[SPARK-12388] change default compression to lz4

[davies]

According the benchmark [1], LZ4-java could be 80% (or 30%) faster than Snappy.

After changing the compressor to LZ4, I saw 20% improvement on end-to-end time for a TPCDS query (Q4).

[1] https://github.com/ning/jvm-compressor-benchmark/wiki

cc @rxin

[rxin]

We also need to update the documentation, don't we?

[JoshRosen]

Note that LZ4 does not support concatenation of serialized streams, meaning that sort-based shuffle won't be able to take advantage of a special spill-merging optimization. This might not make a huge difference in practice compared to the other speed benefits of LZ4.

[SparkQA]

Test build #47876 has finished for PR 10342 at commit 14940de.

• This patch fails from timeout after a configured wait of 250m.
• This patch merges cleanly.
• This patch adds no public classes.

[nongli]

@JoshRosen Can you explain that some more? What are we giving up? Would fixing it if we decide to just to add a simple header that consisted of the lengths of each stream?

[rxin]

I think the main thing this is giving up is that previously we could use file channel's transferTo to copy data directly, but only when the compression blocks can be just concatenated at the byte level. Without that property, we'd have to read them (decompress) and then write them out again.

[davies]

I had patched LZ4BlockInputStream to support concated streams.

[SparkQA]

Test build #47898 has finished for PR 10342 at commit f49f1ef.

• This patch fails from timeout after a configured wait of 250m.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #47908 has finished for PR 10342 at commit 13390e8.

• This patch fails Spark unit tests.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):\n * public final class LZ4BlockInputStream extends FilterInputStream\n

[nongli]

Can you comment what is modified? This is otherwise hard to review/maintain if lz4 updates.

[davies]

See the changes here: davies/lz4-java@cc1fa94

[SparkQA]

Test build #47961 has finished for PR 10342 at commit 83c77f4.

• This patch fails from timeout after a configured wait of 250m.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):\n * public final class LZ4BlockInputStream extends FilterInputStream\n

[SparkQA]

Test build #2231 has finished for PR 10342 at commit 83c77f4.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):\n * public final class LZ4BlockInputStream extends FilterInputStream\n

[davies]

ping @JoshRosen @rxin @nongli

[nongli]

LGTM

[rxin]

One thing I noticed when I was reviewing another pr just now: in the future we should also include test cases when pulling in 3rd party code. Otherwise it's pretty difficult to ensure correctness if we need to change the file.