[
https://issues.apache.org/jira/browse/DRILL-8139?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17493991#comment-17493991
]
ASF GitHub Bot commented on DRILL-8139:
---------------------------------------
vdiravka commented on a change in pull request #2463:
URL: https://github.com/apache/drill/pull/2463#discussion_r809090421
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/store/parquet/compression/AirliftBytesInputCompressor.java
##########
@@ -159,13 +160,13 @@ public void decompress(ByteBuffer input, int
compressedSize, ByteBuffer output,
@Override
public void release() {
- logger.debug(
- "will release {} allocated buffers.",
- this.allocatedBuffers.size()
- );
+ int bufCount = allocatedBuffers.size();
- while (!this.allocatedBuffers.isEmpty()) {
- this.allocator.release(allocatedBuffers.pop());
+ // LIFO release order to try to reduce memory fragmentation.
+ while (!allocatedBuffers.isEmpty()) {
+ allocator.release(allocatedBuffers.pop());
}
+
+ logger.debug("released {} allocated buffers", bufCount);
Review comment:
We possibly can assert or log the actual `allocatedBuffers.size()` to
make sure allocated buffers are released
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/store/parquet/compression/DrillCompressionCodecFactory.java
##########
@@ -50,17 +52,31 @@
// The set of codecs to be handled by aircompressor
private static final Set<CompressionCodecName> AIRCOMPRESSOR_CODECS = new
HashSet<>(
- Arrays.asList(CompressionCodecName.LZ4, CompressionCodecName.LZO,
- CompressionCodecName.SNAPPY, CompressionCodecName.ZSTD));
+ Arrays.asList(
+ CompressionCodecName.LZ4,
+ CompressionCodecName.LZO,
+ CompressionCodecName.SNAPPY,
+ CompressionCodecName.ZSTD
+ )
+ );
- // pool of reused aircompressor compressors (parquet-mr's factory has its
own)
+ // pool of reusable thread-safe aircompressor compressors (parquet-mr's
factory
Review comment:
```suggestion
// pool of reusable thread-safe aircompressor compressors (parquet-mr's
factory has its own)
```
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/store/parquet/compression/DrillCompressionCodecFactory.java
##########
@@ -50,17 +52,31 @@
// The set of codecs to be handled by aircompressor
private static final Set<CompressionCodecName> AIRCOMPRESSOR_CODECS = new
HashSet<>(
- Arrays.asList(CompressionCodecName.LZ4, CompressionCodecName.LZO,
- CompressionCodecName.SNAPPY, CompressionCodecName.ZSTD));
+ Arrays.asList(
+ CompressionCodecName.LZ4,
+ CompressionCodecName.LZO,
+ CompressionCodecName.SNAPPY,
+ CompressionCodecName.ZSTD
+ )
+ );
- // pool of reused aircompressor compressors (parquet-mr's factory has its
own)
+ // pool of reusable thread-safe aircompressor compressors (parquet-mr's
factory
+ // has its own)
private final Map<CompressionCodecName, AirliftBytesInputCompressor>
airCompressors = new HashMap<>();
// fallback parquet-mr compression codec factory
- private CompressionCodecFactory parqCodecFactory;
+ private final CompressionCodecFactory parqCodecFactory;
// direct memory allocator to be used during (de)compression
- private ByteBufferAllocator allocator;
+ private final ByteBufferAllocator allocator;
+
+ // Start: members for working around a gzip concurrency bug c.f. DRILL-8139
+ private final Deque<CompressionCodecFactory> singleUseFactories;
+
Review comment:
```suggestion
```
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/store/parquet/compression/DrillCompressionCodecFactory.java
##########
@@ -80,38 +99,56 @@ public DrillCompressionCodecFactory(Configuration config,
ByteBufferAllocator al
}
@Override
- public BytesInputCompressor getCompressor(CompressionCodecName codecName) {
+ public synchronized BytesInputCompressor getCompressor(CompressionCodecName
codecName) {
if (AIRCOMPRESSOR_CODECS.contains(codecName)) {
return airCompressors.computeIfAbsent(
codecName,
c -> new AirliftBytesInputCompressor(codecName, allocator)
);
+ } else if (codecName != CompressionCodecName.SNAPPY) {
+ // Work around PARQUET-2126: construct a new codec factory every time to
+ // avoid a concurrrency bug c.f. DRILL-8139. Remove once PARQUET-2126 is
+ // fixed. Snappy is immune because of the thread safety in the Xerial
lib.
+ CompressionCodecFactory ccf =
CodecFactory.createDirectCodecFactory(config, allocator, pageSize);
+ // hold onto a reference for later release()
+ singleUseFactories.add(ccf);
+ return ccf.getCompressor(codecName);
} else {
return parqCodecFactory.getCompressor(codecName);
}
}
@Override
- public BytesInputDecompressor getDecompressor(CompressionCodecName
codecName) {
+ public synchronized BytesInputDecompressor
getDecompressor(CompressionCodecName codecName) {
if (AIRCOMPRESSOR_CODECS.contains(codecName)) {
return airCompressors.computeIfAbsent(
codecName,
c -> new AirliftBytesInputCompressor(codecName, allocator)
);
+ } else if (codecName != CompressionCodecName.SNAPPY) {
Review comment:
Looks like existed `else` block starts to be never reached:
When `codecName` is SNAPPY or other from `AIRCOMPRESSOR_CODECS` we will pass
through the first condition
and for all other cases we will pass through the new `else if` block. If
that by design the logic can be simplified: `else if` -> `else`. `return
parqCodecFactory.getDecompressor(codecName);` possibly can be commented.
The same for `#getCompressor` method
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/store/parquet/compression/DrillCompressionCodecFactory.java
##########
@@ -80,38 +99,56 @@ public DrillCompressionCodecFactory(Configuration config,
ByteBufferAllocator al
}
@Override
- public BytesInputCompressor getCompressor(CompressionCodecName codecName) {
+ public synchronized BytesInputCompressor getCompressor(CompressionCodecName
codecName) {
if (AIRCOMPRESSOR_CODECS.contains(codecName)) {
return airCompressors.computeIfAbsent(
codecName,
c -> new AirliftBytesInputCompressor(codecName, allocator)
);
+ } else if (codecName != CompressionCodecName.SNAPPY) {
+ // Work around PARQUET-2126: construct a new codec factory every time to
+ // avoid a concurrrency bug c.f. DRILL-8139. Remove once PARQUET-2126 is
Review comment:
```suggestion
// avoid a concurrrency bug c.f. DRILL-8139. TODO: Remove once
PARQUET-2126 is
Write Preview
```
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/store/parquet/compression/DrillCompressionCodecFactory.java
##########
@@ -50,17 +52,31 @@
// The set of codecs to be handled by aircompressor
private static final Set<CompressionCodecName> AIRCOMPRESSOR_CODECS = new
HashSet<>(
- Arrays.asList(CompressionCodecName.LZ4, CompressionCodecName.LZO,
- CompressionCodecName.SNAPPY, CompressionCodecName.ZSTD));
+ Arrays.asList(
+ CompressionCodecName.LZ4,
+ CompressionCodecName.LZO,
+ CompressionCodecName.SNAPPY,
+ CompressionCodecName.ZSTD
+ )
+ );
- // pool of reused aircompressor compressors (parquet-mr's factory has its
own)
+ // pool of reusable thread-safe aircompressor compressors (parquet-mr's
factory
+ // has its own)
private final Map<CompressionCodecName, AirliftBytesInputCompressor>
airCompressors = new HashMap<>();
// fallback parquet-mr compression codec factory
- private CompressionCodecFactory parqCodecFactory;
+ private final CompressionCodecFactory parqCodecFactory;
// direct memory allocator to be used during (de)compression
- private ByteBufferAllocator allocator;
+ private final ByteBufferAllocator allocator;
+
+ // Start: members for working around a gzip concurrency bug c.f. DRILL-8139
+ private final Deque<CompressionCodecFactory> singleUseFactories;
+
+ private final Configuration config;
+
Review comment:
```suggestion
```
--
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: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
> Data corruption and occasional segfaults querying Parquet/gzip under the
> async column reader and sync page reader
> -----------------------------------------------------------------------------------------------------------------
>
> Key: DRILL-8139
> URL: https://issues.apache.org/jira/browse/DRILL-8139
> Project: Apache Drill
> Issue Type: Bug
> Components: Storage - Parquet
> Affects Versions: 1.19.0
> Reporter: James Turton
> Assignee: James Turton
> Priority: Blocker
> Attachments: recording.mp4
>
>
> Update: PARQUET-2126 has been opened to describe the underlying thread safety
> problem in parquet-mr. The rdblue/brotli-codec is also affected.
> In previously released versions of Drill, back to at least 1.17, this bug
> only appears under the combination of the async column reader and the _sync_
> page reader, as per the reproduction script below. In master, the bug
> appears under the async column reader and both the sync and async page
> readers.
> {code:java}
> set `store.parquet.compression` = 'gzip';
> drop table if exists dfs.tmp.m;
> create table dfs.tmp.m as select * from cp.`tpch/supplier.parquet`;
> set `store.parquet.reader.pagereader.async` = false;
> set `store.parquet.reader.columnreader.async` = true;
> select * from dfs.tmp.m order by s_suppkey; – repeat this last query and
> watch the returned data. Eventually you will also failed queries or JVM
> crashes
> {code}
>
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
