Reduce the size of the bufio.Reader for message header parsing.

[thomas-long-f3]

When profiling a service that consumes messages from NATS Jetstream, we observed a lot of allocations from the Go client. The culprit is this line, by default bufio.NewReader() allocates a 4 KiB buffer. This PR reduces the buffer to 128, this seems to be a good middle-ground based on benchmarks below.

64

name                         old time/op    new time/op    delta
HeaderDecode/Small_-_25-8      1.44µs ±17%    0.31µs ±15%  -78.66%  (p=0.000 n=23+23)
HeaderDecode/Medium_-_141-8    2.46µs ±29%    0.95µs ±13%  -61.57%  (p=0.000 n=24+24)
HeaderDecode/Large_-_368-8     2.88µs ±22%    1.42µs ± 6%  -50.59%  (p=0.000 n=25+25)

name                         old alloc/op   new alloc/op   delta
HeaderDecode/Small_-_25-8      4.58kB ± 0%    0.55kB ± 0%  -87.96%  (p=0.000 n=25+25)
HeaderDecode/Medium_-_141-8    4.94kB ± 0%    0.91kB ± 0%  -81.55%  (p=0.000 n=25+25)
HeaderDecode/Large_-_368-8     5.49kB ± 0%    1.84kB ± 0%  -66.47%  (p=0.000 n=25+25)

name                         old allocs/op  new allocs/op  delta
HeaderDecode/Small_-_25-8        7.00 ± 0%      7.00 ± 0%     ~     (all equal)
HeaderDecode/Medium_-_141-8      21.0 ± 0%      21.0 ± 0%     ~     (all equal)
HeaderDecode/Large_-_368-8       27.0 ± 0%      31.0 ± 0%  +14.81%  (p=0.000 n=25+25)

128

name                         old time/op    new time/op    delta
HeaderDecode/Small_-_25-8      1.44µs ±17%    0.29µs ± 3%  -79.96%  (p=0.000 n=23+24)
HeaderDecode/Medium_-_141-8    2.46µs ±29%    0.90µs ± 1%  -63.29%  (p=0.000 n=24+20)
HeaderDecode/Large_-_368-8     2.88µs ±22%    1.20µs ± 8%  -58.47%  (p=0.000 n=25+23)

name                         old alloc/op   new alloc/op   delta
HeaderDecode/Small_-_25-8      4.58kB ± 0%    0.62kB ± 0%  -86.56%  (p=0.000 n=25+25)
HeaderDecode/Medium_-_141-8    4.94kB ± 0%    0.98kB ± 0%  -80.26%  (p=0.000 n=25+25)
HeaderDecode/Large_-_368-8     5.49kB ± 0%    1.52kB ± 0%  -72.30%  (p=0.000 n=25+25)

name                         old allocs/op  new allocs/op  delta
HeaderDecode/Small_-_25-8        7.00 ± 0%      7.00 ± 0%     ~     (all equal)
HeaderDecode/Medium_-_141-8      21.0 ± 0%      21.0 ± 0%     ~     (all equal)
HeaderDecode/Large_-_368-8       27.0 ± 0%      27.0 ± 0%     ~     (all equal)

256

name                         old time/op    new time/op    delta
HeaderDecode/Small_-_25-8      1.44µs ±17%    0.52µs ±17%  -63.63%  (p=0.000 n=23+21)
HeaderDecode/Medium_-_141-8    2.46µs ±29%    1.03µs ±19%  -58.22%  (p=0.000 n=24+23)
HeaderDecode/Large_-_368-8     2.88µs ±22%    1.27µs ±23%  -55.77%  (p=0.000 n=25+22)

name                         old alloc/op   new alloc/op   delta
HeaderDecode/Small_-_25-8      4.58kB ± 0%    0.74kB ± 0%  -83.77%  (p=0.000 n=25+25)
HeaderDecode/Medium_-_141-8    4.94kB ± 0%    1.10kB ± 0%  -77.67%  (p=0.000 n=25+25)
HeaderDecode/Large_-_368-8     5.49kB ± 0%    1.65kB ± 0%  -69.97%  (p=0.000 n=25+25)

name                         old allocs/op  new allocs/op  delta
HeaderDecode/Small_-_25-8        7.00 ± 0%      7.00 ± 0%     ~     (all equal)
HeaderDecode/Medium_-_141-8      21.0 ± 0%      21.0 ± 0%     ~     (all equal)
HeaderDecode/Large_-_368-8       27.0 ± 0%      27.0 ± 0%     ~     (all equal)

[coveralls]

Coverage: 85.493% (+0.002%) from 85.492% when pulling 7f3b22e on thomas-long-f3:decode-header-allocations into eada5ec on nats-io:main.

[piotrpio]

LGTM!