1use std::path::PathBuf;
16use std::sync::Arc;
17use std::time::Instant;
18
19use clap::Parser;
20use colored::Colorize;
21use datanode::config::RegionEngineConfig;
22use datanode::store;
23use futures::stream;
24use mito2::access_layer::{
25 AccessLayer, AccessLayerRef, Metrics, OperationType, SstWriteRequest, WriteType,
26};
27use mito2::cache::{CacheManager, CacheManagerRef};
28use mito2::config::{FulltextIndexConfig, MitoConfig, Mode};
29use mito2::read::FlatSource;
30use mito2::sst::FormatType;
31use mito2::sst::file::{FileHandle, FileMeta};
32use mito2::sst::file_purger::{FilePurger, FilePurgerRef};
33use mito2::sst::index::intermediate::IntermediateManager;
34use mito2::sst::index::puffin_manager::PuffinManagerFactory;
35use mito2::sst::parquet::reader::ParquetReaderBuilder;
36use mito2::sst::parquet::{PARQUET_METADATA_KEY, WriteOptions};
37use mito2::worker::write_cache_from_config;
38use object_store::ObjectStore;
39use parquet::file::metadata::{FooterTail, KeyValue};
40use regex::Regex;
41use snafu::OptionExt;
42use store_api::metadata::{RegionMetadata, RegionMetadataRef};
43use store_api::path_utils::region_name;
44use store_api::region_request::PathType;
45use store_api::storage::FileId;
46
47use crate::datanode::{StorageConfig, StorageConfigWrapper};
48use crate::error;
49
50#[derive(Debug, Parser)]
52pub struct ObjbenchCommand {
53 #[clap(long, value_name = "FILE")]
55 pub config: PathBuf,
56
57 #[clap(long, value_name = "PATH")]
59 pub source: String,
60
61 #[clap(short, long, default_value_t = false)]
63 pub verbose: bool,
64
65 #[clap(long, value_name = "FILE")]
67 pub pprof_file: Option<PathBuf>,
68}
69
70pub(super) fn parse_config(
71 config_path: &PathBuf,
72) -> error::Result<(
73 StorageConfig,
74 MitoConfig,
75 common_wal::config::DatanodeWalConfig,
76)> {
77 let cfg_str = std::fs::read_to_string(config_path).map_err(|e| {
78 error::IllegalConfigSnafu {
79 msg: format!("failed to read config {}: {e}", config_path.display()),
80 }
81 .build()
82 })?;
83
84 let store_cfg: StorageConfigWrapper = toml::from_str(&cfg_str).map_err(|e| {
85 error::IllegalConfigSnafu {
86 msg: format!("failed to parse config {}: {e}", config_path.display()),
87 }
88 .build()
89 })?;
90
91 let wal_config = store_cfg.wal;
92 let storage_config = store_cfg.storage;
93 let mito_engine_config = store_cfg
94 .region_engine
95 .into_iter()
96 .filter_map(|c| {
97 if let RegionEngineConfig::Mito(mito) = c {
98 Some(mito)
99 } else {
100 None
101 }
102 })
103 .next()
104 .with_context(|| error::IllegalConfigSnafu {
105 msg: format!("Engine config not found in {:?}", config_path),
106 })?;
107 Ok((storage_config, mito_engine_config, wal_config))
108}
109
110impl ObjbenchCommand {
111 pub async fn run(&self) -> error::Result<()> {
112 if self.verbose {
113 common_telemetry::init_default_ut_logging();
114 }
115
116 println!("{}", "Starting objbench with config:".cyan().bold());
117
118 let (store_cfg, mut mito_engine_config, _wal_config) = parse_config(&self.config)?;
120
121 let object_store = build_object_store(&store_cfg).await?;
122 println!("{} Object store initialized", "✓".green());
123
124 let components = parse_file_dir_components(&self.source)?;
126 println!(
127 "{} Source path parsed: {}, components: {:?}",
128 "✓".green(),
129 self.source,
130 components
131 );
132
133 println!("{}", "Loading parquet metadata...".yellow());
135 let file_size = object_store
136 .stat(&self.source)
137 .await
138 .map_err(|e| {
139 error::IllegalConfigSnafu {
140 msg: format!("stat failed: {e}"),
141 }
142 .build()
143 })?
144 .content_length();
145 let parquet_meta = load_parquet_metadata(object_store.clone(), &self.source, file_size)
146 .await
147 .map_err(|e| {
148 error::IllegalConfigSnafu {
149 msg: format!("read parquet metadata failed: {e}"),
150 }
151 .build()
152 })?;
153
154 let region_meta = extract_region_metadata(&self.source, &parquet_meta)?;
155 let num_rows = parquet_meta.file_metadata().num_rows() as u64;
156 let num_row_groups = parquet_meta.num_row_groups() as u64;
157 let max_row_group_uncompressed_size: u64 = parquet_meta
158 .row_groups()
159 .iter()
160 .map(|rg| {
161 rg.columns()
162 .iter()
163 .map(|c| c.uncompressed_size() as u64)
164 .sum::<u64>()
165 })
166 .max()
167 .unwrap_or(0);
168
169 println!(
170 "{} Metadata loaded - rows: {}, size: {} bytes",
171 "✓".green(),
172 num_rows,
173 file_size
174 );
175
176 let file_meta = FileMeta {
178 region_id: region_meta.region_id,
179 file_id: components.file_id,
180 time_range: Default::default(),
181 level: 0,
182 file_size,
183 max_row_group_uncompressed_size,
184 available_indexes: Default::default(),
185 indexes: Default::default(),
186 index_file_size: 0,
187 index_version: 0,
188 num_rows,
189 num_row_groups,
190 sequence: None,
191 partition_expr: None,
192 num_series: 0,
193 ..Default::default()
194 };
195 let src_handle = FileHandle::new(file_meta, new_noop_file_purger());
196
197 let table_dir = components.table_dir();
199 let (src_access_layer, cache_manager) = build_access_layer_simple(
200 &components,
201 object_store.clone(),
202 &mut mito_engine_config,
203 &store_cfg.data_home,
204 )
205 .await?;
206 let reader_build_start = Instant::now();
207
208 let reader = ParquetReaderBuilder::new(
209 table_dir,
210 components.path_type,
211 src_handle.clone(),
212 object_store.clone(),
213 )
214 .expected_metadata(Some(region_meta.clone()))
215 .build()
216 .await
217 .map_err(|e| {
218 error::IllegalConfigSnafu {
219 msg: format!("build reader failed: {e:?}"),
220 }
221 .build()
222 })?;
223 let reader = reader.ok_or_else(|| {
224 error::IllegalConfigSnafu {
225 msg: format!(
226 "build reader returned no readable rows for source file {}",
227 src_handle.file_id()
228 ),
229 }
230 .build()
231 })?;
232
233 let reader_build_elapsed = reader_build_start.elapsed();
234 let total_rows = reader.parquet_metadata().file_metadata().num_rows();
235 println!("{} Reader built in {:?}", "✓".green(), reader_build_elapsed);
236 let reader_stream = Box::pin(stream::try_unfold(reader, |mut reader| async move {
237 let batch = reader.next_record_batch().await?;
238 Ok(batch.map(|batch| (batch, reader)))
239 }));
240
241 let fulltext_index_config = FulltextIndexConfig {
243 create_on_compaction: Mode::Disable,
244 ..Default::default()
245 };
246
247 let write_req = SstWriteRequest {
248 op_type: OperationType::Flush,
249 metadata: region_meta,
250 source: FlatSource::Stream(reader_stream),
251 cache_manager,
252 storage: None,
253 max_sequence: None,
254 sst_write_format: FormatType::PrimaryKey,
255 index_options: Default::default(),
256 index_config: mito_engine_config.index.clone(),
257 inverted_index_config: MitoConfig::default().inverted_index,
258 fulltext_index_config,
259 bloom_filter_index_config: MitoConfig::default().bloom_filter_index,
260 #[cfg(feature = "vector_index")]
261 vector_index_config: Default::default(),
262 };
263
264 println!("{}", "Writing SST...".yellow());
266
267 #[cfg(unix)]
269 let profiler_guard = if self.pprof_file.is_some() {
270 println!("{} Starting profiling...", "⚡".yellow());
271 Some(
272 pprof::ProfilerGuardBuilder::default()
273 .frequency(99)
274 .blocklist(&["libc", "libgcc", "pthread", "vdso"])
275 .build()
276 .map_err(|e| {
277 error::IllegalConfigSnafu {
278 msg: format!("Failed to start profiler: {e}"),
279 }
280 .build()
281 })?,
282 )
283 } else {
284 None
285 };
286
287 #[cfg(not(unix))]
288 if self.pprof_file.is_some() {
289 eprintln!(
290 "{}: Profiling is not supported on this platform",
291 "Warning".yellow()
292 );
293 }
294
295 let write_start = Instant::now();
296 let mut metrics = Metrics::new(WriteType::Flush);
297 let infos = src_access_layer
298 .write_sst(write_req, &WriteOptions::default(), &mut metrics)
299 .await
300 .map_err(|e| {
301 error::IllegalConfigSnafu {
302 msg: format!("write_sst failed: {e:?}"),
303 }
304 .build()
305 })?;
306
307 let write_elapsed = write_start.elapsed();
308
309 #[cfg(unix)]
311 if let (Some(guard), Some(pprof_file)) = (profiler_guard, &self.pprof_file) {
312 println!("{} Generating flamegraph...", "🔥".yellow());
313 match guard.report().build() {
314 Ok(report) => {
315 let mut flamegraph_data = Vec::new();
316 if let Err(e) = report.flamegraph(&mut flamegraph_data) {
317 println!("{}: Failed to generate flamegraph: {}", "Error".red(), e);
318 } else if let Err(e) = std::fs::write(pprof_file, flamegraph_data) {
319 println!(
320 "{}: Failed to write flamegraph to {}: {}",
321 "Error".red(),
322 pprof_file.display(),
323 e
324 );
325 } else {
326 println!(
327 "{} Flamegraph saved to {}",
328 "✓".green(),
329 pprof_file.display().to_string().cyan()
330 );
331 }
332 }
333 Err(e) => {
334 println!("{}: Failed to generate pprof report: {}", "Error".red(), e);
335 }
336 }
337 }
338 assert_eq!(infos.len(), 1);
339 let dst_file_id = infos[0].file_id;
340 let dst_file_path = format!("{}/{}.parquet", components.region_dir(), dst_file_id);
341 let mut dst_index_path = None;
342 if infos[0].index_metadata.file_size > 0 {
343 dst_index_path = Some(format!(
344 "{}/index/{}.puffin",
345 components.region_dir(),
346 dst_file_id
347 ));
348 }
349
350 println!("\n{} {}", "Write complete!".green().bold(), "✓".green());
352 println!(" {}: {}", "Destination file".bold(), dst_file_path.cyan());
353 println!(" {}: {}", "Rows".bold(), total_rows.to_string().cyan());
354 println!(
355 " {}: {}",
356 "File size".bold(),
357 format!("{} bytes", file_size).cyan()
358 );
359 println!(
360 " {}: {:?}",
361 "Reader build time".bold(),
362 reader_build_elapsed
363 );
364 println!(" {}: {:?}", "Total time".bold(), write_elapsed);
365
366 println!(" {}: {:?}", "Metrics".bold(), metrics,);
368
369 println!(" {}: {:?}", "Index".bold(), infos[0].index_metadata);
371
372 println!("\n{}", "Cleaning up...".yellow());
374 object_store.delete(&dst_file_path).await.map_err(|e| {
375 error::IllegalConfigSnafu {
376 msg: format!("Failed to delete dest file {}: {}", dst_file_path, e),
377 }
378 .build()
379 })?;
380 println!("{} Temporary file {} deleted", "✓".green(), dst_file_path);
381
382 if let Some(index_path) = dst_index_path {
383 object_store.delete(&index_path).await.map_err(|e| {
384 error::IllegalConfigSnafu {
385 msg: format!("Failed to delete dest index file {}: {}", index_path, e),
386 }
387 .build()
388 })?;
389 println!(
390 "{} Temporary index file {} deleted",
391 "✓".green(),
392 index_path
393 );
394 }
395
396 println!("\n{}", "Benchmark completed successfully!".green().bold());
397 Ok(())
398 }
399}
400
401#[derive(Debug)]
402struct FileDirComponents {
403 catalog: String,
404 schema: String,
405 table_id: u32,
406 region_sequence: u32,
407 path_type: PathType,
408 file_id: FileId,
409}
410
411impl FileDirComponents {
412 fn table_dir(&self) -> String {
413 format!("data/{}/{}/{}", self.catalog, self.schema, self.table_id)
414 }
415
416 fn region_dir(&self) -> String {
417 let region_name = region_name(self.table_id, self.region_sequence);
418 match self.path_type {
419 PathType::Bare => {
420 format!(
421 "data/{}/{}/{}/{}",
422 self.catalog, self.schema, self.table_id, region_name
423 )
424 }
425 PathType::Data => {
426 format!(
427 "data/{}/{}/{}/{}/data",
428 self.catalog, self.schema, self.table_id, region_name
429 )
430 }
431 PathType::Metadata => {
432 format!(
433 "data/{}/{}/{}/{}/metadata",
434 self.catalog, self.schema, self.table_id, region_name
435 )
436 }
437 }
438 }
439}
440
441fn parse_file_dir_components(path: &str) -> error::Result<FileDirComponents> {
442 let pattern =
444 r"^data/([^/]+)/([^/]+)/([^/]+)/([^/]+)_([^/]+)(?:/data|/metadata)?/(.+).parquet$";
445
446 let re = Regex::new(pattern).expect("Invalid regex pattern");
448
449 let path_type = if path.contains("/data/") {
451 PathType::Data
452 } else if path.contains("/metadata/") {
453 PathType::Metadata
454 } else {
455 PathType::Bare
456 };
457
458 let components = (|| {
460 let captures = re.captures(path)?;
461 if captures.len() != 7 {
462 return None;
463 }
464 let mut components = FileDirComponents {
465 catalog: "".to_string(),
466 schema: "".to_string(),
467 table_id: 0,
468 region_sequence: 0,
469 path_type,
470 file_id: FileId::default(),
471 };
472 components.catalog = captures.get(1)?.as_str().to_string();
474 components.schema = captures.get(2)?.as_str().to_string();
475 components.table_id = captures[3].parse().ok()?;
476 components.region_sequence = captures[5].parse().ok()?;
477 let file_id_str = &captures[6];
478 components.file_id = FileId::parse_str(file_id_str).ok()?;
479 Some(components)
480 })();
481 components.context(error::IllegalConfigSnafu {
482 msg: format!("Expect valid source file path, got: {}", path),
483 })
484}
485
486fn extract_region_metadata(
487 file_path: &str,
488 meta: &parquet::file::metadata::ParquetMetaData,
489) -> error::Result<RegionMetadataRef> {
490 let kvs: Option<&Vec<KeyValue>> = meta.file_metadata().key_value_metadata();
491 let Some(kvs) = kvs else {
492 return Err(error::IllegalConfigSnafu {
493 msg: format!("{file_path}: missing parquet key_value metadata"),
494 }
495 .build());
496 };
497 let json = kvs
498 .iter()
499 .find(|kv| kv.key == PARQUET_METADATA_KEY)
500 .and_then(|kv| kv.value.as_ref())
501 .ok_or_else(|| {
502 error::IllegalConfigSnafu {
503 msg: format!("{file_path}: key {PARQUET_METADATA_KEY} not found or empty"),
504 }
505 .build()
506 })?;
507 let region: RegionMetadata = RegionMetadata::from_json(json).map_err(|e| {
508 error::IllegalConfigSnafu {
509 msg: format!("invalid region metadata json: {e}"),
510 }
511 .build()
512 })?;
513 Ok(Arc::new(region))
514}
515
516pub(super) async fn build_object_store(sc: &StorageConfig) -> error::Result<ObjectStore> {
517 store::new_object_store(sc.store.clone(), &sc.data_home)
518 .await
519 .map_err(|e| {
520 error::IllegalConfigSnafu {
521 msg: format!("Failed to build object store: {e:?}"),
522 }
523 .build()
524 })
525}
526
527async fn build_access_layer_simple(
528 components: &FileDirComponents,
529 object_store: ObjectStore,
530 config: &mut MitoConfig,
531 data_home: &str,
532) -> error::Result<(AccessLayerRef, CacheManagerRef)> {
533 let _ = config.index.sanitize(data_home, &config.inverted_index);
534 let puffin_manager = PuffinManagerFactory::new(
535 &config.index.aux_path,
536 config.index.staging_size.as_bytes(),
537 Some(config.index.write_buffer_size.as_bytes() as _),
538 config.index.staging_ttl,
539 )
540 .await
541 .map_err(|e| {
542 error::IllegalConfigSnafu {
543 msg: format!("Failed to build access layer: {e:?}"),
544 }
545 .build()
546 })?;
547
548 let intermediate_manager = IntermediateManager::init_fs(&config.index.aux_path)
549 .await
550 .map_err(|e| {
551 error::IllegalConfigSnafu {
552 msg: format!("Failed to build IntermediateManager: {e:?}"),
553 }
554 .build()
555 })?
556 .with_buffer_size(Some(config.index.write_buffer_size.as_bytes() as _));
557
558 let cache_manager =
559 build_cache_manager(config, puffin_manager.clone(), intermediate_manager.clone()).await?;
560 let layer = AccessLayer::new(
561 components.table_dir(),
562 components.path_type,
563 object_store,
564 puffin_manager,
565 intermediate_manager,
566 );
567 Ok((Arc::new(layer), cache_manager))
568}
569
570async fn build_cache_manager(
571 config: &MitoConfig,
572 puffin_manager: PuffinManagerFactory,
573 intermediate_manager: IntermediateManager,
574) -> error::Result<CacheManagerRef> {
575 let write_cache = write_cache_from_config(config, puffin_manager, intermediate_manager)
576 .await
577 .map_err(|e| {
578 error::IllegalConfigSnafu {
579 msg: format!("Failed to build write cache: {e:?}"),
580 }
581 .build()
582 })?;
583 let cache_manager = Arc::new(
584 CacheManager::builder()
585 .sst_meta_cache_size(config.sst_meta_cache_size.as_bytes())
586 .vector_cache_size(config.vector_cache_size.as_bytes())
587 .page_cache_size(config.page_cache_size.as_bytes())
588 .selector_result_cache_size(config.selector_result_cache_size.as_bytes())
589 .index_metadata_size(config.index.metadata_cache_size.as_bytes())
590 .index_content_size(config.index.content_cache_size.as_bytes())
591 .index_content_page_size(config.index.content_cache_page_size.as_bytes())
592 .index_result_cache_size(config.index.result_cache_size.as_bytes())
593 .puffin_metadata_size(config.index.metadata_cache_size.as_bytes())
594 .write_cache(write_cache)
595 .build(),
596 );
597 Ok(cache_manager)
598}
599
600fn new_noop_file_purger() -> FilePurgerRef {
601 #[derive(Debug)]
602 struct Noop;
603 impl FilePurger for Noop {
604 fn remove_file(&self, _file_meta: FileMeta, _is_delete: bool, _index_outdated: bool) {}
605 }
606 Arc::new(Noop)
607}
608
609async fn load_parquet_metadata(
610 object_store: ObjectStore,
611 path: &str,
612 file_size: u64,
613) -> Result<parquet::file::metadata::ParquetMetaData, Box<dyn std::error::Error + Send + Sync>> {
614 use parquet::file::FOOTER_SIZE;
615 use parquet::file::metadata::ParquetMetaDataReader;
616 let actual_size = if file_size == 0 {
617 object_store.stat(path).await?.content_length()
618 } else {
619 file_size
620 };
621 if actual_size < FOOTER_SIZE as u64 {
622 return Err("file too small".into());
623 }
624 let prefetch: u64 = 64 * 1024;
625 let start = actual_size.saturating_sub(prefetch);
626 let buffer = object_store
627 .read_with(path)
628 .range(start..actual_size)
629 .await?
630 .to_vec();
631 let buffer_len = buffer.len();
632 let mut footer = [0; 8];
633 footer.copy_from_slice(&buffer[buffer_len - FOOTER_SIZE..]);
634 let footer = FooterTail::try_new(&footer)?;
635 let metadata_len = footer.metadata_length() as u64;
636 if actual_size - (FOOTER_SIZE as u64) < metadata_len {
637 return Err("invalid footer/metadata length".into());
638 }
639 if (metadata_len as usize) <= buffer_len - FOOTER_SIZE {
640 let metadata_start = buffer_len - metadata_len as usize - FOOTER_SIZE;
641 let meta = ParquetMetaDataReader::decode_metadata(
642 &buffer[metadata_start..buffer_len - FOOTER_SIZE],
643 )?;
644 Ok(meta)
645 } else {
646 let metadata_start = actual_size - metadata_len - FOOTER_SIZE as u64;
647 let data = object_store
648 .read_with(path)
649 .range(metadata_start..(actual_size - FOOTER_SIZE as u64))
650 .await?
651 .to_vec();
652 let meta = ParquetMetaDataReader::decode_metadata(&data)?;
653 Ok(meta)
654 }
655}
656
657#[cfg(test)]
658mod tests {
659 use std::path::PathBuf;
660 use std::str::FromStr;
661
662 use common_base::readable_size::ReadableSize;
663 use store_api::region_request::PathType;
664
665 use crate::datanode::objbench::{parse_config, parse_file_dir_components};
666
667 #[test]
668 fn test_parse_dir() {
669 let meta_path = "data/greptime/public/1024/1024_0000000000/metadata/00020380-009c-426d-953e-b4e34c15af34.parquet";
670 let c = parse_file_dir_components(meta_path).unwrap();
671 assert_eq!(
672 c.file_id.to_string(),
673 "00020380-009c-426d-953e-b4e34c15af34"
674 );
675 assert_eq!(c.catalog, "greptime");
676 assert_eq!(c.schema, "public");
677 assert_eq!(c.table_id, 1024);
678 assert_eq!(c.region_sequence, 0);
679 assert_eq!(c.path_type, PathType::Metadata);
680
681 let c = parse_file_dir_components(
682 "data/greptime/public/1024/1024_0000000000/data/00020380-009c-426d-953e-b4e34c15af34.parquet",
683 ).unwrap();
684 assert_eq!(
685 c.file_id.to_string(),
686 "00020380-009c-426d-953e-b4e34c15af34"
687 );
688 assert_eq!(c.catalog, "greptime");
689 assert_eq!(c.schema, "public");
690 assert_eq!(c.table_id, 1024);
691 assert_eq!(c.region_sequence, 0);
692 assert_eq!(c.path_type, PathType::Data);
693
694 let c = parse_file_dir_components(
695 "data/greptime/public/1024/1024_0000000000/00020380-009c-426d-953e-b4e34c15af34.parquet",
696 ).unwrap();
697 assert_eq!(
698 c.file_id.to_string(),
699 "00020380-009c-426d-953e-b4e34c15af34"
700 );
701 assert_eq!(c.catalog, "greptime");
702 assert_eq!(c.schema, "public");
703 assert_eq!(c.table_id, 1024);
704 assert_eq!(c.region_sequence, 0);
705 assert_eq!(c.path_type, PathType::Bare);
706 }
707
708 #[test]
709 fn test_parse_config() {
710 let path = "../../config/datanode.example.toml";
711 let (storage, engine, _wal) = parse_config(&PathBuf::from_str(path).unwrap()).unwrap();
712 assert_eq!(storage.data_home, "./greptimedb_data");
713 assert_eq!(engine.index.staging_size, ReadableSize::gb(2));
714 }
715}