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            max_sequence: None,
204            sst_write_format: FormatType::PrimaryKey,
205            index_options: Default::default(),
206            index_config: mito_engine_config.index.clone(),
207            inverted_index_config: MitoConfig::default().inverted_index,
208            fulltext_index_config,
209            bloom_filter_index_config: MitoConfig::default().bloom_filter_index,
210            preserve_row_sequence: false,
211        };
212
213        // Write SST
214        println!("{}", "Writing SST...".yellow());
215
216        // Start profiling if pprof_file is specified
217        #[cfg(unix)]
218        let profiler_guard = if self.pprof_file.is_some() {
219            println!("{} Starting profiling...", "⚡".yellow());
220            Some(
221                pprof::ProfilerGuardBuilder::default()
222                    .frequency(99)
223                    .blocklist(&["libc", "libgcc", "pthread", "vdso"])
224                    .build()
225                    .map_err(|e| {
226                        error::IllegalConfigSnafu {
227                            msg: format!("Failed to start profiler: {e}"),
228                        }
229                        .build()
230                    })?,
231            )
232        } else {
233            None
234        };
235
236        #[cfg(not(unix))]
237        if self.pprof_file.is_some() {
238            eprintln!(
239                "{}: Profiling is not supported on this platform",
240                "Warning".yellow()
241            );
242        }
243
244        let write_start = Instant::now();
245        let mut metrics = Metrics::new(WriteType::Flush);
246        let infos = src_access_layer
247            .write_sst(write_req, &WriteOptions::default(), &mut metrics)
248            .await
249            .map_err(|e| {
250                error::IllegalConfigSnafu {
251                    msg: format!("write_sst failed: {e:?}"),
252                }
253                .build()
254            })?;
255
256        let write_elapsed = write_start.elapsed();
257
258        // Stop profiling and generate flamegraph if enabled
259        #[cfg(unix)]
260        if let (Some(guard), Some(pprof_file)) = (profiler_guard, &self.pprof_file) {
261            println!("{} Generating flamegraph...", "🔥".yellow());
262            match guard.report().build() {
263                Ok(report) => {
264                    let mut flamegraph_data = Vec::new();
265                    if let Err(e) = report.flamegraph(&mut flamegraph_data) {
266                        println!("{}: Failed to generate flamegraph: {}", "Error".red(), e);
267                    } else if let Err(e) = std::fs::write(pprof_file, flamegraph_data) {
268                        println!(
269                            "{}: Failed to write flamegraph to {}: {}",
270                            "Error".red(),
271                            pprof_file.display(),
272                            e
273                        );
274                    } else {
275                        println!(
276                            "{} Flamegraph saved to {}",
277                            "✓".green(),
278                            pprof_file.display().to_string().cyan()
279                        );
280                    }
281                }
282                Err(e) => {
283                    println!("{}: Failed to generate pprof report: {}", "Error".red(), e);
284                }
285            }
286        }
287        assert_eq!(infos.len(), 1);
288        let dst_file_id = infos[0].file_id;
289        let dst_file_path = format!("{}/{}.parquet", components.region_dir(), dst_file_id);
290        let mut dst_index_path = None;
291        if infos[0].index_metadata.file_size > 0 {
292            dst_index_path = Some(format!(
293                "{}/index/{}.puffin",
294                components.region_dir(),
295                dst_file_id
296            ));
297        }
298
299        // Report results with ANSI colors
300        println!("\n{} {}", "Write complete!".green().bold(), "✓".green());
301        println!("  {}: {}", "Destination file".bold(), dst_file_path.cyan());
302        println!("  {}: {}", "Rows".bold(), total_rows.to_string().cyan());
303        println!(
304            "  {}: {}",
305            "File size".bold(),
306            format!("{} bytes", file_size).cyan()
307        );
308        println!(
309            "  {}: {:?}",
310            "Reader build time".bold(),
311            reader_build_elapsed
312        );
313        println!("  {}: {:?}", "Total time".bold(), write_elapsed);
314
315        // Print metrics in a formatted way
316        println!("  {}: {:?}", "Metrics".bold(), metrics,);
317
318        // Print infos
319        println!("  {}: {:?}", "Index".bold(), infos[0].index_metadata);
320
321        // Cleanup
322        println!("\n{}", "Cleaning up...".yellow());
323        object_store.delete(&dst_file_path).await.map_err(|e| {
324            error::IllegalConfigSnafu {
325                msg: format!("Failed to delete dest file {}: {}", dst_file_path, e),
326            }
327            .build()
328        })?;
329        println!("{} Temporary file {} deleted", "✓".green(), dst_file_path);
330
331        if let Some(index_path) = dst_index_path {
332            object_store.delete(&index_path).await.map_err(|e| {
333                error::IllegalConfigSnafu {
334                    msg: format!("Failed to delete dest index file {}: {}", index_path, e),
335                }
336                .build()
337            })?;
338            println!(
339                "{} Temporary index file {} deleted",
340                "✓".green(),
341                index_path
342            );
343        }
344
345        println!("\n{}", "Benchmark completed successfully!".green().bold());
346        Ok(())
347    }
348}
349
350#[derive(Debug)]
351struct FileDirComponents {
352    catalog: String,
353    schema: String,
354    table_id: u32,
355    region_sequence: u32,
356    path_type: PathType,
357    file_id: FileId,
358}
359
360impl FileDirComponents {
361    fn table_dir(&self) -> String {
362        format!("data/{}/{}/{}", self.catalog, self.schema, self.table_id)
363    }
364
365    fn region_dir(&self) -> String {
366        let region_name = region_name(self.table_id, self.region_sequence);
367        match self.path_type {
368            PathType::Bare => {
369                format!(
370                    "data/{}/{}/{}/{}",
371                    self.catalog, self.schema, self.table_id, region_name
372                )
373            }
374            PathType::Data => {
375                format!(
376                    "data/{}/{}/{}/{}/data",
377                    self.catalog, self.schema, self.table_id, region_name
378                )
379            }
380            PathType::Metadata => {
381                format!(
382                    "data/{}/{}/{}/{}/metadata",
383                    self.catalog, self.schema, self.table_id, region_name
384                )
385            }
386        }
387    }
388}
389
390fn parse_file_dir_components(path: &str) -> error::Result<FileDirComponents> {
391    // Define the regex pattern to match all three path styles
392    let pattern =
393        r"^data/([^/]+)/([^/]+)/([^/]+)/([^/]+)_([^/]+)(?:/data|/metadata)?/(.+).parquet$";
394
395    // Compile the regex
396    let re = Regex::new(pattern).expect("Invalid regex pattern");
397
398    // Determine the path type
399    let path_type = if path.contains("/data/") {
400        PathType::Data
401    } else if path.contains("/metadata/") {
402        PathType::Metadata
403    } else {
404        PathType::Bare
405    };
406
407    // Try to match the path
408    let components = (|| {
409        let captures = re.captures(path)?;
410        if captures.len() != 7 {
411            return None;
412        }
413        let mut components = FileDirComponents {
414            catalog: "".to_string(),
415            schema: "".to_string(),
416            table_id: 0,
417            region_sequence: 0,
418            path_type,
419            file_id: FileId::default(),
420        };
421        // Extract the components
422        components.catalog = captures.get(1)?.as_str().to_string();
423        components.schema = captures.get(2)?.as_str().to_string();
424        components.table_id = captures[3].parse().ok()?;
425        components.region_sequence = captures[5].parse().ok()?;
426        let file_id_str = &captures[6];
427        components.file_id = FileId::parse_str(file_id_str).ok()?;
428        Some(components)
429    })();
430    components.context(error::IllegalConfigSnafu {
431        msg: format!("Expect valid source file path, got: {}", path),
432    })
433}
434
435async fn build_access_layer_simple(
436    components: &FileDirComponents,
437    object_store: ObjectStore,
438    config: &mut MitoConfig,
439    data_home: &str,
440) -> error::Result<(AccessLayerRef, CacheManagerRef)> {
441    let _ = config.index.sanitize(data_home, &config.inverted_index);
442    let puffin_manager = PuffinManagerFactory::new(
443        &config.index.aux_path,
444        config.index.staging_size.as_bytes(),
445        Some(config.index.write_buffer_size.as_bytes() as _),
446        config.index.staging_ttl,
447    )
448    .await
449    .map_err(|e| {
450        error::IllegalConfigSnafu {
451            msg: format!("Failed to build access layer: {e:?}"),
452        }
453        .build()
454    })?;
455
456    let intermediate_manager = IntermediateManager::init_fs(&config.index.aux_path)
457        .await
458        .map_err(|e| {
459            error::IllegalConfigSnafu {
460                msg: format!("Failed to build IntermediateManager: {e:?}"),
461            }
462            .build()
463        })?
464        .with_buffer_size(Some(config.index.write_buffer_size.as_bytes() as _));
465
466    let cache_manager =
467        build_cache_manager(config, puffin_manager.clone(), intermediate_manager.clone()).await?;
468    let layer = AccessLayer::new(
469        components.table_dir(),
470        components.path_type,
471        object_store,
472        puffin_manager,
473        intermediate_manager,
474    );
475    Ok((Arc::new(layer), cache_manager))
476}
477
478async fn build_cache_manager(
479    config: &MitoConfig,
480    puffin_manager: PuffinManagerFactory,
481    intermediate_manager: IntermediateManager,
482) -> error::Result<CacheManagerRef> {
483    let write_cache = write_cache_from_config(config, puffin_manager, intermediate_manager, None)
484        .await
485        .map_err(|e| {
486            error::IllegalConfigSnafu {
487                msg: format!("Failed to build write cache: {e:?}"),
488            }
489            .build()
490        })?;
491    let cache_manager = Arc::new(
492        CacheManager::builder()
493            .sst_meta_cache_size(config.sst_meta_cache_size.as_bytes())
494            .vector_cache_size(config.vector_cache_size.as_bytes())
495            .page_cache_size(config.page_cache_size.as_bytes())
496            .selector_result_cache_size(config.selector_result_cache_size.as_bytes())
497            .range_result_cache_size(config.range_result_cache_size.as_bytes())
498            .prefilter_result_cache_size(config.prefilter_result_cache_size.as_bytes())
499            .index_metadata_size(config.index.metadata_cache_size.as_bytes())
500            .index_content_size(config.index.content_cache_size.as_bytes())
501            .index_content_page_size(config.index.content_cache_page_size.as_bytes())
502            .index_result_cache_size(config.index.result_cache_size.as_bytes())
503            .puffin_metadata_size(config.index.metadata_cache_size.as_bytes())
504            .write_cache(write_cache)
505            .build(),
506    );
507    Ok(cache_manager)
508}
509
510fn new_noop_file_purger() -> FilePurgerRef {
511    #[derive(Debug)]
512    struct Noop;
513    impl FilePurger for Noop {
514        fn remove_file(&self, _file_meta: FileMeta, _is_delete: bool, _index_outdated: bool) {}
515    }
516    Arc::new(Noop)
517}
518
519async fn load_parquet_metadata(
520    object_store: ObjectStore,
521    path: &str,
522    file_size: u64,
523) -> Result<parquet::file::metadata::ParquetMetaData, Box<dyn std::error::Error + Send + Sync>> {
524    use parquet::file::FOOTER_SIZE;
525    use parquet::file::metadata::ParquetMetaDataReader;
526    let actual_size = if file_size == 0 {
527        object_store.stat(path).await?.content_length()
528    } else {
529        file_size
530    };
531    if actual_size < FOOTER_SIZE as u64 {
532        return Err("file too small".into());
533    }
534    let prefetch: u64 = 64 * 1024;
535    let start = actual_size.saturating_sub(prefetch);
536    let buffer = object_store
537        .read_with(path)
538        .range(start..actual_size)
539        .await?
540        .to_vec();
541    let buffer_len = buffer.len();
542    let mut footer = [0; 8];
543    footer.copy_from_slice(&buffer[buffer_len - FOOTER_SIZE..]);
544    let footer = FooterTail::try_new(&footer)?;
545    let metadata_len = footer.metadata_length() as u64;
546    if actual_size - (FOOTER_SIZE as u64) < metadata_len {
547        return Err("invalid footer/metadata length".into());
548    }
549    if (metadata_len as usize) <= buffer_len - FOOTER_SIZE {
550        let metadata_start = buffer_len - metadata_len as usize - FOOTER_SIZE;
551        let meta = ParquetMetaDataReader::decode_metadata(
552            &buffer[metadata_start..buffer_len - FOOTER_SIZE],
553        )?;
554        Ok(meta)
555    } else {
556        let metadata_start = actual_size - metadata_len - FOOTER_SIZE as u64;
557        let data = object_store
558            .read_with(path)
559            .range(metadata_start..(actual_size - FOOTER_SIZE as u64))
560            .await?
561            .to_vec();
562        let meta = ParquetMetaDataReader::decode_metadata(&data)?;
563        Ok(meta)
564    }
565}
566
567#[cfg(test)]
568mod tests {
569    use std::path::PathBuf;
570    use std::str::FromStr;
571
572    use common_base::readable_size::ReadableSize;
573    use store_api::region_request::PathType;
574
575    use crate::datanode::objbench::parse_file_dir_components;
576    use crate::datanode::tool_util::parse_config;
577
578    #[test]
579    fn test_parse_dir() {
580        let meta_path = "data/greptime/public/1024/1024_0000000000/metadata/00020380-009c-426d-953e-b4e34c15af34.parquet";
581        let c = parse_file_dir_components(meta_path).unwrap();
582        assert_eq!(
583            c.file_id.to_string(),
584            "00020380-009c-426d-953e-b4e34c15af34"
585        );
586        assert_eq!(c.catalog, "greptime");
587        assert_eq!(c.schema, "public");
588        assert_eq!(c.table_id, 1024);
589        assert_eq!(c.region_sequence, 0);
590        assert_eq!(c.path_type, PathType::Metadata);
591
592        let c = parse_file_dir_components(
593            "data/greptime/public/1024/1024_0000000000/data/00020380-009c-426d-953e-b4e34c15af34.parquet",
594        ).unwrap();
595        assert_eq!(
596            c.file_id.to_string(),
597            "00020380-009c-426d-953e-b4e34c15af34"
598        );
599        assert_eq!(c.catalog, "greptime");
600        assert_eq!(c.schema, "public");
601        assert_eq!(c.table_id, 1024);
602        assert_eq!(c.region_sequence, 0);
603        assert_eq!(c.path_type, PathType::Data);
604
605        let c = parse_file_dir_components(
606            "data/greptime/public/1024/1024_0000000000/00020380-009c-426d-953e-b4e34c15af34.parquet",
607        ).unwrap();
608        assert_eq!(
609            c.file_id.to_string(),
610            "00020380-009c-426d-953e-b4e34c15af34"
611        );
612        assert_eq!(c.catalog, "greptime");
613        assert_eq!(c.schema, "public");
614        assert_eq!(c.table_id, 1024);
615        assert_eq!(c.region_sequence, 0);
616        assert_eq!(c.path_type, PathType::Bare);
617    }
618
619    #[test]
620    fn test_parse_config() {
621        let path = "../../config/datanode.example.toml";
622        let (storage, engine, _wal) = parse_config(&PathBuf::from_str(path).unwrap()).unwrap();
623        assert_eq!(storage.data_home, "./greptimedb_data");
624        assert_eq!(engine.index.staging_size, ReadableSize::gb(2));
625    }
626}