[ https://issues.apache.org/jira/browse/CASSANDRA-8765?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14312392#comment-14312392 ]
Robert Stupp commented on CASSANDRA-8765: ----------------------------------------- Might be a chance to re-validate performance of snappy(native) vs. snappy(java) vs. gzip. As far as I remember from another project, the speedup of snappy(native) vs. gzip (using pooled in/deflaters) was very marginal (different use case, of course). > Switch to a pure Java implementation of Snappy to reduce compatibility issues > ----------------------------------------------------------------------------- > > Key: CASSANDRA-8765 > URL: https://issues.apache.org/jira/browse/CASSANDRA-8765 > Project: Cassandra > Issue Type: Improvement > Reporter: Michaël Figuière > Priority: Minor > > Since it has been introduced as a dependency, the Xerial's Snappy library, > which relies on a few native libraries for performance purpose, has caused > several compatibility issue or unexpected behaviors. > Dain's implementation (https://github.com/dain/snappy) is written in pure > Java using some {{Unsafe}} memory access for secure its performance. > According to the benchmark published by its author > (https://github.com/dain/snappy#block-compress), it ends up being often > faster than the JNI implementation, but without the compatibility issues. -- This message was sent by Atlassian JIRA (v6.3.4#6332)