1use std::num::NonZero;
16use std::sync::Arc;
17use std::time::Duration;
18
19use api::v1::region::compact_request;
20use common_meta::key::SchemaMetadataManagerRef;
21use common_telemetry::{info, warn};
22use common_time::TimeToLive;
23use either::Either;
24use itertools::Itertools;
25use object_store::manager::ObjectStoreManagerRef;
26use partition::expr::PartitionExpr;
27use serde::{Deserialize, Serialize};
28use snafu::{OptionExt, ResultExt};
29use store_api::metadata::RegionMetadataRef;
30use store_api::region_request::PathType;
31use store_api::storage::RegionId;
32
33use crate::access_layer::{
34 AccessLayer, AccessLayerRef, Metrics, OperationType, SstWriteRequest, WriteType,
35};
36use crate::cache::{CacheManager, CacheManagerRef};
37use crate::compaction::picker::{PickerOutput, new_picker};
38use crate::compaction::{CompactionOutput, CompactionSstReaderBuilder, find_dynamic_options};
39use crate::config::MitoConfig;
40use crate::error::{
41 EmptyRegionDirSnafu, InvalidPartitionExprSnafu, JoinSnafu, ObjectStoreNotFoundSnafu, Result,
42};
43use crate::manifest::action::{RegionEdit, RegionMetaAction, RegionMetaActionList};
44use crate::manifest::manager::{RegionManifestManager, RegionManifestOptions};
45use crate::metrics;
46use crate::read::FlatSource;
47use crate::region::options::RegionOptions;
48use crate::region::version::VersionRef;
49use crate::region::{ManifestContext, RegionLeaderState, RegionRoleState};
50use crate::schedule::scheduler::LocalScheduler;
51use crate::sst::FormatType;
52use crate::sst::file::FileMeta;
53use crate::sst::file_purger::LocalFilePurger;
54use crate::sst::index::intermediate::IntermediateManager;
55use crate::sst::index::puffin_manager::PuffinManagerFactory;
56use crate::sst::location::region_dir_from_table_dir;
57use crate::sst::parquet::WriteOptions;
58use crate::sst::version::{SstVersion, SstVersionRef};
59
60#[derive(Clone)]
62pub struct CompactionVersion {
63 pub(crate) metadata: RegionMetadataRef,
68 pub(crate) options: RegionOptions,
70 pub(crate) ssts: SstVersionRef,
72 pub(crate) compaction_time_window: Option<Duration>,
74}
75
76impl From<VersionRef> for CompactionVersion {
77 fn from(value: VersionRef) -> Self {
78 Self {
79 metadata: value.metadata.clone(),
80 options: value.options.clone(),
81 ssts: value.ssts.clone(),
82 compaction_time_window: value.compaction_time_window,
83 }
84 }
85}
86
87#[derive(Clone)]
90pub struct CompactionRegion {
91 pub region_id: RegionId,
92 pub region_options: RegionOptions,
93
94 pub(crate) engine_config: Arc<MitoConfig>,
95 pub(crate) region_metadata: RegionMetadataRef,
96 pub(crate) cache_manager: CacheManagerRef,
97 pub access_layer: AccessLayerRef,
99 pub(crate) manifest_ctx: Arc<ManifestContext>,
100 pub(crate) current_version: CompactionVersion,
101 pub(crate) file_purger: Option<Arc<LocalFilePurger>>,
102 pub(crate) ttl: Option<TimeToLive>,
103
104 pub max_parallelism: usize,
109}
110
111#[derive(Debug, Clone)]
113pub struct OpenCompactionRegionRequest {
114 pub region_id: RegionId,
115 pub table_dir: String,
116 pub path_type: PathType,
117 pub region_options: RegionOptions,
118 pub max_parallelism: usize,
119}
120
121pub async fn open_compaction_region(
124 req: &OpenCompactionRegionRequest,
125 mito_config: &MitoConfig,
126 object_store_manager: ObjectStoreManagerRef,
127 ttl_provider: Either<TimeToLive, SchemaMetadataManagerRef>,
128) -> Result<CompactionRegion> {
129 let object_store = {
130 let name = &req.region_options.storage;
131 if let Some(name) = name {
132 object_store_manager
133 .find(name)
134 .with_context(|| ObjectStoreNotFoundSnafu {
135 object_store: name.clone(),
136 })?
137 } else {
138 object_store_manager.default_object_store()
139 }
140 };
141
142 let access_layer = {
143 let puffin_manager_factory = PuffinManagerFactory::new(
144 &mito_config.index.aux_path,
145 mito_config.index.staging_size.as_bytes(),
146 Some(mito_config.index.write_buffer_size.as_bytes() as _),
147 mito_config.index.staging_ttl,
148 )
149 .await?;
150 let intermediate_manager =
151 IntermediateManager::init_fs(mito_config.index.aux_path.clone()).await?;
152
153 Arc::new(AccessLayer::new(
154 &req.table_dir,
155 req.path_type,
156 object_store.clone(),
157 puffin_manager_factory,
158 intermediate_manager,
159 ))
160 };
161
162 let manifest_manager = {
163 let region_dir = region_dir_from_table_dir(&req.table_dir, req.region_id, req.path_type);
164 let region_manifest_options =
165 RegionManifestOptions::new(mito_config, ®ion_dir, object_store);
166
167 RegionManifestManager::open(region_manifest_options, &Default::default())
168 .await?
169 .with_context(|| EmptyRegionDirSnafu {
170 region_id: req.region_id,
171 region_dir: region_dir_from_table_dir(&req.table_dir, req.region_id, req.path_type),
172 })?
173 };
174
175 let manifest = manifest_manager.manifest();
176 let region_metadata = manifest.metadata.clone();
177 let manifest_ctx = Arc::new(ManifestContext::new(
178 manifest_manager,
179 RegionRoleState::Leader(RegionLeaderState::Writable),
180 ));
181
182 let file_purger = {
183 let purge_scheduler = Arc::new(LocalScheduler::new(mito_config.max_background_purges));
184 Arc::new(LocalFilePurger::new(
185 purge_scheduler.clone(),
186 access_layer.clone(),
187 None,
188 ))
189 };
190
191 let current_version = {
192 let mut ssts = SstVersion::new();
193 ssts.add_files(file_purger.clone(), manifest.files.values().cloned());
194 CompactionVersion {
195 metadata: region_metadata.clone(),
196 options: req.region_options.clone(),
197 ssts: Arc::new(ssts),
198 compaction_time_window: manifest.compaction_time_window,
199 }
200 };
201
202 let ttl = match ttl_provider {
203 Either::Left(ttl) => ttl,
205 Either::Right(schema_metadata_manager) => {
207 let (_, ttl) = find_dynamic_options(
208 req.region_id.table_id(),
209 &req.region_options,
210 &schema_metadata_manager,
211 )
212 .await
213 .unwrap_or_else(|e| {
214 warn!(e; "Failed to get ttl for region: {}", region_metadata.region_id);
215 (
216 crate::region::options::CompactionOptions::default(),
217 TimeToLive::default(),
218 )
219 });
220 ttl
221 }
222 };
223
224 Ok(CompactionRegion {
225 region_id: req.region_id,
226 region_options: req.region_options.clone(),
227 engine_config: Arc::new(mito_config.clone()),
228 region_metadata: region_metadata.clone(),
229 cache_manager: Arc::new(CacheManager::default()),
230 access_layer,
231 manifest_ctx,
232 current_version,
233 file_purger: Some(file_purger),
234 ttl: Some(ttl),
235 max_parallelism: req.max_parallelism,
236 })
237}
238
239impl CompactionRegion {
240 pub fn file_purger(&self) -> Option<Arc<LocalFilePurger>> {
242 self.file_purger.clone()
243 }
244
245 pub async fn stop_purger_scheduler(&self) -> Result<()> {
247 if let Some(file_purger) = &self.file_purger {
248 file_purger.stop_scheduler().await
249 } else {
250 Ok(())
251 }
252 }
253}
254
255#[derive(Default, Clone, Debug, Serialize, Deserialize)]
257pub struct MergeOutput {
258 pub files_to_add: Vec<FileMeta>,
259 pub files_to_remove: Vec<FileMeta>,
260 pub compaction_time_window: Option<i64>,
261}
262
263impl MergeOutput {
264 pub fn is_empty(&self) -> bool {
265 self.files_to_add.is_empty() && self.files_to_remove.is_empty()
266 }
267
268 pub fn input_file_size(&self) -> u64 {
269 self.files_to_remove.iter().map(|f| f.file_size).sum()
270 }
271
272 pub fn output_file_size(&self) -> u64 {
273 self.files_to_add.iter().map(|f| f.file_size).sum()
274 }
275}
276
277#[async_trait::async_trait]
279pub trait Compactor: Send + Sync + 'static {
280 async fn merge_ssts(
282 &self,
283 compaction_region: &CompactionRegion,
284 picker_output: PickerOutput,
285 ) -> Result<MergeOutput>;
286
287 async fn update_manifest(
289 &self,
290 compaction_region: &CompactionRegion,
291 merge_output: MergeOutput,
292 ) -> Result<RegionEdit>;
293
294 async fn compact(
296 &self,
297 compaction_region: &CompactionRegion,
298 compact_request_options: compact_request::Options,
299 ) -> Result<()>;
300}
301
302pub struct DefaultCompactor;
304
305impl DefaultCompactor {
306 async fn merge_single_output(
308 compaction_region: CompactionRegion,
309 output: CompactionOutput,
310 write_opts: WriteOptions,
311 ) -> Result<Vec<FileMeta>> {
312 let region_id = compaction_region.region_id;
313 let storage = compaction_region.region_options.storage.clone();
314 let index_options = compaction_region
315 .current_version
316 .options
317 .index_options
318 .clone();
319 let append_mode = compaction_region.current_version.options.append_mode;
320 let merge_mode = compaction_region.current_version.options.merge_mode();
321 let flat_format = compaction_region
322 .region_options
323 .sst_format
324 .map(|format| format == FormatType::Flat)
325 .unwrap_or(compaction_region.engine_config.default_flat_format);
326
327 let index_config = compaction_region.engine_config.index.clone();
328 let inverted_index_config = compaction_region.engine_config.inverted_index.clone();
329 let fulltext_index_config = compaction_region.engine_config.fulltext_index.clone();
330 let bloom_filter_index_config = compaction_region.engine_config.bloom_filter_index.clone();
331 #[cfg(feature = "vector_index")]
332 let vector_index_config = compaction_region.engine_config.vector_index.clone();
333
334 let input_file_names = output
335 .inputs
336 .iter()
337 .map(|f| f.file_id().to_string())
338 .join(",");
339 let max_sequence = output
340 .inputs
341 .iter()
342 .map(|f| f.meta_ref().sequence)
343 .max()
344 .flatten();
345 let builder = CompactionSstReaderBuilder {
346 metadata: compaction_region.region_metadata.clone(),
347 sst_layer: compaction_region.access_layer.clone(),
348 cache: compaction_region.cache_manager.clone(),
349 inputs: &output.inputs,
350 append_mode,
351 filter_deleted: output.filter_deleted,
352 time_range: output.output_time_range,
353 merge_mode,
354 };
355 let reader = builder.build_flat_sst_reader().await?;
356 let source = FlatSource::Stream(reader);
357 let mut metrics = Metrics::new(WriteType::Compaction);
358 let region_metadata = compaction_region.region_metadata.clone();
359 let sst_infos = compaction_region
360 .access_layer
361 .write_sst(
362 SstWriteRequest {
363 op_type: OperationType::Compact,
364 metadata: region_metadata.clone(),
365 source,
366 cache_manager: compaction_region.cache_manager.clone(),
367 storage,
368 max_sequence: max_sequence.map(NonZero::get),
369 sst_write_format: if flat_format {
370 FormatType::Flat
371 } else {
372 FormatType::PrimaryKey
373 },
374 index_options,
375 index_config,
376 inverted_index_config,
377 fulltext_index_config,
378 bloom_filter_index_config,
379 #[cfg(feature = "vector_index")]
380 vector_index_config,
381 },
382 &write_opts,
383 &mut metrics,
384 )
385 .await?;
386 let partition_expr = match ®ion_metadata.partition_expr {
388 None => None,
389 Some(json_str) if json_str.is_empty() => None,
390 Some(json_str) => PartitionExpr::from_json_str(json_str).with_context(|_| {
391 InvalidPartitionExprSnafu {
392 expr: json_str.clone(),
393 }
394 })?,
395 };
396
397 let output_files = sst_infos
398 .into_iter()
399 .map(|sst_info| FileMeta {
400 region_id,
401 file_id: sst_info.file_id,
402 time_range: sst_info.time_range,
403 level: output.output_level,
404 file_size: sst_info.file_size,
405 max_row_group_uncompressed_size: sst_info.max_row_group_uncompressed_size,
406 available_indexes: sst_info.index_metadata.build_available_indexes(),
407 indexes: sst_info.index_metadata.build_indexes(),
408 index_file_size: sst_info.index_metadata.file_size,
409 index_version: 0,
410 num_rows: sst_info.num_rows as u64,
411 num_row_groups: sst_info.num_row_groups,
412 sequence: max_sequence,
413 partition_expr: partition_expr.clone(),
414 num_series: sst_info.num_series,
415 })
416 .collect::<Vec<_>>();
417 let output_file_names = output_files.iter().map(|f| f.file_id.to_string()).join(",");
418 info!(
419 "Region {} compaction inputs: [{}], outputs: [{}], flat_format: {}, metrics: {:?}",
420 region_id, input_file_names, output_file_names, flat_format, metrics
421 );
422 metrics.observe();
423 Ok(output_files)
424 }
425}
426
427#[async_trait::async_trait]
428impl Compactor for DefaultCompactor {
429 async fn merge_ssts(
430 &self,
431 compaction_region: &CompactionRegion,
432 mut picker_output: PickerOutput,
433 ) -> Result<MergeOutput> {
434 let mut futs = Vec::with_capacity(picker_output.outputs.len());
435 let mut compacted_inputs =
436 Vec::with_capacity(picker_output.outputs.iter().map(|o| o.inputs.len()).sum());
437 let internal_parallelism = compaction_region.max_parallelism.max(1);
438 let compaction_time_window = picker_output.time_window_size;
439
440 for output in picker_output.outputs.drain(..) {
441 let inputs_to_remove: Vec<_> =
442 output.inputs.iter().map(|f| f.meta_ref().clone()).collect();
443 compacted_inputs.extend(inputs_to_remove.iter().cloned());
444 let write_opts = WriteOptions {
445 write_buffer_size: compaction_region.engine_config.sst_write_buffer_size,
446 max_file_size: picker_output.max_file_size,
447 ..Default::default()
448 };
449 futs.push(Self::merge_single_output(
450 compaction_region.clone(),
451 output,
452 write_opts,
453 ));
454 }
455 let mut output_files = Vec::with_capacity(futs.len());
456 while !futs.is_empty() {
457 let mut task_chunk = Vec::with_capacity(internal_parallelism);
458 for _ in 0..internal_parallelism {
459 if let Some(task) = futs.pop() {
460 task_chunk.push(common_runtime::spawn_compact(task));
461 }
462 }
463 let metas = futures::future::try_join_all(task_chunk)
464 .await
465 .context(JoinSnafu)?
466 .into_iter()
467 .collect::<Result<Vec<Vec<_>>>>()?;
468 output_files.extend(metas.into_iter().flatten());
469 }
470
471 let mut inputs: Vec<_> = compacted_inputs.into_iter().collect();
474 inputs.extend(
475 picker_output
476 .expired_ssts
477 .iter()
478 .map(|f| f.meta_ref().clone()),
479 );
480
481 Ok(MergeOutput {
482 files_to_add: output_files,
483 files_to_remove: inputs,
484 compaction_time_window: Some(compaction_time_window),
485 })
486 }
487
488 async fn update_manifest(
489 &self,
490 compaction_region: &CompactionRegion,
491 merge_output: MergeOutput,
492 ) -> Result<RegionEdit> {
493 let edit = RegionEdit {
495 files_to_add: merge_output.files_to_add,
496 files_to_remove: merge_output.files_to_remove,
497 timestamp_ms: Some(chrono::Utc::now().timestamp_millis()),
499 compaction_time_window: merge_output
500 .compaction_time_window
501 .map(|seconds| Duration::from_secs(seconds as u64)),
502 flushed_entry_id: None,
503 flushed_sequence: None,
504 committed_sequence: None,
505 };
506
507 let action_list = RegionMetaActionList::with_action(RegionMetaAction::Edit(edit.clone()));
508 compaction_region
510 .manifest_ctx
511 .update_manifest(RegionLeaderState::Writable, action_list, false)
512 .await?;
513
514 Ok(edit)
515 }
516
517 async fn compact(
520 &self,
521 compaction_region: &CompactionRegion,
522 compact_request_options: compact_request::Options,
523 ) -> Result<()> {
524 let picker_output = {
525 let picker_output = new_picker(
526 &compact_request_options,
527 &compaction_region.region_options.compaction,
528 compaction_region.region_options.append_mode,
529 None,
530 )
531 .pick(compaction_region);
532
533 if let Some(picker_output) = picker_output {
534 picker_output
535 } else {
536 info!(
537 "No files to compact for region_id: {}",
538 compaction_region.region_id
539 );
540 return Ok(());
541 }
542 };
543
544 let merge_output = self.merge_ssts(compaction_region, picker_output).await?;
545 if merge_output.is_empty() {
546 info!(
547 "No files to compact for region_id: {}",
548 compaction_region.region_id
549 );
550 return Ok(());
551 }
552
553 metrics::COMPACTION_INPUT_BYTES.inc_by(merge_output.input_file_size() as f64);
554 metrics::COMPACTION_OUTPUT_BYTES.inc_by(merge_output.output_file_size() as f64);
555 self.update_manifest(compaction_region, merge_output)
556 .await?;
557
558 Ok(())
559 }
560}