danny0405 commented on code in PR #19685:
URL: https://github.com/apache/hudi/pull/19685#discussion_r3879158960
##########
hudi-client/hudi-client-common/src/main/java/org/apache/hudi/config/HoodieWriteConfig.java:
##########
@@ -1429,6 +1433,40 @@ public static HoodieWriteConfig.Builder newBuilder() {
return new Builder();
}
+ @VisibleForTesting
+ static String getDefaultParquetCompressionCodec(EngineType engineType) {
+ switch (engineType) {
+ case FLINK:
+ return ZSTD_COMPRESSION_CODEC;
+ case SPARK:
+ // Spark 3.5 and newer use ZSTD. Spark 3.3 and 3.4 retain GZIP because
their
+ // non-vectorized file-group reader uses parquet-java 1.12.x and can
leak off-heap
+ // memory when reading ZSTD files:
https://issues.apache.org/jira/browse/PARQUET-2160.
+ Option<String> sparkVersion = getSparkRuntimeVersion();
+ return sparkVersion.isPresent()
+ && StringUtils.compareVersions(sparkVersion.get(),
MIN_SPARK_VERSION_WITH_ZSTD_DEFAULT) >= 0
+ ? ZSTD_COMPRESSION_CODEC : GZIP_COMPRESSION_CODEC;
+ default:
+ // The Java client does not own its Parquet runtime: Parquet
dependencies are provided by
+ // the embedding application, and older Parquet versions use Hadoop
native ZSTD rather than
+ // zstd-jni. For example, the recommended Kafka HDFS Connector 10.1.0
uses Parquet 1.11.1,
+ // which risks leaking memory when reading ZSTD-compressed files. Keep
GZIP as the portable
+ // default across supported Java deployments.
+ return GZIP_COMPRESSION_CODEC;
Review Comment:
gzip or snappy
--
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]
