improve performance and correctness

* properly handle rebalances
* simplify code
* return on NewBroker instance and not interface

Signed-off-by: Vasiliy Tolstov <v.tolstov@unistack.org>

event.go

@@ -0,0 +1,42 @@
+package kgo
+
+import (
+	"sync"
+
+	"go.unistack.org/micro/v3/broker"
+)
+
+type event struct {
+	topic string
+	err   error
+	sync.RWMutex
+	msg *broker.Message
+	ack bool
+}
+
+func (p *event) Topic() string {
+	return p.topic
+}
+
+func (p *event) Message() *broker.Message {
+	return p.msg
+}
+
+func (p *event) Ack() error {
+	p.ack = true
+	return nil
+}
+
+func (p *event) Error() error {
+	return p.err
+}
+
+func (p *event) SetError(err error) {
+	p.err = err
+}
+
+var eventPool = sync.Pool{
+	New: func() interface{} {
+		return &event{msg: &broker.Message{}}
+	},
+}

kgo.go

@@ -3,26 +3,56 @@ package kgo // import "go.unistack.org/micro-broker-kgo/v3"
 
 import (
 	"context"
+	"errors"
 	"fmt"
+	"math/rand"
 	"strings"
 	"sync"
 	"time"
 
-	"github.com/twmb/franz-go/pkg/kerr"
+	"github.com/twmb/franz-go/pkg/kgo"
-	kgo "github.com/twmb/franz-go/pkg/kgo"
-	"github.com/twmb/franz-go/pkg/kmsg"
-	"github.com/twmb/franz-go/pkg/kversion"
 	"go.unistack.org/micro/v3/broker"
-	"go.unistack.org/micro/v3/logger"
 	"go.unistack.org/micro/v3/metadata"
-	"go.unistack.org/micro/v3/util/id"
+	id "go.unistack.org/micro/v3/util/id"
 	mrand "go.unistack.org/micro/v3/util/rand"
 )
 
-var _ broker.Broker = &kBroker{}
+var _ broker.Broker = &Broker{}
 
-type kBroker struct {
+var ErrLostMessage = errors.New("message not marked for offsets commit and will be lost in next iteration")
-	writer    *kgo.Client // used only to push messages
+
+var DefaultRetryBackoffFn = func() func(int) time.Duration {
+	var rngMu sync.Mutex
+	rng := rand.New(rand.NewSource(time.Now().UnixNano()))
+	return func(fails int) time.Duration {
+		const (
+			min = 100 * time.Millisecond
+			max = time.Second
+		)
+		if fails <= 0 {
+			return min
+		}
+		if fails > 10 {
+			return max
+		}
+
+		backoff := min * time.Duration(1<<(fails-1))
+
+		rngMu.Lock()
+		jitter := 0.8 + 0.4*rng.Float64()
+		rngMu.Unlock()
+
+		backoff = time.Duration(float64(backoff) * jitter)
+
+		if backoff > max {
+			return max
+		}
+		return backoff
+	}
+}()
+
+type Broker struct {
+	c         *kgo.Client
 	kopts     []kgo.Opt
 	connected bool
 	init      bool
@@ -31,79 +61,34 @@ type kBroker struct {
 	subs []*subscriber
 }
 
-type subscriber struct {
+func (k *Broker) Address() string {
-	reader       *kgo.Client // used only to pull messages
-	topic        string
-	opts         broker.SubscribeOptions
-	kopts        broker.Options
-	handler      broker.Handler
-	batchhandler broker.BatchHandler
-	closed       bool
-	done         chan struct{}
-	consumers    map[string]map[int32]worker
-	sync.RWMutex
-}
-
-type publication struct {
-	topic string
-	err   error
-	sync.RWMutex
-	msg *broker.Message
-	ack bool
-}
-
-func (p *publication) Topic() string {
-	return p.topic
-}
-
-func (p *publication) Message() *broker.Message {
-	return p.msg
-}
-
-func (p *publication) Ack() error {
-	p.ack = true
-	return nil
-}
-
-func (p *publication) Error() error {
-	return p.err
-}
-
-func (p *publication) SetError(err error) {
-	p.err = err
-}
-
-func (s *subscriber) Options() broker.SubscribeOptions {
-	return s.opts
-}
-
-func (s *subscriber) Topic() string {
-	return s.topic
-}
-
-func (s *subscriber) Unsubscribe(ctx context.Context) error {
-	if s.closed {
-		return nil
-	}
-	select {
-	case <-ctx.Done():
-		return ctx.Err()
-	default:
-		close(s.done)
-		s.closed = true
-	}
-	return nil
-}
-
-func (k *kBroker) Address() string {
 	return strings.Join(k.opts.Addrs, ",")
 }
 
-func (k *kBroker) Name() string {
+func (k *Broker) Name() string {
 	return k.opts.Name
 }
 
-func (k *kBroker) Connect(ctx context.Context) error {
+func (k *Broker) connect(ctx context.Context, opts ...kgo.Opt) (*kgo.Client, error) {
+	var c *kgo.Client
+	var err error
+
+	select {
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	default:
+		c, err = kgo.NewClient(opts...)
+		if err == nil {
+			err = c.Ping(ctx) // check connectivity to cluster
+		}
+		if err != nil {
+			return nil, err
+		}
+	}
+	return c, nil
+}
+
+func (k *Broker) Connect(ctx context.Context) error {
 	k.RLock()
 	if k.connected {
 		k.RUnlock()
@@ -116,51 +101,20 @@ func (k *kBroker) Connect(ctx context.Context) error {
 		nctx = ctx
 	}
 
-	kaddrs := k.opts.Addrs
+	c, err := k.connect(nctx, k.kopts...)
-
-	// shuffle addrs
-	var rng mrand.Rand
-	rng.Shuffle(len(kaddrs), func(i, j int) {
-		kaddrs[i], kaddrs[j] = kaddrs[j], kaddrs[i]
-	})
-
-	kopts := append(k.kopts, kgo.SeedBrokers(kaddrs...))
-
-	select {
-	case <-nctx.Done():
-		return nctx.Err()
-	default:
-		c, err := kgo.NewClient(kopts...)
 	if err != nil {
 		return err
 	}
 
-		// Request versions in order to guess Kafka Cluster version
-		versionsReq := kmsg.NewApiVersionsRequest()
-		versionsRes, err := versionsReq.RequestWith(ctx, c)
-		if err != nil {
-			return fmt.Errorf("failed to request api versions: %w", err)
-		}
-		err = kerr.ErrorForCode(versionsRes.ErrorCode)
-		if err != nil {
-			return fmt.Errorf("failed to request api versions. Inner kafka error: %w", err)
-		}
-		versions := kversion.FromApiVersionsResponse(versionsRes)
-
-		if k.opts.Logger.V(logger.InfoLevel) {
-			logger.Infof(ctx, "[kgo] connected to to kafka cluster version %v", versions.VersionGuess())
-		}
-
 	k.Lock()
+	k.c = c
 	k.connected = true
-		k.writer = c
 	k.Unlock()
-	}
 
 	return nil
 }
 
-func (k *kBroker) Disconnect(ctx context.Context) error {
+func (k *Broker) Disconnect(ctx context.Context) error {
 	k.RLock()
 	if !k.connected {
 		k.RUnlock()
@@ -168,14 +122,13 @@ func (k *kBroker) Disconnect(ctx context.Context) error {
 	}
 	k.RUnlock()
 
-	k.Lock()
-	defer k.Unlock()
-
 	nctx := k.opts.Context
 	if ctx != nil {
 		nctx = ctx
 	}
 
+	k.Lock()
+	defer k.Unlock()
 	select {
 	case <-nctx.Done():
 		return nctx.Err()
@@ -185,20 +138,20 @@ func (k *kBroker) Disconnect(ctx context.Context) error {
 				return err
 			}
 		}
-		k.writer.Close()
+		if k.c != nil {
+			k.c.CloseAllowingRebalance()
+			// k.c.Close()
+		}
 	}
 
 	k.connected = false
 	return nil
 }
 
-func (k *kBroker) Init(opts ...broker.Option) error {
+func (k *Broker) Init(opts ...broker.Option) error {
 	k.Lock()
 	defer k.Unlock()
 
-	if len(opts) == 0 && k.init {
-		return nil
-	}
 	for _, o := range opts {
 		o(&k.opts)
 	}
@@ -222,33 +175,38 @@ func (k *kBroker) Init(opts ...broker.Option) error {
 		}
 	}
 
-	//	kgo.RecordPartitioner(),
-
-	k.init = true
-
 	return nil
 }
 
-func (k *kBroker) Options() broker.Options {
+func (k *Broker) Options() broker.Options {
 	return k.opts
 }
 
-func (k *kBroker) BatchPublish(ctx context.Context, msgs []*broker.Message, opts ...broker.PublishOption) error {
+func (k *Broker) BatchPublish(ctx context.Context, msgs []*broker.Message, opts ...broker.PublishOption) error {
 	return k.publish(ctx, msgs, opts...)
 }
 
-func (k *kBroker) Publish(ctx context.Context, topic string, msg *broker.Message, opts ...broker.PublishOption) error {
+func (k *Broker) Publish(ctx context.Context, topic string, msg *broker.Message, opts ...broker.PublishOption) error {
 	msg.Header.Set(metadata.HeaderTopic, topic)
 	return k.publish(ctx, []*broker.Message{msg}, opts...)
 }
 
-func (k *kBroker) publish(ctx context.Context, msgs []*broker.Message, opts ...broker.PublishOption) error {
+func (k *Broker) publish(ctx context.Context, msgs []*broker.Message, opts ...broker.PublishOption) error {
 	k.RLock()
 	if !k.connected {
 		k.RUnlock()
-		return broker.ErrNotConnected
+		k.Lock()
+		c, err := k.connect(ctx, k.kopts...)
+		if err != nil {
+			k.Unlock()
+			return err
+		}
+		k.c = c
+		k.connected = true
+		k.Unlock()
 	}
 	k.RUnlock()
+
 	options := broker.NewPublishOptions(opts...)
 	records := make([]*kgo.Record, 0, len(msgs))
 	var errs []string
@@ -264,13 +222,13 @@ func (k *kBroker) publish(ctx context.Context, msgs []*broker.Message, opts ...b
 	for _, msg := range msgs {
 		rec := &kgo.Record{Key: key}
 		rec.Topic, _ = msg.Header.Get(metadata.HeaderTopic)
-		if k.opts.Codec.String() == "noop" {
+		if options.BodyOnly {
+			rec.Value = msg.Body
+		} else if k.opts.Codec.String() == "noop" {
 			rec.Value = msg.Body
 			for k, v := range msg.Header {
 				rec.Headers = append(rec.Headers, kgo.RecordHeader{Key: k, Value: []byte(v)})
 			}
-		} else if options.BodyOnly {
-			rec.Value = msg.Body
 		} else {
 			rec.Value, err = k.opts.Codec.Marshal(msg)
 			if err != nil {
@@ -280,7 +238,7 @@ func (k *kBroker) publish(ctx context.Context, msgs []*broker.Message, opts ...b
 		records = append(records, rec)
 	}
 
-	results := k.writer.ProduceSync(ctx, records...)
+	results := k.c.ProduceSync(ctx, records...)
 	for _, result := range results {
 		if result.Err != nil {
 			errs = append(errs, result.Err.Error())
@@ -294,53 +252,11 @@ func (k *kBroker) publish(ctx context.Context, msgs []*broker.Message, opts ...b
 	return nil
 }
 
-type mlogger struct {
+func (k *Broker) BatchSubscribe(ctx context.Context, topic string, handler broker.BatchHandler, opts ...broker.SubscribeOption) (broker.Subscriber, error) {
-	l   logger.Logger
-	ctx context.Context
-}
-
-func (l *mlogger) Log(lvl kgo.LogLevel, msg string, args ...interface{}) {
-	var mlvl logger.Level
-	switch lvl {
-	case kgo.LogLevelNone:
-		return
-	case kgo.LogLevelError:
-		mlvl = logger.ErrorLevel
-	case kgo.LogLevelWarn:
-		mlvl = logger.WarnLevel
-	case kgo.LogLevelInfo:
-		mlvl = logger.InfoLevel
-	case kgo.LogLevelDebug:
-		mlvl = logger.DebugLevel
-	default:
-		return
-	}
-	fields := make(map[string]interface{}, int(len(args)/2))
-	for i := 0; i < len(args)/2; i += 2 {
-		fields[fmt.Sprintf("%v", args[i])] = args[i+1]
-	}
-	l.l.Fields(fields).Log(l.ctx, mlvl, msg)
-}
-
-func (l *mlogger) Level() kgo.LogLevel {
-	switch l.l.Options().Level {
-	case logger.ErrorLevel:
-		return kgo.LogLevelError
-	case logger.WarnLevel:
-		return kgo.LogLevelWarn
-	case logger.InfoLevel:
-		return kgo.LogLevelInfo
-	case logger.DebugLevel, logger.TraceLevel:
-		return kgo.LogLevelDebug
-	}
-	return kgo.LogLevelNone
-}
-
-func (k *kBroker) BatchSubscribe(ctx context.Context, topic string, handler broker.BatchHandler, opts ...broker.SubscribeOption) (broker.Subscriber, error) {
 	return nil, nil
 }
 
-func (k *kBroker) Subscribe(ctx context.Context, topic string, handler broker.Handler, opts ...broker.SubscribeOption) (broker.Subscriber, error) {
+func (k *Broker) Subscribe(ctx context.Context, topic string, handler broker.Handler, opts ...broker.SubscribeOption) (broker.Subscriber, error) {
 	options := broker.NewSubscribeOptions(opts...)
 
 	if options.Group == "" {
@@ -351,54 +267,47 @@ func (k *kBroker) Subscribe(ctx context.Context, topic string, handler broker.Ha
 		options.Group = uid
 	}
 
-	kaddrs := k.opts.Addrs
+	commitInterval := DefaultCommitInterval
-
-	// shuffle addrs
-	var rng mrand.Rand
-	rng.Shuffle(len(kaddrs), func(i, j int) {
-		kaddrs[i], kaddrs[j] = kaddrs[j], kaddrs[i]
-	})
-
-	td := DefaultCommitInterval
 	if k.opts.Context != nil {
 		if v, ok := k.opts.Context.Value(commitIntervalKey{}).(time.Duration); ok && v > 0 {
-			td = v
+			commitInterval = v
 		}
 	}
 
 	sub := &subscriber{
 		topic:     topic,
-		done:      make(chan struct{}),
 		opts:      options,
 		handler:   handler,
 		kopts:     k.opts,
-		consumers: make(map[string]map[int32]worker),
+		consumers: make(map[tp]*consumer),
+		done:      make(chan struct{}),
 	}
 
 	kopts := append(k.kopts,
-		kgo.SeedBrokers(kaddrs...),
 		kgo.ConsumerGroup(options.Group),
 		kgo.ConsumeTopics(topic),
 		kgo.ConsumeResetOffset(kgo.NewOffset().AtStart()),
 		kgo.FetchMaxWait(1*time.Second),
-		// kgo.KeepControlRecords(),
+		kgo.AutoCommitInterval(commitInterval),
-		kgo.Balancers(kgo.CooperativeStickyBalancer(), kgo.StickyBalancer()),
-		kgo.FetchIsolationLevel(kgo.ReadUncommitted()),
-		kgo.WithHooks(&metrics{meter: k.opts.Meter}),
-		kgo.AutoCommitMarks(),
-		kgo.AutoCommitInterval(td),
 		kgo.OnPartitionsAssigned(sub.assigned),
 		kgo.OnPartitionsRevoked(sub.revoked),
-		kgo.OnPartitionsLost(sub.revoked),
+		kgo.OnPartitionsLost(sub.lost),
+		kgo.AutoCommitMarks(),
 	)
 
-	reader, err := kgo.NewClient(kopts...)
+	if options.Context != nil {
+		if v, ok := options.Context.Value(optionsKey{}).([]kgo.Opt); ok && len(v) > 0 {
+			kopts = append(kopts, v...)
+		}
+	}
+
+	c, err := k.connect(ctx, kopts...)
 	if err != nil {
 		return nil, err
 	}
 
-	sub.reader = reader
+	sub.c = c
-	go sub.run(ctx)
+	go sub.poll(ctx)
 
 	k.Lock()
 	k.subs = append(k.subs, sub)
@@ -406,45 +315,32 @@ func (k *kBroker) Subscribe(ctx context.Context, topic string, handler broker.Ha
 	return sub, nil
 }
 
-func (k *kBroker) String() string {
+func (k *Broker) String() string {
 	return "kgo"
 }
 
-func NewBroker(opts ...broker.Option) *kBroker {
+func NewBroker(opts ...broker.Option) *Broker {
+	rand.Seed(time.Now().Unix())
 	options := broker.NewOptions(opts...)
-	if options.Codec.String() != "noop" {
+
-		options.Logger.Infof(options.Context, "broker codec not noop, disable plain kafka headers usage")
+	kaddrs := options.Addrs
-	}
+	// shuffle addrs
+	var rng mrand.Rand
+	rng.Shuffle(len(kaddrs), func(i, j int) {
+		kaddrs[i], kaddrs[j] = kaddrs[j], kaddrs[i]
+	})
 	kopts := []kgo.Opt{
+		kgo.DialTimeout(3 * time.Second),
 		kgo.DisableIdempotentWrite(),
 		kgo.ProducerBatchCompression(kgo.NoCompression()),
 		kgo.WithLogger(&mlogger{l: options.Logger, ctx: options.Context}),
-		kgo.RetryBackoffFn(
+		// kgo.WithLogger(kgo.BasicLogger(os.Stderr, kgo.LogLevelDebug, func() string { return time.Now().Format(time.StampMilli) })),
-			func() func(int) time.Duration {
+		kgo.WithHooks(&metrics{meter: options.Meter}),
-				var rng mrand.Rand
+		kgo.SeedBrokers(kaddrs...),
-				return func(fails int) time.Duration {
+		kgo.RetryBackoffFn(DefaultRetryBackoffFn),
-					const (
+		kgo.BlockRebalanceOnPoll(),
-						min = 250 * time.Millisecond
+		kgo.Balancers(kgo.CooperativeStickyBalancer()),
-						max = 2 * time.Second
+		kgo.FetchIsolationLevel(kgo.ReadUncommitted()),
-					)
-					if fails <= 0 {
-						return min
-					}
-					if fails > 10 {
-						return max
-					}
-
-					backoff := min * time.Duration(1<<(fails-1))
-					jitter := 0.8 + 0.4*rng.Float64()
-					backoff = time.Duration(float64(backoff) * jitter)
-
-					if backoff > max {
-						return max
-					}
-					return backoff
-				}
-			}(),
-		),
 	}
 
 	if options.Context != nil {
@@ -453,7 +349,7 @@ func NewBroker(opts ...broker.Option) *kBroker {
 		}
 	}
 
-	return &kBroker{
+	return &Broker{
 		opts:  options,
 		kopts: kopts,
 	}

kgo_test.go

@@ -28,6 +28,23 @@ var bm = &broker.Message{
 	Body:   []byte(`"body"`),
 }
 
+func TestConnect(t *testing.T) {
+	var addrs []string
+	ctx := context.TODO()
+	b := kgo.NewBroker(
+		broker.Addrs(addrs...),
+		kgo.CommitInterval(5*time.Second),
+		kgo.Options(kg.ClientID("test"), kg.FetchMaxBytes(10*1024*1024)),
+	)
+	if err := b.Init(); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := b.Connect(ctx); err != nil {
+		t.Fatal(err)
+	}
+}
+
 func TestPubSub(t *testing.T) {
 	if tr := os.Getenv("INTEGRATION_TESTS"); len(tr) > 0 {
 		t.Skip()

logger.go

@@ -0,0 +1,55 @@
+package kgo
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/twmb/franz-go/pkg/kgo"
+	"go.unistack.org/micro/v3/logger"
+)
+
+type mlogger struct {
+	l   logger.Logger
+	ctx context.Context
+}
+
+func (l *mlogger) Log(lvl kgo.LogLevel, msg string, args ...interface{}) {
+	var mlvl logger.Level
+	switch lvl {
+	case kgo.LogLevelNone:
+		return
+	case kgo.LogLevelError:
+		mlvl = logger.ErrorLevel
+	case kgo.LogLevelWarn:
+		mlvl = logger.WarnLevel
+	case kgo.LogLevelInfo:
+		mlvl = logger.InfoLevel
+	case kgo.LogLevelDebug:
+		mlvl = logger.DebugLevel
+	default:
+		return
+	}
+	if len(args) > 0 {
+		fields := make(map[string]interface{}, int(len(args)/2))
+		for i := 0; i <= len(args)/2; i += 2 {
+			fields[fmt.Sprintf("%v", args[i])] = args[i+1]
+		}
+		l.l.Fields(fields).Log(l.ctx, mlvl, msg)
+	} else {
+		l.l.Log(l.ctx, mlvl, msg)
+	}
+}
+
+func (l *mlogger) Level() kgo.LogLevel {
+	switch l.l.Options().Level {
+	case logger.ErrorLevel:
+		return kgo.LogLevelError
+	case logger.WarnLevel:
+		return kgo.LogLevelWarn
+	case logger.InfoLevel:
+		return kgo.LogLevelInfo
+	case logger.DebugLevel, logger.TraceLevel:
+		return kgo.LogLevelDebug
+	}
+	return kgo.LogLevelNone
+}

options.go

@@ -4,10 +4,9 @@ import (
 	"context"
 	"time"
 
-	kgo "github.com/twmb/franz-go/pkg/kgo"
+	"github.com/twmb/franz-go/pkg/kgo"
 	"go.unistack.org/micro/v3/broker"
 	"go.unistack.org/micro/v3/client"
-	"go.unistack.org/micro/v3/server"
 )
 
 // DefaultCommitInterval specifies how fast send commit offsets to kafka
@@ -49,7 +48,7 @@ func Options(opts ...kgo.Opt) broker.Option {
 	}
 }
 
-// SubscribeOptions pass additional options to broker
+// SubscribeOptions pass additional options to broker in Subscribe
 func SubscribeOptions(opts ...kgo.Opt) broker.SubscribeOption {
 	return func(o *broker.SubscribeOptions) {
 		if o.Context == nil {
@@ -64,21 +63,6 @@ func SubscribeOptions(opts ...kgo.Opt) broker.SubscribeOption {
 	}
 }
 
-// SubscriberOptions pass additional options to broker
-func SubscriberOptions(opts ...kgo.Opt) server.SubscriberOption {
-	return func(o *server.SubscriberOptions) {
-		if o.Context == nil {
-			o.Context = context.Background()
-		}
-		options, ok := o.Context.Value(optionsKey{}).([]kgo.Opt)
-		if !ok {
-			options = make([]kgo.Opt, 0, len(opts))
-		}
-		options = append(options, opts...)
-		o.Context = context.WithValue(o.Context, optionsKey{}, options)
-	}
-}
-
 type commitIntervalKey struct{}
 
 // CommitInterval specifies interval to send commits
@@ -86,7 +70,7 @@ func CommitInterval(td time.Duration) broker.Option {
 	return broker.SetOption(commitIntervalKey{}, td)
 }
 
-var DefaultSubscribeMaxInflight = 1000
+var DefaultSubscribeMaxInflight = 10
 
 type subscribeMaxInflightKey struct{}
 

subscriber.go

@@ -0,0 +1,229 @@
+package kgo
+
+import (
+	"context"
+	"sync"
+
+	"github.com/twmb/franz-go/pkg/kgo"
+	"go.unistack.org/micro/v3/broker"
+	"go.unistack.org/micro/v3/logger"
+	"go.unistack.org/micro/v3/metadata"
+)
+
+type tp struct {
+	t string
+	p int32
+}
+
+type consumer struct {
+	c         *kgo.Client
+	topic     string
+	partition int32
+	opts      broker.SubscribeOptions
+	kopts     broker.Options
+	handler   broker.Handler
+	quit      chan struct{}
+	done      chan struct{}
+	recs      chan kgo.FetchTopicPartition
+}
+
+type subscriber struct {
+	c         *kgo.Client
+	topic     string
+	opts      broker.SubscribeOptions
+	kopts     broker.Options
+	handler   broker.Handler
+	closed    bool
+	done      chan struct{}
+	consumers map[tp]*consumer
+	sync.RWMutex
+}
+
+func (s *subscriber) Options() broker.SubscribeOptions {
+	return s.opts
+}
+
+func (s *subscriber) Topic() string {
+	return s.topic
+}
+
+func (s *subscriber) Unsubscribe(ctx context.Context) error {
+	if s.closed {
+		return nil
+	}
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	default:
+		close(s.done)
+		s.closed = true
+	}
+	return nil
+}
+
+func (s *subscriber) poll(ctx context.Context) {
+	maxInflight := DefaultSubscribeMaxInflight
+	if s.opts.Context != nil {
+		if n, ok := s.opts.Context.Value(subscribeMaxInflightKey{}).(int); n > 0 && ok {
+			maxInflight = n
+		}
+	}
+	for {
+		select {
+		case <-ctx.Done():
+			s.c.Close()
+			return
+		case <-s.done:
+			s.c.Close()
+			return
+		default:
+			fetches := s.c.PollRecords(ctx, maxInflight)
+			if fetches.IsClientClosed() {
+				s.kopts.Logger.Errorf(ctx, "[kgo] client closed")
+				s.closed = true
+				return
+			}
+			fetches.EachError(func(t string, p int32, err error) {
+				s.kopts.Logger.Fatalf(ctx, "[kgo] fetch topic %s partition %d err: %v", t, p, err)
+			})
+
+			fetches.EachPartition(func(p kgo.FetchTopicPartition) {
+				tp := tp{p.Topic, p.Partition}
+				s.consumers[tp].recs <- p
+			})
+			s.c.AllowRebalance()
+		}
+	}
+}
+
+func (s *subscriber) killConsumers(ctx context.Context, lost map[string][]int32) {
+	var wg sync.WaitGroup
+	defer wg.Wait()
+
+	for topic, partitions := range lost {
+		for _, partition := range partitions {
+			tp := tp{topic, partition}
+			pc := s.consumers[tp]
+			delete(s.consumers, tp)
+			close(pc.quit)
+			s.kopts.Logger.Debugf(ctx, "[kgo] waiting for work to finish topic %s partition %d", topic, partition)
+			wg.Add(1)
+			go func() { <-pc.done; wg.Done() }()
+		}
+	}
+}
+
+func (s *subscriber) lost(ctx context.Context, _ *kgo.Client, lost map[string][]int32) {
+	s.kopts.Logger.Debugf(ctx, "[kgo] lost %#+v", lost)
+	s.killConsumers(ctx, lost)
+}
+
+func (s *subscriber) revoked(ctx context.Context, c *kgo.Client, revoked map[string][]int32) {
+	s.kopts.Logger.Debugf(ctx, "[kgo] revoked %#+v", revoked)
+	s.killConsumers(ctx, revoked)
+	if err := c.CommitMarkedOffsets(ctx); err != nil {
+		s.kopts.Logger.Errorf(ctx, "[kgo] revoked CommitMarkedOffsets err: %v", err)
+	}
+}
+
+func (s *subscriber) assigned(_ context.Context, c *kgo.Client, assigned map[string][]int32) {
+	for topic, partitions := range assigned {
+		for _, partition := range partitions {
+			pc := &consumer{
+				c:         c,
+				topic:     topic,
+				partition: partition,
+
+				quit:    make(chan struct{}),
+				done:    make(chan struct{}),
+				recs:    make(chan kgo.FetchTopicPartition, 4),
+				handler: s.handler,
+				kopts:   s.kopts,
+				opts:    s.opts,
+			}
+			s.consumers[tp{topic, partition}] = pc
+			go pc.consume()
+		}
+	}
+}
+
+func (pc *consumer) consume() {
+	defer close(pc.done)
+	pc.kopts.Logger.Debugf(pc.kopts.Context, "starting, topic %s partition %d", pc.topic, pc.partition)
+	defer pc.kopts.Logger.Debugf(pc.kopts.Context, "killing, topic %s partition %d", pc.topic, pc.partition)
+
+	eh := pc.kopts.ErrorHandler
+	if pc.opts.ErrorHandler != nil {
+		eh = pc.opts.ErrorHandler
+	}
+
+	for {
+		select {
+		case <-pc.quit:
+			return
+		case p := <-pc.recs:
+			for _, record := range p.Records {
+				p := eventPool.Get().(*event)
+				p.msg.Header = nil
+				p.msg.Body = nil
+				p.topic = record.Topic
+				p.err = nil
+				p.ack = false
+				if pc.kopts.Codec.String() == "noop" {
+					p.msg.Header = metadata.New(len(record.Headers))
+					for _, hdr := range record.Headers {
+						p.msg.Header.Set(hdr.Key, string(hdr.Value))
+					}
+					p.msg.Body = record.Value
+				} else if pc.opts.BodyOnly {
+					p.msg.Body = record.Value
+				} else {
+					if err := pc.kopts.Codec.Unmarshal(record.Value, p.msg); err != nil {
+						p.err = err
+						p.msg.Body = record.Value
+						if eh != nil {
+							_ = eh(p)
+							if p.ack {
+								pc.c.MarkCommitRecords(record)
+							} else {
+								eventPool.Put(p)
+								pc.kopts.Logger.Infof(pc.kopts.Context, "[kgo] ErrLostMessage wtf?")
+								return
+							}
+							eventPool.Put(p)
+							continue
+						} else {
+							if pc.kopts.Logger.V(logger.ErrorLevel) {
+								pc.kopts.Logger.Errorf(pc.kopts.Context, "[kgo]: failed to unmarshal: %v", err)
+							}
+						}
+						eventPool.Put(p)
+						pc.kopts.Logger.Infof(pc.kopts.Context, "[kgo] Unmarshal err not handled wtf?")
+						return
+					}
+				}
+				err := pc.handler(p)
+				if err == nil && pc.opts.AutoAck {
+					p.ack = true
+				} else if err != nil {
+					p.err = err
+					if eh != nil {
+						_ = eh(p)
+					} else {
+						if pc.kopts.Logger.V(logger.ErrorLevel) {
+							pc.kopts.Logger.Errorf(pc.kopts.Context, "[kgo]: subscriber error: %v", err)
+						}
+					}
+				}
+				if p.ack {
+					eventPool.Put(p)
+					pc.c.MarkCommitRecords(record)
+				} else {
+					eventPool.Put(p)
+					pc.kopts.Logger.Fatalf(pc.kopts.Context, "[kgo] ErrLostMessage wtf?")
+					return
+				}
+			}
+		}
+	}
+}

util.go

@@ -1,244 +0,0 @@
-package kgo
-
-import (
-	"context"
-	"errors"
-	"sync"
-
-	kgo "github.com/twmb/franz-go/pkg/kgo"
-	"go.unistack.org/micro/v3/broker"
-	"go.unistack.org/micro/v3/logger"
-	"go.unistack.org/micro/v3/metadata"
-)
-
-var ErrLostMessage = errors.New("message not marked for offsets commit and will be lost in next iteration")
-
-var pPool = sync.Pool{
-	New: func() interface{} {
-		return &publication{msg: broker.NewMessage("")}
-	},
-}
-
-type worker struct {
-	done         chan struct{}
-	recs         chan []*kgo.Record
-	cherr        chan error
-	handler      broker.Handler
-	batchHandler broker.BatchHandler
-	opts         broker.SubscribeOptions
-	kopts        broker.Options
-	tpmap        map[string][]int32
-	maxInflight  int
-	reader       *kgo.Client
-	ctx          context.Context
-}
-
-func (s *subscriber) run(ctx context.Context) {
-	for {
-		select {
-		case <-ctx.Done():
-			return
-		case <-s.kopts.Context.Done():
-			return
-		default:
-			fetches := s.reader.PollFetches(ctx)
-			if fetches.IsClientClosed() {
-				// TODO: fatal ?
-				return
-			}
-			if len(fetches.Errors()) > 0 {
-				for _, err := range fetches.Errors() {
-					s.kopts.Logger.Fatalf(ctx, "fetch err topic %s partition %d: %v", err.Topic, err.Partition, err.Err)
-				}
-				// TODO: fatal ?
-				return
-			}
-
-			fetches.EachPartition(func(p kgo.FetchTopicPartition) {
-				s.Lock()
-				consumers := s.consumers[p.Topic]
-				s.Unlock()
-				if consumers == nil {
-					return
-				}
-				w, ok := consumers[p.Partition]
-				if !ok {
-					return
-				}
-				select {
-				case err := <-w.cherr:
-					s.kopts.Logger.Fatalf(ctx, "handle err: %v", err)
-					return
-				case w.recs <- p.Records:
-				case <-w.done:
-				}
-			})
-		}
-	}
-}
-
-func (s *subscriber) assigned(ctx context.Context, _ *kgo.Client, assigned map[string][]int32) {
-	maxInflight := DefaultSubscribeMaxInflight
-
-	if s.opts.Context != nil {
-		if n, ok := s.opts.Context.Value(subscribeMaxInflightKey{}).(int); n > 0 && ok {
-			maxInflight = n
-		}
-	}
-
-	s.Lock()
-	for topic, partitions := range assigned {
-		if s.consumers[topic] == nil {
-			s.consumers[topic] = make(map[int32]worker)
-		}
-		for _, partition := range partitions {
-			w := worker{
-				done:         make(chan struct{}),
-				recs:         make(chan []*kgo.Record),
-				cherr:        make(chan error),
-				kopts:        s.kopts,
-				opts:         s.opts,
-				ctx:          ctx,
-				tpmap:        map[string][]int32{topic: []int32{partition}},
-				reader:       s.reader,
-				handler:      s.handler,
-				batchHandler: s.batchhandler,
-				maxInflight:  maxInflight,
-			}
-			s.consumers[topic][partition] = w
-			go w.handle()
-		}
-	}
-	s.Unlock()
-}
-
-func (s *subscriber) revoked(_ context.Context, _ *kgo.Client, revoked map[string][]int32) {
-	s.Lock()
-	for topic, partitions := range revoked {
-		ptopics := s.consumers[topic]
-		for _, partition := range partitions {
-			w := ptopics[partition]
-			delete(ptopics, partition)
-			if len(ptopics) == 0 {
-				delete(s.consumers, topic)
-			}
-			close(w.done)
-		}
-	}
-	s.Unlock()
-}
-
-func (w *worker) handle() {
-	var err error
-
-	eh := w.kopts.ErrorHandler
-	if w.opts.ErrorHandler != nil {
-		eh = w.opts.ErrorHandler
-	}
-
-	paused := false
-	for {
-		select {
-		case <-w.ctx.Done():
-			w.cherr <- w.ctx.Err()
-			return
-		case <-w.done:
-			return
-		case recs := <-w.recs:
-			if len(recs) >= w.maxInflight {
-				paused = true
-				w.reader.PauseFetchPartitions(w.tpmap)
-			}
-			for _, record := range recs {
-				p := pPool.Get().(*publication)
-				p.msg.Header = nil
-				p.msg.Body = nil
-				p.topic = record.Topic
-				p.err = nil
-				p.ack = false
-				if w.opts.BodyOnly {
-					p.msg.Body = record.Value
-					if l := len(record.Headers); l > 0 {
-						if p.msg.Header == nil {
-							p.msg.Header = metadata.New(l)
-						}
-						for _, h := range record.Headers {
-							p.msg.Header.Set(h.Key, string(h.Value))
-						}
-					}
-				} else if w.kopts.Codec.String() == "noop" {
-					p.msg.Body = record.Value
-					p.msg.Header = metadata.New(len(record.Headers))
-					for _, h := range record.Headers {
-						p.msg.Header.Set(h.Key, string(h.Value))
-					}
-				} else {
-					if err := w.kopts.Codec.Unmarshal(record.Value, p.msg); err != nil {
-						p.err = err
-						p.msg.Body = record.Value
-						if l := len(record.Headers); l > 0 {
-							if p.msg.Header == nil {
-								p.msg.Header = metadata.New(l)
-							}
-							for _, h := range record.Headers {
-								p.msg.Header.Set(h.Key, string(h.Value))
-							}
-						}
-						if eh != nil {
-							_ = eh(p)
-							if p.ack {
-								w.reader.MarkCommitRecords(record)
-							} else {
-								w.cherr <- ErrLostMessage
-								pPool.Put(p)
-								return
-							}
-							pPool.Put(p)
-							continue
-						} else {
-							if w.kopts.Logger.V(logger.ErrorLevel) {
-								w.kopts.Logger.Errorf(w.kopts.Context, "[kgo]: failed to unmarshal: %v", err)
-							}
-						}
-						pPool.Put(p)
-						w.cherr <- err
-						return
-					}
-					if l := len(record.Headers); l > 0 {
-						if p.msg.Header == nil {
-							p.msg.Header = metadata.New(l)
-						}
-						for _, h := range record.Headers {
-							p.msg.Header.Set(h.Key, string(h.Value))
-						}
-					}
-				}
-				err = w.handler(p)
-				if err == nil && w.opts.AutoAck {
-					p.ack = true
-				} else if err != nil {
-					p.err = err
-					if eh != nil {
-						_ = eh(p)
-					} else {
-						if w.kopts.Logger.V(logger.ErrorLevel) {
-							w.kopts.Logger.Errorf(w.kopts.Context, "[kgo]: subscriber error: %v", err)
-						}
-					}
-				}
-				if p.ack {
-					pPool.Put(p)
-					w.reader.MarkCommitRecords(record)
-				} else {
-					pPool.Put(p)
-					w.cherr <- ErrLostMessage
-					return
-				}
-			}
-			if paused {
-				paused = false
-				w.reader.ResumeFetchPartitions(w.tpmap)
-			}
-		}
-	}
-}