You are viewing a plain text version of this content. The canonical link for it is here.
Posted to github@arrow.apache.org by "devinjdangelo (via GitHub)" <gi...@apache.org> on 2024/03/22 13:47:25 UTC
[PR] fix: parallel parquet can underflow when max_record_batch_rows < execution.batch_size [arrow-datafusion]
devinjdangelo opened a new pull request, #9737:
URL: https://github.com/apache/arrow-datafusion/pull/9737
## Which issue does this PR close?
Closes #9736
## Rationale for this change
See issue
## What changes are included in this PR?
Parallel parquet writer can now handle the case when max_record_batch_rows < execution.batch_size by iteratively splitting the record batch rather than assuming it only needs to be split once.
## Are these changes tested?
todo...
## Are there any user-facing changes?
Just bugfix
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] fix: parallel parquet can underflow when max_record_batch_rows < execution.batch_size [arrow-datafusion]
Posted by "devinjdangelo (via GitHub)" <gi...@apache.org>.
devinjdangelo commented on code in PR #9737:
URL: https://github.com/apache/arrow-datafusion/pull/9737#discussion_r1536033942
##########
datafusion/core/src/dataframe/parquet.rs:
##########
@@ -166,4 +166,47 @@ mod tests {
Ok(())
}
+
+ #[tokio::test]
+ async fn write_parquet_with_small_rg_size() -> Result<()> {
+ let mut test_df = test_util::test_table().await?;
+ // make the test data larger so there are multiple batches
+ for _ in 0..7 {
+ test_df = test_df.clone().union(test_df)?;
Review Comment:
We should be able to trigger the issue with less data by lowering `execution.batch_size` to something small like 10 rows.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] fix: parallel parquet can underflow when max_record_batch_rows < execution.batch_size [arrow-datafusion]
Posted by "devinjdangelo (via GitHub)" <gi...@apache.org>.
devinjdangelo commented on code in PR #9737:
URL: https://github.com/apache/arrow-datafusion/pull/9737#discussion_r1536286168
##########
datafusion/core/src/dataframe/parquet.rs:
##########
@@ -166,4 +166,47 @@ mod tests {
Ok(())
}
+
+ #[tokio::test]
+ async fn write_parquet_with_small_rg_size() -> Result<()> {
+ let mut test_df = test_util::test_table().await?;
+ // make the test data larger so there are multiple batches
+ for _ in 0..7 {
+ test_df = test_df.clone().union(test_df)?;
Review Comment:
The test with a batch size of 10 still panics on main and passes in this PR, but runs in 0.41 seconds.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] fix: parallel parquet can underflow when max_record_batch_rows < execution.batch_size [arrow-datafusion]
Posted by "devinjdangelo (via GitHub)" <gi...@apache.org>.
devinjdangelo commented on code in PR #9737:
URL: https://github.com/apache/arrow-datafusion/pull/9737#discussion_r1536035101
##########
datafusion/core/src/dataframe/parquet.rs:
##########
@@ -166,4 +166,47 @@ mod tests {
Ok(())
}
+
+ #[tokio::test]
+ async fn write_parquet_with_small_rg_size() -> Result<()> {
+ let mut test_df = test_util::test_table().await?;
+ // make the test data larger so there are multiple batches
+ for _ in 0..7 {
+ test_df = test_df.clone().union(test_df)?;
+ }
+ let output_path = "file://local/test.parquet";
+
+ for rg_size in (1..7).step_by(5) {
Review Comment:
Yes, [1, 6] is all I meant and would be clearer... I originally wanted to loop over more rg sizes but the test was slow. If we streamline the test, we could actually range over more values here.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] fix: parallel parquet can underflow when max_record_batch_rows < execution.batch_size [arrow-datafusion]
Posted by "devinjdangelo (via GitHub)" <gi...@apache.org>.
devinjdangelo commented on code in PR #9737:
URL: https://github.com/apache/arrow-datafusion/pull/9737#discussion_r1536286778
##########
datafusion/core/src/dataframe/parquet.rs:
##########
@@ -166,4 +166,47 @@ mod tests {
Ok(())
}
+
+ #[tokio::test]
+ async fn write_parquet_with_small_rg_size() -> Result<()> {
+ let mut test_df = test_util::test_table().await?;
+ // make the test data larger so there are multiple batches
+ for _ in 0..7 {
+ test_df = test_df.clone().union(test_df)?;
+ }
+ let output_path = "file://local/test.parquet";
+
+ for rg_size in (1..7).step_by(5) {
+ let df = test_df.clone();
+ let tmp_dir = TempDir::new()?;
+ let local = Arc::new(LocalFileSystem::new_with_prefix(&tmp_dir)?);
+ let local_url = Url::parse("file://local").unwrap();
+ let ctx = &test_df.session_state;
+ ctx.runtime_env().register_object_store(&local_url, local);
+ let mut options = TableParquetOptions::default();
+ options.global.max_row_group_size = rg_size;
+ options.global.allow_single_file_parallelism = true;
+ df.write_parquet(
+ output_path,
+ DataFrameWriteOptions::new().with_single_file_output(true),
+ Some(options),
+ )
+ .await?;
+
+ // Check that file actually used the correct rg size
+ let file = std::fs::File::open(tmp_dir.into_path().join("test.parquet"))?;
Review Comment:
I fixed this in the new test and a preexisting one.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] fix: parallel parquet can underflow when max_record_batch_rows < execution.batch_size [arrow-datafusion]
Posted by "alamb (via GitHub)" <gi...@apache.org>.
alamb commented on code in PR #9737:
URL: https://github.com/apache/arrow-datafusion/pull/9737#discussion_r1536615561
##########
datafusion/core/src/dataframe/parquet.rs:
##########
@@ -150,7 +152,7 @@ mod tests {
.await?;
// Check that file actually used the specified compression
- let file = std::fs::File::open(tmp_dir.into_path().join("test.parquet"))?;
+ let file = std::fs::File::open(tmp_dir.path().join("test.parquet"))?;
Review Comment:
👍
##########
datafusion/core/src/dataframe/parquet.rs:
##########
@@ -150,7 +152,7 @@ mod tests {
.await?;
// Check that file actually used the specified compression
- let file = std::fs::File::open(tmp_dir.into_path().join("test.parquet"))?;
+ let file = std::fs::File::open(tmp_dir.path().join("test.parquet"))?;
Review Comment:
thank you for the driveby cleanup
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] fix: parallel parquet can underflow when max_record_batch_rows < execution.batch_size [arrow-datafusion]
Posted by "alamb (via GitHub)" <gi...@apache.org>.
alamb commented on code in PR #9737:
URL: https://github.com/apache/arrow-datafusion/pull/9737#discussion_r1536022590
##########
datafusion/core/src/dataframe/parquet.rs:
##########
@@ -166,4 +166,47 @@ mod tests {
Ok(())
}
+
+ #[tokio::test]
+ async fn write_parquet_with_small_rg_size() -> Result<()> {
+ let mut test_df = test_util::test_table().await?;
+ // make the test data larger so there are multiple batches
+ for _ in 0..7 {
+ test_df = test_df.clone().union(test_df)?;
+ }
+ let output_path = "file://local/test.parquet";
+
+ for rg_size in (1..7).step_by(5) {
+ let df = test_df.clone();
+ let tmp_dir = TempDir::new()?;
+ let local = Arc::new(LocalFileSystem::new_with_prefix(&tmp_dir)?);
+ let local_url = Url::parse("file://local").unwrap();
+ let ctx = &test_df.session_state;
+ ctx.runtime_env().register_object_store(&local_url, local);
+ let mut options = TableParquetOptions::default();
+ options.global.max_row_group_size = rg_size;
+ options.global.allow_single_file_parallelism = true;
+ df.write_parquet(
+ output_path,
+ DataFrameWriteOptions::new().with_single_file_output(true),
+ Some(options),
+ )
+ .await?;
+
+ // Check that file actually used the correct rg size
+ let file = std::fs::File::open(tmp_dir.into_path().join("test.parquet"))?;
Review Comment:
Calling `into_path` here I think means the file won't be cleaned up
I think calling `path()` would ensure the file is cleaned up
##########
datafusion/core/src/dataframe/parquet.rs:
##########
@@ -166,4 +166,47 @@ mod tests {
Ok(())
}
+
+ #[tokio::test]
+ async fn write_parquet_with_small_rg_size() -> Result<()> {
+ let mut test_df = test_util::test_table().await?;
+ // make the test data larger so there are multiple batches
+ for _ in 0..7 {
+ test_df = test_df.clone().union(test_df)?;
+ }
+ let output_path = "file://local/test.parquet";
+
+ for rg_size in (1..7).step_by(5) {
Review Comment:
My reading of the [docs](https://doc.rust-lang.org/std/iter/trait.Iterator.html#method.step_by) and my [playground experiments](https://play.rust-lang.org/?version=stable&mode=debug&edition=2021&gist=2ff81b06e8447214338607095c726d7c) suggests this is the same as `[1, 6]` -- is that the intent? Or did you mean `1, 5, 10, 15, 20, 25, 30, 35`?
##########
datafusion/core/src/dataframe/parquet.rs:
##########
@@ -166,4 +166,47 @@ mod tests {
Ok(())
}
+
+ #[tokio::test]
+ async fn write_parquet_with_small_rg_size() -> Result<()> {
+ let mut test_df = test_util::test_table().await?;
+ // make the test data larger so there are multiple batches
+ for _ in 0..7 {
+ test_df = test_df.clone().union(test_df)?;
Review Comment:
When I ran this test it takes more than 22 seconds on my laptop. I wonder if we really need to generate so much data -- maybe we can try slicing up the batch (or else maybe use larger rg_sizes)
```shell
$ cargo test --lib -p datafusion -- write_parquet_with_small_rg_size
...
Finished test [unoptimized + debuginfo] target(s) in 0.16s
Running unittests src/lib.rs (target/debug/deps/datafusion-4cbfc61ad6017be4)
running 1 test
test dataframe::parquet::tests::write_parquet_with_small_rg_size ... ok
test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 651 filtered out; finished in 22.31s
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] fix: parallel parquet can underflow when max_record_batch_rows < execution.batch_size [arrow-datafusion]
Posted by "devinjdangelo (via GitHub)" <gi...@apache.org>.
devinjdangelo commented on code in PR #9737:
URL: https://github.com/apache/arrow-datafusion/pull/9737#discussion_r1536285765
##########
datafusion/core/src/dataframe/parquet.rs:
##########
@@ -166,4 +166,47 @@ mod tests {
Ok(())
}
+
+ #[tokio::test]
+ async fn write_parquet_with_small_rg_size() -> Result<()> {
+ let mut test_df = test_util::test_table().await?;
+ // make the test data larger so there are multiple batches
+ for _ in 0..7 {
+ test_df = test_df.clone().union(test_df)?;
+ }
+ let output_path = "file://local/test.parquet";
+
+ for rg_size in (1..7).step_by(5) {
Review Comment:
This now loops over 0..10 with datafusion.execution.batch_size set to 10.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
Re: [PR] fix: parallel parquet can underflow when max_record_batch_rows < execution.batch_size [arrow-datafusion]
Posted by "alamb (via GitHub)" <gi...@apache.org>.
alamb merged PR #9737:
URL: https://github.com/apache/arrow-datafusion/pull/9737
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: github-unsubscribe@arrow.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org