iceberg/spec/manifest/data_file.rs
1// Licensed to the Apache Software Foundation (ASF) under one
2// or more contributor license agreements. See the NOTICE file
3// distributed with this work for additional information
4// regarding copyright ownership. The ASF licenses this file
5// to you under the Apache License, Version 2.0 (the
6// "License"); you may not use this file except in compliance
7// with the License. You may obtain a copy of the License at
8//
9// http://www.apache.org/licenses/LICENSE-2.0
10//
11// Unless required by applicable law or agreed to in writing,
12// software distributed under the License is distributed on an
13// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
14// KIND, either express or implied. See the License for the
15// specific language governing permissions and limitations
16// under the License.
17
18use std::collections::HashMap;
19use std::io::{Read, Write};
20use std::str::FromStr;
21
22use apache_avro::{Reader as AvroReader, from_value, to_value};
23use serde_derive::{Deserialize, Serialize};
24use serde_with::{DeserializeFromStr, SerializeDisplay};
25
26use super::_serde::DataFileSerde;
27use super::{
28 Datum, FormatVersion, Schema, data_file_schema_v1, data_file_schema_v2, data_file_schema_v3,
29};
30use crate::Error;
31use crate::avro::new_writer;
32use crate::error::{Result, invalid_data};
33use crate::spec::{
34 DEFAULT_PARTITION_SPEC_ID, Literal, NestedFieldRef, PrimitiveType, Struct, StructType, Type,
35};
36
37/// Data file carries data file path, partition tuple, metrics, …
38#[derive(Debug, PartialEq, Clone, Eq, Builder)]
39pub struct DataFile {
40 /// field id: 134
41 ///
42 /// Type of content stored by the data file: data, equality deletes,
43 /// or position deletes (all v1 files are data files)
44 pub(crate) content: DataContentType,
45 /// field id: 100
46 ///
47 /// Full URI for the file with FS scheme
48 pub(crate) file_path: String,
49 /// field id: 101
50 ///
51 /// String file format name, `avro`, `orc`, `parquet`, or `puffin`
52 pub(crate) file_format: DataFileFormat,
53 /// field id: 102
54 ///
55 /// Partition data tuple, schema based on the partition spec output using
56 /// partition field ids for the struct field ids
57 #[builder(default = "Struct::empty()")]
58 pub(crate) partition: Struct,
59 /// field id: 103
60 ///
61 /// Number of records in this file, or the cardinality of a deletion vector
62 pub(crate) record_count: u64,
63 /// field id: 104
64 ///
65 /// Total file size in bytes
66 pub(crate) file_size_in_bytes: u64,
67 /// field id: 108
68 /// key field id: 117
69 /// value field id: 118
70 ///
71 /// Map from column id to the total size on disk of all regions that
72 /// store the column. Does not include bytes necessary to read other
73 /// columns, like footers. Leave null for row-oriented formats (Avro)
74 #[builder(default)]
75 pub(crate) column_sizes: HashMap<i32, u64>,
76 /// field id: 109
77 /// key field id: 119
78 /// value field id: 120
79 ///
80 /// Map from column id to number of values in the column (including null
81 /// and NaN values)
82 #[builder(default)]
83 pub(crate) value_counts: HashMap<i32, u64>,
84 /// field id: 110
85 /// key field id: 121
86 /// value field id: 122
87 ///
88 /// Map from column id to number of null values in the column
89 #[builder(default)]
90 pub(crate) null_value_counts: HashMap<i32, u64>,
91 /// field id: 137
92 /// key field id: 138
93 /// value field id: 139
94 ///
95 /// Map from column id to number of NaN values in the column
96 #[builder(default)]
97 pub(crate) nan_value_counts: HashMap<i32, u64>,
98 /// field id: 125
99 /// key field id: 126
100 /// value field id: 127
101 ///
102 /// Map from column id to lower bound in the column serialized as binary.
103 /// Each value must be less than or equal to all non-null, non-NaN values
104 /// in the column for the file.
105 ///
106 /// Reference:
107 ///
108 /// - [Binary single-value serialization](https://iceberg.apache.org/spec/#binary-single-value-serialization)
109 #[builder(default)]
110 pub(crate) lower_bounds: HashMap<i32, Datum>,
111 /// field id: 128
112 /// key field id: 129
113 /// value field id: 130
114 ///
115 /// Map from column id to upper bound in the column serialized as binary.
116 /// Each value must be greater than or equal to all non-null, non-Nan
117 /// values in the column for the file.
118 ///
119 /// Reference:
120 ///
121 /// - [Binary single-value serialization](https://iceberg.apache.org/spec/#binary-single-value-serialization)
122 #[builder(default)]
123 pub(crate) upper_bounds: HashMap<i32, Datum>,
124 /// field id: 131
125 ///
126 /// Implementation-specific key metadata for encryption
127 #[builder(default)]
128 pub(crate) key_metadata: Option<Vec<u8>>,
129 /// field id: 132
130 /// element field id: 133
131 ///
132 /// Split offsets for the data file. For example, all row group offsets
133 /// in a Parquet file. Must be sorted ascending. Optional field that
134 /// should be serialized as null when not present.
135 #[builder(default)]
136 pub(crate) split_offsets: Option<Vec<i64>>,
137 /// field id: 135
138 /// element field id: 136
139 ///
140 /// Field ids used to determine row equality in equality delete files.
141 /// Required when content is EqualityDeletes and should be null
142 /// otherwise. Fields with ids listed in this column must be present
143 /// in the delete file
144 #[builder(default)]
145 pub(crate) equality_ids: Option<Vec<i32>>,
146 /// field id: 140
147 ///
148 /// ID representing sort order for this file.
149 ///
150 /// If sort order ID is missing or unknown, then the order is assumed to
151 /// be unsorted. Only data files and equality delete files should be
152 /// written with a non-null order id. Position deletes are required to be
153 /// sorted by file and position, not a table order, and should set sort
154 /// order id to null. Readers must ignore sort order id for position
155 /// delete files.
156 #[builder(default, setter(strip_option))]
157 pub(crate) sort_order_id: Option<i32>,
158 /// field id: 142
159 ///
160 /// The _row_id for the first row in the data file.
161 /// For more details, refer to <https://github.com/apache/iceberg/blob/main/format/spec.md#first-row-id-inheritance>
162 #[builder(default)]
163 pub(crate) first_row_id: Option<i64>,
164 /// This field is not included in spec. It is just store in memory representation used
165 /// in process.
166 #[builder(default = "DEFAULT_PARTITION_SPEC_ID")]
167 pub(crate) partition_spec_id: i32,
168 /// field id: 143
169 ///
170 /// Fully qualified location (URI with FS scheme) of a data file that all deletes reference.
171 /// Position delete metadata can use `referenced_data_file` when all deletes tracked by the
172 /// entry are in a single data file. Setting the referenced file is required for deletion vectors.
173 #[builder(default)]
174 pub(crate) referenced_data_file: Option<String>,
175 /// field: 144
176 ///
177 /// The offset in the file where the content starts.
178 /// The `content_offset` and `content_size_in_bytes` fields are used to reference a specific blob
179 /// for direct access to a deletion vector. For deletion vectors, these values are required and must
180 /// exactly match the `offset` and `length` stored in the Puffin footer for the deletion vector blob.
181 #[builder(default)]
182 pub(crate) content_offset: Option<i64>,
183 /// field: 145
184 ///
185 /// The length of a referenced content stored in the file; required if `content_offset` is present
186 #[builder(default)]
187 pub(crate) content_size_in_bytes: Option<i64>,
188}
189
190fn promote_partition_value(
191 value: Option<&Literal>,
192 current_field: &NestedFieldRef,
193 written_field: &NestedFieldRef,
194) -> Option<Literal> {
195 match value {
196 Some(Literal::Primitive(primitive)) => {
197 let promoted = match current_field.field_type.as_primitive_type() {
198 // The source column was dropped, so there is no current type to widen to.
199 // Widen a numeric to its promotion ceiling so this value and its deletes
200 // normalize the same way; keep every other type (e.g. a `date`, stored as
201 // an `int`) at its written type so deletes that skip normalization match.
202 Some(PrimitiveType::Unknown) => {
203 match written_field.field_type.as_primitive_type() {
204 Some(PrimitiveType::Int | PrimitiveType::Float) => primitive.widen(),
205 _ => None,
206 }
207 }
208 Some(target) => primitive.promote_to(target),
209 None => None,
210 };
211 promoted.map(Literal::Primitive).or_else(|| value.cloned())
212 }
213 other => other.cloned(),
214 }
215}
216
217impl DataFile {
218 /// Returns this file's partition tuple normalized to the current schema: a field
219 /// promoted (`int` -> `long`, `float` -> `double`) since the manifest was written is
220 /// widened to its `current` type; a dropped numeric source, which has no `current`
221 /// type, is widened to its promotion ceiling by its `written` type; all other values
222 /// are copied through. Column bounds are widened separately, lazily, where they are
223 /// compared.
224 ///
225 /// `current` and `written` are the file's spec resolved against the current schema and
226 /// against the manifest's schema, so they match the stored tuple in length.
227 ///
228 /// # Errors
229 /// Returns `DataInvalid` if the stored tuple and `current` have different lengths; the
230 /// lengths should always match, since both come from the same spec.
231 pub(crate) fn promoted_partition(
232 &self,
233 current: &StructType,
234 written: &StructType,
235 ) -> Result<Struct> {
236 if self.partition.fields().len() != current.fields().len() {
237 return Err(invalid_data!(
238 "Partition tuple has {} values but the spec partition type has {} fields",
239 self.partition.fields().len(),
240 current.fields().len()
241 ));
242 }
243
244 Ok(self
245 .partition
246 .iter()
247 .zip(current.fields())
248 .zip(written.fields())
249 .map(|((value, current_field), written_field)| {
250 promote_partition_value(value, current_field, written_field)
251 })
252 .collect())
253 }
254
255 /// Get the content type of the data file (data, equality deletes, or position deletes)
256 pub fn content_type(&self) -> DataContentType {
257 self.content
258 }
259 /// Get the file path as full URI with FS scheme
260 pub fn file_path(&self) -> &str {
261 &self.file_path
262 }
263 /// Get the file format of the file (avro, orc or parquet).
264 pub fn file_format(&self) -> DataFileFormat {
265 self.file_format
266 }
267 /// Get the partition values of the file.
268 pub fn partition(&self) -> &Struct {
269 &self.partition
270 }
271 /// Get the record count in the data file.
272 pub fn record_count(&self) -> u64 {
273 self.record_count
274 }
275 /// Get the file size in bytes.
276 pub fn file_size_in_bytes(&self) -> u64 {
277 self.file_size_in_bytes
278 }
279 /// Get the column sizes.
280 /// Map from column id to the total size on disk of all regions that
281 /// store the column. Does not include bytes necessary to read other
282 /// columns, like footers. Null for row-oriented formats (Avro)
283 pub fn column_sizes(&self) -> &HashMap<i32, u64> {
284 &self.column_sizes
285 }
286 /// Get the columns value counts for the data file.
287 /// Map from column id to number of values in the column (including null
288 /// and NaN values)
289 pub fn value_counts(&self) -> &HashMap<i32, u64> {
290 &self.value_counts
291 }
292 /// Get the null value counts of the data file.
293 /// Map from column id to number of null values in the column
294 pub fn null_value_counts(&self) -> &HashMap<i32, u64> {
295 &self.null_value_counts
296 }
297 /// Get the nan value counts of the data file.
298 /// Map from column id to number of NaN values in the column
299 pub fn nan_value_counts(&self) -> &HashMap<i32, u64> {
300 &self.nan_value_counts
301 }
302 /// Get the lower bounds of the data file values per column.
303 /// Map from column id to lower bound in the column serialized as binary.
304 pub fn lower_bounds(&self) -> &HashMap<i32, Datum> {
305 &self.lower_bounds
306 }
307 /// Get the upper bounds of the data file values per column.
308 /// Map from column id to upper bound in the column serialized as binary.
309 pub fn upper_bounds(&self) -> &HashMap<i32, Datum> {
310 &self.upper_bounds
311 }
312 /// Get the Implementation-specific key metadata for the data file.
313 pub fn key_metadata(&self) -> Option<&[u8]> {
314 self.key_metadata.as_deref()
315 }
316 /// Get the split offsets of the data file.
317 /// For example, all row group offsets in a Parquet file.
318 /// Returns `None` if no split offsets are present.
319 pub fn split_offsets(&self) -> Option<&[i64]> {
320 self.split_offsets.as_deref()
321 }
322 /// Get the equality ids of the data file.
323 /// Field ids used to determine row equality in equality delete files.
324 /// null when content is not EqualityDeletes.
325 pub fn equality_ids(&self) -> Option<Vec<i32>> {
326 self.equality_ids.clone()
327 }
328 /// Get the first row id in the data file.
329 pub fn first_row_id(&self) -> Option<i64> {
330 self.first_row_id
331 }
332 /// Get the sort order id of the data file.
333 /// Only data files and equality delete files should be
334 /// written with a non-null order id. Position deletes are required to be
335 /// sorted by file and position, not a table order, and should set sort
336 /// order id to null. Readers must ignore sort order id for position
337 /// delete files.
338 pub fn sort_order_id(&self) -> Option<i32> {
339 self.sort_order_id
340 }
341 /// Get the fully qualified referenced location for the corresponding data file.
342 /// Positional delete files could have the field set, and deletion vectors must the field set.
343 pub fn referenced_data_file(&self) -> Option<String> {
344 self.referenced_data_file.clone()
345 }
346 /// Get the offset in the file where the blob content starts.
347 /// Only meaningful for puffin blobs, and required for deletion vectors.
348 pub fn content_offset(&self) -> Option<i64> {
349 self.content_offset
350 }
351 /// Get the length of a puffin blob.
352 /// Only meaningful for puffin blobs, and required for deletion vectors.
353 pub fn content_size_in_bytes(&self) -> Option<i64> {
354 self.content_size_in_bytes
355 }
356}
357
358/// Convert data files to avro bytes and write to writer.
359/// Return the bytes written.
360pub fn write_data_files_to_avro<W: Write>(
361 writer: &mut W,
362 data_files: impl IntoIterator<Item = DataFile>,
363 partition_type: &StructType,
364 version: FormatVersion,
365) -> Result<usize> {
366 let avro_schema = match version {
367 FormatVersion::V1 => data_file_schema_v1(partition_type).unwrap(),
368 FormatVersion::V2 => data_file_schema_v2(partition_type).unwrap(),
369 FormatVersion::V3 => data_file_schema_v3(partition_type).unwrap(),
370 };
371 let mut writer = new_writer(&avro_schema, writer)?;
372
373 // Wrap once and reuse across files so the field-name lookup is not rebuilt
374 // for each data file.
375 let partition_struct_type = Type::Struct(partition_type.clone());
376 for data_file in data_files {
377 let value = to_value(DataFileSerde::try_from(
378 data_file,
379 &partition_struct_type,
380 version,
381 )?)?
382 .resolve(&avro_schema)?;
383 writer.append_value(value)?;
384 }
385
386 Ok(writer.flush()?)
387}
388
389/// Parse data files from avro bytes.
390pub fn read_data_files_from_avro<R: Read>(
391 reader: &mut R,
392 schema: &Schema,
393 partition_spec_id: i32,
394 partition_type: &StructType,
395 version: FormatVersion,
396) -> Result<Vec<DataFile>> {
397 let avro_schema = match version {
398 FormatVersion::V1 => data_file_schema_v1(partition_type).unwrap(),
399 FormatVersion::V2 => data_file_schema_v2(partition_type).unwrap(),
400 FormatVersion::V3 => data_file_schema_v3(partition_type).unwrap(),
401 };
402
403 let reader = AvroReader::builder(reader)
404 .reader_schema(&avro_schema)
405 .build()?;
406 // Wrap once and reuse across files so the field-name lookup is not rebuilt
407 // for each data file.
408 let partition_struct_type = Type::Struct(partition_type.clone());
409 reader
410 .into_iter()
411 .map(|value| {
412 from_value::<DataFileSerde>(&value?)?.try_into(
413 partition_spec_id,
414 &partition_struct_type,
415 schema,
416 )
417 })
418 .collect::<Result<Vec<_>>>()
419}
420
421/// Type of content stored by the data file: data, equality deletes, or
422/// position deletes (all v1 files are data files)
423#[derive(Debug, PartialEq, Eq, Clone, Copy, Serialize, Deserialize, Default)]
424pub enum DataContentType {
425 /// value: 0
426 #[default]
427 Data = 0,
428 /// value: 1
429 PositionDeletes = 1,
430 /// value: 2
431 EqualityDeletes = 2,
432}
433
434impl TryFrom<i32> for DataContentType {
435 type Error = Error;
436
437 fn try_from(v: i32) -> Result<DataContentType> {
438 match v {
439 0 => Ok(DataContentType::Data),
440 1 => Ok(DataContentType::PositionDeletes),
441 2 => Ok(DataContentType::EqualityDeletes),
442 _ => Err(invalid_data!("data content type {v} is invalid")),
443 }
444 }
445}
446
447/// Format of this data.
448#[derive(Debug, PartialEq, Eq, Clone, Copy, SerializeDisplay, DeserializeFromStr)]
449pub enum DataFileFormat {
450 /// Avro file format: <https://avro.apache.org/>
451 Avro,
452 /// Orc file format: <https://orc.apache.org/>
453 Orc,
454 /// Parquet file format: <https://parquet.apache.org/>
455 Parquet,
456 /// Puffin file format: <https://iceberg.apache.org/puffin-spec/>
457 Puffin,
458}
459
460impl FromStr for DataFileFormat {
461 type Err = Error;
462
463 fn from_str(s: &str) -> Result<Self> {
464 match s.to_lowercase().as_str() {
465 "avro" => Ok(Self::Avro),
466 "orc" => Ok(Self::Orc),
467 "parquet" => Ok(Self::Parquet),
468 "puffin" => Ok(Self::Puffin),
469 _ => Err(invalid_data!("Unsupported data file format: {s}")),
470 }
471 }
472}
473
474impl std::fmt::Display for DataFileFormat {
475 fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
476 match self {
477 DataFileFormat::Avro => write!(f, "avro"),
478 DataFileFormat::Orc => write!(f, "orc"),
479 DataFileFormat::Parquet => write!(f, "parquet"),
480 DataFileFormat::Puffin => write!(f, "puffin"),
481 }
482 }
483}
484
485#[cfg(test)]
486mod test {
487 use crate::ErrorKind;
488 use crate::spec::{
489 DataContentType, DataFileBuilder, DataFileFormat, Literal, NestedField, PrimitiveType,
490 Struct, StructType, Type,
491 };
492
493 #[test]
494 fn test_data_content_type_default() {
495 assert_eq!(DataContentType::default(), DataContentType::Data);
496 }
497
498 #[test]
499 fn test_data_content_type_default_value() {
500 assert_eq!(DataContentType::default() as i32, 0);
501 }
502
503 #[test]
504 fn promoted_partition_widens_int_and_float() {
505 // Partition columns promoted int -> long and float -> double; a file
506 // written before the promotion still carries the narrower types.
507 let current = StructType::new(vec![
508 NestedField::required(1000, "b", Type::Primitive(PrimitiveType::Long)).into(),
509 NestedField::required(1001, "c", Type::Primitive(PrimitiveType::Double)).into(),
510 ]);
511 let written = StructType::new(vec![
512 NestedField::required(1000, "b", Type::Primitive(PrimitiveType::Int)).into(),
513 NestedField::required(1001, "c", Type::Primitive(PrimitiveType::Float)).into(),
514 ]);
515
516 let data_file = DataFileBuilder::default()
517 .content(DataContentType::Data)
518 .file_path("x.parquet".to_string())
519 .file_format(DataFileFormat::Parquet)
520 .partition(Struct::from_iter([
521 Some(Literal::int(7)),
522 Some(Literal::float(1.5_f32)),
523 ]))
524 .record_count(1)
525 .file_size_in_bytes(1)
526 .partition_spec_id(0)
527 .build()
528 .unwrap();
529
530 assert_eq!(
531 data_file.promoted_partition(¤t, &written).unwrap(),
532 Struct::from_iter([Some(Literal::long(7)), Some(Literal::double(1.5))])
533 );
534 }
535
536 #[test]
537 fn promoted_partition_widens_dropped_numeric_source() {
538 // A dropped partition source resolves to `Unknown`, so there is no current
539 // type; a numeric still widens to its ceiling by its written type, keeping it
540 // aligned with deletes written while the source still existed.
541 let current = StructType::new(vec![
542 NestedField::required(1000, "b", Type::Primitive(PrimitiveType::Unknown)).into(),
543 ]);
544 let written = StructType::new(vec![
545 NestedField::required(1000, "b", Type::Primitive(PrimitiveType::Int)).into(),
546 ]);
547
548 let data_file = DataFileBuilder::default()
549 .content(DataContentType::Data)
550 .file_path("x.parquet".to_string())
551 .file_format(DataFileFormat::Parquet)
552 .partition(Struct::from_iter([Some(Literal::int(7))]))
553 .record_count(1)
554 .file_size_in_bytes(1)
555 .partition_spec_id(0)
556 .build()
557 .unwrap();
558
559 assert_eq!(
560 data_file.promoted_partition(¤t, &written).unwrap(),
561 Struct::from_iter([Some(Literal::long(7))])
562 );
563 }
564
565 #[test]
566 fn promoted_partition_preserves_dropped_date_source() {
567 // A `date` is stored as an `int`, so widening a dropped date the way a dropped
568 // int widens would turn it into a `long` and diverge from delete manifests that
569 // skip normalization. A dropped non-numeric source is copied through unchanged.
570 let current = StructType::new(vec![
571 NestedField::required(1000, "d", Type::Primitive(PrimitiveType::Unknown)).into(),
572 ]);
573 let written = StructType::new(vec![
574 NestedField::required(1000, "d", Type::Primitive(PrimitiveType::Date)).into(),
575 ]);
576
577 let data_file = DataFileBuilder::default()
578 .content(DataContentType::Data)
579 .file_path("x.parquet".to_string())
580 .file_format(DataFileFormat::Parquet)
581 .partition(Struct::from_iter([Some(Literal::date(20_000))]))
582 .record_count(1)
583 .file_size_in_bytes(1)
584 .partition_spec_id(0)
585 .build()
586 .unwrap();
587
588 assert_eq!(
589 data_file.promoted_partition(¤t, &written).unwrap(),
590 data_file.partition
591 );
592 }
593
594 #[test]
595 fn promoted_partition_leaves_matching_types_unchanged() {
596 let partition_type = StructType::new(vec![
597 NestedField::required(1000, "b", Type::Primitive(PrimitiveType::Long)).into(),
598 ]);
599
600 let data_file = DataFileBuilder::default()
601 .content(DataContentType::Data)
602 .file_path("x.parquet".to_string())
603 .file_format(DataFileFormat::Parquet)
604 .partition(Struct::from_iter([Some(Literal::long(7))]))
605 .record_count(1)
606 .file_size_in_bytes(1)
607 .partition_spec_id(0)
608 .build()
609 .unwrap();
610
611 assert_eq!(
612 data_file
613 .promoted_partition(&partition_type, &partition_type)
614 .unwrap(),
615 data_file.partition
616 );
617 }
618
619 #[test]
620 fn promoted_partition_rejects_tuple_length_mismatch() {
621 let data_file = |values: Vec<Option<Literal>>| {
622 DataFileBuilder::default()
623 .content(DataContentType::Data)
624 .file_path("x.parquet".to_string())
625 .file_format(DataFileFormat::Parquet)
626 .partition(Struct::from_iter(values))
627 .record_count(1)
628 .file_size_in_bytes(1)
629 .partition_spec_id(0)
630 .build()
631 .unwrap()
632 };
633 let one_field = StructType::new(vec![
634 NestedField::required(1000, "b", Type::Primitive(PrimitiveType::Long)).into(),
635 ]);
636 let two_fields = StructType::new(vec![
637 NestedField::required(1000, "b", Type::Primitive(PrimitiveType::Long)).into(),
638 NestedField::required(1001, "c", Type::Primitive(PrimitiveType::Long)).into(),
639 ]);
640
641 let longer = data_file(vec![Some(Literal::int(7)), Some(Literal::int(8))]);
642 assert_eq!(
643 longer
644 .promoted_partition(&one_field, &one_field)
645 .unwrap_err()
646 .kind(),
647 ErrorKind::DataInvalid
648 );
649
650 let shorter = data_file(vec![Some(Literal::int(7))]);
651 assert_eq!(
652 shorter
653 .promoted_partition(&two_fields, &two_fields)
654 .unwrap_err()
655 .kind(),
656 ErrorKind::DataInvalid
657 );
658 }
659}