Line data Source code
1 : use std::collections::{HashMap, HashSet};
2 :
3 : use crate::checks::{
4 : branch_cleanup_and_check_errors, list_timeline_blobs, BlobDataParseResult,
5 : RemoteTimelineBlobData, TenantObjectListing, TimelineAnalysis,
6 : };
7 : use crate::metadata_stream::{stream_tenant_timelines, stream_tenants};
8 : use crate::{init_remote, BucketConfig, NodeKind, RootTarget, TenantShardTimelineId};
9 : use futures_util::{StreamExt, TryStreamExt};
10 : use pageserver::tenant::remote_timeline_client::remote_layer_path;
11 : use pageserver_api::controller_api::MetadataHealthUpdateRequest;
12 : use pageserver_api::shard::TenantShardId;
13 : use remote_storage::GenericRemoteStorage;
14 : use serde::Serialize;
15 : use tracing::{info_span, Instrument};
16 : use utils::id::TenantId;
17 : use utils::shard::ShardCount;
18 :
19 : #[derive(Serialize, Default)]
20 : pub struct MetadataSummary {
21 : tenant_count: usize,
22 : timeline_count: usize,
23 : timeline_shard_count: usize,
24 : /// Tenant-shard timeline (key) mapping to errors. The key has to be a string because it will be serialized to a JSON.
25 : /// The key is generated using `TenantShardTimelineId::to_string()`.
26 : with_errors: HashMap<String, Vec<String>>,
27 : /// Tenant-shard timeline (key) mapping to warnings. The key has to be a string because it will be serialized to a JSON.
28 : /// The key is generated using `TenantShardTimelineId::to_string()`.
29 : with_warnings: HashMap<String, Vec<String>>,
30 : with_orphans: HashSet<TenantShardTimelineId>,
31 : indices_by_version: HashMap<usize, usize>,
32 :
33 : #[serde(skip)]
34 : pub(crate) healthy_tenant_shards: HashSet<TenantShardId>,
35 : #[serde(skip)]
36 : pub(crate) unhealthy_tenant_shards: HashSet<TenantShardId>,
37 : }
38 :
39 : impl MetadataSummary {
40 0 : fn new() -> Self {
41 0 : Self::default()
42 0 : }
43 :
44 0 : fn update_data(&mut self, data: &RemoteTimelineBlobData) {
45 0 : self.timeline_shard_count += 1;
46 : if let BlobDataParseResult::Parsed {
47 0 : index_part,
48 : index_part_generation: _,
49 : s3_layers: _,
50 : index_part_last_modified_time: _,
51 : index_part_snapshot_time: _,
52 0 : } = &data.blob_data
53 0 : {
54 0 : *self
55 0 : .indices_by_version
56 0 : .entry(index_part.version())
57 0 : .or_insert(0) += 1;
58 0 : }
59 0 : }
60 :
61 0 : fn update_analysis(
62 0 : &mut self,
63 0 : id: &TenantShardTimelineId,
64 0 : analysis: &TimelineAnalysis,
65 0 : verbose: bool,
66 0 : ) {
67 0 : if analysis.is_healthy() {
68 0 : self.healthy_tenant_shards.insert(id.tenant_shard_id);
69 0 : } else {
70 0 : self.healthy_tenant_shards.remove(&id.tenant_shard_id);
71 0 : self.unhealthy_tenant_shards.insert(id.tenant_shard_id);
72 0 : }
73 :
74 0 : if !analysis.errors.is_empty() {
75 0 : let entry = self.with_errors.entry(id.to_string()).or_default();
76 0 : if verbose {
77 0 : entry.extend(analysis.errors.iter().cloned());
78 0 : }
79 0 : }
80 :
81 0 : if !analysis.warnings.is_empty() {
82 0 : let entry = self.with_warnings.entry(id.to_string()).or_default();
83 0 : if verbose {
84 0 : entry.extend(analysis.warnings.iter().cloned());
85 0 : }
86 0 : }
87 0 : }
88 :
89 0 : fn notify_timeline_orphan(&mut self, ttid: &TenantShardTimelineId) {
90 0 : self.with_orphans.insert(*ttid);
91 0 : }
92 :
93 : /// Long-form output for printing at end of a scan
94 0 : pub fn summary_string(&self) -> String {
95 0 : let version_summary: String = itertools::join(
96 0 : self.indices_by_version
97 0 : .iter()
98 0 : .map(|(k, v)| format!("{k}: {v}")),
99 0 : ", ",
100 0 : );
101 0 :
102 0 : format!(
103 0 : "Tenants: {}
104 0 : Timelines: {}
105 0 : Timeline-shards: {}
106 0 : With errors: {}
107 0 : With warnings: {}
108 0 : With orphan layers: {}
109 0 : Index versions: {version_summary}
110 0 : ",
111 0 : self.tenant_count,
112 0 : self.timeline_count,
113 0 : self.timeline_shard_count,
114 0 : self.with_errors.len(),
115 0 : self.with_warnings.len(),
116 0 : self.with_orphans.len(),
117 0 : )
118 0 : }
119 :
120 0 : pub fn is_fatal(&self) -> bool {
121 0 : !self.with_errors.is_empty()
122 0 : }
123 :
124 0 : pub fn is_empty(&self) -> bool {
125 0 : self.timeline_shard_count == 0
126 0 : }
127 :
128 0 : pub fn build_health_update_request(&self) -> MetadataHealthUpdateRequest {
129 0 : MetadataHealthUpdateRequest {
130 0 : healthy_tenant_shards: self.healthy_tenant_shards.clone(),
131 0 : unhealthy_tenant_shards: self.unhealthy_tenant_shards.clone(),
132 0 : }
133 0 : }
134 : }
135 :
136 : /// Scan the pageserver metadata in an S3 bucket, reporting errors and statistics.
137 0 : pub async fn scan_pageserver_metadata(
138 0 : bucket_config: BucketConfig,
139 0 : tenant_ids: Vec<TenantShardId>,
140 0 : verbose: bool,
141 0 : ) -> anyhow::Result<MetadataSummary> {
142 0 : let (remote_client, target) = init_remote(bucket_config, NodeKind::Pageserver).await?;
143 :
144 0 : let tenants = if tenant_ids.is_empty() {
145 0 : futures::future::Either::Left(stream_tenants(&remote_client, &target))
146 : } else {
147 0 : futures::future::Either::Right(futures::stream::iter(tenant_ids.into_iter().map(Ok)))
148 : };
149 :
150 : // How many tenants to process in parallel. We need to be mindful of pageservers
151 : // accessing the same per tenant prefixes, so use a lower setting than pageservers.
152 : const CONCURRENCY: usize = 32;
153 :
154 : // Generate a stream of TenantTimelineId
155 0 : let timelines = tenants.map_ok(|t| stream_tenant_timelines(&remote_client, &target, t));
156 0 : let timelines = timelines.try_buffered(CONCURRENCY);
157 0 : let timelines = timelines.try_flatten();
158 :
159 : // Generate a stream of S3TimelineBlobData
160 0 : async fn report_on_timeline(
161 0 : remote_client: &GenericRemoteStorage,
162 0 : target: &RootTarget,
163 0 : ttid: TenantShardTimelineId,
164 0 : ) -> anyhow::Result<(TenantShardTimelineId, RemoteTimelineBlobData)> {
165 0 : let data = list_timeline_blobs(remote_client, ttid, target).await?;
166 0 : Ok((ttid, data))
167 0 : }
168 0 : let timelines = timelines.map_ok(|ttid| report_on_timeline(&remote_client, &target, ttid));
169 0 : let mut timelines = std::pin::pin!(timelines.try_buffered(CONCURRENCY));
170 0 :
171 0 : // We must gather all the TenantShardTimelineId->S3TimelineBlobData for each tenant, because different
172 0 : // shards in the same tenant might refer to one anothers' keys if a shard split has happened.
173 0 :
174 0 : let mut tenant_id = None;
175 0 : let mut tenant_objects = TenantObjectListing::default();
176 0 : let mut tenant_timeline_results = Vec::new();
177 :
178 0 : async fn analyze_tenant(
179 0 : remote_client: &GenericRemoteStorage,
180 0 : tenant_id: TenantId,
181 0 : summary: &mut MetadataSummary,
182 0 : mut tenant_objects: TenantObjectListing,
183 0 : timelines: Vec<(TenantShardTimelineId, RemoteTimelineBlobData)>,
184 0 : highest_shard_count: ShardCount,
185 0 : verbose: bool,
186 0 : ) {
187 0 : summary.tenant_count += 1;
188 0 :
189 0 : let mut timeline_ids = HashSet::new();
190 0 : let mut timeline_generations = HashMap::new();
191 0 : for (ttid, data) in timelines {
192 0 : async {
193 0 : if ttid.tenant_shard_id.shard_count == highest_shard_count {
194 : // Only analyze `TenantShardId`s with highest shard count.
195 :
196 : // Stash the generation of each timeline, for later use identifying orphan layers
197 : if let BlobDataParseResult::Parsed {
198 0 : index_part,
199 0 : index_part_generation,
200 : s3_layers: _,
201 : index_part_last_modified_time: _,
202 : index_part_snapshot_time: _,
203 0 : } = &data.blob_data
204 : {
205 0 : if index_part.deleted_at.is_some() {
206 : // skip deleted timeline.
207 0 : tracing::info!(
208 0 : "Skip analysis of {} b/c timeline is already deleted",
209 : ttid
210 : );
211 0 : return;
212 0 : }
213 0 : timeline_generations.insert(ttid, *index_part_generation);
214 0 : }
215 :
216 : // Apply checks to this timeline shard's metadata, and in the process update `tenant_objects`
217 : // reference counts for layers across the tenant.
218 0 : let analysis = branch_cleanup_and_check_errors(
219 0 : remote_client,
220 0 : &ttid,
221 0 : &mut tenant_objects,
222 0 : None,
223 0 : None,
224 0 : Some(data),
225 0 : )
226 0 : .await;
227 0 : summary.update_analysis(&ttid, &analysis, verbose);
228 0 :
229 0 : timeline_ids.insert(ttid.timeline_id);
230 : } else {
231 0 : tracing::info!(
232 0 : "Skip analysis of {} b/c a lower shard count than {}",
233 : ttid,
234 : highest_shard_count.0,
235 : );
236 : }
237 0 : }
238 : .instrument(
239 0 : info_span!("analyze-timeline", shard = %ttid.tenant_shard_id.shard_slug(), timeline = %ttid.timeline_id),
240 : )
241 0 : .await
242 : }
243 :
244 0 : summary.timeline_count += timeline_ids.len();
245 :
246 : // Identifying orphan layers must be done on a tenant-wide basis, because individual
247 : // shards' layers may be referenced by other shards.
248 : //
249 : // Orphan layers are not a corruption, and not an indication of a problem. They are just
250 : // consuming some space in remote storage, and may be cleaned up at leisure.
251 0 : for (shard_index, timeline_id, layer_file, generation) in tenant_objects.get_orphans() {
252 0 : let ttid = TenantShardTimelineId {
253 0 : tenant_shard_id: TenantShardId {
254 0 : tenant_id,
255 0 : shard_count: shard_index.shard_count,
256 0 : shard_number: shard_index.shard_number,
257 0 : },
258 0 : timeline_id,
259 0 : };
260 :
261 0 : if let Some(timeline_generation) = timeline_generations.get(&ttid) {
262 0 : if &generation >= timeline_generation {
263 : // Candidate orphan layer is in the current or future generation relative
264 : // to the index we read for this timeline shard, so its absence from the index
265 : // doesn't make it an orphan: more likely, it is a case where the layer was
266 : // uploaded, but the index referencing the layer wasn't written yet.
267 0 : continue;
268 0 : }
269 0 : }
270 :
271 0 : let orphan_path = remote_layer_path(
272 0 : &tenant_id,
273 0 : &timeline_id,
274 0 : shard_index,
275 0 : &layer_file,
276 0 : generation,
277 0 : );
278 0 :
279 0 : tracing::info!("Orphan layer detected: {orphan_path}");
280 :
281 0 : summary.notify_timeline_orphan(&ttid);
282 : }
283 0 : }
284 :
285 : // Iterate through all the timeline results. These are in key-order, so
286 : // all results for the same tenant will be adjacent. We accumulate these,
287 : // and then call `analyze_tenant` to flush, when we see the next tenant ID.
288 0 : let mut summary = MetadataSummary::new();
289 0 : let mut highest_shard_count = ShardCount::MIN;
290 0 : while let Some(i) = timelines.next().await {
291 0 : let (ttid, data) = i?;
292 0 : summary.update_data(&data);
293 0 :
294 0 : match tenant_id {
295 0 : Some(prev_tenant_id) => {
296 0 : if prev_tenant_id != ttid.tenant_shard_id.tenant_id {
297 : // New tenant: analyze this tenant's timelines, clear accumulated tenant_timeline_results
298 0 : let tenant_objects = std::mem::take(&mut tenant_objects);
299 0 : let timelines = std::mem::take(&mut tenant_timeline_results);
300 0 : analyze_tenant(
301 0 : &remote_client,
302 0 : prev_tenant_id,
303 0 : &mut summary,
304 0 : tenant_objects,
305 0 : timelines,
306 0 : highest_shard_count,
307 0 : verbose,
308 0 : )
309 0 : .instrument(info_span!("analyze-tenant", tenant = %prev_tenant_id))
310 0 : .await;
311 0 : tenant_id = Some(ttid.tenant_shard_id.tenant_id);
312 0 : highest_shard_count = ttid.tenant_shard_id.shard_count;
313 0 : } else {
314 0 : highest_shard_count = highest_shard_count.max(ttid.tenant_shard_id.shard_count);
315 0 : }
316 : }
317 0 : None => {
318 0 : tenant_id = Some(ttid.tenant_shard_id.tenant_id);
319 0 : highest_shard_count = highest_shard_count.max(ttid.tenant_shard_id.shard_count);
320 0 : }
321 : }
322 :
323 0 : match &data.blob_data {
324 : BlobDataParseResult::Parsed {
325 : index_part: _,
326 0 : index_part_generation: _index_part_generation,
327 0 : s3_layers,
328 0 : index_part_last_modified_time: _,
329 0 : index_part_snapshot_time: _,
330 0 : } => {
331 0 : tenant_objects.push(ttid, s3_layers.clone());
332 0 : }
333 0 : BlobDataParseResult::Relic => (),
334 : BlobDataParseResult::Incorrect {
335 : errors: _,
336 0 : s3_layers,
337 0 : } => {
338 0 : tenant_objects.push(ttid, s3_layers.clone());
339 0 : }
340 : }
341 0 : tenant_timeline_results.push((ttid, data));
342 : }
343 :
344 0 : if !tenant_timeline_results.is_empty() {
345 0 : let tenant_id = tenant_id.expect("Must be set if results are present");
346 0 : analyze_tenant(
347 0 : &remote_client,
348 0 : tenant_id,
349 0 : &mut summary,
350 0 : tenant_objects,
351 0 : tenant_timeline_results,
352 0 : highest_shard_count,
353 0 : verbose,
354 0 : )
355 0 : .instrument(info_span!("analyze-tenant", tenant = %tenant_id))
356 0 : .await;
357 0 : }
358 :
359 0 : Ok(summary)
360 0 : }
|
