Line data Source code
1 : use std::collections::{HashMap, HashSet};
2 :
3 : use anyhow::Context;
4 : use aws_sdk_s3::Client;
5 : use pageserver::tenant::layer_map::LayerMap;
6 : use pageserver::tenant::remote_timeline_client::index::LayerFileMetadata;
7 : use pageserver_api::shard::ShardIndex;
8 : use tracing::{error, info, warn};
9 : use utils::generation::Generation;
10 : use utils::id::TimelineId;
11 :
12 : use crate::cloud_admin_api::BranchData;
13 : use crate::metadata_stream::stream_listing;
14 : use crate::{download_object_with_retries, RootTarget, TenantShardTimelineId};
15 : use futures_util::StreamExt;
16 : use pageserver::tenant::remote_timeline_client::{parse_remote_index_path, remote_layer_path};
17 : use pageserver::tenant::storage_layer::LayerName;
18 : use pageserver::tenant::IndexPart;
19 : use remote_storage::RemotePath;
20 :
21 : pub(crate) struct TimelineAnalysis {
22 : /// Anomalies detected
23 : pub(crate) errors: Vec<String>,
24 :
25 : /// Healthy-but-noteworthy, like old-versioned structures that are readable but
26 : /// worth reporting for awareness that we must not remove that old version decoding
27 : /// yet.
28 : pub(crate) warnings: Vec<String>,
29 :
30 : /// Keys not referenced in metadata: candidates for removal, but NOT NECESSARILY: beware
31 : /// of races between reading the metadata and reading the objects.
32 : pub(crate) garbage_keys: Vec<String>,
33 : }
34 :
35 : impl TimelineAnalysis {
36 0 : fn new() -> Self {
37 0 : Self {
38 0 : errors: Vec::new(),
39 0 : warnings: Vec::new(),
40 0 : garbage_keys: Vec::new(),
41 0 : }
42 0 : }
43 :
44 : /// Whether a timeline is healthy.
45 0 : pub(crate) fn is_healthy(&self) -> bool {
46 0 : self.errors.is_empty() && self.warnings.is_empty()
47 0 : }
48 : }
49 :
50 0 : pub(crate) async fn branch_cleanup_and_check_errors(
51 0 : s3_client: &Client,
52 0 : target: &RootTarget,
53 0 : id: &TenantShardTimelineId,
54 0 : tenant_objects: &mut TenantObjectListing,
55 0 : s3_active_branch: Option<&BranchData>,
56 0 : console_branch: Option<BranchData>,
57 0 : s3_data: Option<S3TimelineBlobData>,
58 0 : ) -> TimelineAnalysis {
59 0 : let mut result = TimelineAnalysis::new();
60 0 :
61 0 : info!("Checking timeline {id}");
62 :
63 0 : if let Some(s3_active_branch) = s3_active_branch {
64 0 : info!(
65 0 : "Checking console status for timeline for branch {:?}/{:?}",
66 : s3_active_branch.project_id, s3_active_branch.id
67 : );
68 0 : match console_branch {
69 0 : Some(_) => {result.errors.push(format!("Timeline has deleted branch data in the console (id = {:?}, project_id = {:?}), recheck whether it got removed during the check",
70 0 : s3_active_branch.id, s3_active_branch.project_id))
71 : },
72 : None => {
73 0 : result.errors.push(format!("Timeline has no branch data in the console (id = {:?}, project_id = {:?}), recheck whether it got removed during the check",
74 0 : s3_active_branch.id, s3_active_branch.project_id))
75 : }
76 : };
77 0 : }
78 :
79 0 : match s3_data {
80 0 : Some(s3_data) => {
81 0 : result.garbage_keys.extend(s3_data.unknown_keys);
82 0 :
83 0 : match s3_data.blob_data {
84 : BlobDataParseResult::Parsed {
85 0 : index_part,
86 0 : index_part_generation: _index_part_generation,
87 0 : s3_layers: _s3_layers,
88 0 : } => {
89 0 : if !IndexPart::KNOWN_VERSIONS.contains(&index_part.version()) {
90 0 : result
91 0 : .errors
92 0 : .push(format!("index_part.json version: {}", index_part.version()))
93 0 : }
94 :
95 0 : let mut newest_versions = IndexPart::KNOWN_VERSIONS.iter().rev().take(2);
96 0 : if !newest_versions.any(|ip| ip == &index_part.version()) {
97 0 : info!(
98 0 : "index_part.json version is not latest: {}",
99 0 : index_part.version()
100 : );
101 0 : }
102 :
103 0 : if index_part.metadata.disk_consistent_lsn()
104 0 : != index_part.duplicated_disk_consistent_lsn()
105 : {
106 : // Tech debt: let's get rid of one of these, they are redundant
107 : // https://github.com/neondatabase/neon/issues/8343
108 0 : result.errors.push(format!(
109 0 : "Mismatching disk_consistent_lsn in TimelineMetadata ({}) and in the index_part ({})",
110 0 : index_part.metadata.disk_consistent_lsn(),
111 0 : index_part.duplicated_disk_consistent_lsn(),
112 0 : ))
113 0 : }
114 :
115 0 : if index_part.layer_metadata.is_empty() {
116 0 : if index_part.metadata.ancestor_timeline().is_none() {
117 0 : // The initial timeline with no ancestor should ALWAYS have layers.
118 0 : result.errors.push(
119 0 : "index_part.json has no layers (ancestor_timeline=None)"
120 0 : .to_string(),
121 0 : );
122 0 : } else {
123 : // Not an error, can happen for branches with zero writes, but notice that
124 0 : info!("index_part.json has no layers (ancestor_timeline exists)");
125 : }
126 0 : }
127 :
128 0 : for (layer, metadata) in index_part.layer_metadata {
129 0 : if metadata.file_size == 0 {
130 0 : result.errors.push(format!(
131 0 : "index_part.json contains a layer {} that has 0 size in its layer metadata", layer,
132 0 : ))
133 0 : }
134 :
135 0 : if !tenant_objects.check_ref(id.timeline_id, &layer, &metadata) {
136 0 : let path = remote_layer_path(
137 0 : &id.tenant_shard_id.tenant_id,
138 0 : &id.timeline_id,
139 0 : metadata.shard,
140 0 : &layer,
141 0 : metadata.generation,
142 0 : );
143 :
144 : // HEAD request used here to address a race condition when an index was uploaded concurrently
145 : // with our scan. We check if the object is uploaded to S3 after taking the listing snapshot.
146 0 : let response = s3_client
147 0 : .head_object()
148 0 : .bucket(target.bucket_name())
149 0 : .key(path.get_path().as_str())
150 0 : .send()
151 0 : .await;
152 :
153 0 : if response.is_err() {
154 : // Object is not present.
155 0 : let is_l0 = LayerMap::is_l0(layer.key_range());
156 0 :
157 0 : let msg = format!(
158 0 : "index_part.json contains a layer {}{} (shard {}) that is not present in remote storage (layer_is_l0: {})",
159 0 : layer,
160 0 : metadata.generation.get_suffix(),
161 0 : metadata.shard,
162 0 : is_l0,
163 0 : );
164 0 :
165 0 : if is_l0 {
166 0 : result.warnings.push(msg);
167 0 : } else {
168 0 : result.errors.push(msg);
169 0 : }
170 0 : }
171 0 : }
172 : }
173 : }
174 0 : BlobDataParseResult::Relic => {}
175 0 : BlobDataParseResult::Incorrect(parse_errors) => result.errors.extend(
176 0 : parse_errors
177 0 : .into_iter()
178 0 : .map(|error| format!("parse error: {error}")),
179 0 : ),
180 : }
181 : }
182 0 : None => result
183 0 : .errors
184 0 : .push("Timeline has no data on S3 at all".to_string()),
185 : }
186 :
187 0 : if result.errors.is_empty() {
188 0 : info!("No check errors found");
189 : } else {
190 0 : warn!("Timeline metadata errors: {0:?}", result.errors);
191 : }
192 :
193 0 : if !result.warnings.is_empty() {
194 0 : warn!("Timeline metadata warnings: {0:?}", result.warnings);
195 0 : }
196 :
197 0 : if !result.garbage_keys.is_empty() {
198 0 : error!(
199 0 : "The following keys should be removed from S3: {0:?}",
200 : result.garbage_keys
201 : )
202 0 : }
203 :
204 0 : result
205 0 : }
206 :
207 : #[derive(Default)]
208 : pub(crate) struct LayerRef {
209 : ref_count: usize,
210 : }
211 :
212 : /// Top-level index of objects in a tenant. This may be used by any shard-timeline within
213 : /// the tenant to query whether an object exists.
214 : #[derive(Default)]
215 : pub(crate) struct TenantObjectListing {
216 : shard_timelines: HashMap<(ShardIndex, TimelineId), HashMap<(LayerName, Generation), LayerRef>>,
217 : }
218 :
219 : impl TenantObjectListing {
220 : /// Having done an S3 listing of the keys within a timeline prefix, merge them into the overall
221 : /// list of layer keys for the Tenant.
222 0 : pub(crate) fn push(
223 0 : &mut self,
224 0 : ttid: TenantShardTimelineId,
225 0 : layers: HashSet<(LayerName, Generation)>,
226 0 : ) {
227 0 : let shard_index = ShardIndex::new(
228 0 : ttid.tenant_shard_id.shard_number,
229 0 : ttid.tenant_shard_id.shard_count,
230 0 : );
231 0 : let replaced = self.shard_timelines.insert(
232 0 : (shard_index, ttid.timeline_id),
233 0 : layers
234 0 : .into_iter()
235 0 : .map(|l| (l, LayerRef::default()))
236 0 : .collect(),
237 0 : );
238 0 :
239 0 : assert!(
240 0 : replaced.is_none(),
241 0 : "Built from an S3 object listing, which should never repeat a key"
242 : );
243 0 : }
244 :
245 : /// Having loaded a timeline index, check if a layer referenced by the index exists. If it does,
246 : /// the layer's refcount will be incremented. Later, after calling this for all references in all indices
247 : /// in a tenant, orphan layers may be detected by their zero refcounts.
248 : ///
249 : /// Returns true if the layer exists
250 0 : pub(crate) fn check_ref(
251 0 : &mut self,
252 0 : timeline_id: TimelineId,
253 0 : layer_file: &LayerName,
254 0 : metadata: &LayerFileMetadata,
255 0 : ) -> bool {
256 0 : let Some(shard_tl) = self.shard_timelines.get_mut(&(metadata.shard, timeline_id)) else {
257 0 : return false;
258 : };
259 :
260 0 : let Some(layer_ref) = shard_tl.get_mut(&(layer_file.clone(), metadata.generation)) else {
261 0 : return false;
262 : };
263 :
264 0 : layer_ref.ref_count += 1;
265 0 :
266 0 : true
267 0 : }
268 :
269 0 : pub(crate) fn get_orphans(&self) -> Vec<(ShardIndex, TimelineId, LayerName, Generation)> {
270 0 : let mut result = Vec::new();
271 0 : for ((shard_index, timeline_id), layers) in &self.shard_timelines {
272 0 : for ((layer_file, generation), layer_ref) in layers {
273 0 : if layer_ref.ref_count == 0 {
274 0 : result.push((*shard_index, *timeline_id, layer_file.clone(), *generation))
275 0 : }
276 : }
277 : }
278 :
279 0 : result
280 0 : }
281 : }
282 :
283 : #[derive(Debug)]
284 : pub(crate) struct S3TimelineBlobData {
285 : pub(crate) blob_data: BlobDataParseResult,
286 :
287 : // Index objects that were not used when loading `blob_data`, e.g. those from old generations
288 : pub(crate) unused_index_keys: Vec<String>,
289 :
290 : // Objects whose keys were not recognized at all, i.e. not layer files, not indices
291 : pub(crate) unknown_keys: Vec<String>,
292 : }
293 :
294 : #[derive(Debug)]
295 : pub(crate) enum BlobDataParseResult {
296 : Parsed {
297 : index_part: Box<IndexPart>,
298 : index_part_generation: Generation,
299 : s3_layers: HashSet<(LayerName, Generation)>,
300 : },
301 : /// The remains of a deleted Timeline (i.e. an initdb archive only)
302 : Relic,
303 : Incorrect(Vec<String>),
304 : }
305 :
306 0 : pub(crate) fn parse_layer_object_name(name: &str) -> Result<(LayerName, Generation), String> {
307 0 : match name.rsplit_once('-') {
308 : // FIXME: this is gross, just use a regex?
309 0 : Some((layer_filename, gen)) if gen.len() == 8 => {
310 0 : let layer = layer_filename.parse::<LayerName>()?;
311 0 : let gen =
312 0 : Generation::parse_suffix(gen).ok_or("Malformed generation suffix".to_string())?;
313 0 : Ok((layer, gen))
314 : }
315 0 : _ => Ok((name.parse::<LayerName>()?, Generation::none())),
316 : }
317 0 : }
318 :
319 0 : pub(crate) async fn list_timeline_blobs(
320 0 : s3_client: &Client,
321 0 : id: TenantShardTimelineId,
322 0 : s3_root: &RootTarget,
323 0 : ) -> anyhow::Result<S3TimelineBlobData> {
324 0 : let mut s3_layers = HashSet::new();
325 0 :
326 0 : let mut errors = Vec::new();
327 0 : let mut unknown_keys = Vec::new();
328 0 :
329 0 : let mut timeline_dir_target = s3_root.timeline_root(&id);
330 0 : timeline_dir_target.delimiter = String::new();
331 0 :
332 0 : let mut index_part_keys: Vec<String> = Vec::new();
333 0 : let mut initdb_archive: bool = false;
334 0 :
335 0 : let mut stream = std::pin::pin!(stream_listing(s3_client, &timeline_dir_target));
336 0 : while let Some(obj) = stream.next().await {
337 0 : let obj = obj?;
338 0 : let key = obj.key();
339 0 :
340 0 : let blob_name = key.strip_prefix(&timeline_dir_target.prefix_in_bucket);
341 0 : match blob_name {
342 0 : Some(name) if name.starts_with("index_part.json") => {
343 0 : tracing::debug!("Index key {key}");
344 0 : index_part_keys.push(key.to_owned())
345 : }
346 0 : Some("initdb.tar.zst") => {
347 0 : tracing::debug!("initdb archive {key}");
348 0 : initdb_archive = true;
349 : }
350 0 : Some("initdb-preserved.tar.zst") => {
351 0 : tracing::info!("initdb archive preserved {key}");
352 : }
353 0 : Some(maybe_layer_name) => match parse_layer_object_name(maybe_layer_name) {
354 0 : Ok((new_layer, gen)) => {
355 0 : tracing::debug!("Parsed layer key: {} {:?}", new_layer, gen);
356 0 : s3_layers.insert((new_layer, gen));
357 : }
358 0 : Err(e) => {
359 0 : tracing::info!("Error parsing key {maybe_layer_name}");
360 0 : errors.push(
361 0 : format!("S3 list response got an object with key {key} that is not a layer name: {e}"),
362 0 : );
363 0 : unknown_keys.push(key.to_string());
364 : }
365 : },
366 : None => {
367 0 : tracing::warn!("Unknown key {}", key);
368 0 : errors.push(format!("S3 list response got an object with odd key {key}"));
369 0 : unknown_keys.push(key.to_string());
370 : }
371 : }
372 : }
373 :
374 0 : if index_part_keys.is_empty() && s3_layers.is_empty() && initdb_archive {
375 0 : tracing::debug!(
376 0 : "Timeline is empty apart from initdb archive: expected post-deletion state."
377 : );
378 0 : return Ok(S3TimelineBlobData {
379 0 : blob_data: BlobDataParseResult::Relic,
380 0 : unused_index_keys: index_part_keys,
381 0 : unknown_keys: Vec::new(),
382 0 : });
383 0 : }
384 :
385 : // Choose the index_part with the highest generation
386 0 : let (index_part_object, index_part_generation) = match index_part_keys
387 0 : .iter()
388 0 : .filter_map(|key| {
389 0 : // Stripping the index key to the last part, because RemotePath doesn't
390 0 : // like absolute paths, and depending on prefix_in_bucket it's possible
391 0 : // for the keys we read back to start with a slash.
392 0 : let basename = key.rsplit_once('/').unwrap().1;
393 0 : parse_remote_index_path(RemotePath::from_string(basename).unwrap()).map(|g| (key, g))
394 0 : })
395 0 : .max_by_key(|i| i.1)
396 0 : .map(|(k, g)| (k.clone(), g))
397 : {
398 0 : Some((key, gen)) => (Some(key), gen),
399 : None => {
400 : // Legacy/missing case: one or zero index parts, which did not have a generation
401 0 : (index_part_keys.pop(), Generation::none())
402 : }
403 : };
404 :
405 0 : match index_part_object.as_ref() {
406 0 : Some(selected) => index_part_keys.retain(|k| k != selected),
407 0 : None => {
408 0 : errors.push("S3 list response got no index_part.json file".to_string());
409 0 : }
410 : }
411 :
412 0 : if let Some(index_part_object_key) = index_part_object.as_ref() {
413 0 : let index_part_bytes = download_object_with_retries(
414 0 : s3_client,
415 0 : &timeline_dir_target.bucket_name,
416 0 : index_part_object_key,
417 0 : )
418 0 : .await
419 0 : .context("index_part.json download")?;
420 :
421 0 : match serde_json::from_slice(&index_part_bytes) {
422 0 : Ok(index_part) => {
423 0 : return Ok(S3TimelineBlobData {
424 0 : blob_data: BlobDataParseResult::Parsed {
425 0 : index_part: Box::new(index_part),
426 0 : index_part_generation,
427 0 : s3_layers,
428 0 : },
429 0 : unused_index_keys: index_part_keys,
430 0 : unknown_keys,
431 0 : })
432 : }
433 0 : Err(index_parse_error) => errors.push(format!(
434 0 : "index_part.json body parsing error: {index_parse_error}"
435 0 : )),
436 : }
437 0 : }
438 :
439 0 : if errors.is_empty() {
440 0 : errors.push(
441 0 : "Unexpected: no errors did not lead to a successfully parsed blob return".to_string(),
442 0 : );
443 0 : }
444 :
445 0 : Ok(S3TimelineBlobData {
446 0 : blob_data: BlobDataParseResult::Incorrect(errors),
447 0 : unused_index_keys: index_part_keys,
448 0 : unknown_keys,
449 0 : })
450 0 : }
|
