[
https://issues.apache.org/jira/browse/HADOOP-19855?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18071429#comment-18071429
]
ASF GitHub Bot commented on HADOOP-19855:
-----------------------------------------
pan3793 commented on code in PR #8399:
URL: https://github.com/apache/hadoop/pull/8399#discussion_r3039415975
##########
hadoop-common-project/hadoop-common/src/main/java/org/apache/hadoop/io/compress/zstd/ZStandardCompressor.java:
##########
@@ -210,30 +196,49 @@ public int compress(byte[] b, int off, int len) throws
IOException {
return n;
}
- // Re-initialize the output direct buffer
- compressedDirectBuf.rewind();
- compressedDirectBuf.limit(directBufferSize);
+ boolean allConsumed = (uncompressedDirectBufLen - uncompressedDirectBufOff
<= 0);
+ // Use END only when finish=true, no more user data, and all direct-buffer
+ // data consumed (mirrors ZSTD_endStream); otherwise FLUSH (mirrors
+ // ZSTD_compressStream + ZSTD_flushStream).
+ boolean shouldEnd = finish && userBufLen == 0 && allConsumed;
+ if (!allConsumed || shouldEnd) {
+ // Re-initialize the output direct buffer
Review Comment:
changed to always call `FLUSH`/`END` to drain any internally buffered bytes,
even with no input, which mirrors how old native code always calls
`ZSTD_flushStream`
> Use zstd-jni in ZStandardCodec
> ------------------------------
>
> Key: HADOOP-19855
> URL: https://issues.apache.org/jira/browse/HADOOP-19855
> Project: Hadoop Common
> Issue Type: Improvement
> Components: compress
> Affects Versions: 3.5.0
> Reporter: Cheng Pan
> Priority: Major
> Labels: pull-request-available
>
> In Hadoop, we use native libs for zstd codec which has several disadvantages:
> * It requires native *libhadoop* and *libzstd* to be installed in system
> {*}LD_LIBRARY_PATH{*}, and they have to be installed separately on each node
> of the clusters, container images, or local test environments which adds huge
> complexities from deployment point of view. In some environments, it requires
> compiling the natives from sources which is non-trivial. Also, this approach
> is platform dependent; the binary may not work in different platform, so it
> requires recompilation.
> * It requires extra configuration of *java.library.path* to load the
> natives, and it results higher application deployment and maintenance cost
> for users.
> Projects such as *Spark* and *Parquet* use
> [zstd-jni|https://github.com/luben/zstd-jni] which is JNI-based
> implementation. It contains native binaries for Linux, Mac, and IBM in jar
> file, and it can automatically load the native binaries into JVM from jar
> without any setup.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
