[jira] [Commented] (PARQUET-1641) Parquet pages for different columns cannot be read in parallel

["Samarth Jain (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/PARQUET-1641?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16911737#comment-16911737 ] 

Samarth Jain commented on PARQUET-1641:
---------------------------------------

Note that I ran into this issue when files were compressed using GZIP. Reading the code, it seems like BuiltInGzipDecompressor is not meant to be pooled. But the CodeFactory$HeapBytesDecompressor is still sort of pooling it. 

> Parquet pages for different columns cannot be read in parallel 
> ---------------------------------------------------------------
>
>                 Key: PARQUET-1641
>                 URL: https://issues.apache.org/jira/browse/PARQUET-1641
>             Project: Parquet
>          Issue Type: Improvement
>            Reporter: Samarth Jain
>            Priority: Major
>
> All ColumnChunkPageReader instances use the same decompressor. 
> [https://github.com/apache/parquet-mr/blob/master/parquet-hadoop/src/main/java/org/apache/parquet/hadoop/ParquetFileReader.java#L1286]
> {code:java}
> BytesInputDecompressor decompressor = options.getCodecFactory().getDecompressor(descriptor.metadata.getCodec());
> return new ColumnChunkPageReader(decompressor, pagesInChunk, dictionaryPage);
> {code}
> The CodecFactory caches the decompressors for every codec type returning the same instance on every getCompressor(codecName) call. See the caching happening here:
> [https://github.com/apache/parquet-mr/blob/master/parquet-hadoop/src/main/java/org/apache/parquet/hadoop/CodecFactory.java#L197]
> {code:java}
> @Override
>  public BytesDecompressor getDecompressor(CompressionCodecName codecName) {
>     BytesDecompressor decomp = decompressors.get(codecName);
>     if (decomp == null){ 
>        decomp = createDecompressor(codecName); decompressors.put(codecName, decomp); 
>     }
>     return decomp;
>  }
>  
> {code}
>  
> If multiple threads try to read the pages belonging to different columns, they run into thread
> safety issues. This issue prevents increasing the throughput at which applications can read parquet data by parallelizing page reads. 



--
This message was sent by Atlassian Jira
(v8.3.2#803003)