Push `ChunkReader` into `SerializedPageReader` (#2463) #2464

tustvold · 2022-08-16T15:52:06Z

Which issue does this PR close?

Rationale for this change

This pushes ChunkReader into SerializedPageReader. This not only simplifies the code, but will allow removing the InMemoryColumnChunkReader in async_reader.rs and implementing page skipping the absence of a PageIndex #2460

What changes are included in this PR?

Are there any user-facing changes?

This changes the interface of SerializedPageReader, which is public, albeit a very low-level API

tustvold · 2022-08-16T15:53:24Z

parquet/src/file/serialized_reader.rs

-                    let row_count = calculate_row_count(
-                        offset_index,
-                        *seen_num_data_pages,
-                        self.total_num_values,


This was actually previously incorrect, it was using total_num_values, instead of total_row_count. We don't currently support skipping with lists, #2122 but this would cause issues if we did

tustvold · 2022-08-16T16:08:42Z

parquet/src/file/serialized_reader.rs

+/// Reads a [`PageHeader`] from the provided [`Read`] returning the number of bytes read
+fn read_page_header_len<T: Read>(input: &mut T) -> Result<(usize, PageHeader)> {
+    /// A wrapper around a [`std::io::Read`] that keeps track of the bytes read
+    struct TrackedRead<R> {


This is kind of ugly, but the length of the header isn't stored anywhere

tustvold · 2022-08-16T16:09:31Z

parquet/src/file/serialized_reader.rs

-                    Ok(())
-                }
+        match &mut self.state {
+            SerializedPageReaderState::Values {..} =>{ Err(general_err!("Must set page_offset_index when using skip_next_page in SerializedPageReader.")) },


It should now be possible to read the header, and then skip by just incrementing the offset @Ted-Jiang

Cool ! 👍

tustvold · 2022-08-16T16:10:37Z

parquet/src/file/writer.rs

+                .build()
+                .unwrap();
+
+            let desc = ColumnDescriptor::new(Arc::new(t), 0, 0, ColumnPath::new(vec![]));


This is somewhat cumbersome, but I opted to take ColumnChunkMetadata in the constructor as it is more future-proof, not to mention slightly more obvious what the values are. In particular it avoids ambiguity over whether values includes nulls (it does).

tustvold · 2022-08-16T16:11:39Z

parquet/src/file/serialized_reader.rs

+
+                    let page_len = front.compressed_page_size as usize;
+
+                    // TODO: Add ChunkReader get_bytes to potentially avoid copy


This is what will allow removing InMemoryRowGroupReader and just using SerializedPageReader. I intend to do this as a follow up

You mean InMemoryColumnChunkReader?

If the idea that we make InMemoryColumnChunk impl ChunkReader?

I think we could just replace InMemoryColumnChunk with SerializedPageReader<Bytes>

Cool, I was working on #2426. If you want I can rebase onto this branch and do that as part of that ticket

That would be fantastic if you could 👍

alamb

Looks reasonable to me but I have not been down in this code for a while. I would defer to @thinkharderdev @Ted-Jiang @nevi-me or @sunchao

Thanks for pushing this along @tustvold 👍

alamb · 2022-08-16T21:10:57Z

parquet/src/column/writer/mod.rs

-            7,
-            Compression::UNCOMPRESSED,
-            Type::INT32,
+            Arc::new(Bytes::from(buf)),


I thought Bytes were already ref counted -- is there any need to wrap this in an additional Arc?

Because the signature needs Arc<T: ChunkReader> because ChunkReader doesn't impl Clone

alamb · 2022-08-16T21:15:51Z

parquet/src/file/serialized_reader.rs

-    // The file source buffer which references exactly the bytes for the column trunk
-    // to be read by this page reader.
-    buf: SerializedPages<T>,
+pub struct SerializedPageReader<R: ChunkReader> {


This is a non trivial change, right? To make this different than std::io::Read? But on the other hand the SerializedFileReader isn't changed -- https://docs.rs/parquet/20.0.0/parquet/file/index.html#example-of-reading-an-existing-file

Maybe to ease the transition we can add some docstring example showing how to create a ChunkReader from a std::io::Read? Or maybe it doesn't matter.

I think we should be sensitive and over communicate a change like this in the SerializedPageReader.

@zeevm do you have any comments on this change (especially in regards to your comments on #2394 (comment))?

This is a non trivial change, right?

It brings SerializedPageReader into line with all the other readers, so I'm not really sure I agree that this is a major change. I suspect almost all users are using RowGroupReader::get_column_page_reader and not calling this constructor. Tbh I'm not entirely sure why this method is even public... Perhaps I should take the opportunity to make it crate private as part of slowly reducing the amount of implementation detail that leaks out of the crate?

Perhaps I should take the opportunity to make it crate private as part of slowly reducing the amount of implementation detail that leaks out of the crate?

github codesearch https://cs.github.com/?q=SerializedPageReader%20language%3ARust&scopeName=All%20repos&scope= seems to suggest that most of the uses of this structure are forks of the arrow-rs codebase in various states of divergence.

Not that github codesearch would find all the possible issues, but it is a good sanity check that this isn't widel used

thinkharderdev

LGTM!

Ted-Jiang · 2022-08-17T03:17:36Z

parquet/src/file/serialized_reader.rs

-                    Ok(())
-                }
+        match &mut self.state {
+            SerializedPageReaderState::Values {..} =>{ Err(general_err!("Must set page_offset_index when using skip_next_page in SerializedPageReader.")) },


Cool ! 👍

Ted-Jiang · 2022-08-17T03:20:22Z

parquet/src/file/serialized_reader.rs

-        Ok(result)
-    }
+        let decompressor = create_codec(meta.compression())?;
+        let (start, len) = meta.byte_range();


i find

/// Returns the offset and length in bytes of the column chunk within the file pub fn byte_range(&self) -> (u64, u64) { let col_start = match self.dictionary_page_offset() { Some(dictionary_page_offset) => dictionary_page_offset, None => self.data_page_offset(), }; let col_len = self.compressed_size(); assert!( col_start >= 0 && col_len >= 0, "column start and length should not be negative" ); (col_start as u64, col_len as u64) }

It may return the dictionary_page_start, which is not right in checking wether is dictPage

That is intentional, we want it to return the dictionary_page_start if there is one as we use this to construct the dictionary page location.

Oh! forgot there is no dict_offset in locations 😂

Ted-Jiang

LGTM! @tustvold Thanks a lot 😊 !
This change simplifies the code and help me implementing the skip row without pageIndex in SerializedPageReader.
I think the situation in skip row without pageIndex, is not coverd in fuzz test. I will working one this file one issue #2475

tustvold · 2022-08-17T09:01:12Z

parquet/src/file/serialized_reader.rs

-        Ok(result)
-    }
+        let decompressor = create_codec(meta.compression())?;
+        let (start, len) = meta.byte_range();


This is not correct, as it will potentially miss the dictionary page. I originally tried to use the presence of the dictionary_page_offset to infer the existence of a dictionary page, but this isn't consistently set. In particular "alltypes_tiny_pages.parquet" does not set the dictionary page offset, and yet the first page is a dictionary page

ursabot · 2022-08-17T09:11:40Z

Benchmark runs are scheduled for baseline = ecc6210 and contender = 63ab69e. 63ab69e is a master commit associated with this PR. Results will be available as each benchmark for each run completes.
Conbench compare runs links:
[Skipped ⚠️ Benchmarking of arrow-rs-commits is not supported on ec2-t3-xlarge-us-east-2] ec2-t3-xlarge-us-east-2
[Skipped ⚠️ Benchmarking of arrow-rs-commits is not supported on test-mac-arm] test-mac-arm
[Skipped ⚠️ Benchmarking of arrow-rs-commits is not supported on ursa-i9-9960x] ursa-i9-9960x
[Skipped ⚠️ Benchmarking of arrow-rs-commits is not supported on ursa-thinkcentre-m75q] ursa-thinkcentre-m75q
Buildkite builds:
Supported benchmarks:
ec2-t3-xlarge-us-east-2: Supported benchmark langs: Python, R. Runs only benchmarks with cloud = True
test-mac-arm: Supported benchmark langs: C++, Python, R
ursa-i9-9960x: Supported benchmark langs: Python, R, JavaScript
ursa-thinkcentre-m75q: Supported benchmark langs: C++, Java

Push ChunkReader into SerializedPageReader (apache#2463)

857f801

tustvold added the api-change Changes to the arrow API label Aug 16, 2022

github-actions bot added the parquet Changes to the parquet crate label Aug 16, 2022

tustvold commented Aug 16, 2022

View reviewed changes

alamb reviewed Aug 16, 2022

View reviewed changes

thinkharderdev mentioned this pull request Aug 16, 2022

Use Parquet OffsetIndex to prune IO with RowSelection #2473

Merged

thinkharderdev approved these changes Aug 16, 2022

View reviewed changes

Ted-Jiang reviewed Aug 17, 2022

View reviewed changes

Ted-Jiang approved these changes Aug 17, 2022

View reviewed changes

tustvold commented Aug 17, 2022

View reviewed changes

tustvold merged commit 63ab69e into apache:master Aug 17, 2022

alamb changed the title ~~Push ChunkReader into SerializedPageReader (#2463)~~ Push ChunkReader into SerializedPageReader (#2463) Aug 17, 2022

tustvold mentioned this pull request Apr 24, 2023

Improve Documentation of Parquet ChunkReader #4118

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Push `ChunkReader` into `SerializedPageReader` (#2463) #2464

Push `ChunkReader` into `SerializedPageReader` (#2463) #2464

tustvold commented Aug 16, 2022

tustvold Aug 16, 2022

tustvold Aug 16, 2022

tustvold Aug 16, 2022

Ted-Jiang Aug 17, 2022

tustvold Aug 16, 2022

tustvold Aug 16, 2022 •

edited

Loading

thinkharderdev Aug 16, 2022

tustvold Aug 16, 2022 •

edited

Loading

thinkharderdev Aug 16, 2022

tustvold Aug 16, 2022

alamb left a comment

alamb Aug 16, 2022

tustvold Aug 16, 2022

alamb Aug 16, 2022

tustvold Aug 16, 2022 •

edited

Loading

alamb Aug 16, 2022

thinkharderdev left a comment

Ted-Jiang Aug 17, 2022

Ted-Jiang Aug 17, 2022

tustvold Aug 17, 2022

Ted-Jiang Aug 17, 2022

Ted-Jiang left a comment

tustvold Aug 17, 2022

ursabot commented Aug 17, 2022


		let page_len = front.compressed_page_size as usize;

		// TODO: Add ChunkReader get_bytes to potentially avoid copy

Push ChunkReader into SerializedPageReader (#2463) #2464

Push ChunkReader into SerializedPageReader (#2463) #2464

Conversation

tustvold commented Aug 16, 2022

Which issue does this PR close?

Rationale for this change

What changes are included in this PR?

Are there any user-facing changes?

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tustvold Aug 16, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tustvold Aug 16, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

alamb left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tustvold Aug 16, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

thinkharderdev left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Ted-Jiang left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ursabot commented Aug 17, 2022

Push `ChunkReader` into `SerializedPageReader` (#2463) #2464

Push `ChunkReader` into `SerializedPageReader` (#2463) #2464

tustvold Aug 16, 2022 •

edited

Loading

tustvold Aug 16, 2022 •

edited

Loading

tustvold Aug 16, 2022 •

edited

Loading