1pub(crate) mod file_stream;
16
17use std::collections::HashSet;
18use std::pin::Pin;
19use std::sync::Arc;
20use std::task::{Context, Poll};
21
22use common_datasource::object_store::build_backend;
23use common_error::ext::BoxedError;
24use common_recordbatch::adapter::RecordBatchMetrics;
25use common_recordbatch::error::{CastVectorSnafu, ExternalSnafu, Result as RecordBatchResult};
26use common_recordbatch::{OrderOption, RecordBatch, RecordBatchStream, SendableRecordBatchStream};
27use datafusion::logical_expr::utils as df_logical_expr_utils;
28use datafusion_expr::expr::Expr;
29use datatypes::prelude::ConcreteDataType;
30use datatypes::schema::{ColumnSchema, Schema, SchemaRef};
31use datatypes::vectors::VectorRef;
32use futures::Stream;
33use snafu::{ensure, OptionExt, ResultExt};
34use store_api::storage::ScanRequest;
35
36use self::file_stream::ScanPlanConfig;
37use crate::error::{
38 BuildBackendSnafu, CreateDefaultSnafu, ExtractColumnFromFilterSnafu,
39 MissingColumnNoDefaultSnafu, ProjectSchemaSnafu, ProjectionOutOfBoundsSnafu, Result,
40};
41use crate::region::FileRegion;
42
43impl FileRegion {
44 pub fn query(&self, request: ScanRequest) -> Result<SendableRecordBatchStream> {
45 let store = build_backend(&self.url, &self.options).context(BuildBackendSnafu)?;
46
47 let file_projection = self.projection_pushdown_to_file(&request.projection)?;
48 let file_filters = self.filters_pushdown_to_file(&request.filters)?;
49 let file_schema = Arc::new(Schema::new(self.file_options.file_column_schemas.clone()));
50
51 let file_stream = file_stream::create_stream(
52 &self.format,
53 &ScanPlanConfig {
54 file_schema,
55 files: &self.file_options.files,
56 projection: file_projection.as_ref(),
57 filters: &file_filters,
58 limit: request.limit,
59 store,
60 },
61 )?;
62
63 let scan_schema = self.scan_schema(&request.projection)?;
64
65 Ok(Box::pin(FileToScanRegionStream::new(
66 scan_schema,
67 file_stream,
68 )))
69 }
70
71 fn projection_pushdown_to_file(
72 &self,
73 req_projection: &Option<Vec<usize>>,
74 ) -> Result<Option<Vec<usize>>> {
75 let Some(scan_projection) = req_projection.as_ref() else {
76 return Ok(None);
77 };
78
79 let file_column_schemas = &self.file_options.file_column_schemas;
80 let mut file_projection = Vec::with_capacity(scan_projection.len());
81 for column_index in scan_projection {
82 ensure!(
83 *column_index < self.metadata.schema.num_columns(),
84 ProjectionOutOfBoundsSnafu {
85 column_index: *column_index,
86 bounds: self.metadata.schema.num_columns()
87 }
88 );
89
90 let column_name = self.metadata.schema.column_name_by_index(*column_index);
91 let file_column_index = file_column_schemas
92 .iter()
93 .position(|c| c.name == column_name);
94 if let Some(file_column_index) = file_column_index {
95 file_projection.push(file_column_index);
96 }
97 }
98 Ok(Some(file_projection))
99 }
100
101 fn filters_pushdown_to_file(&self, scan_filters: &[Expr]) -> Result<Vec<Expr>> {
104 let mut file_filters = Vec::with_capacity(scan_filters.len());
105
106 let file_column_names = self
107 .file_options
108 .file_column_schemas
109 .iter()
110 .map(|c| &c.name)
111 .collect::<HashSet<_>>();
112
113 let mut aux_column_set = HashSet::new();
114 for scan_filter in scan_filters {
115 df_logical_expr_utils::expr_to_columns(scan_filter, &mut aux_column_set)
116 .context(ExtractColumnFromFilterSnafu)?;
117
118 let all_file_columns = aux_column_set
119 .iter()
120 .all(|column_in_expr| file_column_names.contains(&column_in_expr.name));
121 if all_file_columns {
122 file_filters.push(scan_filter.clone());
123 }
124 aux_column_set.clear();
125 }
126 Ok(file_filters)
127 }
128
129 fn scan_schema(&self, req_projection: &Option<Vec<usize>>) -> Result<SchemaRef> {
130 let schema = if let Some(indices) = req_projection {
131 Arc::new(
132 self.metadata
133 .schema
134 .try_project(indices)
135 .context(ProjectSchemaSnafu)?,
136 )
137 } else {
138 self.metadata.schema.clone()
139 };
140
141 Ok(schema)
142 }
143}
144
145struct FileToScanRegionStream {
146 scan_schema: SchemaRef,
147 file_stream: SendableRecordBatchStream,
148}
149
150impl RecordBatchStream for FileToScanRegionStream {
151 fn schema(&self) -> SchemaRef {
152 self.scan_schema.clone()
153 }
154
155 fn output_ordering(&self) -> Option<&[OrderOption]> {
156 None
157 }
158
159 fn metrics(&self) -> Option<RecordBatchMetrics> {
160 None
161 }
162}
163
164impl Stream for FileToScanRegionStream {
165 type Item = RecordBatchResult<RecordBatch>;
166
167 fn poll_next(mut self: Pin<&mut Self>, ctx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
168 match Pin::new(&mut self.file_stream).poll_next(ctx) {
169 Poll::Pending => Poll::Pending,
170 Poll::Ready(Some(file_record_batch)) => {
171 let file_record_batch = file_record_batch?;
172 let scan_record_batch = if self.schema_eq(&file_record_batch) {
173 Ok(file_record_batch)
174 } else {
175 self.convert_record_batch(&file_record_batch)
176 };
177
178 Poll::Ready(Some(scan_record_batch))
179 }
180 Poll::Ready(None) => Poll::Ready(None),
181 }
182 }
183}
184
185impl FileToScanRegionStream {
186 fn new(scan_schema: SchemaRef, file_stream: SendableRecordBatchStream) -> Self {
187 Self {
188 scan_schema,
189 file_stream,
190 }
191 }
192
193 fn schema_eq(&self, file_record_batch: &RecordBatch) -> bool {
194 self.scan_schema
195 .column_schemas()
196 .iter()
197 .all(|scan_column_schema| {
198 file_record_batch
199 .column_by_name(&scan_column_schema.name)
200 .map(|rb| rb.data_type() == scan_column_schema.data_type)
201 .unwrap_or_default()
202 })
203 }
204
205 fn convert_record_batch(
212 &self,
213 file_record_batch: &RecordBatch,
214 ) -> RecordBatchResult<RecordBatch> {
215 let file_row_count = file_record_batch.num_rows();
216 let columns = self
217 .scan_schema
218 .column_schemas()
219 .iter()
220 .map(|scan_column_schema| {
221 let file_column = file_record_batch.column_by_name(&scan_column_schema.name);
222 if let Some(file_column) = file_column {
223 Self::cast_column_type(file_column, &scan_column_schema.data_type)
224 } else {
225 Self::backfill_column(scan_column_schema, file_row_count)
226 }
227 })
228 .collect::<RecordBatchResult<Vec<_>>>()?;
229
230 RecordBatch::new(self.scan_schema.clone(), columns)
231 }
232
233 fn cast_column_type(
234 source_column: &VectorRef,
235 target_data_type: &ConcreteDataType,
236 ) -> RecordBatchResult<VectorRef> {
237 if &source_column.data_type() == target_data_type {
238 Ok(source_column.clone())
239 } else {
240 source_column
241 .cast(target_data_type)
242 .context(CastVectorSnafu {
243 from_type: source_column.data_type(),
244 to_type: target_data_type.clone(),
245 })
246 }
247 }
248
249 fn backfill_column(
250 column_schema: &ColumnSchema,
251 num_rows: usize,
252 ) -> RecordBatchResult<VectorRef> {
253 Self::create_default_vector(column_schema, num_rows)
254 .map_err(BoxedError::new)
255 .context(ExternalSnafu)
256 }
257
258 fn create_default_vector(column_schema: &ColumnSchema, num_rows: usize) -> Result<VectorRef> {
259 column_schema
260 .create_default_vector(num_rows)
261 .with_context(|_| CreateDefaultSnafu {
262 column: column_schema.name.clone(),
263 })?
264 .with_context(|| MissingColumnNoDefaultSnafu {
265 column: column_schema.name.clone(),
266 })
267 }
268}