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 : with_errors: HashSet<TenantShardTimelineId>,
25 : with_warnings: HashSet<TenantShardTimelineId>,
26 : with_orphans: HashSet<TenantShardTimelineId>,
27 : indices_by_version: HashMap<usize, usize>,
28 :
29 : #[serde(skip)]
30 : pub(crate) healthy_tenant_shards: HashSet<TenantShardId>,
31 : #[serde(skip)]
32 : pub(crate) unhealthy_tenant_shards: HashSet<TenantShardId>,
33 : }
34 :
35 : impl MetadataSummary {
36 0 : fn new() -> Self {
37 0 : Self::default()
38 0 : }
39 :
40 0 : fn update_data(&mut self, data: &RemoteTimelineBlobData) {
41 0 : self.timeline_shard_count += 1;
42 : if let BlobDataParseResult::Parsed {
43 0 : index_part,
44 : index_part_generation: _,
45 : s3_layers: _,
46 0 : } = &data.blob_data
47 0 : {
48 0 : *self
49 0 : .indices_by_version
50 0 : .entry(index_part.version())
51 0 : .or_insert(0) += 1;
52 0 : }
53 0 : }
54 :
55 0 : fn update_analysis(&mut self, id: &TenantShardTimelineId, analysis: &TimelineAnalysis) {
56 0 : if analysis.is_healthy() {
57 0 : self.healthy_tenant_shards.insert(id.tenant_shard_id);
58 0 : } else {
59 0 : self.healthy_tenant_shards.remove(&id.tenant_shard_id);
60 0 : self.unhealthy_tenant_shards.insert(id.tenant_shard_id);
61 0 : }
62 :
63 0 : if !analysis.errors.is_empty() {
64 0 : self.with_errors.insert(*id);
65 0 : }
66 :
67 0 : if !analysis.warnings.is_empty() {
68 0 : self.with_warnings.insert(*id);
69 0 : }
70 0 : }
71 :
72 0 : fn notify_timeline_orphan(&mut self, ttid: &TenantShardTimelineId) {
73 0 : self.with_orphans.insert(*ttid);
74 0 : }
75 :
76 : /// Long-form output for printing at end of a scan
77 0 : pub fn summary_string(&self) -> String {
78 0 : let version_summary: String = itertools::join(
79 0 : self.indices_by_version
80 0 : .iter()
81 0 : .map(|(k, v)| format!("{k}: {v}")),
82 0 : ", ",
83 0 : );
84 0 :
85 0 : format!(
86 0 : "Tenants: {}
87 0 : Timelines: {}
88 0 : Timeline-shards: {}
89 0 : With errors: {}
90 0 : With warnings: {}
91 0 : With orphan layers: {}
92 0 : Index versions: {version_summary}
93 0 : ",
94 0 : self.tenant_count,
95 0 : self.timeline_count,
96 0 : self.timeline_shard_count,
97 0 : self.with_errors.len(),
98 0 : self.with_warnings.len(),
99 0 : self.with_orphans.len(),
100 0 : )
101 0 : }
102 :
103 0 : pub fn is_fatal(&self) -> bool {
104 0 : !self.with_errors.is_empty()
105 0 : }
106 :
107 0 : pub fn is_empty(&self) -> bool {
108 0 : self.timeline_shard_count == 0
109 0 : }
110 :
111 0 : pub fn build_health_update_request(&self) -> MetadataHealthUpdateRequest {
112 0 : MetadataHealthUpdateRequest {
113 0 : healthy_tenant_shards: self.healthy_tenant_shards.clone(),
114 0 : unhealthy_tenant_shards: self.unhealthy_tenant_shards.clone(),
115 0 : }
116 0 : }
117 : }
118 :
119 : /// Scan the pageserver metadata in an S3 bucket, reporting errors and statistics.
120 0 : pub async fn scan_pageserver_metadata(
121 0 : bucket_config: BucketConfig,
122 0 : tenant_ids: Vec<TenantShardId>,
123 0 : ) -> anyhow::Result<MetadataSummary> {
124 0 : let (remote_client, target) = init_remote(bucket_config, NodeKind::Pageserver).await?;
125 :
126 0 : let tenants = if tenant_ids.is_empty() {
127 0 : futures::future::Either::Left(stream_tenants(&remote_client, &target))
128 : } else {
129 0 : futures::future::Either::Right(futures::stream::iter(tenant_ids.into_iter().map(Ok)))
130 : };
131 :
132 : // How many tenants to process in parallel. We need to be mindful of pageservers
133 : // accessing the same per tenant prefixes, so use a lower setting than pageservers.
134 : const CONCURRENCY: usize = 32;
135 :
136 : // Generate a stream of TenantTimelineId
137 0 : let timelines = tenants.map_ok(|t| stream_tenant_timelines(&remote_client, &target, t));
138 0 : let timelines = timelines.try_buffered(CONCURRENCY);
139 0 : let timelines = timelines.try_flatten();
140 :
141 : // Generate a stream of S3TimelineBlobData
142 0 : async fn report_on_timeline(
143 0 : remote_client: &GenericRemoteStorage,
144 0 : target: &RootTarget,
145 0 : ttid: TenantShardTimelineId,
146 0 : ) -> anyhow::Result<(TenantShardTimelineId, RemoteTimelineBlobData)> {
147 0 : let data = list_timeline_blobs(remote_client, ttid, target).await?;
148 0 : Ok((ttid, data))
149 0 : }
150 0 : let timelines = timelines.map_ok(|ttid| report_on_timeline(&remote_client, &target, ttid));
151 0 : let mut timelines = std::pin::pin!(timelines.try_buffered(CONCURRENCY));
152 0 :
153 0 : // We must gather all the TenantShardTimelineId->S3TimelineBlobData for each tenant, because different
154 0 : // shards in the same tenant might refer to one anothers' keys if a shard split has happened.
155 0 :
156 0 : let mut tenant_id = None;
157 0 : let mut tenant_objects = TenantObjectListing::default();
158 0 : let mut tenant_timeline_results = Vec::new();
159 :
160 0 : async fn analyze_tenant(
161 0 : remote_client: &GenericRemoteStorage,
162 0 : tenant_id: TenantId,
163 0 : summary: &mut MetadataSummary,
164 0 : mut tenant_objects: TenantObjectListing,
165 0 : timelines: Vec<(TenantShardTimelineId, RemoteTimelineBlobData)>,
166 0 : highest_shard_count: ShardCount,
167 0 : ) {
168 0 : summary.tenant_count += 1;
169 0 :
170 0 : let mut timeline_ids = HashSet::new();
171 0 : let mut timeline_generations = HashMap::new();
172 0 : for (ttid, data) in timelines {
173 0 : async {
174 0 : if ttid.tenant_shard_id.shard_count == highest_shard_count {
175 : // Only analyze `TenantShardId`s with highest shard count.
176 :
177 : // Stash the generation of each timeline, for later use identifying orphan layers
178 : if let BlobDataParseResult::Parsed {
179 0 : index_part,
180 0 : index_part_generation,
181 0 : s3_layers: _s3_layers,
182 0 : } = &data.blob_data
183 : {
184 0 : if index_part.deleted_at.is_some() {
185 : // skip deleted timeline.
186 0 : tracing::info!(
187 0 : "Skip analysis of {} b/c timeline is already deleted",
188 : ttid
189 : );
190 0 : return;
191 0 : }
192 0 : timeline_generations.insert(ttid, *index_part_generation);
193 0 : }
194 :
195 : // Apply checks to this timeline shard's metadata, and in the process update `tenant_objects`
196 : // reference counts for layers across the tenant.
197 0 : let analysis = branch_cleanup_and_check_errors(
198 0 : remote_client,
199 0 : &ttid,
200 0 : &mut tenant_objects,
201 0 : None,
202 0 : None,
203 0 : Some(data),
204 0 : )
205 0 : .await;
206 0 : summary.update_analysis(&ttid, &analysis);
207 0 :
208 0 : timeline_ids.insert(ttid.timeline_id);
209 : } else {
210 0 : tracing::info!(
211 0 : "Skip analysis of {} b/c a lower shard count than {}",
212 : ttid,
213 : highest_shard_count.0,
214 : );
215 : }
216 0 : }
217 : .instrument(
218 0 : info_span!("analyze-timeline", shard = %ttid.tenant_shard_id.shard_slug(), timeline = %ttid.timeline_id),
219 : )
220 0 : .await
221 : }
222 :
223 0 : summary.timeline_count += timeline_ids.len();
224 :
225 : // Identifying orphan layers must be done on a tenant-wide basis, because individual
226 : // shards' layers may be referenced by other shards.
227 : //
228 : // Orphan layers are not a corruption, and not an indication of a problem. They are just
229 : // consuming some space in remote storage, and may be cleaned up at leisure.
230 0 : for (shard_index, timeline_id, layer_file, generation) in tenant_objects.get_orphans() {
231 0 : let ttid = TenantShardTimelineId {
232 0 : tenant_shard_id: TenantShardId {
233 0 : tenant_id,
234 0 : shard_count: shard_index.shard_count,
235 0 : shard_number: shard_index.shard_number,
236 0 : },
237 0 : timeline_id,
238 0 : };
239 :
240 0 : if let Some(timeline_generation) = timeline_generations.get(&ttid) {
241 0 : if &generation >= timeline_generation {
242 : // Candidate orphan layer is in the current or future generation relative
243 : // to the index we read for this timeline shard, so its absence from the index
244 : // doesn't make it an orphan: more likely, it is a case where the layer was
245 : // uploaded, but the index referencing the layer wasn't written yet.
246 0 : continue;
247 0 : }
248 0 : }
249 :
250 0 : let orphan_path = remote_layer_path(
251 0 : &tenant_id,
252 0 : &timeline_id,
253 0 : shard_index,
254 0 : &layer_file,
255 0 : generation,
256 0 : );
257 0 :
258 0 : tracing::info!("Orphan layer detected: {orphan_path}");
259 :
260 0 : summary.notify_timeline_orphan(&ttid);
261 : }
262 0 : }
263 :
264 : // Iterate through all the timeline results. These are in key-order, so
265 : // all results for the same tenant will be adjacent. We accumulate these,
266 : // and then call `analyze_tenant` to flush, when we see the next tenant ID.
267 0 : let mut summary = MetadataSummary::new();
268 0 : let mut highest_shard_count = ShardCount::MIN;
269 0 : while let Some(i) = timelines.next().await {
270 0 : let (ttid, data) = i?;
271 0 : summary.update_data(&data);
272 0 :
273 0 : match tenant_id {
274 0 : None => {
275 0 : tenant_id = Some(ttid.tenant_shard_id.tenant_id);
276 0 : highest_shard_count = highest_shard_count.max(ttid.tenant_shard_id.shard_count);
277 0 : }
278 0 : Some(prev_tenant_id) => {
279 0 : if prev_tenant_id != ttid.tenant_shard_id.tenant_id {
280 : // New tenant: analyze this tenant's timelines, clear accumulated tenant_timeline_results
281 0 : let tenant_objects = std::mem::take(&mut tenant_objects);
282 0 : let timelines = std::mem::take(&mut tenant_timeline_results);
283 0 : analyze_tenant(
284 0 : &remote_client,
285 0 : prev_tenant_id,
286 0 : &mut summary,
287 0 : tenant_objects,
288 0 : timelines,
289 0 : highest_shard_count,
290 0 : )
291 0 : .instrument(info_span!("analyze-tenant", tenant = %prev_tenant_id))
292 0 : .await;
293 0 : tenant_id = Some(ttid.tenant_shard_id.tenant_id);
294 0 : highest_shard_count = ttid.tenant_shard_id.shard_count;
295 0 : } else {
296 0 : highest_shard_count = highest_shard_count.max(ttid.tenant_shard_id.shard_count);
297 0 : }
298 : }
299 : }
300 :
301 0 : match &data.blob_data {
302 : BlobDataParseResult::Parsed {
303 0 : index_part: _index_part,
304 0 : index_part_generation: _index_part_generation,
305 0 : s3_layers,
306 0 : } => {
307 0 : tenant_objects.push(ttid, s3_layers.clone());
308 0 : }
309 0 : BlobDataParseResult::Relic => (),
310 : BlobDataParseResult::Incorrect {
311 : errors: _,
312 0 : s3_layers,
313 0 : } => {
314 0 : tenant_objects.push(ttid, s3_layers.clone());
315 0 : }
316 : }
317 0 : tenant_timeline_results.push((ttid, data));
318 : }
319 :
320 0 : let tenant_id = tenant_id.expect("Must be set if results are present");
321 0 :
322 0 : if !tenant_timeline_results.is_empty() {
323 0 : analyze_tenant(
324 0 : &remote_client,
325 0 : tenant_id,
326 0 : &mut summary,
327 0 : tenant_objects,
328 0 : tenant_timeline_results,
329 0 : highest_shard_count,
330 0 : )
331 0 : .instrument(info_span!("analyze-tenant", tenant = %tenant_id))
332 0 : .await;
333 0 : }
334 :
335 0 : Ok(summary)
336 0 : }
|
