Line data Source code
1 : //! A DeltaLayer represents a collection of WAL records or page images in a range of
2 : //! LSNs, and in a range of Keys. It is stored on a file on disk.
3 : //!
4 : //! Usually a delta layer only contains differences, in the form of WAL records
5 : //! against a base LSN. However, if a relation extended or a whole new relation
6 : //! is created, there would be no base for the new pages. The entries for them
7 : //! must be page images or WAL records with the 'will_init' flag set, so that
8 : //! they can be replayed without referring to an older page version.
9 : //!
10 : //! The delta files are stored in `timelines/<timeline_id>` directory. Currently,
11 : //! there are no subdirectories, and each delta file is named like this:
12 : //!
13 : //! ```text
14 : //! <key start>-<key end>__<start LSN>-<end LSN>
15 : //! ```
16 : //!
17 : //! For example:
18 : //!
19 : //! ```text
20 : //! 000000067F000032BE0000400000000020B6-000000067F000032BE0000400000000030B6__000000578C6B29-0000000057A50051
21 : //! ```
22 : //!
23 : //! Every delta file consists of three parts: "summary", "values", and
24 : //! "index". The summary is a fixed size header at the beginning of the file,
25 : //! and it contains basic information about the layer, and offsets to the other
26 : //! parts. The "index" is a B-tree, mapping from Key and LSN to an offset in the
27 : //! "values" part. The actual page images and WAL records are stored in the
28 : //! "values" part.
29 : //!
30 : use crate::config::PageServerConf;
31 : use crate::context::{PageContentKind, RequestContext, RequestContextBuilder};
32 : use crate::page_cache::{self, FileId, PAGE_SZ};
33 : use crate::repository::{Key, Value, KEY_SIZE};
34 : use crate::tenant::blob_io::BlobWriter;
35 : use crate::tenant::block_io::{BlockBuf, BlockCursor, BlockLease, BlockReader, FileBlockReader};
36 : use crate::tenant::disk_btree::{DiskBtreeBuilder, DiskBtreeReader, VisitDirection};
37 : use crate::tenant::storage_layer::{Layer, ValueReconstructResult, ValueReconstructState};
38 : use crate::tenant::timeline::GetVectoredError;
39 : use crate::tenant::vectored_blob_io::{
40 : BlobFlag, MaxVectoredReadBytes, VectoredBlobReader, VectoredRead, VectoredReadPlanner,
41 : };
42 : use crate::tenant::{PageReconstructError, Timeline};
43 : use crate::virtual_file::{self, VirtualFile};
44 : use crate::{walrecord, TEMP_FILE_SUFFIX};
45 : use crate::{DELTA_FILE_MAGIC, STORAGE_FORMAT_VERSION};
46 : use anyhow::{anyhow, bail, ensure, Context, Result};
47 : use bytes::BytesMut;
48 : use camino::{Utf8Path, Utf8PathBuf};
49 : use futures::StreamExt;
50 : use itertools::Itertools;
51 : use pageserver_api::keyspace::KeySpace;
52 : use pageserver_api::models::LayerAccessKind;
53 : use pageserver_api::shard::TenantShardId;
54 : use rand::{distributions::Alphanumeric, Rng};
55 : use serde::{Deserialize, Serialize};
56 : use std::fs::File;
57 : use std::io::SeekFrom;
58 : use std::ops::Range;
59 : use std::os::unix::fs::FileExt;
60 : use std::str::FromStr;
61 : use std::sync::Arc;
62 : use tokio::sync::OnceCell;
63 : use tracing::*;
64 :
65 : use utils::{
66 : bin_ser::BeSer,
67 : id::{TenantId, TimelineId},
68 : lsn::Lsn,
69 : };
70 :
71 : use super::{
72 : AsLayerDesc, LayerAccessStats, LayerName, PersistentLayerDesc, ResidentLayer,
73 : ValuesReconstructState,
74 : };
75 :
76 : ///
77 : /// Header stored in the beginning of the file
78 : ///
79 : /// After this comes the 'values' part, starting on block 1. After that,
80 : /// the 'index' starts at the block indicated by 'index_start_blk'
81 : ///
82 986 : #[derive(Debug, Serialize, Deserialize, PartialEq, Eq)]
83 : pub struct Summary {
84 : /// Magic value to identify this as a neon delta file. Always DELTA_FILE_MAGIC.
85 : pub magic: u16,
86 : pub format_version: u16,
87 :
88 : pub tenant_id: TenantId,
89 : pub timeline_id: TimelineId,
90 : pub key_range: Range<Key>,
91 : pub lsn_range: Range<Lsn>,
92 :
93 : /// Block number where the 'index' part of the file begins.
94 : pub index_start_blk: u32,
95 : /// Block within the 'index', where the B-tree root page is stored
96 : pub index_root_blk: u32,
97 : }
98 :
99 : impl From<&DeltaLayer> for Summary {
100 0 : fn from(layer: &DeltaLayer) -> Self {
101 0 : Self::expected(
102 0 : layer.desc.tenant_shard_id.tenant_id,
103 0 : layer.desc.timeline_id,
104 0 : layer.desc.key_range.clone(),
105 0 : layer.desc.lsn_range.clone(),
106 0 : )
107 0 : }
108 : }
109 :
110 : impl Summary {
111 986 : pub(super) fn expected(
112 986 : tenant_id: TenantId,
113 986 : timeline_id: TimelineId,
114 986 : keys: Range<Key>,
115 986 : lsns: Range<Lsn>,
116 986 : ) -> Self {
117 986 : Self {
118 986 : magic: DELTA_FILE_MAGIC,
119 986 : format_version: STORAGE_FORMAT_VERSION,
120 986 :
121 986 : tenant_id,
122 986 : timeline_id,
123 986 : key_range: keys,
124 986 : lsn_range: lsns,
125 986 :
126 986 : index_start_blk: 0,
127 986 : index_root_blk: 0,
128 986 : }
129 986 : }
130 : }
131 :
132 : // Flag indicating that this version initialize the page
133 : const WILL_INIT: u64 = 1;
134 :
135 : /// Struct representing reference to BLOB in layers. Reference contains BLOB
136 : /// offset, and for WAL records it also contains `will_init` flag. The flag
137 : /// helps to determine the range of records that needs to be applied, without
138 : /// reading/deserializing records themselves.
139 0 : #[derive(Debug, Serialize, Deserialize, Copy, Clone)]
140 : pub struct BlobRef(pub u64);
141 :
142 : impl BlobRef {
143 240506 : pub fn will_init(&self) -> bool {
144 240506 : (self.0 & WILL_INIT) != 0
145 240506 : }
146 :
147 4439450 : pub fn pos(&self) -> u64 {
148 4439450 : self.0 >> 1
149 4439450 : }
150 :
151 6450860 : pub fn new(pos: u64, will_init: bool) -> BlobRef {
152 6450860 : let mut blob_ref = pos << 1;
153 6450860 : if will_init {
154 6449624 : blob_ref |= WILL_INIT;
155 6449624 : }
156 6450860 : BlobRef(blob_ref)
157 6450860 : }
158 : }
159 :
160 : pub const DELTA_KEY_SIZE: usize = KEY_SIZE + 8;
161 : struct DeltaKey([u8; DELTA_KEY_SIZE]);
162 :
163 : /// This is the key of the B-tree index stored in the delta layer. It consists
164 : /// of the serialized representation of a Key and LSN.
165 : impl DeltaKey {
166 2064208 : fn from_slice(buf: &[u8]) -> Self {
167 2064208 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
168 2064208 : bytes.copy_from_slice(buf);
169 2064208 : DeltaKey(bytes)
170 2064208 : }
171 :
172 6696623 : fn from_key_lsn(key: &Key, lsn: Lsn) -> Self {
173 6696623 : let mut bytes: [u8; DELTA_KEY_SIZE] = [0u8; DELTA_KEY_SIZE];
174 6696623 : key.write_to_byte_slice(&mut bytes[0..KEY_SIZE]);
175 6696623 : bytes[KEY_SIZE..].copy_from_slice(&u64::to_be_bytes(lsn.0));
176 6696623 : DeltaKey(bytes)
177 6696623 : }
178 :
179 2064208 : fn key(&self) -> Key {
180 2064208 : Key::from_slice(&self.0)
181 2064208 : }
182 :
183 2064208 : fn lsn(&self) -> Lsn {
184 2064208 : Lsn(u64::from_be_bytes(self.0[KEY_SIZE..].try_into().unwrap()))
185 2064208 : }
186 :
187 311182 : fn extract_lsn_from_buf(buf: &[u8]) -> Lsn {
188 311182 : let mut lsn_buf = [0u8; 8];
189 311182 : lsn_buf.copy_from_slice(&buf[KEY_SIZE..]);
190 311182 : Lsn(u64::from_be_bytes(lsn_buf))
191 311182 : }
192 : }
193 :
194 : /// This is used only from `pagectl`. Within pageserver, all layers are
195 : /// [`crate::tenant::storage_layer::Layer`], which can hold a [`DeltaLayerInner`].
196 : pub struct DeltaLayer {
197 : path: Utf8PathBuf,
198 : pub desc: PersistentLayerDesc,
199 : access_stats: LayerAccessStats,
200 : inner: OnceCell<Arc<DeltaLayerInner>>,
201 : }
202 :
203 : impl std::fmt::Debug for DeltaLayer {
204 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
205 0 : use super::RangeDisplayDebug;
206 0 :
207 0 : f.debug_struct("DeltaLayer")
208 0 : .field("key_range", &RangeDisplayDebug(&self.desc.key_range))
209 0 : .field("lsn_range", &self.desc.lsn_range)
210 0 : .field("file_size", &self.desc.file_size)
211 0 : .field("inner", &self.inner)
212 0 : .finish()
213 0 : }
214 : }
215 :
216 : /// `DeltaLayerInner` is the in-memory data structure associated with an on-disk delta
217 : /// file.
218 : pub struct DeltaLayerInner {
219 : // values copied from summary
220 : index_start_blk: u32,
221 : index_root_blk: u32,
222 :
223 : file: VirtualFile,
224 : file_id: FileId,
225 :
226 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
227 : }
228 :
229 : impl std::fmt::Debug for DeltaLayerInner {
230 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
231 0 : f.debug_struct("DeltaLayerInner")
232 0 : .field("index_start_blk", &self.index_start_blk)
233 0 : .field("index_root_blk", &self.index_root_blk)
234 0 : .finish()
235 0 : }
236 : }
237 :
238 : /// Boilerplate to implement the Layer trait, always use layer_desc for persistent layers.
239 : impl std::fmt::Display for DeltaLayer {
240 0 : fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
241 0 : write!(f, "{}", self.layer_desc().short_id())
242 0 : }
243 : }
244 :
245 : impl AsLayerDesc for DeltaLayer {
246 0 : fn layer_desc(&self) -> &PersistentLayerDesc {
247 0 : &self.desc
248 0 : }
249 : }
250 :
251 : impl DeltaLayer {
252 0 : pub(crate) async fn dump(&self, verbose: bool, ctx: &RequestContext) -> Result<()> {
253 0 : self.desc.dump();
254 0 :
255 0 : if !verbose {
256 0 : return Ok(());
257 0 : }
258 :
259 0 : let inner = self.load(LayerAccessKind::Dump, ctx).await?;
260 :
261 0 : inner.dump(ctx).await
262 0 : }
263 :
264 1318 : fn temp_path_for(
265 1318 : conf: &PageServerConf,
266 1318 : tenant_shard_id: &TenantShardId,
267 1318 : timeline_id: &TimelineId,
268 1318 : key_start: Key,
269 1318 : lsn_range: &Range<Lsn>,
270 1318 : ) -> Utf8PathBuf {
271 1318 : let rand_string: String = rand::thread_rng()
272 1318 : .sample_iter(&Alphanumeric)
273 1318 : .take(8)
274 1318 : .map(char::from)
275 1318 : .collect();
276 1318 :
277 1318 : conf.timeline_path(tenant_shard_id, timeline_id)
278 1318 : .join(format!(
279 1318 : "{}-XXX__{:016X}-{:016X}.{}.{}",
280 1318 : key_start,
281 1318 : u64::from(lsn_range.start),
282 1318 : u64::from(lsn_range.end),
283 1318 : rand_string,
284 1318 : TEMP_FILE_SUFFIX,
285 1318 : ))
286 1318 : }
287 :
288 : ///
289 : /// Open the underlying file and read the metadata into memory, if it's
290 : /// not loaded already.
291 : ///
292 0 : async fn load(
293 0 : &self,
294 0 : access_kind: LayerAccessKind,
295 0 : ctx: &RequestContext,
296 0 : ) -> Result<&Arc<DeltaLayerInner>> {
297 0 : self.access_stats.record_access(access_kind, ctx);
298 0 : // Quick exit if already loaded
299 0 : self.inner
300 0 : .get_or_try_init(|| self.load_inner(ctx))
301 0 : .await
302 0 : .with_context(|| format!("Failed to load delta layer {}", self.path()))
303 0 : }
304 :
305 0 : async fn load_inner(&self, ctx: &RequestContext) -> Result<Arc<DeltaLayerInner>> {
306 0 : let path = self.path();
307 :
308 0 : let loaded = DeltaLayerInner::load(&path, None, None, ctx)
309 0 : .await
310 0 : .and_then(|res| res)?;
311 :
312 : // not production code
313 0 : let actual_layer_name = LayerName::from_str(path.file_name().unwrap()).unwrap();
314 0 : let expected_layer_name = self.layer_desc().layer_name();
315 0 :
316 0 : if actual_layer_name != expected_layer_name {
317 0 : println!("warning: filename does not match what is expected from in-file summary");
318 0 : println!("actual: {:?}", actual_layer_name.to_string());
319 0 : println!("expected: {:?}", expected_layer_name.to_string());
320 0 : }
321 :
322 0 : Ok(Arc::new(loaded))
323 0 : }
324 :
325 : /// Create a DeltaLayer struct representing an existing file on disk.
326 : ///
327 : /// This variant is only used for debugging purposes, by the 'pagectl' binary.
328 0 : pub fn new_for_path(path: &Utf8Path, file: File) -> Result<Self> {
329 0 : let mut summary_buf = vec![0; PAGE_SZ];
330 0 : file.read_exact_at(&mut summary_buf, 0)?;
331 0 : let summary = Summary::des_prefix(&summary_buf)?;
332 :
333 0 : let metadata = file
334 0 : .metadata()
335 0 : .context("get file metadata to determine size")?;
336 :
337 : // This function is never used for constructing layers in a running pageserver,
338 : // so it does not need an accurate TenantShardId.
339 0 : let tenant_shard_id = TenantShardId::unsharded(summary.tenant_id);
340 0 :
341 0 : Ok(DeltaLayer {
342 0 : path: path.to_path_buf(),
343 0 : desc: PersistentLayerDesc::new_delta(
344 0 : tenant_shard_id,
345 0 : summary.timeline_id,
346 0 : summary.key_range,
347 0 : summary.lsn_range,
348 0 : metadata.len(),
349 0 : ),
350 0 : access_stats: LayerAccessStats::empty_will_record_residence_event_later(),
351 0 : inner: OnceCell::new(),
352 0 : })
353 0 : }
354 :
355 : /// Path to the layer file in pageserver workdir.
356 0 : fn path(&self) -> Utf8PathBuf {
357 0 : self.path.clone()
358 0 : }
359 : }
360 :
361 : /// A builder object for constructing a new delta layer.
362 : ///
363 : /// Usage:
364 : ///
365 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
366 : ///
367 : /// 2. Write the contents by calling `put_value` for every page
368 : /// version to store in the layer.
369 : ///
370 : /// 3. Call `finish`.
371 : ///
372 : struct DeltaLayerWriterInner {
373 : conf: &'static PageServerConf,
374 : pub path: Utf8PathBuf,
375 : timeline_id: TimelineId,
376 : tenant_shard_id: TenantShardId,
377 :
378 : key_start: Key,
379 : lsn_range: Range<Lsn>,
380 :
381 : tree: DiskBtreeBuilder<BlockBuf, DELTA_KEY_SIZE>,
382 :
383 : blob_writer: BlobWriter<true>,
384 : }
385 :
386 : impl DeltaLayerWriterInner {
387 : ///
388 : /// Start building a new delta layer.
389 : ///
390 1318 : async fn new(
391 1318 : conf: &'static PageServerConf,
392 1318 : timeline_id: TimelineId,
393 1318 : tenant_shard_id: TenantShardId,
394 1318 : key_start: Key,
395 1318 : lsn_range: Range<Lsn>,
396 1318 : ctx: &RequestContext,
397 1318 : ) -> anyhow::Result<Self> {
398 1318 : // Create the file initially with a temporary filename. We don't know
399 1318 : // the end key yet, so we cannot form the final filename yet. We will
400 1318 : // rename it when we're done.
401 1318 : //
402 1318 : // Note: This overwrites any existing file. There shouldn't be any.
403 1318 : // FIXME: throw an error instead?
404 1318 : let path =
405 1318 : DeltaLayer::temp_path_for(conf, &tenant_shard_id, &timeline_id, key_start, &lsn_range);
406 :
407 1318 : let mut file = VirtualFile::create(&path, ctx).await?;
408 : // make room for the header block
409 1318 : file.seek(SeekFrom::Start(PAGE_SZ as u64)).await?;
410 1318 : let blob_writer = BlobWriter::new(file, PAGE_SZ as u64);
411 1318 :
412 1318 : // Initialize the b-tree index builder
413 1318 : let block_buf = BlockBuf::new();
414 1318 : let tree_builder = DiskBtreeBuilder::new(block_buf);
415 1318 :
416 1318 : Ok(Self {
417 1318 : conf,
418 1318 : path,
419 1318 : timeline_id,
420 1318 : tenant_shard_id,
421 1318 : key_start,
422 1318 : lsn_range,
423 1318 : tree: tree_builder,
424 1318 : blob_writer,
425 1318 : })
426 1318 : }
427 :
428 : ///
429 : /// Append a key-value pair to the file.
430 : ///
431 : /// The values must be appended in key, lsn order.
432 : ///
433 2064094 : async fn put_value(
434 2064094 : &mut self,
435 2064094 : key: Key,
436 2064094 : lsn: Lsn,
437 2064094 : val: Value,
438 2064094 : ctx: &RequestContext,
439 2064094 : ) -> anyhow::Result<()> {
440 2064094 : let (_, res) = self
441 2064094 : .put_value_bytes(key, lsn, Value::ser(&val)?, val.will_init(), ctx)
442 1569 : .await;
443 2064094 : res
444 2064094 : }
445 :
446 6450756 : async fn put_value_bytes(
447 6450756 : &mut self,
448 6450756 : key: Key,
449 6450756 : lsn: Lsn,
450 6450756 : val: Vec<u8>,
451 6450756 : will_init: bool,
452 6450756 : ctx: &RequestContext,
453 6450756 : ) -> (Vec<u8>, anyhow::Result<()>) {
454 6450756 : assert!(self.lsn_range.start <= lsn);
455 6450756 : let (val, res) = self.blob_writer.write_blob(val, ctx).await;
456 6450756 : let off = match res {
457 6450756 : Ok(off) => off,
458 0 : Err(e) => return (val, Err(anyhow::anyhow!(e))),
459 : };
460 :
461 6450756 : let blob_ref = BlobRef::new(off, will_init);
462 6450756 :
463 6450756 : let delta_key = DeltaKey::from_key_lsn(&key, lsn);
464 6450756 : let res = self.tree.append(&delta_key.0, blob_ref.0);
465 6450756 : (val, res.map_err(|e| anyhow::anyhow!(e)))
466 6450756 : }
467 :
468 2023972 : fn size(&self) -> u64 {
469 2023972 : self.blob_writer.size() + self.tree.borrow_writer().size()
470 2023972 : }
471 :
472 : ///
473 : /// Finish writing the delta layer.
474 : ///
475 1318 : async fn finish(
476 1318 : self,
477 1318 : key_end: Key,
478 1318 : timeline: &Arc<Timeline>,
479 1318 : ctx: &RequestContext,
480 1318 : ) -> anyhow::Result<ResidentLayer> {
481 1318 : let temp_path = self.path.clone();
482 9462 : let result = self.finish0(key_end, timeline, ctx).await;
483 1318 : if result.is_err() {
484 0 : tracing::info!(%temp_path, "cleaning up temporary file after error during writing");
485 0 : if let Err(e) = std::fs::remove_file(&temp_path) {
486 0 : tracing::warn!(error=%e, %temp_path, "error cleaning up temporary layer file after error during writing");
487 0 : }
488 1318 : }
489 1318 : result
490 1318 : }
491 :
492 1318 : async fn finish0(
493 1318 : self,
494 1318 : key_end: Key,
495 1318 : timeline: &Arc<Timeline>,
496 1318 : ctx: &RequestContext,
497 1318 : ) -> anyhow::Result<ResidentLayer> {
498 1318 : let index_start_blk =
499 1318 : ((self.blob_writer.size() + PAGE_SZ as u64 - 1) / PAGE_SZ as u64) as u32;
500 :
501 1318 : let mut file = self.blob_writer.into_inner(ctx).await?;
502 :
503 : // Write out the index
504 1318 : let (index_root_blk, block_buf) = self.tree.finish()?;
505 1318 : file.seek(SeekFrom::Start(index_start_blk as u64 * PAGE_SZ as u64))
506 0 : .await?;
507 14830 : for buf in block_buf.blocks {
508 13512 : let (_buf, res) = file.write_all(buf, ctx).await;
509 13512 : res?;
510 : }
511 1318 : assert!(self.lsn_range.start < self.lsn_range.end);
512 : // Fill in the summary on blk 0
513 1318 : let summary = Summary {
514 1318 : magic: DELTA_FILE_MAGIC,
515 1318 : format_version: STORAGE_FORMAT_VERSION,
516 1318 : tenant_id: self.tenant_shard_id.tenant_id,
517 1318 : timeline_id: self.timeline_id,
518 1318 : key_range: self.key_start..key_end,
519 1318 : lsn_range: self.lsn_range.clone(),
520 1318 : index_start_blk,
521 1318 : index_root_blk,
522 1318 : };
523 1318 :
524 1318 : let mut buf = Vec::with_capacity(PAGE_SZ);
525 1318 : // TODO: could use smallvec here but it's a pain with Slice<T>
526 1318 : Summary::ser_into(&summary, &mut buf)?;
527 1318 : file.seek(SeekFrom::Start(0)).await?;
528 1318 : let (_buf, res) = file.write_all(buf, ctx).await;
529 1318 : res?;
530 :
531 1318 : let metadata = file
532 1318 : .metadata()
533 663 : .await
534 1318 : .context("get file metadata to determine size")?;
535 :
536 : // 5GB limit for objects without multipart upload (which we don't want to use)
537 : // Make it a little bit below to account for differing GB units
538 : // https://docs.aws.amazon.com/AmazonS3/latest/userguide/upload-objects.html
539 : const S3_UPLOAD_LIMIT: u64 = 4_500_000_000;
540 1318 : ensure!(
541 1318 : metadata.len() <= S3_UPLOAD_LIMIT,
542 0 : "Created delta layer file at {} of size {} above limit {S3_UPLOAD_LIMIT}!",
543 0 : file.path,
544 0 : metadata.len()
545 : );
546 :
547 : // Note: Because we opened the file in write-only mode, we cannot
548 : // reuse the same VirtualFile for reading later. That's why we don't
549 : // set inner.file here. The first read will have to re-open it.
550 :
551 1318 : let desc = PersistentLayerDesc::new_delta(
552 1318 : self.tenant_shard_id,
553 1318 : self.timeline_id,
554 1318 : self.key_start..key_end,
555 1318 : self.lsn_range.clone(),
556 1318 : metadata.len(),
557 1318 : );
558 1318 :
559 1318 : // fsync the file
560 1318 : file.sync_all().await?;
561 :
562 1318 : let layer = Layer::finish_creating(self.conf, timeline, desc, &self.path)?;
563 :
564 1318 : trace!("created delta layer {}", layer.local_path());
565 :
566 1318 : Ok(layer)
567 1318 : }
568 : }
569 :
570 : /// A builder object for constructing a new delta layer.
571 : ///
572 : /// Usage:
573 : ///
574 : /// 1. Create the DeltaLayerWriter by calling DeltaLayerWriter::new(...)
575 : ///
576 : /// 2. Write the contents by calling `put_value` for every page
577 : /// version to store in the layer.
578 : ///
579 : /// 3. Call `finish`.
580 : ///
581 : /// # Note
582 : ///
583 : /// As described in <https://github.com/neondatabase/neon/issues/2650>, it's
584 : /// possible for the writer to drop before `finish` is actually called. So this
585 : /// could lead to odd temporary files in the directory, exhausting file system.
586 : /// This structure wraps `DeltaLayerWriterInner` and also contains `Drop`
587 : /// implementation that cleans up the temporary file in failure. It's not
588 : /// possible to do this directly in `DeltaLayerWriterInner` since `finish` moves
589 : /// out some fields, making it impossible to implement `Drop`.
590 : ///
591 : #[must_use]
592 : pub struct DeltaLayerWriter {
593 : inner: Option<DeltaLayerWriterInner>,
594 : }
595 :
596 : impl DeltaLayerWriter {
597 : ///
598 : /// Start building a new delta layer.
599 : ///
600 1318 : pub async fn new(
601 1318 : conf: &'static PageServerConf,
602 1318 : timeline_id: TimelineId,
603 1318 : tenant_shard_id: TenantShardId,
604 1318 : key_start: Key,
605 1318 : lsn_range: Range<Lsn>,
606 1318 : ctx: &RequestContext,
607 1318 : ) -> anyhow::Result<Self> {
608 1318 : Ok(Self {
609 1318 : inner: Some(
610 1318 : DeltaLayerWriterInner::new(
611 1318 : conf,
612 1318 : timeline_id,
613 1318 : tenant_shard_id,
614 1318 : key_start,
615 1318 : lsn_range,
616 1318 : ctx,
617 1318 : )
618 669 : .await?,
619 : ),
620 : })
621 1318 : }
622 :
623 : ///
624 : /// Append a key-value pair to the file.
625 : ///
626 : /// The values must be appended in key, lsn order.
627 : ///
628 2064094 : pub async fn put_value(
629 2064094 : &mut self,
630 2064094 : key: Key,
631 2064094 : lsn: Lsn,
632 2064094 : val: Value,
633 2064094 : ctx: &RequestContext,
634 2064094 : ) -> anyhow::Result<()> {
635 2064094 : self.inner
636 2064094 : .as_mut()
637 2064094 : .unwrap()
638 2064094 : .put_value(key, lsn, val, ctx)
639 1569 : .await
640 2064094 : }
641 :
642 4386662 : pub async fn put_value_bytes(
643 4386662 : &mut self,
644 4386662 : key: Key,
645 4386662 : lsn: Lsn,
646 4386662 : val: Vec<u8>,
647 4386662 : will_init: bool,
648 4386662 : ctx: &RequestContext,
649 4386662 : ) -> (Vec<u8>, anyhow::Result<()>) {
650 4386662 : self.inner
651 4386662 : .as_mut()
652 4386662 : .unwrap()
653 4386662 : .put_value_bytes(key, lsn, val, will_init, ctx)
654 2950 : .await
655 4386662 : }
656 :
657 2023972 : pub fn size(&self) -> u64 {
658 2023972 : self.inner.as_ref().unwrap().size()
659 2023972 : }
660 :
661 : ///
662 : /// Finish writing the delta layer.
663 : ///
664 1318 : pub(crate) async fn finish(
665 1318 : mut self,
666 1318 : key_end: Key,
667 1318 : timeline: &Arc<Timeline>,
668 1318 : ctx: &RequestContext,
669 1318 : ) -> anyhow::Result<ResidentLayer> {
670 1318 : self.inner
671 1318 : .take()
672 1318 : .unwrap()
673 1318 : .finish(key_end, timeline, ctx)
674 9462 : .await
675 1318 : }
676 : }
677 :
678 : impl Drop for DeltaLayerWriter {
679 1318 : fn drop(&mut self) {
680 1318 : if let Some(inner) = self.inner.take() {
681 0 : // We want to remove the virtual file here, so it's fine to not
682 0 : // having completely flushed unwritten data.
683 0 : let vfile = inner.blob_writer.into_inner_no_flush();
684 0 : vfile.remove();
685 1318 : }
686 1318 : }
687 : }
688 :
689 0 : #[derive(thiserror::Error, Debug)]
690 : pub enum RewriteSummaryError {
691 : #[error("magic mismatch")]
692 : MagicMismatch,
693 : #[error(transparent)]
694 : Other(#[from] anyhow::Error),
695 : }
696 :
697 : impl From<std::io::Error> for RewriteSummaryError {
698 0 : fn from(e: std::io::Error) -> Self {
699 0 : Self::Other(anyhow::anyhow!(e))
700 0 : }
701 : }
702 :
703 : impl DeltaLayer {
704 0 : pub async fn rewrite_summary<F>(
705 0 : path: &Utf8Path,
706 0 : rewrite: F,
707 0 : ctx: &RequestContext,
708 0 : ) -> Result<(), RewriteSummaryError>
709 0 : where
710 0 : F: Fn(Summary) -> Summary,
711 0 : {
712 0 : let mut file = VirtualFile::open_with_options(
713 0 : path,
714 0 : virtual_file::OpenOptions::new().read(true).write(true),
715 0 : ctx,
716 0 : )
717 0 : .await
718 0 : .with_context(|| format!("Failed to open file '{}'", path))?;
719 0 : let file_id = page_cache::next_file_id();
720 0 : let block_reader = FileBlockReader::new(&file, file_id);
721 0 : let summary_blk = block_reader.read_blk(0, ctx).await?;
722 0 : let actual_summary = Summary::des_prefix(summary_blk.as_ref()).context("deserialize")?;
723 0 : if actual_summary.magic != DELTA_FILE_MAGIC {
724 0 : return Err(RewriteSummaryError::MagicMismatch);
725 0 : }
726 0 :
727 0 : let new_summary = rewrite(actual_summary);
728 0 :
729 0 : let mut buf = Vec::with_capacity(PAGE_SZ);
730 0 : // TODO: could use smallvec here, but it's a pain with Slice<T>
731 0 : Summary::ser_into(&new_summary, &mut buf).context("serialize")?;
732 0 : file.seek(SeekFrom::Start(0)).await?;
733 0 : let (_buf, res) = file.write_all(buf, ctx).await;
734 0 : res?;
735 0 : Ok(())
736 0 : }
737 : }
738 :
739 : impl DeltaLayerInner {
740 : /// Returns nested result following Result<Result<_, OpErr>, Critical>:
741 : /// - inner has the success or transient failure
742 : /// - outer has the permanent failure
743 986 : pub(super) async fn load(
744 986 : path: &Utf8Path,
745 986 : summary: Option<Summary>,
746 986 : max_vectored_read_bytes: Option<MaxVectoredReadBytes>,
747 986 : ctx: &RequestContext,
748 986 : ) -> Result<Result<Self, anyhow::Error>, anyhow::Error> {
749 986 : let file = match VirtualFile::open(path, ctx).await {
750 986 : Ok(file) => file,
751 0 : Err(e) => return Ok(Err(anyhow::Error::new(e).context("open layer file"))),
752 : };
753 986 : let file_id = page_cache::next_file_id();
754 986 :
755 986 : let block_reader = FileBlockReader::new(&file, file_id);
756 :
757 986 : let summary_blk = match block_reader.read_blk(0, ctx).await {
758 986 : Ok(blk) => blk,
759 0 : Err(e) => return Ok(Err(anyhow::Error::new(e).context("read first block"))),
760 : };
761 :
762 : // TODO: this should be an assertion instead; see ImageLayerInner::load
763 986 : let actual_summary =
764 986 : Summary::des_prefix(summary_blk.as_ref()).context("deserialize first block")?;
765 :
766 986 : if let Some(mut expected_summary) = summary {
767 : // production code path
768 986 : expected_summary.index_start_blk = actual_summary.index_start_blk;
769 986 : expected_summary.index_root_blk = actual_summary.index_root_blk;
770 986 : // mask out the timeline_id, but still require the layers to be from the same tenant
771 986 : expected_summary.timeline_id = actual_summary.timeline_id;
772 986 :
773 986 : if actual_summary != expected_summary {
774 0 : bail!(
775 0 : "in-file summary does not match expected summary. actual = {:?} expected = {:?}",
776 0 : actual_summary,
777 0 : expected_summary
778 0 : );
779 986 : }
780 0 : }
781 :
782 986 : Ok(Ok(DeltaLayerInner {
783 986 : file,
784 986 : file_id,
785 986 : index_start_blk: actual_summary.index_start_blk,
786 986 : index_root_blk: actual_summary.index_root_blk,
787 986 : max_vectored_read_bytes,
788 986 : }))
789 986 : }
790 :
791 203210 : pub(super) async fn get_value_reconstruct_data(
792 203210 : &self,
793 203210 : key: Key,
794 203210 : lsn_range: Range<Lsn>,
795 203210 : reconstruct_state: &mut ValueReconstructState,
796 203210 : ctx: &RequestContext,
797 203210 : ) -> anyhow::Result<ValueReconstructResult> {
798 203210 : let mut need_image = true;
799 203210 : // Scan the page versions backwards, starting from `lsn`.
800 203210 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
801 203210 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
802 203210 : self.index_start_blk,
803 203210 : self.index_root_blk,
804 203210 : &block_reader,
805 203210 : );
806 203210 : let search_key = DeltaKey::from_key_lsn(&key, Lsn(lsn_range.end.0 - 1));
807 203210 :
808 203210 : let mut offsets: Vec<(Lsn, u64)> = Vec::new();
809 203210 :
810 203210 : tree_reader
811 203210 : .visit(
812 203210 : &search_key.0,
813 203210 : VisitDirection::Backwards,
814 203210 : |key, value| {
815 197168 : let blob_ref = BlobRef(value);
816 197168 : if key[..KEY_SIZE] != search_key.0[..KEY_SIZE] {
817 75437 : return false;
818 121731 : }
819 121731 : let entry_lsn = DeltaKey::extract_lsn_from_buf(key);
820 121731 : if entry_lsn < lsn_range.start {
821 38 : return false;
822 121693 : }
823 121693 : offsets.push((entry_lsn, blob_ref.pos()));
824 121693 :
825 121693 : !blob_ref.will_init()
826 203210 : },
827 203210 : &RequestContextBuilder::extend(ctx)
828 203210 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
829 203210 : .build(),
830 203210 : )
831 20958 : .await?;
832 :
833 203210 : let ctx = &RequestContextBuilder::extend(ctx)
834 203210 : .page_content_kind(PageContentKind::DeltaLayerValue)
835 203210 : .build();
836 203210 :
837 203210 : // Ok, 'offsets' now contains the offsets of all the entries we need to read
838 203210 : let cursor = block_reader.block_cursor();
839 203210 : let mut buf = Vec::new();
840 203218 : for (entry_lsn, pos) in offsets {
841 121693 : cursor
842 121693 : .read_blob_into_buf(pos, &mut buf, ctx)
843 7983 : .await
844 121693 : .with_context(|| {
845 0 : format!("Failed to read blob from virtual file {}", self.file.path)
846 121693 : })?;
847 121693 : let val = Value::des(&buf).with_context(|| {
848 0 : format!(
849 0 : "Failed to deserialize file blob from virtual file {}",
850 0 : self.file.path
851 0 : )
852 121693 : })?;
853 121693 : match val {
854 121685 : Value::Image(img) => {
855 121685 : reconstruct_state.img = Some((entry_lsn, img));
856 121685 : need_image = false;
857 121685 : break;
858 : }
859 8 : Value::WalRecord(rec) => {
860 8 : let will_init = rec.will_init();
861 8 : reconstruct_state.records.push((entry_lsn, rec));
862 8 : if will_init {
863 : // This WAL record initializes the page, so no need to go further back
864 0 : need_image = false;
865 0 : break;
866 8 : }
867 : }
868 : }
869 : }
870 :
871 : // If an older page image is needed to reconstruct the page, let the
872 : // caller know.
873 203210 : if need_image {
874 81525 : Ok(ValueReconstructResult::Continue)
875 : } else {
876 121685 : Ok(ValueReconstructResult::Complete)
877 : }
878 203210 : }
879 :
880 : // Look up the keys in the provided keyspace and update
881 : // the reconstruct state with whatever is found.
882 : //
883 : // If the key is cached, go no further than the cached Lsn.
884 : //
885 : // Currently, the index is visited for each range, but this
886 : // can be further optimised to visit the index only once.
887 154 : pub(super) async fn get_values_reconstruct_data(
888 154 : &self,
889 154 : keyspace: KeySpace,
890 154 : lsn_range: Range<Lsn>,
891 154 : reconstruct_state: &mut ValuesReconstructState,
892 154 : ctx: &RequestContext,
893 154 : ) -> Result<(), GetVectoredError> {
894 154 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
895 154 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
896 154 : self.index_start_blk,
897 154 : self.index_root_blk,
898 154 : block_reader,
899 154 : );
900 154 :
901 154 : let planner = VectoredReadPlanner::new(
902 154 : self.max_vectored_read_bytes
903 154 : .expect("Layer is loaded with max vectored bytes config")
904 154 : .0
905 154 : .into(),
906 154 : );
907 154 :
908 154 : let data_end_offset = self.index_start_offset();
909 :
910 154 : let reads = Self::plan_reads(
911 154 : &keyspace,
912 154 : lsn_range.clone(),
913 154 : data_end_offset,
914 154 : index_reader,
915 154 : planner,
916 154 : reconstruct_state,
917 154 : ctx,
918 154 : )
919 1422 : .await
920 154 : .map_err(GetVectoredError::Other)?;
921 :
922 154 : self.do_reads_and_update_state(reads, reconstruct_state, ctx)
923 8752 : .await;
924 :
925 154 : reconstruct_state.on_lsn_advanced(&keyspace, lsn_range.start);
926 154 :
927 154 : Ok(())
928 154 : }
929 :
930 : /// Load all key-values in the delta layer, should be replaced by an iterator-based interface in the future.
931 : #[cfg(test)]
932 4 : pub(super) async fn load_key_values(
933 4 : &self,
934 4 : ctx: &RequestContext,
935 4 : ) -> anyhow::Result<Vec<(Key, Lsn, Value)>> {
936 4 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
937 4 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
938 4 : self.index_start_blk,
939 4 : self.index_root_blk,
940 4 : block_reader,
941 4 : );
942 4 : let mut result = Vec::new();
943 4 : let mut stream =
944 4 : Box::pin(self.stream_index_forwards(&index_reader, &[0; DELTA_KEY_SIZE], ctx));
945 4 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
946 4 : let cursor = block_reader.block_cursor();
947 4 : let mut buf = Vec::new();
948 14 : while let Some(item) = stream.next().await {
949 10 : let (key, lsn, pos) = item?;
950 : // TODO: dedup code with get_reconstruct_value
951 : // TODO: ctx handling and sharding
952 10 : cursor
953 10 : .read_blob_into_buf(pos.pos(), &mut buf, ctx)
954 2 : .await
955 10 : .with_context(|| {
956 0 : format!("Failed to read blob from virtual file {}", self.file.path)
957 10 : })?;
958 10 : let val = Value::des(&buf).with_context(|| {
959 0 : format!(
960 0 : "Failed to deserialize file blob from virtual file {}",
961 0 : self.file.path
962 0 : )
963 10 : })?;
964 10 : result.push((key, lsn, val));
965 : }
966 4 : Ok(result)
967 4 : }
968 :
969 356 : async fn plan_reads<Reader>(
970 356 : keyspace: &KeySpace,
971 356 : lsn_range: Range<Lsn>,
972 356 : data_end_offset: u64,
973 356 : index_reader: DiskBtreeReader<Reader, DELTA_KEY_SIZE>,
974 356 : mut planner: VectoredReadPlanner,
975 356 : reconstruct_state: &mut ValuesReconstructState,
976 356 : ctx: &RequestContext,
977 356 : ) -> anyhow::Result<Vec<VectoredRead>>
978 356 : where
979 356 : Reader: BlockReader,
980 356 : {
981 356 : let ctx = RequestContextBuilder::extend(ctx)
982 356 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
983 356 : .build();
984 :
985 42623 : for range in keyspace.ranges.iter() {
986 42623 : let mut range_end_handled = false;
987 42623 :
988 42623 : let start_key = DeltaKey::from_key_lsn(&range.start, lsn_range.start);
989 42623 : let index_stream = index_reader.get_stream_from(&start_key.0, &ctx);
990 42623 : let mut index_stream = std::pin::pin!(index_stream);
991 :
992 189665 : while let Some(index_entry) = index_stream.next().await {
993 189451 : let (raw_key, value) = index_entry?;
994 189451 : let key = Key::from_slice(&raw_key[..KEY_SIZE]);
995 189451 : let lsn = DeltaKey::extract_lsn_from_buf(&raw_key);
996 189451 : let blob_ref = BlobRef(value);
997 189451 :
998 189451 : // Lsns are not monotonically increasing across keys, so we don't assert on them.
999 189451 : assert!(key >= range.start);
1000 :
1001 189451 : let outside_lsn_range = !lsn_range.contains(&lsn);
1002 189451 : let below_cached_lsn = reconstruct_state.get_cached_lsn(&key) >= Some(lsn);
1003 :
1004 189451 : let flag = {
1005 189451 : if outside_lsn_range || below_cached_lsn {
1006 70638 : BlobFlag::Ignore
1007 118813 : } else if blob_ref.will_init() {
1008 61165 : BlobFlag::ReplaceAll
1009 : } else {
1010 : // Usual path: add blob to the read
1011 57648 : BlobFlag::None
1012 : }
1013 : };
1014 :
1015 189451 : if key >= range.end || (key.next() == range.end && lsn >= lsn_range.end) {
1016 42409 : planner.handle_range_end(blob_ref.pos());
1017 42409 : range_end_handled = true;
1018 42409 : break;
1019 147042 : } else {
1020 147042 : planner.handle(key, lsn, blob_ref.pos(), flag);
1021 147042 : }
1022 : }
1023 :
1024 42623 : if !range_end_handled {
1025 214 : tracing::debug!("Handling range end fallback at {}", data_end_offset);
1026 214 : planner.handle_range_end(data_end_offset);
1027 42409 : }
1028 : }
1029 :
1030 356 : Ok(planner.finish())
1031 356 : }
1032 :
1033 354 : fn get_min_read_buffer_size(
1034 354 : planned_reads: &[VectoredRead],
1035 354 : read_size_soft_max: usize,
1036 354 : ) -> usize {
1037 36911 : let Some(largest_read) = planned_reads.iter().max_by_key(|read| read.size()) else {
1038 14 : return read_size_soft_max;
1039 : };
1040 :
1041 340 : let largest_read_size = largest_read.size();
1042 340 : if largest_read_size > read_size_soft_max {
1043 : // If the read is oversized, it should only contain one key.
1044 200 : let offenders = largest_read
1045 200 : .blobs_at
1046 200 : .as_slice()
1047 200 : .iter()
1048 200 : .map(|(_, blob_meta)| format!("{}@{}", blob_meta.key, blob_meta.lsn))
1049 200 : .join(", ");
1050 200 : tracing::warn!(
1051 0 : "Oversized vectored read ({} > {}) for keys {}",
1052 : largest_read_size,
1053 : read_size_soft_max,
1054 : offenders
1055 : );
1056 140 : }
1057 :
1058 340 : largest_read_size
1059 354 : }
1060 :
1061 154 : async fn do_reads_and_update_state(
1062 154 : &self,
1063 154 : reads: Vec<VectoredRead>,
1064 154 : reconstruct_state: &mut ValuesReconstructState,
1065 154 : ctx: &RequestContext,
1066 154 : ) {
1067 154 : let vectored_blob_reader = VectoredBlobReader::new(&self.file);
1068 154 : let mut ignore_key_with_err = None;
1069 154 :
1070 154 : let max_vectored_read_bytes = self
1071 154 : .max_vectored_read_bytes
1072 154 : .expect("Layer is loaded with max vectored bytes config")
1073 154 : .0
1074 154 : .into();
1075 154 : let buf_size = Self::get_min_read_buffer_size(&reads, max_vectored_read_bytes);
1076 154 : let mut buf = Some(BytesMut::with_capacity(buf_size));
1077 :
1078 : // Note that reads are processed in reverse order (from highest key+lsn).
1079 : // This is the order that `ReconstructState` requires such that it can
1080 : // track when a key is done.
1081 17187 : for read in reads.into_iter().rev() {
1082 17187 : let res = vectored_blob_reader
1083 17187 : .read_blobs(&read, buf.take().expect("Should have a buffer"), ctx)
1084 8752 : .await;
1085 :
1086 17187 : let blobs_buf = match res {
1087 17187 : Ok(blobs_buf) => blobs_buf,
1088 0 : Err(err) => {
1089 0 : let kind = err.kind();
1090 0 : for (_, blob_meta) in read.blobs_at.as_slice() {
1091 0 : reconstruct_state.on_key_error(
1092 0 : blob_meta.key,
1093 0 : PageReconstructError::from(anyhow!(
1094 0 : "Failed to read blobs from virtual file {}: {}",
1095 0 : self.file.path,
1096 0 : kind
1097 0 : )),
1098 0 : );
1099 0 : }
1100 :
1101 : // We have "lost" the buffer since the lower level IO api
1102 : // doesn't return the buffer on error. Allocate a new one.
1103 0 : buf = Some(BytesMut::with_capacity(buf_size));
1104 0 :
1105 0 : continue;
1106 : }
1107 : };
1108 :
1109 28025 : for meta in blobs_buf.blobs.iter().rev() {
1110 28025 : if Some(meta.meta.key) == ignore_key_with_err {
1111 0 : continue;
1112 28025 : }
1113 28025 :
1114 28025 : let value = Value::des(&blobs_buf.buf[meta.start..meta.end]);
1115 28025 : let value = match value {
1116 28025 : Ok(v) => v,
1117 0 : Err(e) => {
1118 0 : reconstruct_state.on_key_error(
1119 0 : meta.meta.key,
1120 0 : PageReconstructError::from(anyhow!(e).context(format!(
1121 0 : "Failed to deserialize blob from virtual file {}",
1122 0 : self.file.path,
1123 0 : ))),
1124 0 : );
1125 0 :
1126 0 : ignore_key_with_err = Some(meta.meta.key);
1127 0 : continue;
1128 : }
1129 : };
1130 :
1131 : // Invariant: once a key reaches [`ValueReconstructSituation::Complete`]
1132 : // state, no further updates shall be made to it. The call below will
1133 : // panic if the invariant is violated.
1134 28025 : reconstruct_state.update_key(&meta.meta.key, meta.meta.lsn, value);
1135 : }
1136 :
1137 17187 : buf = Some(blobs_buf.buf);
1138 : }
1139 154 : }
1140 :
1141 406 : pub(super) async fn load_keys<'a>(
1142 406 : &'a self,
1143 406 : ctx: &RequestContext,
1144 406 : ) -> Result<Vec<DeltaEntry<'a>>> {
1145 406 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1146 406 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1147 406 : self.index_start_blk,
1148 406 : self.index_root_blk,
1149 406 : block_reader,
1150 406 : );
1151 406 :
1152 406 : let mut all_keys: Vec<DeltaEntry<'_>> = Vec::new();
1153 406 :
1154 406 : tree_reader
1155 406 : .visit(
1156 406 : &[0u8; DELTA_KEY_SIZE],
1157 406 : VisitDirection::Forwards,
1158 2064046 : |key, value| {
1159 2064046 : let delta_key = DeltaKey::from_slice(key);
1160 2064046 : let val_ref = ValueRef {
1161 2064046 : blob_ref: BlobRef(value),
1162 2064046 : reader: BlockCursor::new(crate::tenant::block_io::BlockReaderRef::Adapter(
1163 2064046 : Adapter(self),
1164 2064046 : )),
1165 2064046 : };
1166 2064046 : let pos = BlobRef(value).pos();
1167 2064046 : if let Some(last) = all_keys.last_mut() {
1168 2063640 : // subtract offset of the current and last entries to get the size
1169 2063640 : // of the value associated with this (key, lsn) tuple
1170 2063640 : let first_pos = last.size;
1171 2063640 : last.size = pos - first_pos;
1172 2063640 : }
1173 2064046 : let entry = DeltaEntry {
1174 2064046 : key: delta_key.key(),
1175 2064046 : lsn: delta_key.lsn(),
1176 2064046 : size: pos,
1177 2064046 : val: val_ref,
1178 2064046 : };
1179 2064046 : all_keys.push(entry);
1180 2064046 : true
1181 2064046 : },
1182 406 : &RequestContextBuilder::extend(ctx)
1183 406 : .page_content_kind(PageContentKind::DeltaLayerBtreeNode)
1184 406 : .build(),
1185 406 : )
1186 2170 : .await?;
1187 406 : if let Some(last) = all_keys.last_mut() {
1188 406 : // Last key occupies all space till end of value storage,
1189 406 : // which corresponds to beginning of the index
1190 406 : last.size = self.index_start_offset() - last.size;
1191 406 : }
1192 406 : Ok(all_keys)
1193 406 : }
1194 :
1195 : /// Using the given writer, write out a version which has the earlier Lsns than `until`.
1196 : ///
1197 : /// Return the amount of key value records pushed to the writer.
1198 10 : pub(super) async fn copy_prefix(
1199 10 : &self,
1200 10 : writer: &mut DeltaLayerWriter,
1201 10 : until: Lsn,
1202 10 : ctx: &RequestContext,
1203 10 : ) -> anyhow::Result<usize> {
1204 10 : use crate::tenant::vectored_blob_io::{
1205 10 : BlobMeta, VectoredReadBuilder, VectoredReadExtended,
1206 10 : };
1207 10 : use futures::stream::TryStreamExt;
1208 10 :
1209 10 : #[derive(Debug)]
1210 10 : enum Item {
1211 10 : Actual(Key, Lsn, BlobRef),
1212 10 : Sentinel,
1213 10 : }
1214 10 :
1215 10 : impl From<Item> for Option<(Key, Lsn, BlobRef)> {
1216 70 : fn from(value: Item) -> Self {
1217 70 : match value {
1218 60 : Item::Actual(key, lsn, blob) => Some((key, lsn, blob)),
1219 10 : Item::Sentinel => None,
1220 10 : }
1221 70 : }
1222 10 : }
1223 10 :
1224 10 : impl Item {
1225 70 : fn offset(&self) -> Option<BlobRef> {
1226 70 : match self {
1227 60 : Item::Actual(_, _, blob) => Some(*blob),
1228 10 : Item::Sentinel => None,
1229 10 : }
1230 70 : }
1231 10 :
1232 70 : fn is_last(&self) -> bool {
1233 70 : matches!(self, Item::Sentinel)
1234 70 : }
1235 10 : }
1236 10 :
1237 10 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1238 10 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1239 10 : self.index_start_blk,
1240 10 : self.index_root_blk,
1241 10 : block_reader,
1242 10 : );
1243 10 :
1244 10 : let stream = self.stream_index_forwards(&tree_reader, &[0u8; DELTA_KEY_SIZE], ctx);
1245 60 : let stream = stream.map_ok(|(key, lsn, pos)| Item::Actual(key, lsn, pos));
1246 10 : // put in a sentinel value for getting the end offset for last item, and not having to
1247 10 : // repeat the whole read part
1248 10 : let stream = stream.chain(futures::stream::once(futures::future::ready(Ok(
1249 10 : Item::Sentinel,
1250 10 : ))));
1251 10 : let mut stream = std::pin::pin!(stream);
1252 10 :
1253 10 : let mut prev: Option<(Key, Lsn, BlobRef)> = None;
1254 10 :
1255 10 : let mut read_builder: Option<VectoredReadBuilder> = None;
1256 10 :
1257 10 : let max_read_size = self
1258 10 : .max_vectored_read_bytes
1259 10 : .map(|x| x.0.get())
1260 10 : .unwrap_or(8192);
1261 10 :
1262 10 : let mut buffer = Some(BytesMut::with_capacity(max_read_size));
1263 10 :
1264 10 : // FIXME: buffering of DeltaLayerWriter
1265 10 : let mut per_blob_copy = Vec::new();
1266 10 :
1267 10 : let mut records = 0;
1268 :
1269 80 : while let Some(item) = stream.try_next().await? {
1270 70 : tracing::debug!(?item, "popped");
1271 70 : let offset = item
1272 70 : .offset()
1273 70 : .unwrap_or(BlobRef::new(self.index_start_offset(), false));
1274 :
1275 70 : let actionable = if let Some((key, lsn, start_offset)) = prev.take() {
1276 60 : let end_offset = offset;
1277 60 :
1278 60 : Some((BlobMeta { key, lsn }, start_offset..end_offset))
1279 : } else {
1280 10 : None
1281 : };
1282 :
1283 70 : let is_last = item.is_last();
1284 70 :
1285 70 : prev = Option::from(item);
1286 70 :
1287 70 : let actionable = actionable.filter(|x| x.0.lsn < until);
1288 :
1289 70 : let builder = if let Some((meta, offsets)) = actionable {
1290 : // extend or create a new builder
1291 32 : if read_builder
1292 32 : .as_mut()
1293 32 : .map(|x| x.extend(offsets.start.pos(), offsets.end.pos(), meta))
1294 32 : .unwrap_or(VectoredReadExtended::No)
1295 32 : == VectoredReadExtended::Yes
1296 : {
1297 16 : None
1298 : } else {
1299 16 : read_builder.replace(VectoredReadBuilder::new(
1300 16 : offsets.start.pos(),
1301 16 : offsets.end.pos(),
1302 16 : meta,
1303 16 : max_read_size,
1304 16 : ))
1305 : }
1306 : } else {
1307 : // nothing to do, except perhaps flush any existing for the last element
1308 38 : None
1309 : };
1310 :
1311 : // flush the possible older builder and also the new one if the item was the last one
1312 70 : let builders = builder.into_iter();
1313 70 : let builders = if is_last {
1314 10 : builders.chain(read_builder.take())
1315 : } else {
1316 60 : builders.chain(None)
1317 : };
1318 :
1319 86 : for builder in builders {
1320 16 : let read = builder.build();
1321 16 :
1322 16 : let reader = VectoredBlobReader::new(&self.file);
1323 16 :
1324 16 : let mut buf = buffer.take().unwrap();
1325 16 :
1326 16 : buf.clear();
1327 16 : buf.reserve(read.size());
1328 16 : let res = reader.read_blobs(&read, buf, ctx).await?;
1329 :
1330 48 : for blob in res.blobs {
1331 32 : let key = blob.meta.key;
1332 32 : let lsn = blob.meta.lsn;
1333 32 : let data = &res.buf[blob.start..blob.end];
1334 32 :
1335 32 : #[cfg(debug_assertions)]
1336 32 : Value::des(data)
1337 32 : .with_context(|| {
1338 0 : format!(
1339 0 : "blob failed to deserialize for {}@{}, {}..{}: {:?}",
1340 0 : blob.meta.key,
1341 0 : blob.meta.lsn,
1342 0 : blob.start,
1343 0 : blob.end,
1344 0 : utils::Hex(data)
1345 0 : )
1346 32 : })
1347 32 : .unwrap();
1348 32 :
1349 32 : // is it an image or will_init walrecord?
1350 32 : // FIXME: this could be handled by threading the BlobRef to the
1351 32 : // VectoredReadBuilder
1352 32 : let will_init = crate::repository::ValueBytes::will_init(data)
1353 32 : .inspect_err(|_e| {
1354 0 : #[cfg(feature = "testing")]
1355 0 : tracing::error!(data=?utils::Hex(data), err=?_e, %key, %lsn, "failed to parse will_init out of serialized value");
1356 32 : })
1357 32 : .unwrap_or(false);
1358 32 :
1359 32 : per_blob_copy.clear();
1360 32 : per_blob_copy.extend_from_slice(data);
1361 :
1362 32 : let (tmp, res) = writer
1363 32 : .put_value_bytes(
1364 32 : key,
1365 32 : lsn,
1366 32 : std::mem::take(&mut per_blob_copy),
1367 32 : will_init,
1368 32 : ctx,
1369 32 : )
1370 4 : .await;
1371 32 : per_blob_copy = tmp;
1372 32 :
1373 32 : res?;
1374 :
1375 32 : records += 1;
1376 : }
1377 :
1378 16 : buffer = Some(res.buf);
1379 : }
1380 : }
1381 :
1382 10 : assert!(
1383 10 : read_builder.is_none(),
1384 0 : "with the sentinel above loop should had handled all"
1385 : );
1386 :
1387 10 : Ok(records)
1388 10 : }
1389 :
1390 4 : pub(super) async fn dump(&self, ctx: &RequestContext) -> anyhow::Result<()> {
1391 4 : println!(
1392 4 : "index_start_blk: {}, root {}",
1393 4 : self.index_start_blk, self.index_root_blk
1394 4 : );
1395 4 :
1396 4 : let block_reader = FileBlockReader::new(&self.file, self.file_id);
1397 4 : let tree_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1398 4 : self.index_start_blk,
1399 4 : self.index_root_blk,
1400 4 : block_reader,
1401 4 : );
1402 4 :
1403 4 : tree_reader.dump().await?;
1404 :
1405 4 : let keys = self.load_keys(ctx).await?;
1406 :
1407 8 : async fn dump_blob(val: &ValueRef<'_>, ctx: &RequestContext) -> anyhow::Result<String> {
1408 8 : let buf = val.reader.read_blob(val.blob_ref.pos(), ctx).await?;
1409 8 : let val = Value::des(&buf)?;
1410 8 : let desc = match val {
1411 8 : Value::Image(img) => {
1412 8 : format!(" img {} bytes", img.len())
1413 : }
1414 0 : Value::WalRecord(rec) => {
1415 0 : let wal_desc = walrecord::describe_wal_record(&rec)?;
1416 0 : format!(
1417 0 : " rec {} bytes will_init: {} {}",
1418 0 : buf.len(),
1419 0 : rec.will_init(),
1420 0 : wal_desc
1421 0 : )
1422 : }
1423 : };
1424 8 : Ok(desc)
1425 8 : }
1426 :
1427 12 : for entry in keys {
1428 8 : let DeltaEntry { key, lsn, val, .. } = entry;
1429 8 : let desc = match dump_blob(&val, ctx).await {
1430 8 : Ok(desc) => desc,
1431 0 : Err(err) => {
1432 0 : format!("ERROR: {err}")
1433 : }
1434 : };
1435 8 : println!(" key {key} at {lsn}: {desc}");
1436 8 :
1437 8 : // Print more details about CHECKPOINT records. Would be nice to print details
1438 8 : // of many other record types too, but these are particularly interesting, as
1439 8 : // have a lot of special processing for them in walingest.rs.
1440 8 : use pageserver_api::key::CHECKPOINT_KEY;
1441 8 : use postgres_ffi::CheckPoint;
1442 8 : if key == CHECKPOINT_KEY {
1443 0 : let buf = val.reader.read_blob(val.blob_ref.pos(), ctx).await?;
1444 0 : let val = Value::des(&buf)?;
1445 0 : match val {
1446 0 : Value::Image(img) => {
1447 0 : let checkpoint = CheckPoint::decode(&img)?;
1448 0 : println!(" CHECKPOINT: {:?}", checkpoint);
1449 : }
1450 0 : Value::WalRecord(_rec) => {
1451 0 : println!(" unexpected walrecord value for checkpoint key");
1452 0 : }
1453 : }
1454 8 : }
1455 : }
1456 :
1457 4 : Ok(())
1458 4 : }
1459 :
1460 34 : fn stream_index_forwards<'a, R>(
1461 34 : &'a self,
1462 34 : reader: &'a DiskBtreeReader<R, DELTA_KEY_SIZE>,
1463 34 : start: &'a [u8; DELTA_KEY_SIZE],
1464 34 : ctx: &'a RequestContext,
1465 34 : ) -> impl futures::stream::Stream<
1466 34 : Item = Result<(Key, Lsn, BlobRef), crate::tenant::disk_btree::DiskBtreeError>,
1467 34 : > + 'a
1468 34 : where
1469 34 : R: BlockReader,
1470 34 : {
1471 34 : use futures::stream::TryStreamExt;
1472 34 : let stream = reader.get_stream_from(start, ctx);
1473 162 : stream.map_ok(|(key, value)| {
1474 162 : let key = DeltaKey::from_slice(&key);
1475 162 : let (key, lsn) = (key.key(), key.lsn());
1476 162 : let offset = BlobRef(value);
1477 162 :
1478 162 : (key, lsn, offset)
1479 162 : })
1480 34 : }
1481 :
1482 : /// The file offset to the first block of index.
1483 : ///
1484 : /// The file structure is summary, values, and index. We often need this for the size of last blob.
1485 630 : fn index_start_offset(&self) -> u64 {
1486 630 : let offset = self.index_start_blk as u64 * PAGE_SZ as u64;
1487 630 : let bref = BlobRef(offset);
1488 630 : tracing::debug!(
1489 : index_start_blk = self.index_start_blk,
1490 : offset,
1491 0 : pos = bref.pos(),
1492 0 : "index_start_offset"
1493 : );
1494 630 : offset
1495 630 : }
1496 : }
1497 :
1498 : /// A set of data associated with a delta layer key and its value
1499 : pub struct DeltaEntry<'a> {
1500 : pub key: Key,
1501 : pub lsn: Lsn,
1502 : /// Size of the stored value
1503 : pub size: u64,
1504 : /// Reference to the on-disk value
1505 : pub val: ValueRef<'a>,
1506 : }
1507 :
1508 : /// Reference to an on-disk value
1509 : pub struct ValueRef<'a> {
1510 : blob_ref: BlobRef,
1511 : reader: BlockCursor<'a>,
1512 : }
1513 :
1514 : impl<'a> ValueRef<'a> {
1515 : /// Loads the value from disk
1516 2064038 : pub async fn load(&self, ctx: &RequestContext) -> Result<Value> {
1517 : // theoretically we *could* record an access time for each, but it does not really matter
1518 2064038 : let buf = self.reader.read_blob(self.blob_ref.pos(), ctx).await?;
1519 2064038 : let val = Value::des(&buf)?;
1520 2064038 : Ok(val)
1521 2064038 : }
1522 : }
1523 :
1524 : pub(crate) struct Adapter<T>(T);
1525 :
1526 : impl<T: AsRef<DeltaLayerInner>> Adapter<T> {
1527 2083002 : pub(crate) async fn read_blk(
1528 2083002 : &self,
1529 2083002 : blknum: u32,
1530 2083002 : ctx: &RequestContext,
1531 2083002 : ) -> Result<BlockLease, std::io::Error> {
1532 2083002 : let block_reader = FileBlockReader::new(&self.0.as_ref().file, self.0.as_ref().file_id);
1533 2083002 : block_reader.read_blk(blknum, ctx).await
1534 2083002 : }
1535 : }
1536 :
1537 : impl AsRef<DeltaLayerInner> for DeltaLayerInner {
1538 4166004 : fn as_ref(&self) -> &DeltaLayerInner {
1539 4166004 : self
1540 4166004 : }
1541 : }
1542 :
1543 : impl<'a> pageserver_compaction::interface::CompactionDeltaEntry<'a, Key> for DeltaEntry<'a> {
1544 0 : fn key(&self) -> Key {
1545 0 : self.key
1546 0 : }
1547 0 : fn lsn(&self) -> Lsn {
1548 0 : self.lsn
1549 0 : }
1550 0 : fn size(&self) -> u64 {
1551 0 : self.size
1552 0 : }
1553 : }
1554 :
1555 : #[cfg(test)]
1556 : mod test {
1557 : use std::collections::BTreeMap;
1558 :
1559 : use itertools::MinMaxResult;
1560 : use rand::prelude::{SeedableRng, SliceRandom, StdRng};
1561 : use rand::RngCore;
1562 :
1563 : use super::*;
1564 : use crate::{
1565 : context::DownloadBehavior,
1566 : task_mgr::TaskKind,
1567 : tenant::{disk_btree::tests::TestDisk, harness::TenantHarness},
1568 : DEFAULT_PG_VERSION,
1569 : };
1570 :
1571 : /// Construct an index for a fictional delta layer and and then
1572 : /// traverse in order to plan vectored reads for a query. Finally,
1573 : /// verify that the traversal fed the right index key and value
1574 : /// pairs into the planner.
1575 : #[tokio::test]
1576 2 : async fn test_delta_layer_index_traversal() {
1577 2 : let base_key = Key {
1578 2 : field1: 0,
1579 2 : field2: 1663,
1580 2 : field3: 12972,
1581 2 : field4: 16396,
1582 2 : field5: 0,
1583 2 : field6: 246080,
1584 2 : };
1585 2 :
1586 2 : // Populate the index with some entries
1587 2 : let entries: BTreeMap<Key, Vec<Lsn>> = BTreeMap::from([
1588 2 : (base_key, vec![Lsn(1), Lsn(5), Lsn(25), Lsn(26), Lsn(28)]),
1589 2 : (base_key.add(1), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1590 2 : (base_key.add(2), vec![Lsn(2), Lsn(5), Lsn(10), Lsn(50)]),
1591 2 : (base_key.add(5), vec![Lsn(10), Lsn(15), Lsn(16), Lsn(20)]),
1592 2 : ]);
1593 2 :
1594 2 : let mut disk = TestDisk::default();
1595 2 : let mut writer = DiskBtreeBuilder::<_, DELTA_KEY_SIZE>::new(&mut disk);
1596 2 :
1597 2 : let mut disk_offset = 0;
1598 10 : for (key, lsns) in &entries {
1599 42 : for lsn in lsns {
1600 34 : let index_key = DeltaKey::from_key_lsn(key, *lsn);
1601 34 : let blob_ref = BlobRef::new(disk_offset, false);
1602 34 : writer
1603 34 : .append(&index_key.0, blob_ref.0)
1604 34 : .expect("In memory disk append should never fail");
1605 34 :
1606 34 : disk_offset += 1;
1607 34 : }
1608 2 : }
1609 2 :
1610 2 : // Prepare all the arguments for the call into `plan_reads` below
1611 2 : let (root_offset, _writer) = writer
1612 2 : .finish()
1613 2 : .expect("In memory disk finish should never fail");
1614 2 : let reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(0, root_offset, disk);
1615 2 : let planner = VectoredReadPlanner::new(100);
1616 2 : let mut reconstruct_state = ValuesReconstructState::new();
1617 2 : let ctx = RequestContext::new(TaskKind::UnitTest, DownloadBehavior::Error);
1618 2 :
1619 2 : let keyspace = KeySpace {
1620 2 : ranges: vec![
1621 2 : base_key..base_key.add(3),
1622 2 : base_key.add(3)..base_key.add(100),
1623 2 : ],
1624 2 : };
1625 2 : let lsn_range = Lsn(2)..Lsn(40);
1626 2 :
1627 2 : // Plan and validate
1628 2 : let vectored_reads = DeltaLayerInner::plan_reads(
1629 2 : &keyspace,
1630 2 : lsn_range.clone(),
1631 2 : disk_offset,
1632 2 : reader,
1633 2 : planner,
1634 2 : &mut reconstruct_state,
1635 2 : &ctx,
1636 2 : )
1637 2 : .await
1638 2 : .expect("Read planning should not fail");
1639 2 :
1640 2 : validate(keyspace, lsn_range, vectored_reads, entries);
1641 2 : }
1642 :
1643 2 : fn validate(
1644 2 : keyspace: KeySpace,
1645 2 : lsn_range: Range<Lsn>,
1646 2 : vectored_reads: Vec<VectoredRead>,
1647 2 : index_entries: BTreeMap<Key, Vec<Lsn>>,
1648 2 : ) {
1649 2 : #[derive(Debug, PartialEq, Eq)]
1650 2 : struct BlobSpec {
1651 2 : key: Key,
1652 2 : lsn: Lsn,
1653 2 : at: u64,
1654 2 : }
1655 2 :
1656 2 : let mut planned_blobs = Vec::new();
1657 8 : for read in vectored_reads {
1658 28 : for (at, meta) in read.blobs_at.as_slice() {
1659 28 : planned_blobs.push(BlobSpec {
1660 28 : key: meta.key,
1661 28 : lsn: meta.lsn,
1662 28 : at: *at,
1663 28 : });
1664 28 : }
1665 : }
1666 :
1667 2 : let mut expected_blobs = Vec::new();
1668 2 : let mut disk_offset = 0;
1669 10 : for (key, lsns) in index_entries {
1670 42 : for lsn in lsns {
1671 42 : let key_included = keyspace.ranges.iter().any(|range| range.contains(&key));
1672 34 : let lsn_included = lsn_range.contains(&lsn);
1673 34 :
1674 34 : if key_included && lsn_included {
1675 28 : expected_blobs.push(BlobSpec {
1676 28 : key,
1677 28 : lsn,
1678 28 : at: disk_offset,
1679 28 : });
1680 28 : }
1681 :
1682 34 : disk_offset += 1;
1683 : }
1684 : }
1685 :
1686 2 : assert_eq!(planned_blobs, expected_blobs);
1687 2 : }
1688 :
1689 : mod constants {
1690 : use utils::lsn::Lsn;
1691 :
1692 : /// Offset used by all lsns in this test
1693 : pub(super) const LSN_OFFSET: Lsn = Lsn(0x08);
1694 : /// Number of unique keys including in the test data
1695 : pub(super) const KEY_COUNT: u8 = 60;
1696 : /// Max number of different lsns for each key
1697 : pub(super) const MAX_ENTRIES_PER_KEY: u8 = 20;
1698 : /// Possible value sizes for each key along with a probability weight
1699 : pub(super) const VALUE_SIZES: [(usize, u8); 3] = [(100, 2), (1024, 2), (1024 * 1024, 1)];
1700 : /// Probability that there will be a gap between the current key and the next one (33.3%)
1701 : pub(super) const KEY_GAP_CHANGES: [(bool, u8); 2] = [(true, 1), (false, 2)];
1702 : /// The minimum size of a key range in all the generated reads
1703 : pub(super) const MIN_RANGE_SIZE: i128 = 10;
1704 : /// The number of ranges included in each vectored read
1705 : pub(super) const RANGES_COUNT: u8 = 2;
1706 : /// The number of vectored reads performed
1707 : pub(super) const READS_COUNT: u8 = 100;
1708 : /// Soft max size of a vectored read. Will be violated if we have to read keys
1709 : /// with values larger than the limit
1710 : pub(super) const MAX_VECTORED_READ_BYTES: usize = 64 * 1024;
1711 : }
1712 :
1713 : struct Entry {
1714 : key: Key,
1715 : lsn: Lsn,
1716 : value: Vec<u8>,
1717 : }
1718 :
1719 2 : fn generate_entries(rng: &mut StdRng) -> Vec<Entry> {
1720 2 : let mut current_key = Key::MIN;
1721 2 :
1722 2 : let mut entries = Vec::new();
1723 122 : for _ in 0..constants::KEY_COUNT {
1724 120 : let count = rng.gen_range(1..constants::MAX_ENTRIES_PER_KEY);
1725 120 : let mut lsns_iter =
1726 2260 : std::iter::successors(Some(Lsn(constants::LSN_OFFSET.0 + 0x08)), |lsn| {
1727 2260 : Some(Lsn(lsn.0 + 0x08))
1728 2260 : });
1729 120 : let mut lsns = Vec::new();
1730 2380 : while lsns.len() < count as usize {
1731 2260 : let take = rng.gen_bool(0.5);
1732 2260 : let lsn = lsns_iter.next().unwrap();
1733 2260 : if take {
1734 1112 : lsns.push(lsn);
1735 1148 : }
1736 : }
1737 :
1738 1232 : for lsn in lsns {
1739 1112 : let size = constants::VALUE_SIZES
1740 3336 : .choose_weighted(rng, |item| item.1)
1741 1112 : .unwrap()
1742 1112 : .0;
1743 1112 : let mut buf = vec![0; size];
1744 1112 : rng.fill_bytes(&mut buf);
1745 1112 :
1746 1112 : entries.push(Entry {
1747 1112 : key: current_key,
1748 1112 : lsn,
1749 1112 : value: buf,
1750 1112 : })
1751 : }
1752 :
1753 120 : let gap = constants::KEY_GAP_CHANGES
1754 240 : .choose_weighted(rng, |item| item.1)
1755 120 : .unwrap()
1756 120 : .0;
1757 120 : if gap {
1758 38 : current_key = current_key.add(2);
1759 82 : } else {
1760 82 : current_key = current_key.add(1);
1761 82 : }
1762 : }
1763 :
1764 2 : entries
1765 2 : }
1766 :
1767 : struct EntriesMeta {
1768 : key_range: Range<Key>,
1769 : lsn_range: Range<Lsn>,
1770 : index: BTreeMap<(Key, Lsn), Vec<u8>>,
1771 : }
1772 :
1773 2 : fn get_entries_meta(entries: &[Entry]) -> EntriesMeta {
1774 1112 : let key_range = match entries.iter().minmax_by_key(|e| e.key) {
1775 2 : MinMaxResult::MinMax(min, max) => min.key..max.key.next(),
1776 0 : _ => panic!("More than one entry is always expected"),
1777 : };
1778 :
1779 1112 : let lsn_range = match entries.iter().minmax_by_key(|e| e.lsn) {
1780 2 : MinMaxResult::MinMax(min, max) => min.lsn..Lsn(max.lsn.0 + 1),
1781 0 : _ => panic!("More than one entry is always expected"),
1782 : };
1783 :
1784 2 : let mut index = BTreeMap::new();
1785 1112 : for entry in entries.iter() {
1786 1112 : index.insert((entry.key, entry.lsn), entry.value.clone());
1787 1112 : }
1788 :
1789 2 : EntriesMeta {
1790 2 : key_range,
1791 2 : lsn_range,
1792 2 : index,
1793 2 : }
1794 2 : }
1795 :
1796 200 : fn pick_random_keyspace(rng: &mut StdRng, key_range: &Range<Key>) -> KeySpace {
1797 200 : let start = key_range.start.to_i128();
1798 200 : let end = key_range.end.to_i128();
1799 200 :
1800 200 : let mut keyspace = KeySpace::default();
1801 :
1802 600 : for _ in 0..constants::RANGES_COUNT {
1803 400 : let mut range: Option<Range<Key>> = Option::default();
1804 1244 : while range.is_none() || keyspace.overlaps(range.as_ref().unwrap()) {
1805 844 : let range_start = rng.gen_range(start..end);
1806 844 : let range_end_offset = range_start + constants::MIN_RANGE_SIZE;
1807 844 : if range_end_offset >= end {
1808 100 : range = Some(Key::from_i128(range_start)..Key::from_i128(end));
1809 744 : } else {
1810 744 : let range_end = rng.gen_range((range_start + constants::MIN_RANGE_SIZE)..end);
1811 744 : range = Some(Key::from_i128(range_start)..Key::from_i128(range_end));
1812 744 : }
1813 : }
1814 400 : keyspace.ranges.push(range.unwrap());
1815 : }
1816 :
1817 200 : keyspace
1818 200 : }
1819 :
1820 : #[tokio::test]
1821 2 : async fn test_delta_layer_vectored_read_end_to_end() -> anyhow::Result<()> {
1822 2 : let harness = TenantHarness::create("test_delta_layer_oversized_vectored_read")?;
1823 8 : let (tenant, ctx) = harness.load().await;
1824 2 :
1825 2 : let timeline_id = TimelineId::generate();
1826 2 : let timeline = tenant
1827 2 : .create_test_timeline(timeline_id, constants::LSN_OFFSET, DEFAULT_PG_VERSION, &ctx)
1828 6 : .await?;
1829 2 :
1830 2 : tracing::info!("Generating test data ...");
1831 2 :
1832 2 : let rng = &mut StdRng::seed_from_u64(0);
1833 2 : let entries = generate_entries(rng);
1834 2 : let entries_meta = get_entries_meta(&entries);
1835 2 :
1836 2 : tracing::info!("Done generating {} entries", entries.len());
1837 2 :
1838 2 : tracing::info!("Writing test data to delta layer ...");
1839 2 : let mut writer = DeltaLayerWriter::new(
1840 2 : harness.conf,
1841 2 : timeline_id,
1842 2 : harness.tenant_shard_id,
1843 2 : entries_meta.key_range.start,
1844 2 : entries_meta.lsn_range.clone(),
1845 2 : &ctx,
1846 2 : )
1847 2 : .await?;
1848 2 :
1849 1114 : for entry in entries {
1850 1112 : let (_, res) = writer
1851 1112 : .put_value_bytes(entry.key, entry.lsn, entry.value, false, &ctx)
1852 215 : .await;
1853 1112 : res?;
1854 2 : }
1855 2 :
1856 2 : let resident = writer
1857 2 : .finish(entries_meta.key_range.end, &timeline, &ctx)
1858 5 : .await?;
1859 2 :
1860 2 : let inner = resident.as_delta(&ctx).await?;
1861 2 :
1862 2 : let file_size = inner.file.metadata().await?.len();
1863 2 : tracing::info!(
1864 2 : "Done writing test data to delta layer. Resulting file size is: {}",
1865 2 : file_size
1866 2 : );
1867 2 :
1868 202 : for i in 0..constants::READS_COUNT {
1869 200 : tracing::info!("Doing vectored read {}/{}", i + 1, constants::READS_COUNT);
1870 2 :
1871 200 : let block_reader = FileBlockReader::new(&inner.file, inner.file_id);
1872 200 : let index_reader = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
1873 200 : inner.index_start_blk,
1874 200 : inner.index_root_blk,
1875 200 : block_reader,
1876 200 : );
1877 200 :
1878 200 : let planner = VectoredReadPlanner::new(constants::MAX_VECTORED_READ_BYTES);
1879 200 : let mut reconstruct_state = ValuesReconstructState::new();
1880 200 : let keyspace = pick_random_keyspace(rng, &entries_meta.key_range);
1881 200 : let data_end_offset = inner.index_start_blk as u64 * PAGE_SZ as u64;
1882 2 :
1883 200 : let vectored_reads = DeltaLayerInner::plan_reads(
1884 200 : &keyspace,
1885 200 : entries_meta.lsn_range.clone(),
1886 200 : data_end_offset,
1887 200 : index_reader,
1888 200 : planner,
1889 200 : &mut reconstruct_state,
1890 200 : &ctx,
1891 200 : )
1892 4 : .await?;
1893 2 :
1894 200 : let vectored_blob_reader = VectoredBlobReader::new(&inner.file);
1895 200 : let buf_size = DeltaLayerInner::get_min_read_buffer_size(
1896 200 : &vectored_reads,
1897 200 : constants::MAX_VECTORED_READ_BYTES,
1898 200 : );
1899 200 : let mut buf = Some(BytesMut::with_capacity(buf_size));
1900 2 :
1901 19924 : for read in vectored_reads {
1902 19724 : let blobs_buf = vectored_blob_reader
1903 19724 : .read_blobs(&read, buf.take().expect("Should have a buffer"), &ctx)
1904 10016 : .await?;
1905 57304 : for meta in blobs_buf.blobs.iter() {
1906 57304 : let value = &blobs_buf.buf[meta.start..meta.end];
1907 57304 : assert_eq!(value, entries_meta.index[&(meta.meta.key, meta.meta.lsn)]);
1908 2 : }
1909 2 :
1910 19724 : buf = Some(blobs_buf.buf);
1911 2 : }
1912 2 : }
1913 2 :
1914 2 : Ok(())
1915 2 : }
1916 :
1917 : #[tokio::test]
1918 2 : async fn copy_delta_prefix_smoke() {
1919 2 : use crate::walrecord::NeonWalRecord;
1920 2 : use bytes::Bytes;
1921 2 :
1922 2 : let h = crate::tenant::harness::TenantHarness::create("truncate_delta_smoke").unwrap();
1923 8 : let (tenant, ctx) = h.load().await;
1924 2 : let ctx = &ctx;
1925 2 : let timeline = tenant
1926 2 : .create_test_timeline(TimelineId::generate(), Lsn(0x10), 14, ctx)
1927 5 : .await
1928 2 : .unwrap();
1929 2 :
1930 2 : let initdb_layer = timeline
1931 2 : .layers
1932 2 : .read()
1933 2 : .await
1934 2 : .likely_resident_layers()
1935 2 : .next()
1936 2 : .unwrap();
1937 2 :
1938 2 : {
1939 2 : let mut writer = timeline.writer().await;
1940 2 :
1941 2 : let data = [
1942 2 : (0x20, 12, Value::Image(Bytes::from_static(b"foobar"))),
1943 2 : (
1944 2 : 0x30,
1945 2 : 12,
1946 2 : Value::WalRecord(NeonWalRecord::Postgres {
1947 2 : will_init: false,
1948 2 : rec: Bytes::from_static(b"1"),
1949 2 : }),
1950 2 : ),
1951 2 : (
1952 2 : 0x40,
1953 2 : 12,
1954 2 : Value::WalRecord(NeonWalRecord::Postgres {
1955 2 : will_init: true,
1956 2 : rec: Bytes::from_static(b"2"),
1957 2 : }),
1958 2 : ),
1959 2 : // build an oversized value so we cannot extend and existing read over
1960 2 : // this
1961 2 : (
1962 2 : 0x50,
1963 2 : 12,
1964 2 : Value::WalRecord(NeonWalRecord::Postgres {
1965 2 : will_init: true,
1966 2 : rec: {
1967 2 : let mut buf =
1968 2 : vec![0u8; tenant.conf.max_vectored_read_bytes.0.get() + 1024];
1969 2 : buf.iter_mut()
1970 2 : .enumerate()
1971 264192 : .for_each(|(i, slot)| *slot = (i % 256) as u8);
1972 2 : Bytes::from(buf)
1973 2 : },
1974 2 : }),
1975 2 : ),
1976 2 : // because the oversized read cannot be extended further, we are sure to exercise the
1977 2 : // builder created on the last round with this:
1978 2 : (
1979 2 : 0x60,
1980 2 : 12,
1981 2 : Value::WalRecord(NeonWalRecord::Postgres {
1982 2 : will_init: true,
1983 2 : rec: Bytes::from_static(b"3"),
1984 2 : }),
1985 2 : ),
1986 2 : (
1987 2 : 0x60,
1988 2 : 9,
1989 2 : Value::Image(Bytes::from_static(b"something for a different key")),
1990 2 : ),
1991 2 : ];
1992 2 :
1993 2 : let mut last_lsn = None;
1994 2 :
1995 14 : for (lsn, key, value) in data {
1996 12 : let key = Key::from_i128(key);
1997 12 : writer.put(key, Lsn(lsn), &value, ctx).await.unwrap();
1998 12 : last_lsn = Some(lsn);
1999 2 : }
2000 2 :
2001 2 : writer.finish_write(Lsn(last_lsn.unwrap()));
2002 2 : }
2003 2 : timeline.freeze_and_flush().await.unwrap();
2004 2 :
2005 2 : let new_layer = timeline
2006 2 : .layers
2007 2 : .read()
2008 2 : .await
2009 2 : .likely_resident_layers()
2010 4 : .find(|x| x != &initdb_layer)
2011 2 : .unwrap();
2012 2 :
2013 2 : // create a copy for the timeline, so we don't overwrite the file
2014 2 : let branch = tenant
2015 2 : .branch_timeline_test(&timeline, TimelineId::generate(), None, ctx)
2016 2 : .await
2017 2 : .unwrap();
2018 2 :
2019 2 : assert_eq!(branch.get_ancestor_lsn(), Lsn(0x60));
2020 2 :
2021 2 : // truncating at 0x61 gives us a full copy, otherwise just go backwards until there's just
2022 2 : // a single key
2023 2 :
2024 12 : for truncate_at in [0x61, 0x51, 0x41, 0x31, 0x21] {
2025 10 : let truncate_at = Lsn(truncate_at);
2026 2 :
2027 10 : let mut writer = DeltaLayerWriter::new(
2028 10 : tenant.conf,
2029 10 : branch.timeline_id,
2030 10 : tenant.tenant_shard_id,
2031 10 : Key::MIN,
2032 10 : Lsn(0x11)..truncate_at,
2033 10 : ctx,
2034 10 : )
2035 5 : .await
2036 10 : .unwrap();
2037 2 :
2038 10 : let new_layer = new_layer.download_and_keep_resident().await.unwrap();
2039 10 :
2040 10 : new_layer
2041 10 : .copy_delta_prefix(&mut writer, truncate_at, ctx)
2042 15 : .await
2043 10 : .unwrap();
2044 2 :
2045 24 : let copied_layer = writer.finish(Key::MAX, &branch, ctx).await.unwrap();
2046 10 :
2047 11 : copied_layer.as_delta(ctx).await.unwrap();
2048 10 :
2049 10 : assert_keys_and_values_eq(
2050 10 : new_layer.as_delta(ctx).await.unwrap(),
2051 10 : copied_layer.as_delta(ctx).await.unwrap(),
2052 10 : truncate_at,
2053 10 : ctx,
2054 2 : )
2055 65 : .await;
2056 2 : }
2057 2 : }
2058 :
2059 10 : async fn assert_keys_and_values_eq(
2060 10 : source: &DeltaLayerInner,
2061 10 : truncated: &DeltaLayerInner,
2062 10 : truncated_at: Lsn,
2063 10 : ctx: &RequestContext,
2064 10 : ) {
2065 10 : use futures::future::ready;
2066 10 : use futures::stream::TryStreamExt;
2067 10 :
2068 10 : let start_key = [0u8; DELTA_KEY_SIZE];
2069 10 :
2070 10 : let source_reader = FileBlockReader::new(&source.file, source.file_id);
2071 10 : let source_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2072 10 : source.index_start_blk,
2073 10 : source.index_root_blk,
2074 10 : &source_reader,
2075 10 : );
2076 10 : let source_stream = source.stream_index_forwards(&source_tree, &start_key, ctx);
2077 60 : let source_stream = source_stream.filter(|res| match res {
2078 60 : Ok((_, lsn, _)) => ready(lsn < &truncated_at),
2079 0 : _ => ready(true),
2080 60 : });
2081 10 : let mut source_stream = std::pin::pin!(source_stream);
2082 10 :
2083 10 : let truncated_reader = FileBlockReader::new(&truncated.file, truncated.file_id);
2084 10 : let truncated_tree = DiskBtreeReader::<_, DELTA_KEY_SIZE>::new(
2085 10 : truncated.index_start_blk,
2086 10 : truncated.index_root_blk,
2087 10 : &truncated_reader,
2088 10 : );
2089 10 : let truncated_stream = truncated.stream_index_forwards(&truncated_tree, &start_key, ctx);
2090 10 : let mut truncated_stream = std::pin::pin!(truncated_stream);
2091 10 :
2092 10 : let mut scratch_left = Vec::new();
2093 10 : let mut scratch_right = Vec::new();
2094 :
2095 : loop {
2096 42 : let (src, truncated) = (source_stream.try_next(), truncated_stream.try_next());
2097 42 : let (src, truncated) = tokio::try_join!(src, truncated).unwrap();
2098 42 :
2099 42 : if src.is_none() {
2100 10 : assert!(truncated.is_none());
2101 10 : break;
2102 32 : }
2103 32 :
2104 32 : let (src, truncated) = (src.unwrap(), truncated.unwrap());
2105 32 :
2106 32 : // because we've filtered the source with Lsn, we should always have the same keys from both.
2107 32 : assert_eq!(src.0, truncated.0);
2108 32 : assert_eq!(src.1, truncated.1);
2109 :
2110 : // if this is needed for something else, just drop this assert.
2111 32 : assert!(
2112 32 : src.2.pos() >= truncated.2.pos(),
2113 0 : "value position should not go backwards {} vs. {}",
2114 0 : src.2.pos(),
2115 0 : truncated.2.pos()
2116 : );
2117 :
2118 32 : scratch_left.clear();
2119 32 : let src_cursor = source_reader.block_cursor();
2120 32 : let left = src_cursor.read_blob_into_buf(src.2.pos(), &mut scratch_left, ctx);
2121 32 : scratch_right.clear();
2122 32 : let trunc_cursor = truncated_reader.block_cursor();
2123 32 : let right = trunc_cursor.read_blob_into_buf(truncated.2.pos(), &mut scratch_right, ctx);
2124 :
2125 32 : tokio::try_join!(left, right).unwrap();
2126 32 :
2127 32 : assert_eq!(utils::Hex(&scratch_left), utils::Hex(&scratch_right));
2128 : }
2129 10 : }
2130 : }
|