Skip to main content

cmd/datanode/
objbench.rs

1// Copyright 2023 Greptime Team
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7//     http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15use std::path::PathBuf;
16use std::sync::Arc;
17use std::time::Instant;
18
19use clap::Parser;
20use colored::Colorize;
21use futures::stream;
22use mito2::access_layer::{
23    AccessLayer, AccessLayerRef, Metrics, OperationType, SstWriteRequest, WriteType,
24};
25use mito2::cache::{CacheManager, CacheManagerRef};
26use mito2::config::{FulltextIndexConfig, MitoConfig, Mode};
27use mito2::read::FlatSource;
28use mito2::sst::FormatType;
29use mito2::sst::file::{FileHandle, FileMeta};
30use mito2::sst::file_purger::{FilePurger, FilePurgerRef};
31use mito2::sst::index::intermediate::IntermediateManager;
32use mito2::sst::index::puffin_manager::PuffinManagerFactory;
33use mito2::sst::parquet::WriteOptions;
34use mito2::sst::parquet::reader::ParquetReaderBuilder;
35use mito2::worker::write_cache_from_config;
36use object_store::ObjectStore;
37use parquet::file::metadata::FooterTail;
38use regex::Regex;
39use snafu::OptionExt;
40use store_api::path_utils::region_name;
41use store_api::region_request::PathType;
42use store_api::storage::FileId;
43
44use crate::datanode::tool_util::{
45    build_object_store, extract_region_metadata, max_row_group_uncompressed_size, parse_config,
46};
47use crate::error;
48
49/// Object storage benchmark command
50#[derive(Debug, Parser)]
51pub struct ObjbenchCommand {
52    /// Path to the object-store config file (TOML). Must deserialize into object_store::config::ObjectStoreConfig.
53    #[clap(long, value_name = "FILE")]
54    pub config: PathBuf,
55
56    /// Source SST file path in object-store (e.g. "region_dir/<uuid>.parquet").
57    #[clap(long, value_name = "PATH")]
58    pub source: String,
59
60    /// Verbose output
61    #[clap(short, long, default_value_t = false)]
62    pub verbose: bool,
63
64    /// Output file path for pprof flamegraph (enables profiling)
65    #[clap(long, value_name = "FILE")]
66    pub pprof_file: Option<PathBuf>,
67}
68
69impl ObjbenchCommand {
70    pub async fn run(&self) -> error::Result<()> {
71        if self.verbose {
72            common_telemetry::init_default_ut_logging();
73        }
74
75        println!("{}", "Starting objbench with config:".cyan().bold());
76
77        // Build object store from config
78        let (store_cfg, mut mito_engine_config, _wal_config) = parse_config(&self.config)?;
79
80        let object_store = build_object_store(&store_cfg).await?;
81        println!("{} Object store initialized", "✓".green());
82
83        // Prepare source identifiers
84        let components = parse_file_dir_components(&self.source)?;
85        println!(
86            "{} Source path parsed: {}, components: {:?}",
87            "✓".green(),
88            self.source,
89            components
90        );
91
92        // Load parquet metadata to extract RegionMetadata and file stats
93        println!("{}", "Loading parquet metadata...".yellow());
94        let file_size = object_store
95            .stat(&self.source)
96            .await
97            .map_err(|e| {
98                error::IllegalConfigSnafu {
99                    msg: format!("stat failed: {e}"),
100                }
101                .build()
102            })?
103            .content_length();
104        let parquet_meta = load_parquet_metadata(object_store.clone(), &self.source, file_size)
105            .await
106            .map_err(|e| {
107                error::IllegalConfigSnafu {
108                    msg: format!("read parquet metadata failed: {e}"),
109                }
110                .build()
111            })?;
112
113        let region_meta = extract_region_metadata(&self.source, &parquet_meta)?;
114        let num_rows = parquet_meta.file_metadata().num_rows() as u64;
115        let num_row_groups = parquet_meta.num_row_groups() as u64;
116        let max_row_group_uncompressed_size = max_row_group_uncompressed_size(&parquet_meta);
117
118        println!(
119            "{} Metadata loaded - rows: {}, size: {} bytes",
120            "✓".green(),
121            num_rows,
122            file_size
123        );
124
125        // Build a FileHandle for the source file
126        let file_meta = FileMeta {
127            region_id: region_meta.region_id,
128            file_id: components.file_id,
129            time_range: Default::default(),
130            level: 0,
131            file_size,
132            max_row_group_uncompressed_size,
133            available_indexes: Default::default(),
134            indexes: Default::default(),
135            index_file_size: 0,
136            index_version: 0,
137            num_rows,
138            num_row_groups,
139            sequence: None,
140            partition_expr: None,
141            num_series: 0,
142            ..Default::default()
143        };
144        let src_handle = FileHandle::new(file_meta, new_noop_file_purger());
145
146        // Build the reader for a single file via ParquetReaderBuilder
147        let table_dir = components.table_dir();
148        let (src_access_layer, cache_manager) = build_access_layer_simple(
149            &components,
150            object_store.clone(),
151            &mut mito_engine_config,
152            &store_cfg.data_home,
153        )
154        .await?;
155        let reader_build_start = Instant::now();
156
157        let reader = ParquetReaderBuilder::new(
158            table_dir,
159            components.path_type,
160            src_handle.clone(),
161            object_store.clone(),
162        )
163        .expected_metadata(Some(region_meta.clone()))
164        .build()
165        .await
166        .map_err(|e| {
167            error::IllegalConfigSnafu {
168                msg: format!("build reader failed: {e:?}"),
169            }
170            .build()
171        })?;
172        let reader = reader.ok_or_else(|| {
173            error::IllegalConfigSnafu {
174                msg: format!(
175                    "build reader returned no readable rows for source file {}",
176                    src_handle.file_id()
177                ),
178            }
179            .build()
180        })?;
181
182        let reader_build_elapsed = reader_build_start.elapsed();
183        let total_rows = reader.parquet_metadata().file_metadata().num_rows();
184        println!("{} Reader built in {:?}", "✓".green(), reader_build_elapsed);
185        let reader_stream = Box::pin(stream::try_unfold(reader, |mut reader| async move {
186            let batch = reader.next_record_batch().await?;
187            Ok(batch.map(|batch| (batch, reader)))
188        }));
189
190        // Build write request
191        let fulltext_index_config = FulltextIndexConfig {
192            create_on_compaction: Mode::Disable,
193            ..Default::default()
194        };
195
196        let source =
197            FlatSource::new_stream(region_meta.schema.arrow_schema().clone(), reader_stream);
198        let write_req = SstWriteRequest {
199            op_type: OperationType::Flush,
200            metadata: region_meta,
201            source,
202            cache_manager,
203            storage: None,
204            max_sequence: None,
205            sst_write_format: FormatType::PrimaryKey,
206            index_options: Default::default(),
207            index_config: mito_engine_config.index.clone(),
208            inverted_index_config: MitoConfig::default().inverted_index,
209            fulltext_index_config,
210            bloom_filter_index_config: MitoConfig::default().bloom_filter_index,
211            #[cfg(feature = "vector_index")]
212            vector_index_config: Default::default(),
213        };
214
215        // Write SST
216        println!("{}", "Writing SST...".yellow());
217
218        // Start profiling if pprof_file is specified
219        #[cfg(unix)]
220        let profiler_guard = if self.pprof_file.is_some() {
221            println!("{} Starting profiling...", "⚡".yellow());
222            Some(
223                pprof::ProfilerGuardBuilder::default()
224                    .frequency(99)
225                    .blocklist(&["libc", "libgcc", "pthread", "vdso"])
226                    .build()
227                    .map_err(|e| {
228                        error::IllegalConfigSnafu {
229                            msg: format!("Failed to start profiler: {e}"),
230                        }
231                        .build()
232                    })?,
233            )
234        } else {
235            None
236        };
237
238        #[cfg(not(unix))]
239        if self.pprof_file.is_some() {
240            eprintln!(
241                "{}: Profiling is not supported on this platform",
242                "Warning".yellow()
243            );
244        }
245
246        let write_start = Instant::now();
247        let mut metrics = Metrics::new(WriteType::Flush);
248        let infos = src_access_layer
249            .write_sst(write_req, &WriteOptions::default(), &mut metrics)
250            .await
251            .map_err(|e| {
252                error::IllegalConfigSnafu {
253                    msg: format!("write_sst failed: {e:?}"),
254                }
255                .build()
256            })?;
257
258        let write_elapsed = write_start.elapsed();
259
260        // Stop profiling and generate flamegraph if enabled
261        #[cfg(unix)]
262        if let (Some(guard), Some(pprof_file)) = (profiler_guard, &self.pprof_file) {
263            println!("{} Generating flamegraph...", "🔥".yellow());
264            match guard.report().build() {
265                Ok(report) => {
266                    let mut flamegraph_data = Vec::new();
267                    if let Err(e) = report.flamegraph(&mut flamegraph_data) {
268                        println!("{}: Failed to generate flamegraph: {}", "Error".red(), e);
269                    } else if let Err(e) = std::fs::write(pprof_file, flamegraph_data) {
270                        println!(
271                            "{}: Failed to write flamegraph to {}: {}",
272                            "Error".red(),
273                            pprof_file.display(),
274                            e
275                        );
276                    } else {
277                        println!(
278                            "{} Flamegraph saved to {}",
279                            "✓".green(),
280                            pprof_file.display().to_string().cyan()
281                        );
282                    }
283                }
284                Err(e) => {
285                    println!("{}: Failed to generate pprof report: {}", "Error".red(), e);
286                }
287            }
288        }
289        assert_eq!(infos.len(), 1);
290        let dst_file_id = infos[0].file_id;
291        let dst_file_path = format!("{}/{}.parquet", components.region_dir(), dst_file_id);
292        let mut dst_index_path = None;
293        if infos[0].index_metadata.file_size > 0 {
294            dst_index_path = Some(format!(
295                "{}/index/{}.puffin",
296                components.region_dir(),
297                dst_file_id
298            ));
299        }
300
301        // Report results with ANSI colors
302        println!("\n{} {}", "Write complete!".green().bold(), "✓".green());
303        println!("  {}: {}", "Destination file".bold(), dst_file_path.cyan());
304        println!("  {}: {}", "Rows".bold(), total_rows.to_string().cyan());
305        println!(
306            "  {}: {}",
307            "File size".bold(),
308            format!("{} bytes", file_size).cyan()
309        );
310        println!(
311            "  {}: {:?}",
312            "Reader build time".bold(),
313            reader_build_elapsed
314        );
315        println!("  {}: {:?}", "Total time".bold(), write_elapsed);
316
317        // Print metrics in a formatted way
318        println!("  {}: {:?}", "Metrics".bold(), metrics,);
319
320        // Print infos
321        println!("  {}: {:?}", "Index".bold(), infos[0].index_metadata);
322
323        // Cleanup
324        println!("\n{}", "Cleaning up...".yellow());
325        object_store.delete(&dst_file_path).await.map_err(|e| {
326            error::IllegalConfigSnafu {
327                msg: format!("Failed to delete dest file {}: {}", dst_file_path, e),
328            }
329            .build()
330        })?;
331        println!("{} Temporary file {} deleted", "✓".green(), dst_file_path);
332
333        if let Some(index_path) = dst_index_path {
334            object_store.delete(&index_path).await.map_err(|e| {
335                error::IllegalConfigSnafu {
336                    msg: format!("Failed to delete dest index file {}: {}", index_path, e),
337                }
338                .build()
339            })?;
340            println!(
341                "{} Temporary index file {} deleted",
342                "✓".green(),
343                index_path
344            );
345        }
346
347        println!("\n{}", "Benchmark completed successfully!".green().bold());
348        Ok(())
349    }
350}
351
352#[derive(Debug)]
353struct FileDirComponents {
354    catalog: String,
355    schema: String,
356    table_id: u32,
357    region_sequence: u32,
358    path_type: PathType,
359    file_id: FileId,
360}
361
362impl FileDirComponents {
363    fn table_dir(&self) -> String {
364        format!("data/{}/{}/{}", self.catalog, self.schema, self.table_id)
365    }
366
367    fn region_dir(&self) -> String {
368        let region_name = region_name(self.table_id, self.region_sequence);
369        match self.path_type {
370            PathType::Bare => {
371                format!(
372                    "data/{}/{}/{}/{}",
373                    self.catalog, self.schema, self.table_id, region_name
374                )
375            }
376            PathType::Data => {
377                format!(
378                    "data/{}/{}/{}/{}/data",
379                    self.catalog, self.schema, self.table_id, region_name
380                )
381            }
382            PathType::Metadata => {
383                format!(
384                    "data/{}/{}/{}/{}/metadata",
385                    self.catalog, self.schema, self.table_id, region_name
386                )
387            }
388        }
389    }
390}
391
392fn parse_file_dir_components(path: &str) -> error::Result<FileDirComponents> {
393    // Define the regex pattern to match all three path styles
394    let pattern =
395        r"^data/([^/]+)/([^/]+)/([^/]+)/([^/]+)_([^/]+)(?:/data|/metadata)?/(.+).parquet$";
396
397    // Compile the regex
398    let re = Regex::new(pattern).expect("Invalid regex pattern");
399
400    // Determine the path type
401    let path_type = if path.contains("/data/") {
402        PathType::Data
403    } else if path.contains("/metadata/") {
404        PathType::Metadata
405    } else {
406        PathType::Bare
407    };
408
409    // Try to match the path
410    let components = (|| {
411        let captures = re.captures(path)?;
412        if captures.len() != 7 {
413            return None;
414        }
415        let mut components = FileDirComponents {
416            catalog: "".to_string(),
417            schema: "".to_string(),
418            table_id: 0,
419            region_sequence: 0,
420            path_type,
421            file_id: FileId::default(),
422        };
423        // Extract the components
424        components.catalog = captures.get(1)?.as_str().to_string();
425        components.schema = captures.get(2)?.as_str().to_string();
426        components.table_id = captures[3].parse().ok()?;
427        components.region_sequence = captures[5].parse().ok()?;
428        let file_id_str = &captures[6];
429        components.file_id = FileId::parse_str(file_id_str).ok()?;
430        Some(components)
431    })();
432    components.context(error::IllegalConfigSnafu {
433        msg: format!("Expect valid source file path, got: {}", path),
434    })
435}
436
437async fn build_access_layer_simple(
438    components: &FileDirComponents,
439    object_store: ObjectStore,
440    config: &mut MitoConfig,
441    data_home: &str,
442) -> error::Result<(AccessLayerRef, CacheManagerRef)> {
443    let _ = config.index.sanitize(data_home, &config.inverted_index);
444    let puffin_manager = PuffinManagerFactory::new(
445        &config.index.aux_path,
446        config.index.staging_size.as_bytes(),
447        Some(config.index.write_buffer_size.as_bytes() as _),
448        config.index.staging_ttl,
449    )
450    .await
451    .map_err(|e| {
452        error::IllegalConfigSnafu {
453            msg: format!("Failed to build access layer: {e:?}"),
454        }
455        .build()
456    })?;
457
458    let intermediate_manager = IntermediateManager::init_fs(&config.index.aux_path)
459        .await
460        .map_err(|e| {
461            error::IllegalConfigSnafu {
462                msg: format!("Failed to build IntermediateManager: {e:?}"),
463            }
464            .build()
465        })?
466        .with_buffer_size(Some(config.index.write_buffer_size.as_bytes() as _));
467
468    let cache_manager =
469        build_cache_manager(config, puffin_manager.clone(), intermediate_manager.clone()).await?;
470    let layer = AccessLayer::new(
471        components.table_dir(),
472        components.path_type,
473        object_store,
474        puffin_manager,
475        intermediate_manager,
476    );
477    Ok((Arc::new(layer), cache_manager))
478}
479
480async fn build_cache_manager(
481    config: &MitoConfig,
482    puffin_manager: PuffinManagerFactory,
483    intermediate_manager: IntermediateManager,
484) -> error::Result<CacheManagerRef> {
485    let write_cache = write_cache_from_config(config, puffin_manager, intermediate_manager)
486        .await
487        .map_err(|e| {
488            error::IllegalConfigSnafu {
489                msg: format!("Failed to build write cache: {e:?}"),
490            }
491            .build()
492        })?;
493    let cache_manager = Arc::new(
494        CacheManager::builder()
495            .sst_meta_cache_size(config.sst_meta_cache_size.as_bytes())
496            .vector_cache_size(config.vector_cache_size.as_bytes())
497            .page_cache_size(config.page_cache_size.as_bytes())
498            .selector_result_cache_size(config.selector_result_cache_size.as_bytes())
499            .range_result_cache_size(config.range_result_cache_size.as_bytes())
500            .prefilter_result_cache_size(config.prefilter_result_cache_size.as_bytes())
501            .index_metadata_size(config.index.metadata_cache_size.as_bytes())
502            .index_content_size(config.index.content_cache_size.as_bytes())
503            .index_content_page_size(config.index.content_cache_page_size.as_bytes())
504            .index_result_cache_size(config.index.result_cache_size.as_bytes())
505            .puffin_metadata_size(config.index.metadata_cache_size.as_bytes())
506            .write_cache(write_cache)
507            .build(),
508    );
509    Ok(cache_manager)
510}
511
512fn new_noop_file_purger() -> FilePurgerRef {
513    #[derive(Debug)]
514    struct Noop;
515    impl FilePurger for Noop {
516        fn remove_file(&self, _file_meta: FileMeta, _is_delete: bool, _index_outdated: bool) {}
517    }
518    Arc::new(Noop)
519}
520
521async fn load_parquet_metadata(
522    object_store: ObjectStore,
523    path: &str,
524    file_size: u64,
525) -> Result<parquet::file::metadata::ParquetMetaData, Box<dyn std::error::Error + Send + Sync>> {
526    use parquet::file::FOOTER_SIZE;
527    use parquet::file::metadata::ParquetMetaDataReader;
528    let actual_size = if file_size == 0 {
529        object_store.stat(path).await?.content_length()
530    } else {
531        file_size
532    };
533    if actual_size < FOOTER_SIZE as u64 {
534        return Err("file too small".into());
535    }
536    let prefetch: u64 = 64 * 1024;
537    let start = actual_size.saturating_sub(prefetch);
538    let buffer = object_store
539        .read_with(path)
540        .range(start..actual_size)
541        .await?
542        .to_vec();
543    let buffer_len = buffer.len();
544    let mut footer = [0; 8];
545    footer.copy_from_slice(&buffer[buffer_len - FOOTER_SIZE..]);
546    let footer = FooterTail::try_new(&footer)?;
547    let metadata_len = footer.metadata_length() as u64;
548    if actual_size - (FOOTER_SIZE as u64) < metadata_len {
549        return Err("invalid footer/metadata length".into());
550    }
551    if (metadata_len as usize) <= buffer_len - FOOTER_SIZE {
552        let metadata_start = buffer_len - metadata_len as usize - FOOTER_SIZE;
553        let meta = ParquetMetaDataReader::decode_metadata(
554            &buffer[metadata_start..buffer_len - FOOTER_SIZE],
555        )?;
556        Ok(meta)
557    } else {
558        let metadata_start = actual_size - metadata_len - FOOTER_SIZE as u64;
559        let data = object_store
560            .read_with(path)
561            .range(metadata_start..(actual_size - FOOTER_SIZE as u64))
562            .await?
563            .to_vec();
564        let meta = ParquetMetaDataReader::decode_metadata(&data)?;
565        Ok(meta)
566    }
567}
568
569#[cfg(test)]
570mod tests {
571    use std::path::PathBuf;
572    use std::str::FromStr;
573
574    use common_base::readable_size::ReadableSize;
575    use store_api::region_request::PathType;
576
577    use crate::datanode::objbench::parse_file_dir_components;
578    use crate::datanode::tool_util::parse_config;
579
580    #[test]
581    fn test_parse_dir() {
582        let meta_path = "data/greptime/public/1024/1024_0000000000/metadata/00020380-009c-426d-953e-b4e34c15af34.parquet";
583        let c = parse_file_dir_components(meta_path).unwrap();
584        assert_eq!(
585            c.file_id.to_string(),
586            "00020380-009c-426d-953e-b4e34c15af34"
587        );
588        assert_eq!(c.catalog, "greptime");
589        assert_eq!(c.schema, "public");
590        assert_eq!(c.table_id, 1024);
591        assert_eq!(c.region_sequence, 0);
592        assert_eq!(c.path_type, PathType::Metadata);
593
594        let c = parse_file_dir_components(
595            "data/greptime/public/1024/1024_0000000000/data/00020380-009c-426d-953e-b4e34c15af34.parquet",
596        ).unwrap();
597        assert_eq!(
598            c.file_id.to_string(),
599            "00020380-009c-426d-953e-b4e34c15af34"
600        );
601        assert_eq!(c.catalog, "greptime");
602        assert_eq!(c.schema, "public");
603        assert_eq!(c.table_id, 1024);
604        assert_eq!(c.region_sequence, 0);
605        assert_eq!(c.path_type, PathType::Data);
606
607        let c = parse_file_dir_components(
608            "data/greptime/public/1024/1024_0000000000/00020380-009c-426d-953e-b4e34c15af34.parquet",
609        ).unwrap();
610        assert_eq!(
611            c.file_id.to_string(),
612            "00020380-009c-426d-953e-b4e34c15af34"
613        );
614        assert_eq!(c.catalog, "greptime");
615        assert_eq!(c.schema, "public");
616        assert_eq!(c.table_id, 1024);
617        assert_eq!(c.region_sequence, 0);
618        assert_eq!(c.path_type, PathType::Bare);
619    }
620
621    #[test]
622    fn test_parse_config() {
623        let path = "../../config/datanode.example.toml";
624        let (storage, engine, _wal) = parse_config(&PathBuf::from_str(path).unwrap()).unwrap();
625        assert_eq!(storage.data_home, "./greptimedb_data");
626        assert_eq!(engine.index.staging_size, ReadableSize::gb(2));
627    }
628}