Optimize dynamic unmarshaling

[echlebek]

What is this change?

    Optimize dynamic unmarshaling.
    
    * Use sync.Pool in several places to optimize allocations.
    * Use a wrapper type and json.RawMessage to reduce work needed in
    unmarshaling dynamic attributes.
    * Use a slice instead of a map for getJSONFields.
    
    This commit substantially reduces runtime and allocations for
    dynamic json unmarshaling.
    
    Before:
    [15:34 eric ~/.../sensu-go/types ((6cfa8e40...) $%)]$ go test -bench=. -benchmem
    goos: linux
    goarch: amd64
    pkg: github.com/sensu/sensu-go/types
    BenchmarkCheckRequestMarshal-4            300000              3915 ns/op            3624 B/op         30 allocs/op
    BenchmarkCheckRequestUnmarshal-4           30000             44284 ns/op           39002 B/op        386 allocs/op
    BenchmarkCheckConfigMarshal-4             500000              2399 ns/op            2056 B/op         19 allocs/op
    BenchmarkCheckConfigUnmarshal-4            30000             40347 ns/op           37340 B/op        358 allocs/op
    PASS
    ok      github.com/sensu/sensu-go/types 5.901s
    
    After:
    [15:35 eric ~/.../sensu-go/types (perf/optimize-check-unmarshal $%)]$ go test -bench=. -benchmem
    goos: linux
    goarch: amd64
    pkg: github.com/sensu/sensu-go/types
    BenchmarkCheckRequestMarshal-4            300000              3520 ns/op            3416 B/op         29 allocs/op
    BenchmarkCheckRequestUnmarshal-4          100000             14693 ns/op            5715 B/op        185 allocs/op
    BenchmarkCheckConfigMarshal-4             500000              2204 ns/op            1848 B/op         18 allocs/op
    BenchmarkCheckConfigUnmarshal-4           200000             10317 ns/op            4046 B/op        157 allocs/op
    PASS
    ok      github.com/sensu/sensu-go/types 6.508s
    
    Before:
    [15:32 eric ~/.../types/dynamic ((6cfa8e40...) $%)]$ go test -bench=. -benchmem
    goos: linux
    goarch: amd64
    pkg: github.com/sensu/sensu-go/types/dynamic
    BenchmarkQueryGovaluateSimple-4          1000000              1322 ns/op             320 B/op         18 allocs/op
    BenchmarkQueryGovaluateComplex-4          300000              4431 ns/op            1160 B/op         66 allocs/op
    BenchmarkUnmarshal-4                      200000              9686 ns/op            7656 B/op         82 allocs/op
    BenchmarkMarshal-4                        300000              4783 ns/op            6528 B/op         43 allocs/op
    PASS
    ok      github.com/sensu/sensu-go/types/dynamic 6.277s
    
    After:
    [15:32 eric ~/.../types/dynamic (perf/optimize-check-unmarshal $%)]$ go test -bench=. -benchmem
    goos: linux
    goarch: amd64
    pkg: github.com/sensu/sensu-go/types/dynamic
    BenchmarkQueryGovaluateSimple-4          1000000              1311 ns/op             320 B/op         18 allocs/op
    BenchmarkQueryGovaluateComplex-4          300000              4334 ns/op            1160 B/op         66 allocs/op
    BenchmarkUnmarshal-4                      500000              2752 ns/op            1163 B/op         48 allocs/op
    BenchmarkMarshal-4                        500000              2402 ns/op            5647 B/op         38 allocs/op
    PASS
    ok      github.com/sensu/sensu-go/types/dynamic 5.337s

Why is this change necessary?

Unlike dynamic marshaling, dynamic unmarshaling cannot be avoided. Therefore, we must take steps to optimize it as much as possible to avoid performance problems under load. Dynamic unmarshaling is necessary for every CheckRequest, and cannot be avoided in normal operation. These operations generate many allocations, leading to high GC pressure. This commit substantially reduces these allocations to a more manageable level.

Does your change need a Changelog entry?

Yes

[palourde]

[mercul3s]

[mercul3s]

TIL sync.Pool - that's pretty neat.