Line data Source code
1 : use std::collections::HashMap;
2 : use std::sync::Arc;
3 :
4 : use anyhow::Context;
5 : use async_stream::stream;
6 : use aws_sdk_s3::Client;
7 : use camino::Utf8PathBuf;
8 : use futures::{StreamExt, TryStreamExt};
9 : use pageserver::tenant::IndexPart;
10 : use pageserver::tenant::remote_timeline_client::index::LayerFileMetadata;
11 : use pageserver::tenant::storage_layer::LayerName;
12 : use pageserver_api::shard::TenantShardId;
13 : use remote_storage::{GenericRemoteStorage, S3Config};
14 : use utils::generation::Generation;
15 : use utils::id::TenantId;
16 :
17 : use crate::checks::{BlobDataParseResult, RemoteTimelineBlobData, list_timeline_blobs};
18 : use crate::metadata_stream::{stream_tenant_shards, stream_tenant_timelines};
19 : use crate::{
20 : BucketConfig, NodeKind, RootTarget, TenantShardTimelineId, download_object_to_file_s3,
21 : init_remote, init_remote_s3,
22 : };
23 :
24 : pub struct SnapshotDownloader {
25 : s3_client: Arc<Client>,
26 : s3_root: RootTarget,
27 : bucket_config: BucketConfig,
28 : bucket_config_s3: S3Config,
29 : tenant_id: TenantId,
30 : output_path: Utf8PathBuf,
31 : concurrency: usize,
32 : }
33 :
34 : impl SnapshotDownloader {
35 0 : pub async fn new(
36 0 : bucket_config: BucketConfig,
37 0 : tenant_id: TenantId,
38 0 : output_path: Utf8PathBuf,
39 0 : concurrency: usize,
40 0 : ) -> anyhow::Result<Self> {
41 0 : let bucket_config_s3 = match &bucket_config.0.storage {
42 0 : remote_storage::RemoteStorageKind::AwsS3(config) => config.clone(),
43 0 : _ => panic!("only S3 configuration is supported for snapshot downloading"),
44 : };
45 0 : let (s3_client, s3_root) =
46 0 : init_remote_s3(bucket_config_s3.clone(), NodeKind::Pageserver).await?;
47 0 : Ok(Self {
48 0 : s3_client,
49 0 : s3_root,
50 0 : bucket_config,
51 0 : bucket_config_s3,
52 0 : tenant_id,
53 0 : output_path,
54 0 : concurrency,
55 0 : })
56 0 : }
57 :
58 0 : async fn download_layer(
59 0 : &self,
60 0 : ttid: TenantShardTimelineId,
61 0 : layer_name: LayerName,
62 0 : layer_metadata: LayerFileMetadata,
63 0 : ) -> anyhow::Result<(LayerName, LayerFileMetadata)> {
64 0 : // Note this is local as in a local copy of S3 data, not local as in the pageserver's local format. They use
65 0 : // different layer names (remote-style has the generation suffix)
66 0 : let local_path = self.output_path.join(format!(
67 0 : "{}/timelines/{}/{}{}",
68 0 : ttid.tenant_shard_id,
69 0 : ttid.timeline_id,
70 0 : layer_name,
71 0 : layer_metadata.generation.get_suffix()
72 0 : ));
73 0 :
74 0 : // We should only be called for layers that are owned by the input TTID
75 0 : assert_eq!(layer_metadata.shard, ttid.tenant_shard_id.to_index());
76 :
77 : // Assumption: we always write layer files atomically, and layer files are immutable. Therefore if the file
78 : // already exists on local disk, we assume it is fully correct and skip it.
79 0 : if tokio::fs::try_exists(&local_path).await? {
80 0 : tracing::debug!("{} already exists", local_path);
81 0 : return Ok((layer_name, layer_metadata));
82 : } else {
83 0 : tracing::debug!("{} requires download...", local_path);
84 :
85 0 : let timeline_root = self.s3_root.timeline_root(&ttid);
86 0 : let remote_layer_path = format!(
87 0 : "{}{}{}",
88 0 : timeline_root.prefix_in_bucket,
89 0 : layer_name,
90 0 : layer_metadata.generation.get_suffix()
91 0 : );
92 :
93 : // List versions: the object might be deleted.
94 0 : let versions = self
95 0 : .s3_client
96 0 : .list_object_versions()
97 0 : .bucket(self.bucket_config_s3.bucket_name.clone())
98 0 : .prefix(&remote_layer_path)
99 0 : .send()
100 0 : .await?;
101 0 : let Some(version) = versions.versions.as_ref().and_then(|v| v.first()) else {
102 0 : return Err(anyhow::anyhow!("No versions found for {remote_layer_path}"));
103 : };
104 0 : download_object_to_file_s3(
105 0 : &self.s3_client,
106 0 : &self.bucket_config_s3.bucket_name,
107 0 : &remote_layer_path,
108 0 : version.version_id.as_deref(),
109 0 : &local_path,
110 0 : )
111 0 : .await?;
112 :
113 0 : tracing::debug!("Downloaded successfully to {local_path}");
114 : }
115 :
116 0 : Ok((layer_name, layer_metadata))
117 0 : }
118 :
119 : /// Download many layers belonging to the same TTID, with some concurrency
120 0 : async fn download_layers(
121 0 : &self,
122 0 : ttid: TenantShardTimelineId,
123 0 : layers: Vec<(LayerName, LayerFileMetadata)>,
124 0 : ) -> anyhow::Result<()> {
125 0 : let layer_count = layers.len();
126 0 : tracing::info!("Downloading {} layers for timeline {ttid}...", layer_count);
127 0 : let layers_stream = stream! {
128 0 : for (layer_name, layer_metadata) in layers {
129 0 : yield self.download_layer(ttid, layer_name, layer_metadata);
130 0 : }
131 0 : };
132 0 :
133 0 : tokio::fs::create_dir_all(self.output_path.join(format!(
134 0 : "{}/timelines/{}",
135 0 : ttid.tenant_shard_id, ttid.timeline_id
136 0 : )))
137 0 : .await?;
138 :
139 0 : let layer_results = layers_stream.buffered(self.concurrency);
140 0 : let mut layer_results = std::pin::pin!(layer_results);
141 0 :
142 0 : let mut err = None;
143 0 : let mut download_count = 0;
144 0 : while let Some(i) = layer_results.next().await {
145 0 : download_count += 1;
146 0 : match i {
147 0 : Ok((layer_name, layer_metadata)) => {
148 0 : tracing::info!(
149 0 : "[{download_count}/{layer_count}] OK: {} bytes {ttid} {}",
150 : layer_metadata.file_size,
151 : layer_name
152 : );
153 : }
154 0 : Err(e) => {
155 0 : // Warn and continue: we will download what we can
156 0 : tracing::warn!("Download error: {e}");
157 0 : err = Some(e);
158 : }
159 : }
160 : }
161 0 : if let Some(e) = err {
162 0 : tracing::warn!("Some errors occurred downloading {ttid} layers, last error: {e}");
163 0 : Err(e)
164 : } else {
165 0 : Ok(())
166 : }
167 0 : }
168 :
169 0 : async fn download_timeline(
170 0 : &self,
171 0 : ttid: TenantShardTimelineId,
172 0 : index_part: Box<IndexPart>,
173 0 : index_part_generation: Generation,
174 0 : ancestor_layers: &mut HashMap<TenantShardTimelineId, HashMap<LayerName, LayerFileMetadata>>,
175 0 : ) -> anyhow::Result<()> {
176 0 : let index_bytes = serde_json::to_string(&index_part).unwrap();
177 0 :
178 0 : let layers = index_part
179 0 : .layer_metadata
180 0 : .into_iter()
181 0 : .filter_map(|(layer_name, layer_metadata)| {
182 0 : if layer_metadata.shard.shard_count != ttid.tenant_shard_id.shard_count {
183 : // Accumulate ancestor layers for later download
184 0 : let ancestor_ttid = TenantShardTimelineId::new(
185 0 : TenantShardId {
186 0 : tenant_id: ttid.tenant_shard_id.tenant_id,
187 0 : shard_number: layer_metadata.shard.shard_number,
188 0 : shard_count: layer_metadata.shard.shard_count,
189 0 : },
190 0 : ttid.timeline_id,
191 0 : );
192 0 : let ancestor_ttid_layers = ancestor_layers.entry(ancestor_ttid).or_default();
193 : use std::collections::hash_map::Entry;
194 0 : match ancestor_ttid_layers.entry(layer_name) {
195 0 : Entry::Occupied(entry) => {
196 0 : // Descendent shards that reference a layer from an ancestor should always have matching metadata,
197 0 : // as their siblings, because it is read atomically during a shard split.
198 0 : assert_eq!(entry.get(), &layer_metadata);
199 : }
200 0 : Entry::Vacant(entry) => {
201 0 : entry.insert(layer_metadata);
202 0 : }
203 : }
204 0 : None
205 : } else {
206 0 : Some((layer_name, layer_metadata))
207 : }
208 0 : })
209 0 : .collect();
210 :
211 0 : let download_result = self.download_layers(ttid, layers).await;
212 :
213 : // Write index last, once all the layers it references are downloaded
214 0 : let local_index_path = self.output_path.join(format!(
215 0 : "{}/timelines/{}/index_part.json{}",
216 0 : ttid.tenant_shard_id,
217 0 : ttid.timeline_id,
218 0 : index_part_generation.get_suffix()
219 0 : ));
220 0 : tokio::fs::write(&local_index_path, index_bytes)
221 0 : .await
222 0 : .context("writing index")?;
223 :
224 0 : download_result
225 0 : }
226 :
227 0 : pub async fn download(&self) -> anyhow::Result<()> {
228 0 : let (remote_client, target) =
229 0 : init_remote(self.bucket_config.clone(), NodeKind::Pageserver).await?;
230 :
231 : // Generate a stream of TenantShardId
232 0 : let shards = stream_tenant_shards(&remote_client, &target, self.tenant_id).await?;
233 0 : let shards: Vec<TenantShardId> = shards.try_collect().await?;
234 :
235 : // Only read from shards that have the highest count: avoids redundantly downloading
236 : // from ancestor shards.
237 0 : let Some(shard_count) = shards.iter().map(|s| s.shard_count).max() else {
238 0 : anyhow::bail!("No shards found");
239 : };
240 :
241 : // We will build a collection of layers in anccestor shards to download (this will only
242 : // happen if this tenant has been split at some point)
243 0 : let mut ancestor_layers: HashMap<
244 0 : TenantShardTimelineId,
245 0 : HashMap<LayerName, LayerFileMetadata>,
246 0 : > = Default::default();
247 :
248 0 : for shard in shards.into_iter().filter(|s| s.shard_count == shard_count) {
249 : // Generate a stream of TenantTimelineId
250 0 : let timelines = stream_tenant_timelines(&remote_client, &target, shard).await?;
251 :
252 : // Generate a stream of S3TimelineBlobData
253 0 : async fn load_timeline_index(
254 0 : remote_client: &GenericRemoteStorage,
255 0 : target: &RootTarget,
256 0 : ttid: TenantShardTimelineId,
257 0 : ) -> anyhow::Result<(TenantShardTimelineId, RemoteTimelineBlobData)> {
258 0 : let data = list_timeline_blobs(remote_client, ttid, target).await?;
259 0 : Ok((ttid, data))
260 0 : }
261 0 : let timelines =
262 0 : timelines.map_ok(|ttid| load_timeline_index(&remote_client, &target, ttid));
263 0 : let mut timelines = std::pin::pin!(timelines.try_buffered(8));
264 :
265 0 : while let Some(i) = timelines.next().await {
266 0 : let (ttid, data) = i?;
267 0 : match data.blob_data {
268 : BlobDataParseResult::Parsed {
269 0 : index_part,
270 0 : index_part_generation,
271 0 : s3_layers: _,
272 0 : index_part_last_modified_time: _,
273 0 : index_part_snapshot_time: _,
274 0 : } => {
275 0 : self.download_timeline(
276 0 : ttid,
277 0 : index_part,
278 0 : index_part_generation,
279 0 : &mut ancestor_layers,
280 0 : )
281 0 : .await
282 0 : .context("Downloading timeline")?;
283 : }
284 0 : BlobDataParseResult::Relic => {}
285 : BlobDataParseResult::Incorrect { .. } => {
286 0 : tracing::error!("Bad metadata in timeline {ttid}");
287 : }
288 : };
289 : }
290 : }
291 :
292 0 : for (ttid, layers) in ancestor_layers.into_iter() {
293 0 : tracing::info!(
294 0 : "Downloading {} layers from ancestor timeline {ttid}...",
295 0 : layers.len()
296 : );
297 :
298 0 : self.download_layers(ttid, layers.into_iter().collect())
299 0 : .await?;
300 : }
301 :
302 0 : Ok(())
303 0 : }
304 : }
|
