Performance improvements to StatSerializer

[marcotc]

Similar to #155, this PR reduces the overhead of submitting metrics by dogstatsd-ruby.

Most of the gains came from avoiding successive string resizing performed by String#<<.

Another large gain came from replacing string.dup.gsub! with string.gsub in #formated_name:

require 'benchmark/ips'
a = "abc"
Benchmark.ips do |x|
  x.report('gsub') { a.gsub('a', 'z') }
  x.report('dup.gsub!') { a.dup.gsub!('a', 'z') }
  x.compare!
end

# Comparison:
#        gsub:  1600766.3 i/s
#   dup.gsub!:  1091421.3 i/s - 1.47x  (± 0.00) slower

A smaller gain came from removing the use of tap in two locations:

require 'benchmark/ips'
Benchmark.ips do |x|
  x.report('inline') do
    s = "abc"
    s.tr!('a', 'z')
    s
  end
  x.report('tap') do
    "abc".tap { |s| s.tr!('a', 'z') }
  end
  x.compare!
end

# Comparison:
#   inline:  4300892.6 i/s
#      tap:  3601127.4 i/s - 1.19x  (± 0.00) slower

There were no memory improvements performed.

Benchmark results

Before

Calculating -------------------------------------
             no tags     822.563k (± 1.9%) i/s -      4.159M in   5.058459s
no tags + sample rate    557.104k (± 1.7%) i/s -      2.800M in   5.027398s
           with tags     290.806k (± 1.9%) i/s -      1.474M in   5.070795s
with tags + sample rate  248.824k (± 1.9%) i/s -      1.254M in   5.040946s

Comparison:
                no tags:   822563.2 i/s
  no tags + sample rate:   557104.3 i/s - 1.48x  (± 0.00) slower
              with tags:   290805.8 i/s - 2.83x  (± 0.00) slower
with tags + sample rate:   248823.6 i/s - 3.31x  (± 0.00) slower

After

Calculating -------------------------------------
             no tags       1.098M (± 1.4%) i/s -      5.526M in   5.036068s
no tags + sample rate    707.986k (± 1.3%) i/s -      3.579M in   5.055567s
           with tags     325.535k (± 2.9%) i/s -      1.629M in   5.009379s
with tags + sample rate  275.804k (± 2.6%) i/s -      1.386M in   5.027488s

Comparison:
                no tags:  1097525.6 i/s
  no tags + sample rate:   707986.2 i/s - 1.55x  (± 0.00) slower
              with tags:   325534.7 i/s - 3.37x  (± 0.00) slower
with tags + sample rate:   275803.8 i/s - 3.98x  (± 0.00) slower

Results

Wall time	% improvement
no tags	33.4%
no tags + sample rate	27.0%
with tags	11.9%
with tags + sample rate	10.8%

[marcotc]

Another performance improvement for review when you have a chance, @kbogtob.

[ericmustin]

lgtm, i think using string interpolation should help here, seems like we've preserved all the same logic and case handling

[ericmustin]

👍

[truthbk]

This is interesting given the extra allocation(?) in gsub!, do we know why?

[truthbk]

This looks good to me 👍

(added a question about gsub! but feel free to merge IMO)

[truthbk]

This is interesting given the extra allocation(?) in gsub!, do we know why?