Render marshalSJON/Raw/CSV preallocated bytes buffer

[msaf1980]

$ go test -benchmem -run=^$ -bench ^Benchmark github.com/go-graphite/carbonapi/expr/types | tee marshalJSON2.txt 

$ benchcmp marshalJSON1.txt marshalJSON2.txt 
benchmark                      old ns/op     new ns/op     delta
BenchmarkMarshalJSON-6         9297315       7406833       -20.33%
BenchmarkMarshalJSONLong-6     135895934     121057103     -10.92%
BenchmarkMarshalRaw-6          5717379       5622104       -1.67%
BenchmarkMarshalRawLong-6      94664746      90235233      -4.68%
BenchmarkMarshalCSV-6          38516100      19187483      -50.18%
BenchmarkMarshalCSVLong-6      587776864     320056366     -45.55%

benchmark                      old allocs     new allocs     delta
BenchmarkMarshalJSON-6         39             5              -87.18%
BenchmarkMarshalJSONLong-6     89             43             -51.69%
BenchmarkMarshalRaw-6          29             1              -96.55%
BenchmarkMarshalRawLong-6      42             1              -97.62%
BenchmarkMarshalCSV-6          110037         3              -100.00%
BenchmarkMarshalCSVLong-6      1820049        6              -100.00%

benchmark                      old bytes     new bytes     delta
BenchmarkMarshalJSON-6         8401270       2569729       -69.41%
BenchmarkMarshalJSONLong-6     130283752     27232828      -79.10%
BenchmarkMarshalRaw-6          1981197       2564105       +29.42%
BenchmarkMarshalRawLong-6      41704200      25608206      -38.60%
BenchmarkMarshalCSV-6          23749513      9773060       -58.85%
BenchmarkMarshalCSVLong-6      358308832     288342088     -19.53%

[msaf1980]

Also may be better direct write to http.ResponseWriter instead of use bytes buffer. It's will be require less memory for large metrrics set.

[Civil]

Yeah, that should work as well. I'm not sure it won't have any complications on the behavior, but might improve time to first byte as well for consumers.

[msaf1980]

Yeah, that should work as well. I'm not sure it won't have any complications on the behavior, but might improve time to first byte as well for consumers.

Also I'm not have experience with this. It can produce chunked response (on large metrics set) and need more tests, that no bugs in http clients, grafana, etc.
May be better merge current, and research some later ?