hertz
hertz copied to clipboard
cloudwego
SSE 中 扫描解析 stream的 EOL 逻辑 基于 bytes.IndexByte() 实现 是否最佳
Issue(#1373 #1377)
SSE 中 扫描解析 stream的 EOL 逻辑 基于 bytes.IndexByte() 实现 在 本场景 中 是否最佳？
// file: pkg/protocol/sse/utils.go
// https://html.spec.whatwg.org/multipage/server-sent-events.html#parsing-an-event-stream
// end-of-line = ( cr lf / cr / lf )
func scanEOL(data []byte, atEOF bool) (advance int, token []byte, err error) {
if atEOF && len(data) == 0 {
return 0, nil, nil
}
i := bytes.IndexByte(data, '\r')
j := bytes.IndexByte(data, '\n')
if i >= 0 {
if i+1 == j { // \r\n
return i + 2, data[0:i], nil
}
if j >= 0 { // choose the nearer \r or \n as EOL
if i < j {
return i + 1, data[0:i], nil // \r
}
return j + 1, data[0:j], nil // \n
}
// if ends with '\r', we need to check the next char is NOT '\n' as per spec
// this may cause unexpected blocks on reading more data.
if i < len(data)-1 || atEOF {
return i + 1, data[0:i], nil
}
} else if j >= 0 {
return j + 1, data[0:j], nil
}
if atEOF {
return len(data), data, nil
}
return 0, nil, nil // more data
}
// https://html.spec.whatwg.org/multipage/server-sent-events.html#parsing-an-event-stream
// end-of-line = ( cr lf / cr / lf )
func scanEOL2(data []byte, atEOF bool) (advance int, token []byte, err error) {
size := len(data)
if atEOF && size == 0 {
return
}
for i, c := range data {
switch c {
case '\r': // \r OR \r\n AS EOL
if i+1 < size && data[i+1] == '\n' {
advance, token = i+2, data[:i]
return
}
// if ends with '\r', we need to check the next char is NOT '\n' as per spec
// this may cause unexpected blocks on reading more data.
if i+1 < size || atEOF {
advance, token = i+1, data[:i]
return
}
case '\n': // \n AS EOL
advance, token = i+1, data[:i]
return
default:
// nothing
}
}
if atEOF {
advance, token = size, data
}
return // more data
}
// file: pkg/protocol/sse/utils_test.go
func BenchmarkScanEOL(b *testing.B) {
tests := []struct {
data string
atEOF bool
advance int
token string
}{
{"", true, 0, ""},
{"", false, 0, ""},
{"hello long long long long long long long long long long long long\r\nworld", false, 69, "hello long long long long long long long long long long long long"},
{"hello long long long long long long long long long long long long\rworld", false, 68, "hello long long long long long long long long long long long long"},
{"hello long long long long long long long long long long long long\nworld", false, 68, "hello long long long long long long long long long long long long"},
{"hello long long long long long long long long long long long long world", false, 0, ""},
{"hello long long long long long long long long long long long long world", true, 73, "hello long long long long long long long long long long long long world"},
{"hello\r\nworld", false, 7, "hello"},
{"hello\rworld", false, 6, "hello"},
{"hello\nworld", false, 6, "hello"},
{"hello world", false, 0, ""},
{"hello world", true, 11, "hello world"},
{"\r", false, 0, ""},
{"hello\r", false, 0, ""},
{"hello\r", true, 6, "hello"},
{"\n", false, 1, ""},
{"\r\nhello", false, 2, ""},
{"\r\n", false, 2, ""},
}
for i := 0; i < b.N; i++ {
for _, tc := range tests {
advance, token, _ := scanEOL([]byte(tc.data), tc.atEOF)
if advance != tc.advance || string(token) != tc.token {
b.Fatalf("scanLines(data=%q, atEOF=%v) returns (%d, %q) expect (%d, %q)",
tc.data, tc.atEOF, advance, string(token), tc.advance, tc.token)
}
}
}
}
func BenchmarkScanEOL2(b *testing.B) {
tests := []struct {
data string
atEOF bool
advance int
token string
}{
{"", true, 0, ""},
{"", false, 0, ""},
{"hello long long long long long long long long long long long long\r\nworld", false, 69, "hello long long long long long long long long long long long long"},
{"hello long long long long long long long long long long long long\rworld", false, 68, "hello long long long long long long long long long long long long"},
{"hello long long long long long long long long long long long long\nworld", false, 68, "hello long long long long long long long long long long long long"},
{"hello long long long long long long long long long long long long world", false, 0, ""},
{"hello long long long long long long long long long long long long world", true, 73, "hello long long long long long long long long long long long long world"},
{"hello\r\nworld", false, 7, "hello"},
{"hello\rworld", false, 6, "hello"},
{"hello\nworld", false, 6, "hello"},
{"hello world", false, 0, ""},
{"hello world", true, 11, "hello world"},
{"\r", false, 0, ""},
{"hello\r", false, 0, ""},
{"hello\r", true, 6, "hello"},
{"\n", false, 1, ""},
{"\r\nhello", false, 2, ""},
{"\r\n", false, 2, ""},
}
for i := 0; i < b.N; i++ {
for _, tc := range tests {
advance, token, _ := scanEOL2([]byte(tc.data), tc.atEOF)
if advance != tc.advance || string(token) != tc.token {
b.Fatalf("scanLines(data=%q, atEOF=%v) returns (%d, %q) expect (%d, %q)",
tc.data, tc.atEOF, advance, string(token), tc.advance, tc.token)
}
}
}
}
benchmark
> go test -bench=ScanEOL* -benchmem
goos: linux
goarch: amd64
pkg: github.com/cloudwego/hertz/pkg/protocol/sse
cpu: Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz
BenchmarkScanEOL-8 1581817 681.9 ns/op 400 B/op 5 allocs/op
BenchmarkScanEOL2-8 2307902 515.6 ns/op 0 B/op 0 allocs/op
PASS
ok github.com/cloudwego/hertz/pkg/protocol/sse 3.607s
我是好奇才benchmark了一下，最开始我也以为 既然bytes.IndexByte() 是有指令级优化的，应该会快很多，结果是两种逻辑差不多快； 如果注释掉 这段稍微长一点的文本用例
{"hello long long long long long long long long long long long long\r\nworld", false, 69, "hello long long long long long long long long long long long long"},
{"hello long long long long long long long long long long long long\rworld", false, 68, "hello long long long long long long long long long long long long"},
{"hello long long long long long long long long long long long long\nworld", false, 68, "hello long long long long long long long long long long long long"},
{"hello long long long long long long long long long long long long world", false, 0, ""},
{"hello long long long long long long long long long long long long world", true, 73, "hello long long long long long long long long long long long long world"},
结果是
> go test -bench=ScanEOL* -benchmem
goos: linux
goarch: amd64
pkg: github.com/cloudwego/hertz/pkg/protocol/sse
cpu: Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz
BenchmarkScanEOL-8 4079202 278.3 ns/op 0 B/op 0 allocs/op
BenchmarkScanEOL2-8 6956415 156.8 ns/op 0 B/op 0 allocs/op
PASS
ok github.com/cloudwego/hertz/pkg/protocol/sse 2.875s
go version
> go version
go version go1.24.3 linux/amd64
scanEOL2() 看起来是稍微快一点， 但就是不知道逻辑是否完全符合设计，或者说需要更多的测试用例来检测一下；同时我也比较好奇， 如果 data 文本很大时 比如几kB，结果会怎样？ 有时间的和感兴趣的可以在本机上测试一下；
（本Issue只是一时好奇探索了一下，暂时也没太多时间探索，后面闲了继续看看～)
