fail");
1654 68 :
1655 68 : disk_offset += 1;
1656 68 : }
1657 4 : }
1658 4 :
1659 4 : // Prepare all the arguments for the call into `plan_reads` below
1660 4 : let (root_offset, _writer) = writer
1661 4 : .finish()
1662 4 : .expect("In memory disk finish should never fail");
1663 4 : let reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(0, root_offset, disk);
1664 4 : let planner = VectoredReadPlanner::new(100);
1665 4 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1666 4 :
1667 4 : let keyspace = KeySpace {
1668 4 : ranges: vec![
1669 4 : base_key..base_key.add(3),
1670 4 : base_key.add(3)..base_key.add(100),
1671 4 : ],
1672 4 : };
1673 4 : let lsn_range = Lsn(2)..Lsn(40);
1674 4 :
1675 4 : // Plan and validate
1676 4 : let vectored_reads = DeltaLayerInner::plan_reads(
1677 4 : &keyspace,
1678 4 : lsn_range.clone(),
1679 4 : disk_offset,
1680 4 : reader,
1681 4 : planner,
1682 4 : &ctx,
1683 4 : )
1684 4 : .await
1685 4 : .expect("Read planning should not fail");
1686 4 :
1687 4 : validate(keyspace, lsn_range, vectored_reads, entries);
1688 4 : }
1689 :
1690 4 : fn validate(
1691 4 : keyspace: KeySpace,
1692 4 : lsn_range: Range<Lsn>,
1693 4 : vectored_reads: Vec<VectoredRead>,
1694 4 : index_entries: BTreeMap<Key, Vec<Lsn>>,
1695 4 : ) {
1696 : #[derive(Debug, PartialEq, Eq)]
1697 : struct BlobSpec {
1698 : key: Key,
1699 : lsn: Lsn,
1700 : at: u64,
1701 : }
1702 :
1703 4 : let mut planned_blobs = Vec::new();
1704 60 : for read in vectored_reads {
1705 56 : for (at, meta) in read.blobs_at.as_slice() {
1706 56 : planned_blobs.push(BlobSpec {
1707 56 : key: meta.key,
1708 56 : lsn: meta.lsn,
1709 56 : at: *at,
1710 56 : });
1711 56 : }
1712 : }
1713 :
1714 4 : let mut expected_blobs = Vec::new();
1715 4 : let mut disk_offset = 0;
1716 20 : for (key, lsns) in index_entries {
1717 84 : for lsn in lsns {
1718 84 : let key_included = keyspace.ranges.iter().any(|range| range.contains(&key));
1719 68 : let lsn_included = lsn_range.contains(&lsn);
1720 68 :
1721 68 : if key_included && lsn_included {
1722 56 : expected_blobs.push(BlobSpec {
1723 56 : key,
1724 56 : lsn,
1725 56 : at: disk_offset,
1726 56 : });
1727 56 : }
1728 :
1729 68 : disk_offset += 1;
1730 : }
1731 : }
1732 :
1733 4 : assert_eq!(planned_blobs, expected_blobs);
1734 4 : }
1735 :
1736 : mod constants {
1737 : use utils::lsn::Lsn;
1738 :
1739 : /// Offset used by all lsns in this test
1740 : pub(super) const LSN_OFFSET: Lsn = Lsn(0x08);
1741 : /// Number of unique keys including in the test data
1742 : pub(super) const KEY_COUNT: u8 = 60;
1743 : /// Max number of different lsns for each key
1744 : pub(super) const MAX_ENTRIES_PER_KEY: u8 = 20;
1745 : /// Possible value sizes for each key along with a probability weight
1746 : pub(super) const VALUE_SIZES: [(usize, u8); 3] = [(100, 2), (1024, 2), (1024 * 1024, 1)];
1747 : /// Probability that there will be a gap between the current key and the next one (33.3%)
1748 : pub(super) const KEY_GAP_CHANGES: [(bool, u8); 2] = [(true, 1), (false, 2)];
1749 : /// The minimum size of a key range in all the generated reads
1750 : pub(super) const MIN_RANGE_SIZE: i128 = 10;
1751 : /// The number of ranges included in each vectored read
1752 : pub(super) const RANGES_COUNT: u8 = 2;
1753 : /// The number of vectored reads performed
1754 : pub(super) const READS_COUNT: u8 = 100;
1755 : /// Soft max size of a vectored read. Will be violated if we have to read keys
1756 : /// with values larger than the limit
1757 : pub(super) const MAX_VECTORED_READ_BYTES: usize = 64 * 1024;
1758 : }
1759 :
1760 : struct Entry {
1761 : key: Key,
1762 : lsn: Lsn,
1763 : value: Vec<u8>,
1764 : }
1765 :
1766 4 : fn generate_entries(rng: &mut StdRng) -> Vec<Entry> {
1767 4 : let mut current_key = Key::MIN;
1768 4 :
1769 4 : let mut entries = Vec::new();
1770 244 : for _ in 0..constants::KEY_COUNT {
1771 240 : let count = rng.gen_range(1..constants::MAX_ENTRIES_PER_KEY);
1772 240 : let mut lsns_iter =
1773 4520 : std::iter::successors(Some(Lsn(constants::LSN_OFFSET.0 + 0x08)), |lsn| {
1774 4520 : Some(Lsn(lsn.0 + 0x08))
1775 4520 : });
1776 240 : let mut lsns = Vec::new();
1777 4760 : while lsns.len() < count as usize {
1778 4520 : let take = rng.gen_bool(0.5);
1779 4520 : let lsn = lsns_iter.next().unwrap();
1780 4520 : if take {
1781 2224 : lsns.push(lsn);
1782 2296 : }
1783 : }
1784 :
1785 2464 : for lsn in lsns {
1786 2224 : let size = constants::VALUE_SIZES
1787 6672 : .choose_weighted(rng, |item| item.1)
1788 2224 : .unwrap()
1789 2224 : .0;
1790 2224 : let mut buf = vec![0; size];
1791 2224 : rng.fill_bytes(&mut buf);
1792 2224 :
1793 2224 : entries.push(Entry {
1794 2224 : key: current_key,
1795 2224 : lsn,
1796 2224 : value: buf,
1797 2224 : })
1798 : }
1799 :
1800 240 : let gap = constants::KEY_GAP_CHANGES
1801 480 : .choose_weighted(rng, |item| item.1)
1802 240 : .unwrap()
1803 240 : .0;
1804 240 : if gap {
1805 76 : current_key = current_key.add(2);
1806 164 : } else {
1807 164 : current_key = current_key.add(1);
1808 164 : }
1809 : }
1810 :
1811 4 : entries
1812 4 : }
1813 :
1814 : struct EntriesMeta {
1815 : key_range: Range<Key>,
1816 : lsn_range: Range<Lsn>,
1817 : index: BTreeMap<(Key, Lsn), Vec<u8>>,
1818 : }
1819 :
1820 4 : fn get_entries_meta(entries: &[Entry]) -> EntriesMeta {
1821 2224 : let key_range = match entries.iter().minmax_by_key(|e| e.key) {
1822 4 : MinMaxResult::MinMax(min, max) => min.key..max.key.next(),
1823 0 : _ => panic!("More than one entry is always expected"),
1824 : };
1825 :
1826 2224 : let lsn_range = match entries.iter().minmax_by_key(|e| e.lsn) {
1827 4 : MinMaxResult::MinMax(min, max) => min.lsn..Lsn(max.lsn.0 + 1),
1828 0 : _ => panic!("More than one entry is always expected"),
1829 : };
1830 :
1831 4 : let mut index = BTreeMap::new();
1832 2224 : for entry in entries.iter() {
1833 2224 : index.insert((entry.key, entry.lsn), entry.value.clone());
1834 2224 : }
1835 :
1836 4 : EntriesMeta {
1837 4 : key_range,
1838 4 : lsn_range,
1839 4 : index,
1840 4 : }
1841 4 : }
1842 :
1843 400 : fn pick_random_keyspace(rng: &mut StdRng, key_range: &Range<Key>) -> KeySpace {
1844 400 : let start = key_range.start.to_i128();
1845 400 : let end = key_range.end.to_i128();
1846 400 :
1847 400 : let mut keyspace = KeySpace::default();
1848 :
1849 1200 : for _ in 0..constants::RANGES_COUNT {
1850 800 : let mut range: Option<Range<Key>> = Option::default();
1851 2488 : while range.is_none() || keyspace.overlaps(range.as_ref().unwrap()) {
1852 1688 : let range_start = rng.gen_range(start..end);
1853 1688 : let range_end_offset = range_start + constants::MIN_RANGE_SIZE;
1854 1688 : if range_end_offset >= end {
1855 200 : range = Some(Key::from_i128(range_start)..Key::from_i128(end));
1856 1488 : } else {
1857 1488 : let range_end = rng.gen_range((range_start + constants::MIN_RANGE_SIZE)..end);
1858 1488 : range = Some(Key::from_i128(range_start)..Key::from_i128(range_end));
1859 1488 : }
1860 : }
1861 800 : keyspace.ranges.push(range.unwrap());
1862 : }
1863 :
1864 400 : keyspace
1865 400 : }
1866 :
1867 : #[tokio::test]
1868 4 : async fn test_delta_layer_vectored_read_end_to_end() -> anyhow::Result<()> {
1869 4 : let harness = TenantHarness::create("test_delta_layer_oversized_vectored_read").await?;
1870 4 : let (tenant, ctx) = harness.load().await;
1871 4 :
1872 4 : let timeline_id = TimelineId::generate();
1873 4 : let timeline = tenant
1874 4 : .create_test_timeline(timeline_id, constants::LSN_OFFSET, DEFAULT_PG_VERSION, &ctx)
1875 4 : .await?;
1876 4 :
1877 4 : tracing::info!("Generating test data ...");
1878 4 :
1879 4 : let rng = &mut StdRng::seed_from_u64(0);
1880 4 : let entries = generate_entries(rng);
1881 4 : let entries_meta = get_entries_meta(&entries);
1882 4 :
1883 4 : tracing::info!("Done generating {} entries", entries.len());
1884 4 :
1885 4 : tracing::info!("Writing test data to delta layer ...");
1886 4 : let mut writer = DeltaLayerWriter::new(
1887 4 : harness.conf,
1888 4 : timeline_id,
1889 4 : harness.tenant_shard_id,
1890 4 : entries_meta.key_range.start,
1891 4 : entries_meta.lsn_range.clone(),
1892 4 : &ctx,
1893 4 : )
1894 4 : .await?;
1895 4 :
1896 2228 : for entry in entries {
1897 2224 : let (_, res) = writer
1898 2224 : .put_value_bytes(entry.key, entry.lsn, entry.value.slice_len(), false, &ctx)
1899 2224 : .await;
1900 2224 : res?;
1901 4 : }
1902 4 :
1903 4 : let (desc, path) = writer.finish(entries_meta.key_range.end, &ctx).await?;
1904 4 : let resident = Layer::finish_creating(harness.conf, &timeline, desc, &path)?;
1905 4 :
1906 4 : let inner = resident.get_as_delta(&ctx).await?;
1907 4 :
1908 4 : let file_size = inner.file.metadata().await?.len();
1909 4 : tracing::info!(
1910 4 : "Done writing test data to delta layer. Resulting file size is: {}",
1911 4 : file_size
1912 4 : );
1913 4 :
1914 404 : for i in 0..constants::READS_COUNT {
1915 400 : tracing::info!("Doing vectored read {}/{}", i + 1, constants::READS_COUNT);
1916 4 :
1917 400 : let block_reader = FileBlockReader::new(&inner.file, inner.file_id);
1918 400 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1919 400 : inner.index_start_blk,
1920 400 : inner.index_root_blk,
1921 400 : block_reader,
1922 400 : );
1923 400 :
1924 400 : let planner = VectoredReadPlanner::new(constants::MAX_VECTORED_READ_BYTES);
1925 400 : let keyspace = pick_random_keyspace(rng, &entries_meta.key_range);
1926 400 : let data_end_offset = inner.index_start_blk as u64 * PAGE_SZ as u64;
1927 4 :
1928 400 : let vectored_reads = DeltaLayerInner::plan_reads(
1929 400 : &keyspace,
1930 400 : entries_meta.lsn_range.clone(),
1931 400 : data_end_offset,
1932 400 : index_reader,
1933 400 : planner,
1934 400 : &ctx,
1935 400 : )
1936 400 : .await?;
1937 4 :
1938 400 : let vectored_blob_reader = VectoredBlobReader::new(&inner.file);
1939 400 : let buf_size = DeltaLayerInner::get_min_read_buffer_size(
1940 400 : &vectored_reads,
1941 400 : constants::MAX_VECTORED_READ_BYTES,
1942 400 : );
1943 400 : let mut buf = Some(IoBufferMut::with_capacity(buf_size));
1944 4 :
1945 39848 : for read in vectored_reads {
1946 39448 : let blobs_buf = vectored_blob_reader
1947 39448 : .read_blobs(&read, buf.take().expect("Should have a buffer"), &ctx)
1948 39448 : .await?;
1949 39448 : let view = BufView::new_slice(&blobs_buf.buf);
1950 114608 : for meta in blobs_buf.blobs.iter() {
1951 114608 : let value = meta.read(&view).await?;
1952 114608 : assert_eq!(
1953 114608 : &value[..],
1954 114608 : &entries_meta.index[&(meta.meta.key, meta.meta.lsn)]
1955 114608 : );
1956 4 : }
1957 4 :
1958 39448 : buf = Some(blobs_buf.buf);
1959 4 : }
1960 4 : }
1961 4 :
1962 4 : Ok(())
1963 4 : }
1964 :
1965 : #[tokio::test]
1966 4 : async fn copy_delta_prefix_smoke() {
1967 4 : use bytes::Bytes;
1968 4 : use pageserver_api::record::NeonWalRecord;
1969 4 :
1970 4 : let h = crate::tenant::harness::TenantHarness::create("truncate_delta_smoke")
1971 4 : .await
1972 4 : .unwrap();
1973 4 : let (tenant, ctx) = h.load().await;
1974 4 : let ctx = &ctx;
1975 4 : let timeline = tenant
1976 4 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, ctx)
1977 4 : .await
1978 4 : .unwrap();
1979 4 :
1980 4 : let initdb_layer = timeline
1981 4 : .layers
1982 4 : .read()
1983 4 : .await
1984 4 : .likely_resident_layers()
1985 4 : .next()
1986 4 : .cloned()
1987 4 : .unwrap();
1988 4 :
1989 4 : {
1990 4 : let mut writer = timeline.writer().await;
1991 4 :
1992 4 : let data = [
1993 4 : (0x20, 12, Value::Image(Bytes::from_static(b"foobar"))),
1994 4 : (
1995 4 : 0x30,
1996 4 : 12,
1997 4 : Value::WalRecord(NeonWalRecord::Postgres {
1998 4 : will_init: false,
1999 4 : rec: Bytes::from_static(b"1"),
2000 4 : }),
2001 4 : ),
2002 4 : (
2003 4 : 0x40,
2004 4 : 12,
2005 4 : Value::WalRecord(NeonWalRecord::Postgres {
2006 4 : will_init: true,
2007 4 : rec: Bytes::from_static(b"2"),
2008 4 : }),
2009 4 : ),
2010 4 : // build an oversized value so we cannot extend and existing read over
2011 4 : // this
2012 4 : (
2013 4 : 0x50,
2014 4 : 12,
2015 4 : Value::WalRecord(NeonWalRecord::Postgres {
2016 4 : will_init: true,
2017 4 : rec: {
2018 4 : let mut buf =
2019 4 : vec![0u8; tenant.conf.max_vectored_read_bytes.0.get() + 1024];
2020 4 : buf.iter_mut()
2021 4 : .enumerate()
2022 536576 : .for_each(|(i, slot)| *slot = (i % 256) as u8);
2023 4 : Bytes::from(buf)
2024 4 : },
2025 4 : }),
2026 4 : ),
2027 4 : // because the oversized read cannot be extended further, we are sure to exercise the
2028 4 : // builder created on the last round with this:
2029 4 : (
2030 4 : 0x60,
2031 4 : 12,
2032 4 : Value::WalRecord(NeonWalRecord::Postgres {
2033 4 : will_init: true,
2034 4 : rec: Bytes::from_static(b"3"),
2035 4 : }),
2036 4 : ),
2037 4 : (
2038 4 : 0x60,
2039 4 : 9,
2040 4 : Value::Image(Bytes::from_static(b"something for a different key")),
2041 4 : ),
2042 4 : ];
2043 4 :
2044 4 : let mut last_lsn = None;
2045 4 :
2046 28 : for (lsn, key, value) in data {
2047 24 : let key = Key::from_i128(key);
2048 24 : writer.put(key, Lsn(lsn), &value, ctx).await.unwrap();
2049 24 : last_lsn = Some(lsn);
2050 4 : }
2051 4 :
2052 4 : writer.finish_write(Lsn(last_lsn.unwrap()));
2053 4 : }
2054 4 : timeline.freeze_and_flush().await.unwrap();
2055 4 :
2056 4 : let new_layer = timeline
2057 4 : .layers
2058 4 : .read()
2059 4 : .await
2060 4 : .likely_resident_layers()
2061 5 : .find(|&x| x != &initdb_layer)
2062 4 : .cloned()
2063 4 : .unwrap();
2064 4 :
2065 4 : // create a copy for the timeline, so we don't overwrite the file
2066 4 : let branch = tenant
2067 4 : .branch_timeline_test(&timeline, TimelineId::generate(), None, ctx)
2068 4 : .await
2069 4 : .unwrap();
2070 4 :
2071 4 : assert_eq!(branch.get_ancestor_lsn(), Lsn(0x60));
2072 4 :
2073 4 : // truncating at 0x61 gives us a full copy, otherwise just go backwards until there's just
2074 4 : // a single key
2075 4 :
2076 24 : for truncate_at in [0x61, 0x51, 0x41, 0x31, 0x21] {
2077 20 : let truncate_at = Lsn(truncate_at);
2078 4 :
2079 20 : let mut writer = DeltaLayerWriter::new(
2080 20 : tenant.conf,
2081 20 : branch.timeline_id,
2082 20 : tenant.tenant_shard_id,
2083 20 : Key::MIN,
2084 20 : Lsn(0x11)..truncate_at,
2085 20 : ctx,
2086 20 : )
2087 20 : .await
2088 20 : .unwrap();
2089 4 :
2090 20 : let new_layer = new_layer.download_and_keep_resident().await.unwrap();
2091 20 :
2092 20 : new_layer
2093 20 : .copy_delta_prefix(&mut writer, truncate_at, ctx)
2094 20 : .await
2095 20 : .unwrap();
2096 4 :
2097 20 : let (desc, path) = writer.finish(Key::MAX, ctx).await.unwrap();
2098 20 : let copied_layer = Layer::finish_creating(tenant.conf, &branch, desc, &path).unwrap();
2099 20 :
2100 20 : copied_layer.get_as_delta(ctx).await.unwrap();
2101 20 :
2102 20 : assert_keys_and_values_eq(
2103 20 : new_layer.get_as_delta(ctx).await.unwrap(),
2104 20 : copied_layer.get_as_delta(ctx).await.unwrap(),
2105 20 : truncate_at,
2106 20 : ctx,
2107 20 : )
2108 20 : .await;
2109 4 : }
2110 4 : }
2111 :
2112 20 : async fn assert_keys_and_values_eq(
2113 20 : source: &DeltaLayerInner,
2114 20 : truncated: &DeltaLayerInner,
2115 20 : truncated_at: Lsn,
2116 20 : ctx: &RequestContext,
2117 20 : ) {
2118 : use futures::future::ready;
2119 : use futures::stream::TryStreamExt;
2120 :
2121 20 : let start_key = [0u8; DELTA_KEY_SIZE];
2122 20 :
2123 20 : let source_reader = FileBlockReader::new(&source.file, source.file_id);
2124 20 : let source_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2125 20 : source.index_start_blk,
2126 20 : source.index_root_blk,
2127 20 : &source_reader,
2128 20 : );
2129 20 : let source_stream = source.stream_index_forwards(source_tree, &start_key, ctx);
2130 120 : let source_stream = source_stream.filter(|res| match res {
2131 120 : Ok((_, lsn, _)) => ready(lsn < &truncated_at),
2132 0 : _ => ready(true),
2133 120 : });
2134 20 : let mut source_stream = std::pin::pin!(source_stream);
2135 20 :
2136 20 : let truncated_reader = FileBlockReader::new(&truncated.file, truncated.file_id);
2137 20 : let truncated_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2138 20 : truncated.index_start_blk,
2139 20 : truncated.index_root_blk,
2140 20 : &truncated_reader,
2141 20 : );
2142 20 : let truncated_stream = truncated.stream_index_forwards(truncated_tree, &start_key, ctx);
2143 20 : let mut truncated_stream = std::pin::pin!(truncated_stream);
2144 20 :
2145 20 : let mut scratch_left = Vec::new();
2146 20 : let mut scratch_right = Vec::new();
2147 :
2148 : loop {
2149 84 : let (src, truncated) = (source_stream.try_next(), truncated_stream.try_next());
2150 84 : let (src, truncated) = tokio::try_join!(src, truncated).unwrap();
2151 84 :
2152 84 : if src.is_none() {
2153 20 : assert!(truncated.is_none());
2154 20 : break;
2155 64 : }
2156 64 :
2157 64 : let (src, truncated) = (src.unwrap(), truncated.unwrap());
2158 64 :
2159 64 : // because we've filtered the source with Lsn, we should always have the same keys from both.
2160 64 : assert_eq!(src.0, truncated.0);
2161 64 : assert_eq!(src.1, truncated.1);
2162 :
2163 : // if this is needed for something else, just drop this assert.
2164 64 : assert!(
2165 64 : src.2.pos() >= truncated.2.pos(),
2166 0 : "value position should not go backwards {} vs. {}",
2167 0 : src.2.pos(),
2168 0 : truncated.2.pos()
2169 : );
2170 :
2171 64 : scratch_left.clear();
2172 64 : let src_cursor = source_reader.block_cursor();
2173 64 : let left = src_cursor.read_blob_into_buf(src.2.pos(), &mut scratch_left, ctx);
2174 64 : scratch_right.clear();
2175 64 : let trunc_cursor = truncated_reader.block_cursor();
2176 64 : let right = trunc_cursor.read_blob_into_buf(truncated.2.pos(), &mut scratch_right, ctx);
2177 64 :
2178 64 : tokio::try_join!(left, right).unwrap();
2179 64 :
2180 64 : assert_eq!(utils::Hex(&scratch_left), utils::Hex(&scratch_right));
2181 : }
2182 20 : }
2183 :
2184 36448 : pub(crate) fn sort_delta(
2185 36448 : (k1, l1, _): &(Key, Lsn, Value),
2186 36448 : (k2, l2, _): &(Key, Lsn, Value),
2187 36448 : ) -> std::cmp::Ordering {
2188 36448 : (k1, l1).cmp(&(k2, l2))
2189 36448 : }
2190 :
2191 : #[cfg(feature = "testing")]
2192 188 : pub(crate) fn sort_delta_value(
2193 188 : (k1, l1, v1): &(Key, Lsn, Value),
2194 188 : (k2, l2, v2): &(Key, Lsn, Value),
2195 188 : ) -> std::cmp::Ordering {
2196 188 : let order_1 = if v1.is_image() { 0 } else { 1 };
2197 188 : let order_2 = if v2.is_image() { 0 } else { 1 };
2198 188 : (k1, l1, order_1).cmp(&(k2, l2, order_2))
2199 188 : }
2200 :
2201 44 : pub(crate) async fn produce_delta_layer(
2202 44 : tenant: &Tenant,
2203 44 : tline: &Arc<Timeline>,
2204 44 : mut deltas: Vec<(Key, Lsn, Value)>,
2205 44 : ctx: &RequestContext,
2206 44 : ) -> anyhow::Result<ResidentLayer> {
2207 44 : deltas.sort_by(sort_delta);
2208 44 : let (key_start, _, _) = deltas.first().unwrap();
2209 44 : let (key_max, _, _) = deltas.last().unwrap();
2210 16480 : let lsn_min = deltas.iter().map(|(_, lsn, _)| lsn).min().unwrap();
2211 16480 : let lsn_max = deltas.iter().map(|(_, lsn, _)| lsn).max().unwrap();
2212 44 : let lsn_end = Lsn(lsn_max.0 + 1);
2213 44 : let mut writer = DeltaLayerWriter::new(
2214 44 : tenant.conf,
2215 44 : tline.timeline_id,
2216 44 : tenant.tenant_shard_id,
2217 44 : *key_start,
2218 44 : (*lsn_min)..lsn_end,
2219 44 : ctx,
2220 44 : )
2221 44 : .await?;
2222 44 : let key_end = key_max.next();
2223 :
2224 16524 : for (key, lsn, value) in deltas {
2225 16480 : writer.put_value(key, lsn, value, ctx).await?;
2226 : }
2227 :
2228 44 : let (desc, path) = writer.finish(key_end, ctx).await?;
2229 44 : let delta_layer = Layer::finish_creating(tenant.conf, tline, desc, &path)?;
2230 :
2231 44 : Ok::<_, anyhow::Error>(delta_layer)
2232 44 : }
2233 :
2234 56 : async fn assert_delta_iter_equal(
2235 56 : delta_iter: &mut DeltaLayerIterator<'_>,
2236 56 : expect: &[(Key, Lsn, Value)],
2237 56 : ) {
2238 56 : let mut expect_iter = expect.iter();
2239 : loop {
2240 56056 : let o1 = delta_iter.next().await.unwrap();
2241 56056 : let o2 = expect_iter.next();
2242 56056 : assert_eq!(o1.is_some(), o2.is_some());
2243 56056 : if o1.is_none() && o2.is_none() {
2244 56 : break;
2245 56000 : }
2246 56000 : let (k1, l1, v1) = o1.unwrap();
2247 56000 : let (k2, l2, v2) = o2.unwrap();
2248 56000 : assert_eq!(&k1, k2);
2249 56000 : assert_eq!(l1, *l2);
2250 56000 : assert_eq!(&v1, v2);
2251 : }
2252 56 : }
2253 :
2254 : #[tokio::test]
2255 4 : async fn delta_layer_iterator() {
2256 4 : let harness = TenantHarness::create("delta_layer_iterator").await.unwrap();
2257 4 : let (tenant, ctx) = harness.load().await;
2258 4 :
2259 4 : let tline = tenant
2260 4 : .create_test_timeline(TIMELINE_ID, Lsn(0x10), DEFAULT_PG_VERSION, &ctx)
2261 4 : .await
2262 4 : .unwrap();
2263 4 :
2264 4000 : fn get_key(id: u32) -> Key {
2265 4000 : let mut key = Key::from_hex("000000000033333333444444445500000000").unwrap();
2266 4000 : key.field6 = id;
2267 4000 : key
2268 4000 : }
2269 4 : const N: usize = 1000;
2270 4 : let test_deltas = (0..N)
2271 4000 : .map(|idx| {
2272 4000 : (
2273 4000 : get_key(idx as u32 / 10),
2274 4000 : Lsn(0x10 * ((idx as u64) % 10 + 1)),
2275 4000 : Value::Image(Bytes::from(format!("img{idx:05}"))),
2276 4000 : )
2277 4000 : })
2278 4 : .collect_vec();
2279 4 : let resident_layer = produce_delta_layer(&tenant, &tline, test_deltas.clone(), &ctx)
2280 4 : .await
2281 4 : .unwrap();
2282 4 : let delta_layer = resident_layer.get_as_delta(&ctx).await.unwrap();
2283 12 : for max_read_size in [1, 1024] {
2284 64 : for batch_size in [1, 2, 4, 8, 3, 7, 13] {
2285 56 : println!("running with batch_size={batch_size} max_read_size={max_read_size}");
2286 56 : // Test if the batch size is correctly determined
2287 56 : let mut iter = delta_layer.iter(&ctx);
2288 56 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2289 56 : let mut num_items = 0;
2290 224 : for _ in 0..3 {
2291 168 : iter.next_batch().await.unwrap();
2292 168 : num_items += iter.key_values_batch.len();
2293 168 : if max_read_size == 1 {
2294 4 : // every key should be a batch b/c the value is larger than max_read_size
2295 84 : assert_eq!(iter.key_values_batch.len(), 1);
2296 4 : } else {
2297 84 : assert!(iter.key_values_batch.len() <= batch_size);
2298 4 : }
2299 168 : if num_items >= N {
2300 4 : break;
2301 168 : }
2302 168 : iter.key_values_batch.clear();
2303 4 : }
2304 4 : // Test if the result is correct
2305 56 : let mut iter = delta_layer.iter(&ctx);
2306 56 : iter.planner = StreamingVectoredReadPlanner::new(max_read_size, batch_size);
2307 56 : assert_delta_iter_equal(&mut iter, &test_deltas).await;
2308 4 : }
2309 4 : }
2310 4 : }
2311 : }
|
