apache / apache/arrow-java

[Java] VectorUnloader has multiple memory leak issues

オープン
#1,234 コメント 3 件 リアクション 0 件 担当者 0 名 GitHub で見る
Type: bug
主要言語
Java
スター
94
フォーク
152
平均マージ
3日 16時間
マージ済み PR(30日)
11

説明

### Describe the bug, including details regarding any error messages, version, and platform.

### Problem Description

The VectorUnloader class has serious memory leak issues when serializing Vector data. Any exception thrown during the process will cause previously retained ArrowBuf instances to never be released.
Root Cause
ArrowBuf uses reference counting for memory management. After calling retain(), a corresponding release() must be called to free the buffer. The code has exception-safety defects in multiple locations:

---
#### Bug 1: OOM during compression causes leak
```
Location: appendNodes() method, around line 75
buf.getReferenceManager().retain();
buffers.add(codec.compress(vector.getAllocator(), buf)); // OOM here causes leak
Problem: If codec.compress() throws OutOfMemoryError, retain() has already been called but buffers.add() never executes, making the ArrowBuf unreleasable.
```
---
#### Bug 2: Compression failure in loop causes partial leak
```
Location: appendNodes() method, around lines 78-85
for (ArrowBuf buf : fieldBuffers) {
buf.getReferenceManager().retain();
buffers.add(codec.compress(vector.getAllocator(), buf)); // Any failure leaks all previously retained buffers
}
Problem: Any compression failure in the loop causes all previously retained buffers to leak.
```
---
#### Bug 3: Exception during recursive child vector processing causes leak
```
Location: appendNodes() method, around lines 87-89
for (FieldVector child : vector.getChildrenFromFields()) {
appendNodes(child, nodes, buffers, variadicBufferCounts); // Exception here leaks buffers from current level
}
Problem: An exception during recursive child processing prevents release of buffers already retained at the current level.
```
---
#### Bug 4: Iteration failure in top-level loop causes cumulative leaks
```
Location: getRecordBatch() method, around lines 53-58
for (FieldVector vector : root.getFieldVectors()) {
appendNodes(vector, nodes, buffers, variadicBufferCounts); // Any failure leaks all previously retained buffers
}
Problem: Each iteration retains buffers, and any failure causes cumulative leaks of all previously retained buffers. This is an accumulated leak.
```
---
### Impact
- All code using VectorUnloader.getRecordBatch()
- Leak probability increases significantly when processing large datasets or when compressor runs low on memory
- Leaked ArrowBuf instances persist until process termination
Suggested Fix
1. Use try-finally to ensure release() is called even when exceptions occur
2. Or add buffers to the list before calling retain() and compress()
3. Use transactional error handling for the entire getRecordBatch() operation, rolling back all buffers on failure

コントリビューションガイド

コントリビューションガイドを開く

調査の方向性

VectorUnloader クラスから始め、getRecordBatch() と appendNodes() に焦点を当てます。特に、issue で説明されている retain()、compress()、および再帰的な子処理のパスを確認してください。各箇所での失敗によって、以前の反復処理のバッファを含む、保持されているすべての ArrowBuf が解放されることを検証し、通常のレコードバッチのシリアライズが引き続き機能することを確認してください。

索引モデルが issue の本文から書いたものです。

評価

技術スタック
java
領域
data
issue の種類
バグ
難易度
4/5
見積もり時間
3〜5日
活発さ
静か
明瞭さ
明確に書かれている
初心者へのやさしさ
62/100

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。