GH-3530: Bypass Hadoop codec abstraction to optimize compression performance#3570
Open
iemejia wants to merge 1 commit into
Open
GH-3530: Bypass Hadoop codec abstraction to optimize compression performance#3570iemejia wants to merge 1 commit into
iemejia wants to merge 1 commit into
Conversation
iemejia
force-pushed
the
parquet-perf-v2-par6-compression
branch
from
a413752 to
fe836a6
Compare
…n performance Bypass Hadoop CompressionCodec for Snappy (xerial JNI), LZ4_RAW (airlift), ZSTD (zstd-jni), GZIP (JDK), LZO (airlift), and BROTLI (brotli4j) in both CodecFactory and DirectCodecFactory, eliminating per-page codec pool lookups, stream wrapper allocation, and unnecessary buffer copies. SNAPPY: direct byte-array JNI calls via Snappy.compress/uncompress, avoiding the Hadoop stream abstraction and intermediate direct ByteBuffer copies. ZSTD: replace streaming ZstdOutputStreamNoFinalizer/ZstdInputStreamNoFinalizer with reusable ZstdCompressCtx/ZstdDecompressCtx single-call APIs. Supports compression level and multi-threaded workers configuration. LZ4_RAW: direct airlift Lz4Compressor/Lz4Decompressor with reusable direct ByteBuffers, bypassing Hadoop's NonBlockedCompressor overhead. GZIP: bypass Hadoop's GzipCodec and codec-pool/stream-wrapper overhead with direct JDK GZIPOutputStream/GZIPInputStream. Compression level is read from the existing "zlib.compress.level" Hadoop configuration key. LZO: use aircompressor's LzoHadoopStreams directly, bypassing the GPL-licensed com.hadoop.compression.lzo.LzoCodec. Wire-compatible with Hadoop's LzoCodec. BROTLI: migrate from jbrotli (unmaintained) to brotli4j via reflection, using single-call Encoder.compress/Decoder.decompress byte-array APIs. End-to-end interop tests (TestCompressionInterop) validate that files written with the old Hadoop CompressionCodec path are readable by the new direct path and vice versa, for all 6 codecs including multi-row-group scenarios.
iemejia
force-pushed
the
parquet-perf-v2-par6-compression
branch
from
fe836a6 to
0571b9e
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part of #3530 — Apache Parquet Java Performance Improvements
Summary
Bypass the Hadoop
CompressionCodecabstraction for all six supported codecs, eliminating per-page codec-pool lookups, stream-wrapper allocation, and unnecessary buffer copies in bothCodecFactoryandDirectCodecFactory.SnappyCodecstream wrappersSnappy.compress/uncompressdirect callsLZ4Compressor/LZ4DecompressordirectZstdOutputStreamNoFinalizer/ZstdInputStreamNoFinalizerZstdCompressCtx/ZstdDecompressCtxsingle-call APIsGzipCodecwith codec-pool overheadGZIPOutputStream/GZIPInputStreamdirectcom.hadoop.compression.lzo.LzoCodecLzoHadoopStreams(Apache 2.0, wire-compatible)brotli-codec(jbrotli, 2016, x86-only)brotli4j1.23.0 (10 platforms incl. aarch64, reflection-loaded)Notable side effects:
JMH benchmarks:
CompressionBenchmark,CpuReadBenchmark,CpuWriteBenchmark,FileReadBenchmark,FileWriteBenchmark,ConcurrentReadWriteBenchmark.Benchmark results
Environment: JDK 25.0.3 (Temurin), OpenJDK 64-Bit Server VM, JMH 1.37, Linux x86_64.
End-to-end file write (100K rows, SingleShotTime, ms/op lower is better):
End-to-end file read (ms/op lower is better):
Raw codec throughput (
DirectCodecFactory): Snappy/ZSTD/LZ4/GZIP unchanged (already had native access). Brotli decompression improved 2.3-2.7x (brotli4j >> jbrotli).V2 shows consistently larger speedups than V1 because V2 encoding produces more, smaller pages, meaning more codec invocations per file where the per-invocation Hadoop overhead accumulates.