add pipeline (#74)

pull/87/head
jiankuny 6 years ago committed by GitHub
parent 492b682fe2
commit fcba425d39
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
  1. 3
      pkg/sync/pipeline/README.md
  2. 14
      pkg/sync/pipeline/fanout/README.md
  3. 22
      pkg/sync/pipeline/fanout/example_test.go
  4. 151
      pkg/sync/pipeline/fanout/fanout.go
  5. 30
      pkg/sync/pipeline/fanout/fanout_test.go
  6. 185
      pkg/sync/pipeline/pipeline.go
  7. 132
      pkg/sync/pipeline/pipeline_test.go

@ -0,0 +1,3 @@
# pkg/sync/pipeline
提供内存批量聚合工具

@ -0,0 +1,14 @@
# pkg/sync/pipeline/fanout
功能:
* 支持定义Worker 数量的goroutine,进行消费
* 内部支持的元数据传递(pkg/net/metadata)
示例:
```golang
//名称为cache 执行线程为1 buffer长度为1024
cache := fanout.New("cache", fanout.Worker(1), fanout.Buffer(1024))
cache.Do(c, func(c context.Context) { SomeFunc(c, args...) })
cache.Close()
```

@ -0,0 +1,22 @@
package fanout
import "context"
// addCache 加缓存的例子
func addCache(c context.Context, id, value int) {
// some thing...
}
func Example() {
// 这里只是举个例子 真正使用的时候 应该用bm/rpc 传过来的context
var c = context.Background()
// 新建一个fanout 对象 名称为cache 名称主要用来上报监控和打日志使用 最好不要重复
// (可选参数) worker数量为1 表示后台只有1个线程在工作
// (可选参数) buffer 为1024 表示缓存chan长度为1024 如果chan慢了 再调用Do方法就会报错 设定长度主要为了防止OOM
cache := New("cache", Worker(1), Buffer(1024))
// 需要异步执行的方法
// 这里传进来的c里面的meta信息会被复制 超时会忽略 addCache拿到的context已经没有超时信息了
cache.Do(c, func(c context.Context) { addCache(c, 0, 0) })
// 程序结束的时候关闭fanout 会等待后台线程完成后返回
cache.Close()
}

@ -0,0 +1,151 @@
package fanout
import (
"context"
"errors"
"runtime"
"sync"
"github.com/bilibili/kratos/pkg/log"
"github.com/bilibili/kratos/pkg/net/metadata"
"github.com/bilibili/kratos/pkg/net/trace"
"github.com/bilibili/kratos/pkg/stat/prom"
)
var (
// ErrFull chan full.
ErrFull = errors.New("fanout: chan full")
stats = prom.BusinessInfoCount
traceTags = []trace.Tag{
trace.Tag{Key: trace.TagSpanKind, Value: "background"},
trace.Tag{Key: trace.TagComponent, Value: "sync/pipeline/fanout"},
}
)
type options struct {
worker int
buffer int
}
// Option fanout option
type Option func(*options)
// Worker specifies the worker of fanout
func Worker(n int) Option {
if n <= 0 {
panic("fanout: worker should > 0")
}
return func(o *options) {
o.worker = n
}
}
// Buffer specifies the buffer of fanout
func Buffer(n int) Option {
if n <= 0 {
panic("fanout: buffer should > 0")
}
return func(o *options) {
o.buffer = n
}
}
type item struct {
f func(c context.Context)
ctx context.Context
}
// Fanout async consume data from chan.
type Fanout struct {
name string
ch chan item
options *options
waiter sync.WaitGroup
ctx context.Context
cancel func()
}
// New new a fanout struct.
func New(name string, opts ...Option) *Fanout {
if name == "" {
name = "fanout"
}
o := &options{
worker: 1,
buffer: 1024,
}
for _, op := range opts {
op(o)
}
c := &Fanout{
ch: make(chan item, o.buffer),
name: name,
options: o,
}
c.ctx, c.cancel = context.WithCancel(context.Background())
c.waiter.Add(o.worker)
for i := 0; i < o.worker; i++ {
go c.proc()
}
return c
}
func (c *Fanout) proc() {
defer c.waiter.Done()
for {
select {
case t := <-c.ch:
wrapFunc(t.f)(t.ctx)
stats.State(c.name+"_channel", int64(len(c.ch)))
case <-c.ctx.Done():
return
}
}
}
func wrapFunc(f func(c context.Context)) (res func(context.Context)) {
res = func(ctx context.Context) {
defer func() {
if r := recover(); r != nil {
buf := make([]byte, 64*1024)
buf = buf[:runtime.Stack(buf, false)]
log.Error("panic in fanout proc, err: %s, stack: %s", r, buf)
}
}()
f(ctx)
if tr, ok := trace.FromContext(ctx); ok {
tr.Finish(nil)
}
}
return
}
// Do save a callback func.
func (c *Fanout) Do(ctx context.Context, f func(ctx context.Context)) (err error) {
if f == nil || c.ctx.Err() != nil {
return c.ctx.Err()
}
nakeCtx := metadata.WithContext(ctx)
if tr, ok := trace.FromContext(ctx); ok {
tr = tr.Fork("", "Fanout:Do").SetTag(traceTags...)
nakeCtx = trace.NewContext(nakeCtx, tr)
}
select {
case c.ch <- item{f: f, ctx: nakeCtx}:
default:
err = ErrFull
}
stats.State(c.name+"_channel", int64(len(c.ch)))
return
}
// Close close fanout
func (c *Fanout) Close() error {
if err := c.ctx.Err(); err != nil {
return err
}
c.cancel()
c.waiter.Wait()
return nil
}

@ -0,0 +1,30 @@
package fanout
import (
"context"
"testing"
"time"
)
func TestFanout_Do(t *testing.T) {
ca := New("cache", Worker(1), Buffer(1024))
var run bool
ca.Do(context.Background(), func(c context.Context) {
run = true
panic("error")
})
time.Sleep(time.Millisecond * 50)
t.Log("not panic")
if !run {
t.Fatal("expect run be true")
}
}
func TestFanout_Close(t *testing.T) {
ca := New("cache", Worker(1), Buffer(1024))
ca.Close()
err := ca.Do(context.Background(), func(c context.Context) {})
if err == nil {
t.Fatal("expect get err")
}
}

@ -0,0 +1,185 @@
package pipeline
import (
"context"
"errors"
"sync"
"time"
"github.com/bilibili/kratos/pkg/net/metadata"
xtime "github.com/bilibili/kratos/pkg/time"
)
// ErrFull channel full error
var ErrFull = errors.New("channel full")
type message struct {
key string
value interface{}
}
// Pipeline pipeline struct
type Pipeline struct {
Do func(c context.Context, index int, values map[string][]interface{})
Split func(key string) int
chans []chan *message
mirrorChans []chan *message
config *Config
wait sync.WaitGroup
}
// Config Pipeline config
type Config struct {
// MaxSize merge size
MaxSize int
// Interval merge interval
Interval xtime.Duration
// Buffer channel size
Buffer int
// Worker channel number
Worker int
// Smooth smoothing interval
Smooth bool
}
func (c *Config) fix() {
if c.MaxSize <= 0 {
c.MaxSize = 1000
}
if c.Interval <= 0 {
c.Interval = xtime.Duration(time.Second)
}
if c.Buffer <= 0 {
c.Buffer = 1000
}
if c.Worker <= 0 {
c.Worker = 10
}
}
// NewPipeline new pipline
func NewPipeline(config *Config) (res *Pipeline) {
if config == nil {
config = &Config{}
}
config.fix()
res = &Pipeline{
chans: make([]chan *message, config.Worker),
mirrorChans: make([]chan *message, config.Worker),
config: config,
}
for i := 0; i < config.Worker; i++ {
res.chans[i] = make(chan *message, config.Buffer)
res.mirrorChans[i] = make(chan *message, config.Buffer)
}
return
}
// Start start all mergeproc
func (p *Pipeline) Start() {
if p.Do == nil {
panic("pipeline: do func is nil")
}
if p.Split == nil {
panic("pipeline: split func is nil")
}
var mirror bool
p.wait.Add(len(p.chans) + len(p.mirrorChans))
for i, ch := range p.chans {
go p.mergeproc(mirror, i, ch)
}
mirror = true
for i, ch := range p.mirrorChans {
go p.mergeproc(mirror, i, ch)
}
}
// SyncAdd sync add a value to channal, channel shard in split method
func (p *Pipeline) SyncAdd(c context.Context, key string, value interface{}) {
ch, msg := p.add(c, key, value)
ch <- msg
}
// Add async add a value to channal, channel shard in split method
func (p *Pipeline) Add(c context.Context, key string, value interface{}) (err error) {
ch, msg := p.add(c, key, value)
select {
case ch <- msg:
default:
err = ErrFull
}
return
}
func (p *Pipeline) add(c context.Context, key string, value interface{}) (ch chan *message, m *message) {
shard := p.Split(key) % p.config.Worker
if metadata.String(c, metadata.Mirror) != "" {
ch = p.mirrorChans[shard]
} else {
ch = p.chans[shard]
}
m = &message{key: key, value: value}
return
}
// Close all goroutinue
func (p *Pipeline) Close() (err error) {
for _, ch := range p.chans {
ch <- nil
}
for _, ch := range p.mirrorChans {
ch <- nil
}
p.wait.Wait()
return
}
func (p *Pipeline) mergeproc(mirror bool, index int, ch <-chan *message) {
defer p.wait.Done()
var (
m *message
vals = make(map[string][]interface{}, p.config.MaxSize)
closed bool
count int
inteval = p.config.Interval
oldTicker = true
)
if p.config.Smooth && index > 0 {
inteval = xtime.Duration(int64(index) * (int64(p.config.Interval) / int64(p.config.Worker)))
}
ticker := time.NewTicker(time.Duration(inteval))
for {
select {
case m = <-ch:
if m == nil {
closed = true
break
}
count++
vals[m.key] = append(vals[m.key], m.value)
if count >= p.config.MaxSize {
break
}
continue
case <-ticker.C:
if p.config.Smooth && oldTicker {
ticker.Stop()
ticker = time.NewTicker(time.Duration(p.config.Interval))
oldTicker = false
}
}
if len(vals) > 0 {
ctx := context.Background()
if mirror {
ctx = metadata.NewContext(ctx, metadata.MD{metadata.Mirror: "1"})
}
p.Do(ctx, index, vals)
vals = make(map[string][]interface{}, p.config.MaxSize)
count = 0
}
if closed {
ticker.Stop()
return
}
}
}

@ -0,0 +1,132 @@
package pipeline
import (
"context"
"reflect"
"strconv"
"testing"
"time"
"github.com/bilibili/kratos/pkg/net/metadata"
xtime "github.com/bilibili/kratos/pkg/time"
)
func TestPipeline(t *testing.T) {
conf := &Config{
MaxSize: 3,
Interval: xtime.Duration(time.Millisecond * 20),
Buffer: 3,
Worker: 10,
}
type recv struct {
mirror string
ch int
values map[string][]interface{}
}
var runs []recv
do := func(c context.Context, ch int, values map[string][]interface{}) {
runs = append(runs, recv{
mirror: metadata.String(c, metadata.Mirror),
values: values,
ch: ch,
})
}
split := func(s string) int {
n, _ := strconv.Atoi(s)
return n
}
p := NewPipeline(conf)
p.Do = do
p.Split = split
p.Start()
p.Add(context.Background(), "1", 1)
p.Add(context.Background(), "1", 2)
p.Add(context.Background(), "11", 3)
p.Add(context.Background(), "2", 3)
time.Sleep(time.Millisecond * 60)
mirrorCtx := metadata.NewContext(context.Background(), metadata.MD{metadata.Mirror: "1"})
p.Add(mirrorCtx, "2", 3)
time.Sleep(time.Millisecond * 60)
p.SyncAdd(mirrorCtx, "5", 5)
time.Sleep(time.Millisecond * 60)
p.Close()
expt := []recv{
{
mirror: "",
ch: 1,
values: map[string][]interface{}{
"1": {1, 2},
"11": {3},
},
},
{
mirror: "",
ch: 2,
values: map[string][]interface{}{
"2": {3},
},
},
{
mirror: "1",
ch: 2,
values: map[string][]interface{}{
"2": {3},
},
},
{
mirror: "1",
ch: 5,
values: map[string][]interface{}{
"5": {5},
},
},
}
if !reflect.DeepEqual(runs, expt) {
t.Errorf("expect get %+v,\n got: %+v", expt, runs)
}
}
func TestPipelineSmooth(t *testing.T) {
conf := &Config{
MaxSize: 100,
Interval: xtime.Duration(time.Second),
Buffer: 100,
Worker: 10,
Smooth: true,
}
type result struct {
index int
ts time.Time
}
var results []result
do := func(c context.Context, index int, values map[string][]interface{}) {
results = append(results, result{
index: index,
ts: time.Now(),
})
}
split := func(s string) int {
n, _ := strconv.Atoi(s)
return n
}
p := NewPipeline(conf)
p.Do = do
p.Split = split
p.Start()
for i := 0; i < 10; i++ {
p.Add(context.Background(), strconv.Itoa(i), 1)
}
time.Sleep(time.Millisecond * 1500)
if len(results) != conf.Worker {
t.Errorf("expect results equal worker")
t.FailNow()
}
for i, r := range results {
if i > 0 {
if r.ts.Sub(results[i-1].ts) < time.Millisecond*20 {
t.Errorf("expect runs be smooth")
t.FailNow()
}
}
}
}
Loading…
Cancel
Save