Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
48 changes: 31 additions & 17 deletions parquet/src/arrow/in_memory_row_group.rs
Original file line number Diff line number Diff line change
Expand Up @@ -43,8 +43,9 @@ pub(crate) struct InMemoryRowGroup<'a> {
pub(crate) struct FetchRanges {
/// The byte ranges to fetch
pub(crate) ranges: Vec<Range<u64>>,
/// If `Some`, the start offsets of each page for each column chunk
pub(crate) page_start_offsets: Option<Vec<Vec<u64>>>,
/// If `Some`, the start offsets of each page for each column chunk, or
/// `None` for a column chunk without an offset index (fetched in full)
pub(crate) page_start_offsets: Option<Vec<Option<Vec<u64>>>>,
}

impl InMemoryRowGroup<'_> {
Expand All @@ -70,7 +71,7 @@ impl InMemoryRowGroup<'_> {
// If we have a `RowSelection` and an `OffsetIndex` then only fetch
// pages required for the `RowSelection`
// Consider preallocating outer vec: https://github.com/apache/arrow-rs/issues/8667
let mut page_start_offsets: Vec<Vec<u64>> = vec![];
let mut page_start_offsets: Vec<Option<Vec<u64>>> = vec![];

let ranges = self
.column_chunks
Expand All @@ -88,6 +89,7 @@ impl InMemoryRowGroup<'_> {
let Some(offset_idx) = page_index.offset_index(idx) else {
// No offset index for this column, fetch the entire column
ranges.push(start..start + len);
page_start_offsets.push(None);
return ranges;
};

Expand All @@ -106,7 +108,7 @@ impl InMemoryRowGroup<'_> {
} else {
ranges.extend(selection.scan_ranges(offset_idx.page_locations()));
}
page_start_offsets.push(ranges.iter().map(|range| range.start).collect());
page_start_offsets.push(Some(ranges.iter().map(|range| range.start).collect()));

ranges
})
Expand Down Expand Up @@ -141,7 +143,7 @@ impl InMemoryRowGroup<'_> {
pub(crate) fn fill_column_chunks<I>(
&mut self,
projection: &ProjectionMask,
page_start_offsets: Option<Vec<Vec<u64>>>,
page_start_offsets: Option<Vec<Option<Vec<u64>>>>,
chunk_data: I,
) where
I: IntoIterator<Item = Bytes>,
Expand All @@ -158,20 +160,32 @@ impl InMemoryRowGroup<'_> {
continue;
}

if let Some(offsets) = page_start_offsets.next() {
let mut chunks = Vec::with_capacity(offsets.len());
for _ in 0..offsets.len() {
chunks.push(chunk_data.next().unwrap());
match page_start_offsets.next() {
// No offset index: `fetch_ranges` requested the whole chunk
Some(None) => {
if let Some(data) = chunk_data.next() {
*chunk = Some(Arc::new(ColumnChunkData::Dense {
offset: metadata.column(idx).byte_range().0 as usize,
data,
}));
}
}
Some(Some(offsets)) => {
let mut chunks = Vec::with_capacity(offsets.len());
for _ in 0..offsets.len() {
chunks.push(chunk_data.next().unwrap());
}

*chunk = Some(Arc::new(ColumnChunkData::Sparse {
length: metadata.column(idx).byte_range().1 as usize,
data: offsets
.into_iter()
.map(|x| x as usize)
.zip(chunks)
.collect(),
}))
*chunk = Some(Arc::new(ColumnChunkData::Sparse {
length: metadata.column(idx).byte_range().1 as usize,
data: offsets
.into_iter()
.map(|x| x as usize)
.zip(chunks)
.collect(),
}))
}
None => {}
}
}
} else {
Expand Down
2 changes: 1 addition & 1 deletion parquet/src/arrow/push_decoder/reader_builder/data.rs
Original file line number Diff line number Diff line change
Expand Up @@ -39,7 +39,7 @@ pub(super) struct DataRequest {
ranges: Vec<Range<u64>>,
/// Optional page start offsets for each requested range. This is used
/// to create the relevant InMemoryRowGroup
page_start_offsets: Option<Vec<Vec<u64>>>,
page_start_offsets: Option<Vec<Option<Vec<u64>>>>,
}

impl DataRequest {
Expand Down
1 change: 1 addition & 0 deletions parquet/tests/arrow_reader/mod.rs
Original file line number Diff line number Diff line change
Expand Up @@ -47,6 +47,7 @@ mod invalid_utf8;
mod io;
mod large_string_overflow;
mod parquet_testing;
mod partial_offset_index;
#[cfg(feature = "async")]
mod predicate_cache;
mod row_filter;
Expand Down
216 changes: 216 additions & 0 deletions parquet/tests/arrow_reader/partial_offset_index.rs
Original file line number Diff line number Diff line change
@@ -0,0 +1,216 @@
// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements. See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership. The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License. You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied. See the License for the
// specific language governing permissions and limitations
// under the License.

//! Tests for reading files in which only some column chunks have an offset index

use arrow::compute::kernels::cmp::{gt_eq, lt};
use arrow::compute::{and, concat_batches};
use arrow_array::{ArrayRef, Int32Array, RecordBatch};
use bytes::Bytes;
use parquet::DecodeResult;
use parquet::arrow::arrow_reader::{
ArrowPredicateFn, ArrowReaderBuilder, ArrowReaderMetadata, ArrowReaderOptions,
ParquetRecordBatchReaderBuilder, RowFilter, RowSelection, RowSelector,
};
use parquet::arrow::push_decoder::ParquetPushDecoderBuilder;
use parquet::arrow::{ArrowWriter, ProjectionMask};
use parquet::column::writer::ColumnCloseResult;
use parquet::errors::Result;
use parquet::file::metadata::{PageIndexPolicy, ParquetMetaDataReader};
use parquet::file::properties::WriterProperties;
use parquet::file::writer::SerializedFileWriter;
use std::sync::Arc;

#[test]
fn test_no_offset_index_first_column() {
check(|_row_group, column| column != 0);
}

#[test]
fn test_no_offset_index_middle_column() {
check(|_row_group, column| column != 1);
}

#[test]
fn test_no_offset_index_last_column() {
check(|_row_group, column| column != 2);
}

#[test]
fn test_no_offset_index_row_group() {
check(|row_group, _column| row_group != 1);
}

/// 400 rows in columns "a" (0..400), "b" (400..800) and "c" (800..1200)
fn test_batch() -> RecordBatch {
let column = |start| Arc::new(Int32Array::from_iter_values(start..start + 400)) as ArrayRef;
RecordBatch::try_from_iter([("a", column(0)), ("b", column(400)), ("c", column(800))]).unwrap()
}

/// Returns [`test_batch`] as a file with 2 row groups of 200 rows and data
/// pages of 50 rows. A column chunk has a page index only if
/// `has_index(row_group, column)` is `true`.
fn test_file(has_index: impl Fn(usize, usize) -> bool) -> Bytes {
let batch = test_batch();
let props = WriterProperties::builder()
.set_max_row_group_row_count(Some(200))
.set_data_page_row_count_limit(50)
.set_write_batch_size(50)
.build();
let mut buf = Vec::new();
let mut writer = ArrowWriter::try_new(&mut buf, batch.schema(), Some(props)).unwrap();
writer.write(&batch).unwrap();
writer.close().unwrap();
let source = Bytes::from(buf);

// Copy the column chunks to a new file, without the page index of some chunks
let metadata = ParquetMetaDataReader::new()
.with_page_index_policy(PageIndexPolicy::Required)
.parse_and_finish(&source)
.unwrap();
let page_index = metadata.page_index().unwrap();
let schema = metadata.file_metadata().schema_descr().root_schema_ptr();
let mut buf = Vec::new();
let mut writer = SerializedFileWriter::new(&mut buf, schema, Default::default()).unwrap();
for (row_group, row_group_metadata) in metadata.row_groups().iter().enumerate() {
let mut row_group_writer = writer.next_row_group().unwrap();
for (column, column_metadata) in row_group_metadata.columns().iter().enumerate() {
let keep = has_index(row_group, column);
let close = ColumnCloseResult {
bytes_written: column_metadata.compressed_size() as u64,
rows_written: row_group_metadata.num_rows() as u64,
metadata: column_metadata.clone(),
bloom_filter: None,
column_index: page_index
.column_index(row_group, column)
.filter(|_| keep)
.cloned(),
offset_index: page_index
.offset_index(row_group, column)
.filter(|_| keep)
.cloned(),
};
row_group_writer.append_column(&source, close).unwrap();
}
row_group_writer.close().unwrap();
}
writer.close().unwrap();
Bytes::from(buf)
}

/// Makes the reader select rows 150..250
#[derive(Debug, Clone, Copy)]
enum Trigger {
RowSelection,
RowFilter,
}

fn options() -> ArrowReaderOptions {
ArrowReaderOptions::new().with_page_index_policy(PageIndexPolicy::Optional)
}

fn configure<T>(builder: ArrowReaderBuilder<T>, trigger: Trigger) -> ArrowReaderBuilder<T> {
match trigger {
Trigger::RowSelection => builder.with_row_selection(RowSelection::from(vec![
RowSelector::skip(150),
RowSelector::select(100),
RowSelector::skip(150),
])),
Trigger::RowFilter => {
let mask = ProjectionMask::roots(builder.parquet_schema(), [0]);
let predicate = ArrowPredicateFn::new(mask, |batch: RecordBatch| {
let a = batch.column(0);
and(
&gt_eq(a, &Int32Array::new_scalar(150))?,
&lt(a, &Int32Array::new_scalar(250))?,
)
});
builder.with_row_filter(RowFilter::new(vec![Box::new(predicate)]))
}
}
}

fn read_sync(file: &Bytes, trigger: Trigger) -> Result<Vec<RecordBatch>> {
let builder = ParquetRecordBatchReaderBuilder::try_new_with_options(file.clone(), options())?;
Ok(configure(builder, trigger)
.build()?
.collect::<Result<_, _>>()?)
}

#[cfg(feature = "async")]
fn read_async(file: &Bytes, trigger: Trigger) -> Result<Vec<RecordBatch>> {
use futures::TryStreamExt;
use parquet::arrow::ParquetRecordBatchStreamBuilder;

let runtime = tokio::runtime::Builder::new_current_thread()
.build()
.unwrap();
runtime.block_on(async {
let input = std::io::Cursor::new(file.clone());
let builder = ParquetRecordBatchStreamBuilder::new_with_options(input, options()).await?;
configure(builder, trigger).build()?.try_collect().await
})
}

fn read_push_decoder(file: &Bytes, trigger: Trigger) -> Result<Vec<RecordBatch>> {
let metadata = ArrowReaderMetadata::load(file, options())?;
let builder = ParquetPushDecoderBuilder::new_with_metadata(metadata);
let mut decoder = configure(builder, trigger).build()?;
let mut batches = vec![];
loop {
match decoder.try_decode()? {
DecodeResult::NeedsData(ranges) => {
let data = ranges
.iter()
.map(|r| file.slice(r.start as usize..r.end as usize))
.collect();
decoder.push_ranges(ranges, data)?;
}
DecodeResult::Data(batch) => batches.push(batch),
DecodeResult::Finished => return Ok(batches),
}
}
}

/// Reads rows 150..250 of a file where only the column chunks for which
/// `has_index(row_group, column)` is `true` have an offset index
fn check(has_index: impl Fn(usize, usize) -> bool) {
let file = test_file(&has_index);
let metadata = ArrowReaderMetadata::load(&file, options()).unwrap();
let page_index = metadata.metadata().page_index().unwrap();
for row_group in 0..2 {
for column in 0..3 {
let actual = page_index.offset_index(row_group, column).is_some();
assert_eq!(actual, has_index(row_group, column));
}
}

let expected = test_batch().slice(150, 100);
for trigger in [Trigger::RowSelection, Trigger::RowFilter] {
let results = [
("sync", read_sync(&file, trigger)),
("push decoder", read_push_decoder(&file, trigger)),
#[cfg(feature = "async")]
("async", read_async(&file, trigger)),
];
for (reader, result) in results {
let batches = result.unwrap_or_else(|e| panic!("{reader} reader, {trigger:?}: {e}"));
let actual = concat_batches(&expected.schema(), &batches).unwrap();
assert_eq!(actual, expected, "{reader} reader, {trigger:?}");
}
}
}
Loading