diff --git a/server/avl/norace_test.go b/server/avl/norace_test.go
new file mode 100644
index 00000000..7c18fb70
--- /dev/null
+++ b/server/avl/norace_test.go
@@ -0,0 +1,199 @@
+// Copyright 2023 The NATS Authors
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//go:build !race && !skip_no_race_tests
+// +build !race,!skip_no_race_tests
+
+package avl
+
+import (
+	"flag"
+	"fmt"
+	"math"
+	"math/rand"
+	"runtime"
+	"runtime/debug"
+	"testing"
+	"time"
+)
+
+// Print Results: go test -v  --args --results
+var printResults = flag.Bool("results", false, "Enable Results Logging")
+
+// SequenceSet memory tests vs dmaps.
+func TestNoRaceSeqSetSizeComparison(t *testing.T) {
+	// Create 5M random entries (dupes possible but ok for this test) out of 8M range.
+	num := 5_000_000
+	max := 7_000_000
+
+	seqs := make([]uint64, 0, num)
+	for i := 0; i < num; i++ {
+		n := uint64(rand.Int63n(int64(max + 1)))
+		seqs = append(seqs, n)
+	}
+
+	runtime.GC()
+	// Disable to get stable results.
+	gcp := debug.SetGCPercent(-1)
+	defer debug.SetGCPercent(gcp)
+
+	mem := runtime.MemStats{}
+	runtime.ReadMemStats(&mem)
+	inUseBefore := mem.HeapInuse
+
+	dmap := make(map[uint64]struct{}, num)
+	for _, n := range seqs {
+		dmap[n] = struct{}{}
+	}
+	runtime.ReadMemStats(&mem)
+	dmapUse := mem.HeapInuse - inUseBefore
+	inUseBefore = mem.HeapInuse
+
+	// Now do SequenceSet on same dataset.
+	var sset SequenceSet
+	for _, n := range seqs {
+		sset.Insert(n)
+	}
+
+	runtime.ReadMemStats(&mem)
+	seqSetUse := mem.HeapInuse - inUseBefore
+
+	if seqSetUse > 2*1024*1024 {
+		t.Fatalf("Expected SequenceSet size to be < 2M, got %v", friendlyBytes(seqSetUse))
+	}
+	if seqSetUse*50 > dmapUse {
+		t.Fatalf("Expected SequenceSet to be at least 50x better then dmap approach: %v vs %v",
+			friendlyBytes(seqSetUse),
+			friendlyBytes(dmapUse),
+		)
+	}
+}
+
+func TestNoRaceSeqSetEncodeLarge(t *testing.T) {
+	num := 2_500_000
+	max := 5_000_000
+
+	dmap := make(map[uint64]struct{}, num)
+	var ss SequenceSet
+	for i := 0; i < num; i++ {
+		n := uint64(rand.Int63n(int64(max + 1)))
+		ss.Insert(n)
+		dmap[n] = struct{}{}
+	}
+
+	// Disable to get stable results.
+	gcp := debug.SetGCPercent(-1)
+	defer debug.SetGCPercent(gcp)
+
+	// In general should be about the same, but can see some variability.
+	expected := 500 * time.Microsecond
+
+	start := time.Now()
+	b, err := ss.Encode(nil)
+	require_NoError(t, err)
+
+	if elapsed := time.Since(start); elapsed > expected {
+		t.Fatalf("Expected encode of %d items with encoded size %v to take less than %v, got %v",
+			num, friendlyBytes(len(b)), expected, elapsed)
+	} else {
+		logResults("Encode time for %d items was %v, encoded size is %v\n", num, elapsed, friendlyBytes(len(b)))
+	}
+
+	start = time.Now()
+	ss2, err := Decode(b)
+	require_NoError(t, err)
+	if elapsed := time.Since(start); elapsed > expected {
+		t.Fatalf("Expected decode to take less than %v, got %v", expected, elapsed)
+	} else {
+		logResults("Decode time is %v\n", elapsed)
+	}
+	require_True(t, ss.Nodes() == ss2.Nodes())
+}
+
+func TestNoRaceSeqSetRelativeSpeed(t *testing.T) {
+	// Create 1M random entries (dupes possible but ok for this test) out of 3M range.
+	num := 1_000_000
+	max := 3_000_000
+
+	seqs := make([]uint64, 0, num)
+	for i := 0; i < num; i++ {
+		n := uint64(rand.Int63n(int64(max + 1)))
+		seqs = append(seqs, n)
+	}
+
+	start := time.Now()
+	// Now do SequenceSet on same dataset.
+	var sset SequenceSet
+	for _, n := range seqs {
+		sset.Insert(n)
+	}
+	ssInsertElapsed := time.Since(start)
+	logResults("Inserts SequenceSet: %v for %d items\n", ssInsertElapsed, num)
+
+	start = time.Now()
+	for _, n := range seqs {
+		if ok := sset.Exists(n); !ok {
+			t.Fatalf("Should exist")
+		}
+	}
+	ssLookupElapsed := time.Since(start)
+	logResults("Lookups: %v\n", ssLookupElapsed)
+
+	// Now do a map.
+	dmap := make(map[uint64]struct{})
+	start = time.Now()
+	for _, n := range seqs {
+		dmap[n] = struct{}{}
+	}
+	mapInsertElapsed := time.Since(start)
+	logResults("Inserts Map[uint64]: %v for %d items\n", mapInsertElapsed, num)
+
+	start = time.Now()
+	for _, n := range seqs {
+		if _, ok := dmap[n]; !ok {
+			t.Fatalf("Should exist")
+		}
+	}
+	mapLookupElapsed := time.Since(start)
+	logResults("Lookups: %v\n", mapLookupElapsed)
+
+	// In general we are between 1.5 and 1.75 times slower atm then a straight map.
+	// Let's test an upper bound of 2x for now.
+	if mapInsertElapsed*2 <= ssInsertElapsed {
+		t.Fatalf("Expected SequenceSet insert to be no more than 2x slower (%v vs %v)", mapInsertElapsed, ssInsertElapsed)
+	}
+
+	if mapLookupElapsed*2 <= ssLookupElapsed {
+		t.Fatalf("Expected SequenceSet lookups to be no more than 2x slower (%v vs %v)", mapLookupElapsed, ssLookupElapsed)
+	}
+}
+
+// friendlyBytes returns a string with the given bytes int64
+// represented as a size, such as 1KB, 10MB, etc...
+func friendlyBytes[T int | uint64 | int64](bytes T) string {
+	fbytes := float64(bytes)
+	base := 1024
+	pre := []string{"K", "M", "G", "T", "P", "E"}
+	if fbytes < float64(base) {
+		return fmt.Sprintf("%v B", fbytes)
+	}
+	exp := int(math.Log(fbytes) / math.Log(float64(base)))
+	index := exp - 1
+	return fmt.Sprintf("%.2f %sB", fbytes/math.Pow(float64(base), float64(exp)), pre[index])
+}
+
+func logResults(format string, args ...interface{}) {
+	if *printResults {
+		fmt.Printf(format, args...)
+	}
+}
diff --git a/server/avl/seqset.go b/server/avl/seqset.go
new file mode 100644
index 00000000..c0187706
--- /dev/null
+++ b/server/avl/seqset.go
@@ -0,0 +1,544 @@
+// Copyright 2023 The NATS Authors
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package avl
+
+import (
+	"encoding/binary"
+	"errors"
+	"math/bits"
+	"sort"
+)
+
+// SequenceSet is a memory and encoding optimized set for storing unsigned ints.
+//
+// SequenceSet is ~80-100 times more efficient memory wise than a map[uint64]struct{}.
+// SequenceSet is ~1.75 times slower at inserts than the same map.
+// SequenceSet is not thread safe.
+//
+// We use an AVL tree with nodes that hold bitmasks for set membership.
+//
+// Encoding will convert to a space optimized encoding using bitmasks.
+type SequenceSet struct {
+	root  *node // root node
+	size  int   // number of items
+	nodes int   // number of nodes
+	// Having this here vs on the stack in Insert/Delete
+	// makes a difference in memory usage.
+	changed bool
+}
+
+// Insert will insert the sequence into the set.
+// The tree will be balanced inline.
+func (ss *SequenceSet) Insert(seq uint64) {
+	if ss.root = ss.root.insert(seq, &ss.changed, &ss.nodes); ss.changed {
+		ss.changed = false
+		ss.size++
+	}
+}
+
+// Exists will return true iff the sequence is a member of this set.
+func (ss *SequenceSet) Exists(seq uint64) bool {
+	for n := ss.root; n != nil; {
+		if seq < n.base {
+			n = n.l
+			continue
+		} else if seq >= n.base+numEntries {
+			n = n.r
+			continue
+		}
+		return n.exists(seq)
+	}
+	return false
+}
+
+// Delete will remove the sequence from the set.
+// Wil optionally remove nodes and rebalance.
+func (ss *SequenceSet) Delete(seq uint64) {
+	if ss == nil || ss.root == nil {
+		return
+	}
+	ss.root = ss.root.delete(seq, &ss.changed, &ss.nodes)
+	if ss.changed {
+		ss.changed = false
+		ss.size--
+	}
+}
+
+// Size returns the number of items in the set.
+func (ss *SequenceSet) Size() int {
+	return ss.size
+}
+
+// Nodes returns the number of nodes in the tree.
+func (ss *SequenceSet) Nodes() int {
+	return ss.nodes
+}
+
+// Empty will clear all items from a set.
+func (ss *SequenceSet) Empty() {
+	ss.root = nil
+	ss.size = 0
+	ss.nodes = 0
+}
+
+// IsEmpty is a fast check of the set being empty.
+func (ss *SequenceSet) IsEmpty() bool {
+	if ss == nil || ss.root == nil {
+		return true
+	}
+	return false
+}
+
+// Range will invoke the given function for each item in the set.
+// They will range over the set in ascending order.
+// If the callback returns false we terminate the iteration.
+func (ss *SequenceSet) Range(f func(uint64) bool) {
+	ss.root.iter(f)
+}
+
+// Heights returns the left and right heights of the tree.
+func (ss *SequenceSet) Heights() (l, r int) {
+	if ss.root == nil {
+		return 0, 0
+	}
+	if ss.root.l != nil {
+		l = ss.root.l.h
+	}
+	if ss.root.r != nil {
+		r = ss.root.r.h
+	}
+	return l, r
+}
+
+// MinMax will return the minunum and maximum values in the set.
+func (ss *SequenceSet) MinMax() (min, max uint64) {
+	if ss.root == nil {
+		return 0, 0
+	}
+	for l := ss.root; l != nil; l = l.l {
+		if l.l == nil {
+			min = l.min()
+		}
+	}
+	for r := ss.root; r != nil; r = r.r {
+		if r.r == nil {
+			max = r.max()
+		}
+	}
+	return min, max
+}
+
+func clone(src *node, target **node) {
+	if src == nil {
+		return
+	}
+	n := &node{base: src.base, bits: src.bits, h: src.h}
+	*target = n
+	clone(src.l, &n.l)
+	clone(src.r, &n.r)
+}
+
+// Clone will return a clone of the given SequenceSet.
+func (ss *SequenceSet) Clone() *SequenceSet {
+	if ss == nil {
+		return nil
+	}
+	css := &SequenceSet{nodes: ss.nodes, size: ss.size}
+	clone(ss.root, &css.root)
+
+	return css
+}
+
+// Union will return a union of all sets.
+func Union(ssa ...*SequenceSet) *SequenceSet {
+	if len(ssa) == 0 {
+		return nil
+	}
+	// Sort so we can clone largest.
+	sort.Slice(ssa, func(i, j int) bool { return ssa[i].Size() > ssa[j].Size() })
+	ss := ssa[0].Clone()
+
+	// Insert the rest through range call.
+	for i := 1; i < len(ssa); i++ {
+		ssa[i].Range(func(n uint64) bool {
+			ss.Insert(n)
+			return true
+		})
+	}
+	return ss
+}
+
+const (
+	// Magic is used to identify the encode binary state..
+	magic = uint8(22)
+	// Version
+	version = uint8(1)
+	// hdrLen
+	hdrLen = 2
+	// minimum length of an encoded SequenceSet.
+	minLen = 2 + 4 // magic + version + num nodes.
+)
+
+// EncodeLen returns the bytes needed for encoding.
+func (ss SequenceSet) EncodeLen() int {
+	return minLen + (ss.Nodes() * ((numBuckets+1)*8 + 2))
+}
+
+func (ss SequenceSet) Encode(buf []byte) ([]byte, error) {
+	nn, encLen := ss.Nodes(), ss.EncodeLen()
+
+	if len(buf) < encLen {
+		buf = make([]byte, encLen)
+	}
+
+	// TODO(dlc) - Go 1.19 introduced Append to not have to keep track.
+	// Once 1.20 is out we could change this over.
+	// Also binary.Write() is way slower, do not use.
+
+	var le = binary.LittleEndian
+	buf[0], buf[1] = magic, version
+	i := hdrLen
+	le.PutUint32(buf[i:], uint32(nn))
+	i += 4
+	ss.root.nodeIter(func(n *node) {
+		le.PutUint64(buf[i:], n.base)
+		i += 8
+		for _, b := range n.bits {
+			le.PutUint64(buf[i:], b)
+			i += 8
+		}
+		le.PutUint16(buf[i:], uint16(n.h))
+		i += 2
+	})
+	return buf[:i], nil
+}
+
+// ErrBadEncoding is returned when we can not decode properly.
+var ErrBadEncoding = errors.New("ss: bad encoding")
+
+func Decode(buf []byte) (*SequenceSet, error) {
+	if len(buf) < minLen || buf[0] != magic || buf[1] != version {
+		return nil, ErrBadEncoding
+	}
+
+	var le = binary.LittleEndian
+	index := 2
+	nn := int(le.Uint32(buf[index:]))
+	index += 4
+
+	expectedLen := minLen + (nn * ((numBuckets+1)*8 + 2))
+	if len(buf) != expectedLen {
+		return nil, ErrBadEncoding
+	}
+
+	nodes := make([]node, nn)
+
+	var ss SequenceSet
+	for i := 0; i < nn; i++ {
+		n := &nodes[i]
+		n.base = le.Uint64(buf[index:])
+		index += 8
+		for bi := range n.bits {
+			n.bits[bi] = le.Uint64(buf[index:])
+			index += 8
+		}
+		n.h = int(le.Uint16(buf[index:]))
+		index += 2
+		ss.insertNode(n)
+	}
+
+	return &ss, nil
+}
+
+// insertNode places a decoded node into the tree.
+// These should be done in tree order as defined by Encode()
+// This allows us to not have to calculate height or do rebalancing.
+// So much better performance this way.
+func (ss *SequenceSet) insertNode(n *node) {
+	ss.nodes++
+
+	if ss.root == nil {
+		ss.root = n
+		return
+	}
+	// Walk our way to the insertion point.
+	for p := ss.root; p != nil; {
+		if n.base < p.base {
+			if p.l == nil {
+				p.l = n
+				return
+			}
+			p = p.l
+		} else {
+			if p.r == nil {
+				p.r = n
+				return
+			}
+			p = p.r
+		}
+	}
+}
+
+const (
+	bitsPerBucket = 64 // bits in uint64
+	numBuckets    = 64
+	numEntries    = numBuckets * bitsPerBucket
+)
+
+type node struct {
+	//v dvalue
+	base uint64
+	bits [numBuckets]uint64
+	l    *node
+	r    *node
+	h    int
+}
+
+// Set the proper bit.
+// seq should have already been qualified and inserted should be non nil.
+func (n *node) set(seq uint64, inserted *bool) {
+	seq -= n.base
+	i := seq / bitsPerBucket
+	mask := uint64(1) << (seq % bitsPerBucket)
+	if (n.bits[i] & mask) == 0 {
+		n.bits[i] |= mask
+		*inserted = true
+	}
+}
+
+func (n *node) insert(seq uint64, inserted *bool, nodes *int) *node {
+	if n == nil {
+		base := (seq / numEntries) * numEntries
+		n := &node{base: base, h: 1}
+		n.set(seq, inserted)
+		*nodes++
+		return n
+	}
+
+	if seq < n.base {
+		n.l = n.l.insert(seq, inserted, nodes)
+	} else if seq >= n.base+numEntries {
+		n.r = n.r.insert(seq, inserted, nodes)
+	} else {
+		n.set(seq, inserted)
+	}
+
+	n.h = maxH(n) + 1
+
+	// Don't make a function, impacts performance.
+	if bf := balanceF(n); bf > 1 {
+		// Left unbalanced.
+		if n.l.base+numEntries > seq {
+			return n.rotateR()
+		} else {
+			n.l = n.l.rotateL()
+			return n.rotateR()
+		}
+	} else if bf < -1 {
+		// right unbalanced.
+		if n.r.base+numEntries > seq {
+			n.r = n.r.rotateR()
+			return n.rotateL()
+		} else {
+			return n.rotateL()
+		}
+	}
+	return n
+}
+
+func (n *node) rotateL() *node {
+	r := n.r
+	if r != nil {
+		n.r = r.l
+		r.l = n
+		n.h = maxH(n) + 1
+		r.h = maxH(r) + 1
+	} else {
+		n.r = nil
+		n.h = maxH(n) + 1
+	}
+	return r
+}
+
+func (n *node) rotateR() *node {
+	l := n.l
+	if l != nil {
+		n.l = l.r
+		l.r = n
+		n.h = maxH(n) + 1
+		l.h = maxH(l) + 1
+	} else {
+		n.l = nil
+		n.h = maxH(n) + 1
+	}
+	return l
+}
+
+func balanceF(n *node) int {
+	if n == nil {
+		return 0
+	}
+	var lh, rh int
+	if n.l != nil {
+		lh = n.l.h
+	}
+	if n.r != nil {
+		rh = n.r.h
+	}
+	return lh - rh
+}
+
+func maxH(n *node) int {
+	var lh, rh int
+	if n.l != nil {
+		lh = n.l.h
+	}
+	if n.r != nil {
+		rh = n.r.h
+	}
+	if lh > rh {
+		return lh
+	}
+	return rh
+}
+
+// Clear the proper bit.
+// seq should have already been qualified and deleted should be non nil.
+// Will return true if this node is now empty.
+func (n *node) clear(seq uint64, deleted *bool) bool {
+	seq -= n.base
+	i := seq / bitsPerBucket
+	mask := uint64(1) << (seq % bitsPerBucket)
+	if (n.bits[i] & mask) != 0 {
+		n.bits[i] &^= mask
+		*deleted = true
+	}
+	for _, b := range n.bits {
+		if b != 0 {
+			return false
+		}
+	}
+	return true
+}
+
+func (n *node) delete(seq uint64, deleted *bool, nodes *int) *node {
+	if n == nil {
+		return nil
+	}
+
+	if seq < n.base {
+		n.l = n.l.delete(seq, deleted, nodes)
+	} else if seq >= n.base+numEntries {
+		n.r = n.r.delete(seq, deleted, nodes)
+	} else if empty := n.clear(seq, deleted); empty {
+		*nodes--
+		if nn := n.l; nn == nil {
+			n = n.r
+		} else if nn.r == nil {
+			nn.r = n.r
+			n = nn
+		} else {
+			nn.r.r = n.r
+			n = nn
+		}
+	}
+
+	// Check balance.
+	if bf := balanceF(n); bf > 1 {
+		// Left unbalanced.
+		if n.l.base+numEntries > seq {
+			return n.rotateR()
+		} else {
+			n.l = n.l.rotateL()
+			return n.rotateR()
+		}
+	} else if bf < -1 {
+		// right unbalanced.
+		if n.r.base+numEntries > seq {
+			n.r = n.r.rotateR()
+			return n.rotateL()
+		} else {
+			return n.rotateL()
+		}
+	}
+
+	return n
+}
+
+func (n *node) exists(seq uint64) bool {
+	seq -= n.base
+	i := seq / bitsPerBucket
+	mask := uint64(1) << (seq % bitsPerBucket)
+	return n.bits[i]&mask != 0
+}
+
+// Return minimum sequence in the set.
+// This node can not be empty.
+func (n *node) min() uint64 {
+	for i, b := range n.bits {
+		if b != 0 {
+			return n.base +
+				uint64(i*bitsPerBucket) +
+				uint64(bits.TrailingZeros64(b))
+		}
+	}
+	return 0
+}
+
+// Return maximum sequence in the set.
+// This node can not be empty.
+func (n *node) max() uint64 {
+	for i := numBuckets - 1; i >= 0; i-- {
+		if b := n.bits[i]; b != 0 {
+			return n.base +
+				uint64(i*bitsPerBucket) +
+				uint64(bitsPerBucket-bits.LeadingZeros64(b>>1))
+		}
+	}
+	return 0
+}
+
+// This is done in tree order.
+func (n *node) nodeIter(f func(n *node)) {
+	if n == nil {
+		return
+	}
+	f(n)
+	n.l.nodeIter(f)
+	n.r.nodeIter(f)
+}
+
+// iter will iterate through the set's items in this node.
+// If the supplied function returns false we terminate the iteration.
+func (n *node) iter(f func(uint64) bool) bool {
+	if n == nil {
+		return true
+	}
+
+	if ok := n.l.iter(f); !ok {
+		return false
+	}
+	for num := n.base; num < n.base+numEntries; num++ {
+		if n.exists(num) {
+			if ok := f(num); !ok {
+				return false
+			}
+		}
+	}
+	if ok := n.r.iter(f); !ok {
+		return false
+	}
+
+	return true
+}
diff --git a/server/avl/seqset_test.go b/server/avl/seqset_test.go
new file mode 100644
index 00000000..8a90971d
--- /dev/null
+++ b/server/avl/seqset_test.go
@@ -0,0 +1,233 @@
+// Copyright 2023 The NATS Authors
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package avl
+
+import (
+	"math/rand"
+	"testing"
+)
+
+func TestSeqSetBasics(t *testing.T) {
+	var ss SequenceSet
+
+	seqs := []uint64{22, 222, 2222, 2, 2, 4}
+	for _, seq := range seqs {
+		ss.Insert(seq)
+		require_True(t, ss.Exists(seq))
+	}
+
+	require_True(t, ss.Nodes() == 1)
+	require_True(t, ss.Size() == len(seqs)-1) // We have one dup in there.
+	lh, rh := ss.Heights()
+	require_True(t, lh == 0)
+	require_True(t, rh == 0)
+}
+
+func TestSeqSetLeftLean(t *testing.T) {
+	var ss SequenceSet
+
+	for i := uint64(4 * numEntries); i > 0; i-- {
+		ss.Insert(i)
+	}
+
+	require_True(t, ss.Nodes() == 5)
+	require_True(t, ss.Size() == 4*numEntries)
+	lh, rh := ss.Heights()
+	require_True(t, lh == 2)
+	require_True(t, rh == 1)
+}
+
+func TestSeqSetRightLean(t *testing.T) {
+	var ss SequenceSet
+
+	for i := uint64(0); i < uint64(4*numEntries); i++ {
+		ss.Insert(i)
+	}
+
+	require_True(t, ss.Nodes() == 4)
+	require_True(t, ss.Size() == 4*numEntries)
+	lh, rh := ss.Heights()
+	require_True(t, lh == 1)
+	require_True(t, rh == 2)
+}
+
+func TestSeqSetCorrectness(t *testing.T) {
+	// Generate 100k sequences across 500k range.
+	num := 100_000
+	max := 500_000
+
+	set := make(map[uint64]struct{}, num)
+	var ss SequenceSet
+	for i := 0; i < num; i++ {
+		n := uint64(rand.Int63n(int64(max + 1)))
+		ss.Insert(n)
+		set[n] = struct{}{}
+	}
+
+	for i := uint64(0); i <= uint64(max); i++ {
+		_, exists := set[i]
+		require_True(t, ss.Exists(i) == exists)
+	}
+}
+
+func TestSeqSetRange(t *testing.T) {
+	num := 2*numEntries + 22
+	nums := make([]uint64, 0, num)
+	for i := 0; i < num; i++ {
+		nums = append(nums, uint64(i))
+	}
+	rand.Shuffle(len(nums), func(i, j int) { nums[i], nums[j] = nums[j], nums[i] })
+
+	var ss SequenceSet
+	for _, n := range nums {
+		ss.Insert(n)
+	}
+
+	nums = nums[:0]
+	ss.Range(func(n uint64) bool {
+		nums = append(nums, n)
+		return true
+	})
+	require_True(t, len(nums) == num)
+	for i := uint64(0); i < uint64(num); i++ {
+		require_True(t, nums[i] == i)
+	}
+
+	// Test truncating the range call.
+	nums = nums[:0]
+	ss.Range(func(n uint64) bool {
+		if n >= 10 {
+			return false
+		}
+		nums = append(nums, n)
+		return true
+	})
+	require_True(t, len(nums) == 10)
+	for i := uint64(0); i < 10; i++ {
+		require_True(t, nums[i] == i)
+	}
+}
+
+func TestSeqSetDelete(t *testing.T) {
+	var ss SequenceSet
+
+	// Simple single node.
+	seqs := []uint64{22, 222, 2222, 2, 2, 4}
+	for _, seq := range seqs {
+		ss.Insert(seq)
+	}
+
+	for _, seq := range seqs {
+		ss.Delete(seq)
+		require_True(t, !ss.Exists(seq))
+	}
+	require_True(t, ss.root == nil)
+
+	num := 22*numEntries + 22
+	nums := make([]uint64, 0, num)
+	for i := 0; i < num; i++ {
+		nums = append(nums, uint64(i))
+	}
+	rand.Shuffle(len(nums), func(i, j int) { nums[i], nums[j] = nums[j], nums[i] })
+
+	for _, n := range nums {
+		ss.Insert(n)
+	}
+
+	for _, n := range nums {
+		ss.Delete(n)
+		require_True(t, !ss.Exists(n))
+	}
+	require_True(t, ss.root == nil)
+}
+
+func TestSeqSetMinMax(t *testing.T) {
+	var ss SequenceSet
+
+	// Simple single node.
+	seqs := []uint64{22, 222, 2222, 2, 2, 4}
+	for _, seq := range seqs {
+		ss.Insert(seq)
+	}
+
+	min, max := ss.MinMax()
+	require_True(t, min == 2 && max == 2222)
+
+	// Multi-node
+	ss.Empty()
+
+	num := 22*numEntries + 22
+	nums := make([]uint64, 0, num)
+	for i := 0; i < num; i++ {
+		nums = append(nums, uint64(i))
+	}
+	rand.Shuffle(len(nums), func(i, j int) { nums[i], nums[j] = nums[j], nums[i] })
+	for _, n := range nums {
+		ss.Insert(n)
+	}
+
+	min, max = ss.MinMax()
+	require_True(t, min == 0 && max == uint64(num-1))
+}
+
+func TestSeqSetClone(t *testing.T) {
+	// Generate 100k sequences across 500k range.
+	num := 100_000
+	max := 500_000
+
+	var ss SequenceSet
+	for i := 0; i < num; i++ {
+		ss.Insert(uint64(rand.Int63n(int64(max + 1))))
+	}
+
+	ssc := ss.Clone()
+	require_True(t, ss.Size() == ssc.Size())
+	require_True(t, ss.Nodes() == ssc.Nodes())
+}
+
+func TestSeqSetUnion(t *testing.T) {
+	var ss1, ss2 SequenceSet
+
+	seqs1 := []uint64{22, 222, 2222, 2, 2, 4}
+	for _, seq := range seqs1 {
+		ss1.Insert(seq)
+	}
+
+	seqs2 := []uint64{33, 333, 3333, 3, 33_333, 333_333}
+	for _, seq := range seqs2 {
+		ss2.Insert(seq)
+	}
+
+	ss := Union(&ss1, &ss2)
+	require_True(t, ss.Size() == 11)
+
+	seqs := append(seqs1, seqs2...)
+	for _, n := range seqs {
+		require_True(t, ss.Exists(n))
+	}
+}
+
+func require_NoError(t *testing.T, err error) {
+	t.Helper()
+	if err != nil {
+		t.Fatalf("require no error, but got: %v", err)
+	}
+}
+
+func require_True(t *testing.T, b bool) {
+	t.Helper()
+	if !b {
+		t.Fatalf("require true")
+	}
+}
diff --git a/server/norace_test.go b/server/norace_test.go
index 4804ce78..e6acf5b6 100644
--- a/server/norace_test.go
+++ b/server/norace_test.go
@@ -48,6 +48,7 @@ import (
 
 	"github.com/klauspost/compress/s2"
 	"github.com/nats-io/jwt/v2"
+	"github.com/nats-io/nats-server/v2/server/avl"
 	"github.com/nats-io/nats.go"
 	"github.com/nats-io/nkeys"
 	"github.com/nats-io/nuid"
@@ -6214,6 +6215,93 @@ func TestNoRaceFileStoreStreamMaxAgePerformance(t *testing.T) {
 	fmt.Printf("%.0f msgs/sec\n", float64(num)/elapsed.Seconds())
 }
 
+// SequenceSet memory tests vs dmaps.
+func TestNoRaceSeqSetSizeComparison(t *testing.T) {
+	// Create 5M random entries (dupes possible but ok for this test) out of 8M range.
+	num := 5_000_000
+	max := 7_000_000
+
+	seqs := make([]uint64, 0, num)
+	for i := 0; i < num; i++ {
+		n := uint64(rand.Int63n(int64(max + 1)))
+		seqs = append(seqs, n)
+	}
+
+	runtime.GC()
+	// Disable to get stable results.
+	gcp := debug.SetGCPercent(-1)
+	defer debug.SetGCPercent(gcp)
+
+	mem := runtime.MemStats{}
+	runtime.ReadMemStats(&mem)
+	inUseBefore := mem.HeapInuse
+
+	dmap := make(map[uint64]struct{}, num)
+	for _, n := range seqs {
+		dmap[n] = struct{}{}
+	}
+	runtime.ReadMemStats(&mem)
+	dmapUse := mem.HeapInuse - inUseBefore
+	inUseBefore = mem.HeapInuse
+
+	// Now do SequenceSet on same dataset.
+	var sset avl.SequenceSet
+	for _, n := range seqs {
+		sset.Insert(n)
+	}
+
+	runtime.ReadMemStats(&mem)
+	seqSetUse := mem.HeapInuse - inUseBefore
+
+	if seqSetUse > 2*1024*1024 {
+		t.Fatalf("Expected SequenceSet size to be < 2M, got %v", friendlyBytes(int64(seqSetUse)))
+	}
+	if seqSetUse*50 > dmapUse {
+		t.Fatalf("Expected SequenceSet to be at least 50x better then dmap approach: %v vs %v",
+			friendlyBytes(int64(seqSetUse)),
+			friendlyBytes(int64(dmapUse)),
+		)
+	}
+}
+
+// FilteredState for ">" with large interior deletes was very slow.
+func TestNoRaceFileStoreFilteredStateWithLargeDeletes(t *testing.T) {
+	storeDir := t.TempDir()
+
+	fs, err := newFileStore(
+		FileStoreConfig{StoreDir: storeDir, BlockSize: 4096},
+		StreamConfig{Name: "zzz", Subjects: []string{"foo"}, Storage: FileStorage},
+	)
+	require_NoError(t, err)
+	defer fs.Stop()
+
+	subj, msg := "foo", []byte("Hello World")
+
+	toStore := 500_000
+	for i := 0; i < toStore; i++ {
+		_, _, err := fs.StoreMsg(subj, nil, msg)
+		require_NoError(t, err)
+	}
+
+	// Now delete every other one.
+	for seq := 2; seq <= toStore; seq += 2 {
+		_, err := fs.RemoveMsg(uint64(seq))
+		require_NoError(t, err)
+	}
+
+	runtime.GC()
+	// Disable to get stable results.
+	gcp := debug.SetGCPercent(-1)
+	defer debug.SetGCPercent(gcp)
+
+	start := time.Now()
+	fss := fs.FilteredState(1, _EMPTY_)
+	elapsed := time.Since(start)
+
+	require_True(t, fss.Msgs == uint64(toStore/2))
+	require_True(t, elapsed < 500*time.Microsecond)
+}
+
 // ConsumerInfo seems to being called quite a bit more than we had anticipated.
 // Under certain circumstances, since we reset num pending, this can be very costly.
 // We will use the fast path to alleviate that performance bottleneck but also make