Highlight

Support direct io to bypass page cache and smart io scheduler for stable and perdictable latency
Reduce the memory footprint on blockId bitmap store
Production ready that has been deployed in IQiyi internal Uniffle cluster (stable and high-performance)

What's Changed

feat(direct-io): Support direct io for predictable io performance by @zuston in #19
feat(direct-io): Auto return back to the buffer pool when buffer is useless by @zuston in #20
feat(direct-io): Use take to avoid allocation when releasing ioBuffer by @zuston in #21
feat(debug): Make debug level to 1 for better pprof visualization by @zuston in #22
chore: Update croaring to the version 2.x by @zuston in #23
feat: Purge blockId bitmap on shuffle level deletion by @zuston in #24
feat: Try to merge multi blockIds bitmap into one by shuffleId to reduce memory footprint by @zuston in #25
chore: Allow dead_code and unused import or vars by @zuston in #26
chore: Correct snake case of getHdfsDelegator by @zuston in #27
fix: Correct partition number metric by @zuston in #30
fix: Align with the internal blockId config by @zuston in #33
fix: Only delete hdfs files with owned shuffle-server id by @zuston in #32

Full Changelog: v0.3.1...v0.4.0