Improve ackManager.completeTask performance by two orders of magnitude

[tdeebswihart]

What changed?

I replaced the outstandingTasks map with an ordered treemap and optimized
completeTask to only scan what was necessary to update the ack level.

Why?

The old implementation of completeTask required a full scan of the task
map in order to move the ack level which had terrible performance.
By storing tasks in an ordered set we can limit the scan's size by
stopping at the first unacked task.

This trades addTask performance for completeTask performance but since all
added tasks are presumably completed we should be fine with 1/3 the
performance on addTask for 227x the completeTask performance. With this
change both operations run in about the same amount of time.

Before:

$ go test -bench=AckManager ./service/matching/... -run=FooBarBaz
goos: darwin
goarch: arm64
pkg: go.temporal.io/server/service/matching
BenchmarkAckManager_AddTask-12         	  22768	    52206 ns/op
BenchmarkAckManager_CompleteTask-12    	     38	 29293019 ns/op

After:

$ go test -bench=AckManager ./service/matching -run=FooBarBaz
goos: darwin
goarch: arm64
pkg: go.temporal.io/server/service/matching
BenchmarkAckManager_AddTask-12         	   8127	   147226 ns/op
BenchmarkAckManager_CompleteTask-12    	   8626	   136614 ns/op

How did you test it?

I added both tests and benchmarks to ensure the ackManager worked as before

Potential risks

None.

Is hotfix candidate?

No

[dnr]

Not to be too pedantic since it's already a huge improvement, but I'm curious how the performance would compare to using https://pkg.go.dev/github.com/google/btree, specifically the ordered generic version (NewOrderedG). That's the ordered map for Go I've used a few times in the past

[dnr]

Oh, I see, you picked that one since it's already used in the test. Well, I'm still curious. If it's faster maybe we can change both

[tdeebswihart]

Not to be too pedantic since it's already a huge improvement, but I'm curious how the performance would compare to using https://pkg.go.dev/github.com/google/btree, specifically the ordered generic version (NewOrderedG). That's the ordered map for Go I've used a few times in the past

Looks like this doesn't allow us to store values alongside the keys, hrm. I'll build it with this alongside a map of K->V to see how it performs

[dnr]

Looks like this doesn't allow us to store values alongside the keys, hrm. I'll build it with this alongside a map of K->V to see how it performs

Oh right, I forgot you need the bool. You can totally store values next to the keys, but yeah, you can't use the Ordered one in that case, you need to make a struct with two members and implement a Less

[tdeebswihart]

Not to be too pedantic since it's already a huge improvement, but I'm curious how the performance would compare to using https://pkg.go.dev/github.com/google/btree, specifically the ordered generic version (NewOrderedG). That's the ordered map for Go I've used a few times in the past

The implementation using btree and a second map to actually store the ack status performs far worse than gods' treemap:

$ go test -bench=AckManager ./service/matching -run=FooBarBaz
goos: darwin
goarch: arm64
pkg: go.temporal.io/server/service/matching
BenchmarkAckManager_AddTask-12         	   3495	   296339 ns/op
BenchmarkAckManager_CompleteTask-12    	   3706	   321754 ns/op

Here's the diff if you want to vet it:

diff --git a/service/matching/ack_manager.go b/service/matching/ack_manager.go
index a2d0afc67..25cf972f1 100644
--- a/service/matching/ack_manager.go
+++ b/service/matching/ack_manager.go
@@ -27,7 +27,7 @@ package matching
 import (
 	"sync"
 
-	"github.com/emirpasic/gods/maps/treemap"
+	"github.com/google/btree"
 	"go.uber.org/atomic"
 
 	"go.temporal.io/server/common/log"
@@ -50,9 +50,10 @@ func int64Comparator(a, b interface{}) int {
 // Used to convert out of order acks into ackLevel movement.
 type ackManager struct {
 	sync.RWMutex
-	outstandingTasks *treemap.Map // TaskID->acked
-	readLevel        int64        // Maximum TaskID inserted into outstandingTasks
-	ackLevel         int64        // Maximum TaskID below which all tasks are acked
+	outstandingTasks *btree.BTreeG[int64] // TaskID->acked
+	ackStatus        map[int64]bool
+	readLevel        int64 // Maximum TaskID inserted into outstandingTasks
+	ackLevel         int64 // Maximum TaskID below which all tasks are acked
 	backlogCounter   atomic.Int64
 	logger           log.Logger
 }
@@ -60,7 +61,8 @@ type ackManager struct {
 func newAckManager(logger log.Logger) ackManager {
 	return ackManager{
 		logger:           logger,
-		outstandingTasks: treemap.NewWith(int64Comparator),
+		outstandingTasks: btree.NewOrderedG[int64](2),
+		ackStatus:        make(map[int64]bool),
 		readLevel:        -1,
 		ackLevel:         -1}
 }
@@ -75,10 +77,10 @@ func (m *ackManager) addTask(taskID int64) {
 			tag.ReadLevel(m.readLevel))
 	}
 	m.readLevel = taskID
-	if _, found := m.outstandingTasks.Get(taskID); found {
+	if _, found := m.outstandingTasks.ReplaceOrInsert(taskID); found {
 		m.logger.Fatal("Already present in outstanding tasks", tag.TaskID(taskID))
 	}
-	m.outstandingTasks.Put(taskID, false)
+	m.ackStatus[taskID] = false
 	m.backlogCounter.Inc()
 }
 
@@ -131,14 +133,15 @@ func (m *ackManager) setAckLevel(ackLevel int64) {
 func (m *ackManager) completeTask(taskID int64) int64 {
 	m.Lock()
 	defer m.Unlock()
-	macked, found := m.outstandingTasks.Get(taskID)
-	acked := macked.(bool)
-	if !found || acked {
+	if !m.outstandingTasks.Has(taskID) {
+		return m.ackLevel
+	}
+	acked, _ := m.ackStatus[taskID]
+	if acked {
 		// don't adjust ack level if nothing has changed
 		return m.ackLevel
 	}
-
-	m.outstandingTasks.Put(taskID, true)
+	m.ackStatus[taskID] = true
 	m.backlogCounter.Dec()
 
 	min, _ := m.outstandingTasks.Min()
@@ -147,20 +150,19 @@ func (m *ackManager) completeTask(taskID int64) int64 {
 	}
 
 	// We've acked the minimum task, so should adjust the ack level as far as we can
-	iter := m.outstandingTasks.Iterator()
-	iter.Begin()
 	// deletion invalidates the iterator
 	var toDel []int64
-	for iter.Next() {
-		acked := iter.Value().(bool)
-		if !acked {
-			break
+	m.outstandingTasks.Ascend(func(key int64) bool {
+		if acked := m.ackStatus[key]; !acked {
+			return false
 		}
-		m.ackLevel = iter.Key().(int64)
+		m.ackLevel = key
 		toDel = append(toDel, m.ackLevel)
-	}
+		return true
+	})
 	for i := 0; i < len(toDel); i++ {
-		m.outstandingTasks.Remove(toDel[i])
+		m.outstandingTasks.Delete(toDel[i])
+		delete(m.ackStatus, toDel[i])
 	}
 
 	return m.ackLevel
diff --git a/service/matching/ack_manager.go b/service/matching/ack_manager.go
index a2d0afc67..25cf972f1 100644
--- a/service/matching/ack_manager.go
+++ b/service/matching/ack_manager.go
@@ -27,7 +27,7 @@ package matching
 import (
 	"sync"
 
-	"github.com/emirpasic/gods/maps/treemap"
+	"github.com/google/btree"
 	"go.uber.org/atomic"
 
 	"go.temporal.io/server/common/log"
@@ -50,9 +50,10 @@ func int64Comparator(a, b interface{}) int {
 // Used to convert out of order acks into ackLevel movement.
 type ackManager struct {
 	sync.RWMutex
-	outstandingTasks *treemap.Map // TaskID->acked
-	readLevel        int64        // Maximum TaskID inserted into outstandingTasks
-	ackLevel         int64        // Maximum TaskID below which all tasks are acked
+	outstandingTasks *btree.BTreeG[int64] // TaskID->acked
+	ackStatus        map[int64]bool
+	readLevel        int64 // Maximum TaskID inserted into outstandingTasks
+	ackLevel         int64 // Maximum TaskID below which all tasks are acked
 	backlogCounter   atomic.Int64
 	logger           log.Logger
 }
@@ -60,7 +61,8 @@ type ackManager struct {
 func newAckManager(logger log.Logger) ackManager {
 	return ackManager{
 		logger:           logger,
-		outstandingTasks: treemap.NewWith(int64Comparator),
+		outstandingTasks: btree.NewOrderedG[int64](2),
+		ackStatus:        make(map[int64]bool),
 		readLevel:        -1,
 		ackLevel:         -1}
 }
@@ -75,10 +77,10 @@ func (m *ackManager) addTask(taskID int64) {
 			tag.ReadLevel(m.readLevel))
 	}
 	m.readLevel = taskID
-	if _, found := m.outstandingTasks.Get(taskID); found {
+	if _, found := m.outstandingTasks.ReplaceOrInsert(taskID); found {
 		m.logger.Fatal("Already present in outstanding tasks", tag.TaskID(taskID))
 	}
-	m.outstandingTasks.Put(taskID, false)
+	m.ackStatus[taskID] = false
 	m.backlogCounter.Inc()
 }
 
@@ -131,14 +133,15 @@ func (m *ackManager) setAckLevel(ackLevel int64) {
 func (m *ackManager) completeTask(taskID int64) int64 {
 	m.Lock()
 	defer m.Unlock()
-	macked, found := m.outstandingTasks.Get(taskID)
-	acked := macked.(bool)
-	if !found || acked {
+	if !m.outstandingTasks.Has(taskID) {
+		return m.ackLevel
+	}
+	acked, _ := m.ackStatus[taskID]
+	if acked {
 		// don't adjust ack level if nothing has changed
 		return m.ackLevel
 	}
-
-	m.outstandingTasks.Put(taskID, true)
+	m.ackStatus[taskID] = true
 	m.backlogCounter.Dec()
 
 	min, _ := m.outstandingTasks.Min()
@@ -147,20 +150,19 @@ func (m *ackManager) completeTask(taskID int64) int64 {
 	}
 
 	// We've acked the minimum task, so should adjust the ack level as far as we can
-	iter := m.outstandingTasks.Iterator()
-	iter.Begin()
 	// deletion invalidates the iterator
 	var toDel []int64
-	for iter.Next() {
-		acked := iter.Value().(bool)
-		if !acked {
-			break
+	m.outstandingTasks.Ascend(func(key int64) bool {
+		if acked := m.ackStatus[key]; !acked {
+			return false
 		}
-		m.ackLevel = iter.Key().(int64)
+		m.ackLevel = key
 		toDel = append(toDel, m.ackLevel)
-	}
+		return true
+	})
 	for i := 0; i < len(toDel); i++ {
-		m.outstandingTasks.Remove(toDel[i])
+		m.outstandingTasks.Delete(toDel[i])
+		delete(m.ackStatus, toDel[i])
 	}
 
 	return m.ackLevel

[tdeebswihart]

Looks like this doesn't allow us to store values alongside the keys, hrm. I'll build it with this alongside a map of K->V to see how it performs

Oh right, I forgot you need the bool. You can totally store values next to the keys, but yeah, you can't use the Ordered one in that case, you need to make a struct with two members and implement a Less

Fair. I'll try that and see if it performs better

[dnr]

I see there's a github.com/emirpasic/gods/utils.Int64Comparator.. why copy it here?

[tdeebswihart]

Good question. I copied it from the other place in our codebase that uses the gods library

[dnr]

can we replace them both with that, then?

[tdeebswihart]

Done.

[dnr]

taskID here is int64 and min is interface{}.. I don't think this comparison is valid. actually I'm not sure how the compiler accepts this...

[dnr]

wouldn't it be simpler to write the whole thing (starting from line 146) like this:

for {
	min, acked := m.outstandingTasks.Min()
	if min == nil || !acked.(bool) {
		return m.ackLevel
	}
	m.outstandingTasks.Remove(min)
}

note that we currently never complete tasks out of order

[tdeebswihart]

You're right, I overcomplicated this

[dnr]

can we replace them both with that, then?

[dnr]

let's just get rid of this check? the loop below will take care of it