Taking this out of the proposal process as it does not require api changes.

I think adding a pool for flate is valid given the improvements in performance.

cc @dsnet per https://dev.golang.org/owners

Can i contribute this issue?

can reduce memory pressure slightly more.

How much of the difference would there be between pooling writers and compressors?

Benchmark_gzip

This looks like a custom benchmark, could you provide the code please?

@AlexanderYastrebov
Sure, Absolutely!

I found that depending on compression level, benchmark can be different. so i change benchmark above.

func Benchmark_gzipHuffmanOnly(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, HuffmanOnly)
}

func Benchmark_gzipDefaultCompression(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, DefaultCompression)
}

func Benchmark_gzip0(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, NoCompression)
}

func Benchmark_gzip1(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, 1)
}

func Benchmark_gzip2(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, 2)
}

func Benchmark_gzip3(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, 3)
}

func Benchmark_gzip4(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, 4)
}

func Benchmark_gzip5(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, 5)
}

func Benchmark_gzip6(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, 6)
}

func Benchmark_gzip7(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, 7)
}

func Benchmark_gzip8(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, 8)
}

func Benchmark_gzip9(b *testing.B) {
	b.ReportAllocs()
	b.ResetTimer()

	benchmarkGzipLevel(b, 9)
}

func benchmarkGzipLevel(b *testing.B, level int) {
	for i := 0; i < b.N; i++ {
		var buf bytes.Buffer
		w, _ := NewWriterLevel(&buf, level)
		w.Write([]byte("hello world"))
		w.Flush()
		w.Close()
	}
}

• Huffman

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzipHuffmanOnly
  Benchmark_gzipHuffmanOnly-10    	  788176	      1490 ns/op	     328 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzipHuffmanOnly
  Benchmark_gzipHuffmanOnly-10    	   19839	     56348 ns/op	  737974 B/op	      18 allocs/op

• Default

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzipDefaultCompression
  Benchmark_gzipDefaultCompression-10    	   73125	     15166 ns/op	     315 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzipDefaultCompression
  Benchmark_gzipDefaultCompression-10    	   18640	     59237 ns/op	  814007 B/op	      20 allocs/op

• 0

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip0
  Benchmark_gzip0-10    	 4879906	       238.9 ns/op	     394 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip0
  Benchmark_gzip0-10    	   26554	     44113 ns/op	  733363 B/op	      16 allocs/op

• 1

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip1
  Benchmark_gzip1-10    	 5075827	       237.3 ns/op	     410 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip1
  Benchmark_gzip1-10    	   16958	     71365 ns/op	 1200313 B/op	      19 allocs/op

• 2

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip2
  Benchmark_gzip2-10    	   80785	     14882 ns/op	     324 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip2
  Benchmark_gzip2-10    	   17382	     65414 ns/op	  814008 B/op	      20 allocs/op

• 3

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip3
  Benchmark_gzip3-10    	   77919	     14819 ns/op	     304 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip3
  Benchmark_gzip3-10    	   18144	     64379 ns/op	  814008 B/op	      20 allocs/op

• 4

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip4
  Benchmark_gzip4-10    	   76875	     14987 ns/op	     325 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip4
  Benchmark_gzip4-10    	   18896	     63697 ns/op	  814008 B/op	      20 allocs/op

• 5

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip5
  Benchmark_gzip5-10    	   74888	     14885 ns/op	     304 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip5
  Benchmark_gzip5-10    	   18529	     66656 ns/op	  814008 B/op	      20 allocs/op

• 6

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip6
  Benchmark_gzip6-10    	   79525	     14897 ns/op	     324 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip6
  Benchmark_gzip6-10    	   18442	     63420 ns/op	  814008 B/op	      20 allocs/op

• 7

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip7
  Benchmark_gzip7-10    	   76423	     14833 ns/op	     314 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip7
  Benchmark_gzip7-10    	   18171	     64918 ns/op	  814008 B/op	      20 allocs/op

• 8

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip8
  Benchmark_gzip8-10    	   77782	     14842 ns/op	     325 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip8
  Benchmark_gzip8-10    	   17658	     65884 ns/op	  814009 B/op	      20 allocs/op

• 9

  pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip9
  Benchmark_gzip9-10    	   78577	     14919 ns/op	     324 B/op	       4 allocs/op

  non-pooling

  goos: darwin
  goarch: arm64
  pkg: compress/gzip
  Benchmark_gzip9
  Benchmark_gzip9-10    	   16924	     67334 ns/op	  814009 B/op	      20 allocs/op

I don't think it's wise to add pooling in situations where sync.Pool.Get and the corresponding sync.Pool.Put are outside the control of the package itself.

In this situation, the call to sync.Pool.Get would presumably occur in gzip.Writer.Write, while the associated call to sync.Pool.Put would presumably happen in gzip.Writer.Close. Whether these are ever called is outside of gzip's control. This has at least two issues:

• It assumes that users always call gzip.Writer.Close otherwise we leak the pooled resource, which can affect global performance. Unlike os.File, where not closing the file leaks a file descriptor, there was never a requirement for gzip.Writer.Close to be called. In many code paths where an error may occur, there is no deferred call to gzip.Writer.Close.
• It will hurt use cases that already rely on gzip.Writer.Reset. When gzip.Writer.Close is called, it cannot distinguish between whether the user intends to manually reuse use the writer or whether the writer will shortly be GC-able. Always putting the resource back in the pool and grabbing it back out again later may have pathological effects if the pool is often drained out (perhaps due to the previous effect). A sync.Pool is more expensive than a regular allocation when empty since it needs to do O(n) work checking if other P's have a resource available.

With the current API, pooling should be the responsibility of the user.

OK. I understood. Thank you for answering!