mito2/compaction/
reader.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::collections::HashMap;
16use std::sync::Arc;
17
18use common_time::Timestamp;
19use common_time::range::TimestampRange;
20use common_time::timestamp::TimeUnit;
21use datafusion_common::ScalarValue;
22use datafusion_expr::Expr;
23use datatypes::extension::json::is_json2_extension_type;
24use datatypes::types::json_type::JsonNativeType;
25use parquet::arrow::parquet_to_arrow_schema;
26use parquet::file::metadata::{PageIndexPolicy, ParquetMetaData};
27use snafu::{OptionExt, ResultExt};
28use store_api::metadata::RegionMetadataRef;
29
30use crate::access_layer::AccessLayerRef;
31use crate::cache::{CacheManagerRef, CacheStrategy};
32use crate::error::{
33    DataTypeMismatchSnafu, ParquetToArrowSchemaSnafu, Result, TimeRangePredicateOverflowSnafu,
34};
35use crate::read::FlatSource;
36use crate::read::flat_projection::FlatProjectionMapper;
37use crate::read::read_columns::ReadColumns;
38use crate::read::scan_region::{PredicateGroup, ScanInput};
39use crate::read::seq_scan::SeqScan;
40use crate::region::options::MergeMode;
41use crate::sst::file::FileHandle;
42use crate::sst::parquet::reader::MetadataCacheMetrics;
43
44/// Builders to create [BoxedRecordBatchStream] for compaction.
45pub(crate) struct CompactionSstReaderBuilder<'a> {
46    pub(crate) metadata: RegionMetadataRef,
47    pub(crate) sst_layer: AccessLayerRef,
48    pub(crate) cache: CacheManagerRef,
49    pub(crate) inputs: &'a [FileHandle],
50    pub(crate) append_mode: bool,
51    pub(crate) filter_deleted: bool,
52    pub(crate) time_range: Option<TimestampRange>,
53    pub(crate) merge_mode: MergeMode,
54}
55
56impl CompactionSstReaderBuilder<'_> {
57    /// Build a [FlatSource] that yields Arrow `RecordBatch`s from reading all the input SST files,
58    /// for compaction. The schema of the [FlatSource] is unified.
59    pub(crate) async fn build_flat_sst_reader(self) -> Result<FlatSource> {
60        let scan_input = self.build_scan_input().await?;
61
62        let schema = scan_input.mapper.output_schema();
63        let schema = schema.arrow_schema();
64
65        let stream = SeqScan::new(scan_input)
66            .build_flat_reader_for_compaction()
67            .await?;
68        Ok(FlatSource::new_stream(schema.clone(), stream))
69    }
70
71    async fn build_scan_input(self) -> Result<ScanInput> {
72        let schema = self.metadata.schema.arrow_schema();
73        let parquet_metadata = self.collect_parquet_metadata().await?;
74        let batch_size = crate::batch_size::estimate_batch_size(
75            parquet_metadata
76                .iter()
77                .flat_map(|metadata| metadata.row_groups())
78                .map(|row_group| {
79                    let uncompressed_bytes = row_group
80                        .columns()
81                        .iter()
82                        .map(|column| column.uncompressed_size() as u64)
83                        .sum();
84                    (row_group.num_rows() as u64, uncompressed_bytes)
85                }),
86        );
87        let json_type_hint = if schema.fields().iter().any(is_json2_extension_type) {
88            let mut json_type_hint = schema
89                .fields()
90                .iter()
91                .filter(|&field| is_json2_extension_type(field))
92                .map(|field| (field.name().clone(), JsonNativeType::Null))
93                .collect::<HashMap<_, _>>();
94
95            for metadata in &parquet_metadata {
96                let file_metadata = metadata.file_metadata();
97                let schema = parquet_to_arrow_schema(
98                    file_metadata.schema_descr(),
99                    file_metadata.key_value_metadata(),
100                )
101                .context(ParquetToArrowSchemaSnafu {
102                    file: "compaction input",
103                })?;
104                for field in schema.fields() {
105                    let Some(merged) = json_type_hint.get_mut(field.name()) else {
106                        continue;
107                    };
108
109                    let json_type = JsonNativeType::try_from(field.data_type())
110                        .context(DataTypeMismatchSnafu)?;
111                    merged.merge(&json_type);
112                }
113            }
114
115            Some(json_type_hint)
116        } else {
117            None
118        };
119
120        let projection = (0..self.metadata.column_metadatas.len()).collect();
121        let read_columns = ReadColumns::from_deduped_column_ids(
122            self.metadata.column_metadatas.iter().map(|x| x.column_id),
123        );
124        let mapper = FlatProjectionMapper::new_with_read_columns(
125            &self.metadata,
126            projection,
127            read_columns,
128            json_type_hint.as_ref(),
129        )?;
130
131        let mut scan_input = ScanInput::new(self.sst_layer, mapper)
132            .with_files(self.inputs.to_vec())
133            .with_compaction(true)
134            .with_batch_size(batch_size)
135            .with_append_mode(self.append_mode)
136            // We use special cache strategy for compaction.
137            .with_cache(CacheStrategy::Compaction(self.cache))
138            .with_filter_deleted(self.filter_deleted)
139            // We ignore file not found error during compaction.
140            .with_ignore_file_not_found(true)
141            .with_merge_mode(self.merge_mode);
142
143        // This serves as a workaround of https://github.com/GreptimeTeam/greptimedb/issues/3944
144        // by converting time ranges into predicate.
145        if let Some(time_range) = self.time_range {
146            scan_input =
147                scan_input.with_predicate(time_range_to_predicate(time_range, &self.metadata)?);
148        }
149
150        Ok(scan_input)
151    }
152
153    async fn collect_parquet_metadata(&self) -> Result<Vec<Arc<ParquetMetaData>>> {
154        let mut metadata = Vec::with_capacity(self.inputs.len());
155
156        for file_handle in self.inputs {
157            let file_path =
158                file_handle.file_path(self.sst_layer.table_dir(), self.sst_layer.path_type());
159            let file_size = file_handle.meta_ref().file_size;
160            let parquet_metadata = match self
161                .sst_layer
162                .read_sst(file_handle.clone())
163                .cache(CacheStrategy::Compaction(self.cache.clone()))
164                .read_parquet_metadata(
165                    &file_path,
166                    file_size,
167                    &mut MetadataCacheMetrics::default(),
168                    PageIndexPolicy::default(),
169                )
170                .await
171                .map(|x| x.0.parquet_metadata())
172            {
173                Ok(x) => x,
174                Err(e) if e.is_object_not_found() => continue,
175                Err(e) => return Err(e),
176            };
177            metadata.push(parquet_metadata);
178        }
179        Ok(metadata)
180    }
181}
182
183/// Converts time range to predicates so that rows outside the range will be filtered.
184fn time_range_to_predicate(
185    range: TimestampRange,
186    metadata: &RegionMetadataRef,
187) -> Result<PredicateGroup> {
188    let ts_col = metadata.time_index_column();
189
190    // safety: time index column's type must be a valid timestamp type.
191    let ts_col_unit = ts_col
192        .column_schema
193        .data_type
194        .as_timestamp()
195        .unwrap()
196        .unit();
197
198    let exprs = match (range.start(), range.end()) {
199        (Some(start), Some(end)) => {
200            vec![
201                datafusion_expr::col(ts_col.column_schema.name.clone())
202                    .gt_eq(ts_to_lit(*start, ts_col_unit)?),
203                datafusion_expr::col(ts_col.column_schema.name.clone())
204                    .lt(ts_to_lit(*end, ts_col_unit)?),
205            ]
206        }
207        (Some(start), None) => {
208            vec![
209                datafusion_expr::col(ts_col.column_schema.name.clone())
210                    .gt_eq(ts_to_lit(*start, ts_col_unit)?),
211            ]
212        }
213
214        (None, Some(end)) => {
215            vec![
216                datafusion_expr::col(ts_col.column_schema.name.clone())
217                    .lt(ts_to_lit(*end, ts_col_unit)?),
218            ]
219        }
220        (None, None) => {
221            return Ok(PredicateGroup::default());
222        }
223    };
224
225    let predicate = PredicateGroup::new(metadata, &exprs)?;
226    Ok(predicate)
227}
228
229fn ts_to_lit(ts: Timestamp, ts_col_unit: TimeUnit) -> Result<Expr> {
230    let ts = ts
231        .convert_to(ts_col_unit)
232        .context(TimeRangePredicateOverflowSnafu {
233            timestamp: ts,
234            unit: ts_col_unit,
235        })?;
236    let val = ts.value();
237    let scalar_value = match ts_col_unit {
238        TimeUnit::Second => ScalarValue::TimestampSecond(Some(val), None),
239        TimeUnit::Millisecond => ScalarValue::TimestampMillisecond(Some(val), None),
240        TimeUnit::Microsecond => ScalarValue::TimestampMicrosecond(Some(val), None),
241        TimeUnit::Nanosecond => ScalarValue::TimestampNanosecond(Some(val), None),
242    };
243    Ok(datafusion_expr::lit(scalar_value))
244}
mito2/compaction/reader.rs

mito2/compaction/
reader.rs
