Implementing Snappy Block Encoding

[alanprot]

What this PR does:

This PR is a follow up of #5213 (comment)

This PR is implementing block format snappy encoding:

Benchmark Results:

pkg: github.com/cortexproject/cortex/pkg/util/grpcencoding
cpu: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
BenchmarkCompress/snappy-12         	  452925	      2339 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/snappy-12         	  545708	      2264 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/snappy-12         	  534312	      2230 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/snappy-12         	  538390	      2293 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/snappy-12         	  540451	      2201 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/snappy-block-12   	  536930	      2233 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/snappy-block-12   	  525919	      2306 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/snappy-block-12   	  476133	      2271 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/snappy-block-12   	  518916	      2203 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/snappy-block-12   	  543210	      2187 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/zstd-12           	  518970	      2202 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/zstd-12           	  528393	      2227 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/zstd-12           	  519684	      2455 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/zstd-12           	  507280	      2268 ns/op	      16 B/op	       1 allocs/op
BenchmarkCompress/zstd-12           	  526838	      2234 ns/op	      16 B/op	       1 allocs/op
BenchmarkDecompress/snappy-12       	  115866	     10098 ns/op	   34212 B/op	      12 allocs/op
BenchmarkDecompress/snappy-12       	  117933	     10162 ns/op	   34254 B/op	      12 allocs/op
BenchmarkDecompress/snappy-12       	  107898	     10284 ns/op	   34238 B/op	      12 allocs/op
BenchmarkDecompress/snappy-12       	  116962	     10009 ns/op	   34262 B/op	      12 allocs/op
BenchmarkDecompress/snappy-12       	  115161	     10359 ns/op	   34237 B/op	      12 allocs/op
BenchmarkDecompress/snappy-block-12 	  192163	      6091 ns/op	    9858 B/op	       4 allocs/op
BenchmarkDecompress/snappy-block-12 	  192910	      6224 ns/op	    9859 B/op	       4 allocs/op
BenchmarkDecompress/snappy-block-12 	  191830	      6087 ns/op	    9857 B/op	       4 allocs/op
BenchmarkDecompress/snappy-block-12 	  193527	      6156 ns/op	    9859 B/op	       4 allocs/op
BenchmarkDecompress/snappy-block-12 	  191727	      6156 ns/op	    9858 B/op	       4 allocs/op
BenchmarkDecompress/zstd-12         	  112138	      9586 ns/op	   43905 B/op	      14 allocs/op
BenchmarkDecompress/zstd-12         	  119416	      9455 ns/op	   43903 B/op	      14 allocs/op
BenchmarkDecompress/zstd-12         	  121332	      9560 ns/op	   43903 B/op	      14 allocs/op
BenchmarkDecompress/zstd-12         	  118489	      9505 ns/op	   43903 B/op	      14 allocs/op
BenchmarkDecompress/zstd-12         	  117463	      9563 ns/op	   43902 B/op	      14 allocs/op
PASS
ok  	github.com/cortexproject/cortex/pkg/util/grpcencoding	36.997s

Comparing Snappy x snappy-block (Decompress)

name                  old time/op    new time/op    delta
Decompress/snappy-12    10.2µs ± 2%     6.1µs ± 1%  -39.67%  (p=0.008 n=5+5)

name                  old alloc/op   new alloc/op   delta
Decompress/snappy-12    34.2kB ± 0%     9.9kB ± 0%  -71.21%  (p=0.008 n=5+5)

name                  old allocs/op  new allocs/op  delta
Decompress/snappy-12      12.0 ± 0%       4.0 ± 0%  -66.67%  (p=0.008 n=5+5)

Which issue(s) this PR fixes:
Fixes #

Checklist

• Tests updated
• Documentation added
• CHANGELOG.md updated - the order of entries should be [CHANGE], [FEATURE], [ENHANCEMENT], [BUGFIX]

[alanprot]

@damnever Do you wanna test this change?

[damnever]

@alanprot Great! I may test this patch next week.

[alanprot]

The results seems pretty good. I dont see any change on the CPU and the network is 10X less.

The first deployment of distributors used more CPU because we were changing the len of the dst array to 0, forcing a new array to be allocated on every operation. This was fixed by c1bc632

[alanprot]

@damnever i addressed your comments...

Did you have the chance to test?

[damnever]

@alanprot The result is very good, here is the comparison of the ingesters' CPU usage in our test environment. (snappy/v1.14 vs snappy/#5213 vs snappy-block):

[alanprot]

Nice!! So if i'm reading correcly we have:

snappy/v1.14: 22.5% CPU
snappy/#5213 : 17.9% CPU
snappy-block: 14.6% CPU

Seems that make sense to go ahead with this PR?

@yeya24 WDYT?

Maybe snappy-block is not a good name haha :P

[damnever]

@alanprot You are right, but not in percentage rather CPU usage in seconds.

[alanprot]

Oh i see..

So it reduced the CPU usage in seconds by 45% (22.5*0.65=~14.6)? Nice!
We saw improvements on our test environments as well but not that big. This is good news!

[yeya24]

Nice! This is great improvement.
@alanprot About naming, I don't have strong preference. snappy-block also sounds good to me.