diff --git a/bucket.go b/bucket.go
index 9737128..0b9c17d 100644
--- a/bucket.go
+++ b/bucket.go
@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"errors"
 	"fmt"
-	"sort"
 	"unsafe"
 )
 
@@ -41,10 +40,10 @@ var (
 // Bucket represents a collection of key/value pairs inside the database.
 type Bucket struct {
 	*bucket
-	tx      *Tx
-	buckets map[string]*Bucket
-	nodes   map[pgid]*node
-	pending []*node
+	tx       *Tx
+	buckets  map[string]*Bucket
+	rootNode *node
+	nodes    map[pgid]*node
 }
 
 // bucket represents the on-file representation of a bucket.
@@ -382,76 +381,19 @@ func (b *Bucket) spill() error {
 		c.node().put([]byte(name), []byte(name), value, 0, bucketLeafFlag)
 	}
 
-	// Ignore if there are no nodes to spill.
-	if len(b.nodes) == 0 {
+	// Ignore if there's not a materialized root node.
+	if b.rootNode == nil {
 		return nil
 	}
 
-	// Sort nodes by highest depth first.
-	nodes := make(nodesByDepth, 0, len(b.nodes))
-	for _, n := range b.nodes {
-		nodes = append(nodes, n)
+	// Spill nodes.
+	if err := b.rootNode.spill(); err != nil {
+		return err
 	}
-	sort.Sort(nodes)
-
-	// Spill nodes by deepest first.
-	for i := 0; i < len(nodes); i++ {
-		n := nodes[i]
-
-		// Split nodes into appropriate sized nodes.
-		// The first node in this list will be a reference to n to preserve ancestry.
-		newNodes := n.split(b.tx.db.pageSize)
-		b.pending = newNodes
-
-		// If this is a root node that split then create a parent node.
-		if n.parent == nil && len(newNodes) > 1 {
-			n.parent = &node{bucket: b, isLeaf: false}
-			nodes = append(nodes, n.parent)
-		}
-
-		// Add node's page to the freelist.
-		if n.pgid > 0 {
-			b.tx.db.freelist.free(b.tx.id(), b.tx.page(n.pgid))
-		}
-
-		// Write nodes to dirty pages.
-		for i, newNode := range newNodes {
-			// Allocate contiguous space for the node.
-			p, err := b.tx.allocate((newNode.size() / b.tx.db.pageSize) + 1)
-			if err != nil {
-				return err
-			}
-
-			// Write the node to the page.
-			newNode.write(p)
-			newNode.pgid = p.id
-			newNode.parent = n.parent
-
-			// The first node should use the existing entry, other nodes are inserts.
-			var oldKey []byte
-			if i == 0 {
-				oldKey = n.key
-			} else {
-				oldKey = newNode.inodes[0].key
-			}
-
-			// Update the parent entry.
-			if newNode.parent != nil {
-				newNode.parent.put(oldKey, newNode.inodes[0].key, nil, newNode.pgid, 0)
-			}
-
-			// Update the statistics.
-			b.tx.stats.Spill++
-		}
-
-		b.pending = nil
-	}
-
-	// Clear out nodes now that they are all spilled.
-	b.nodes = make(map[pgid]*node)
+	b.rootNode = b.rootNode.root()
 
 	// Update the root node for this bucket.
-	b.root = nodes[len(nodes)-1].pgid
+	b.root = b.rootNode.pgid
 
 	return nil
 }
@@ -474,10 +416,12 @@ func (b *Bucket) node(pgid pgid, parent *node) *node {
 		return n
 	}
 
-	// Otherwise create a branch and cache it.
+	// Otherwise create a branch node and cache it.
 	n := &node{bucket: b, parent: parent}
-	if n.parent != nil {
-		n.depth = n.parent.depth + 1
+	if parent == nil {
+		b.rootNode = n
+	} else {
+		parent.children = append(parent.children, n)
 	}
 	n.read(b.tx.page(pgid))
 	b.nodes[pgid] = n
@@ -494,16 +438,12 @@ func (b *Bucket) dereference() {
 		n.dereference()
 	}
 
-	for _, n := range b.pending {
-		n.dereference()
-	}
-
 	for _, child := range b.buckets {
 		child.dereference()
 	}
 
 	// Update statistics
-	b.tx.stats.NodeDeref += len(b.nodes) + len(b.pending)
+	b.tx.stats.NodeDeref += len(b.nodes)
 }
 
 // pageNode returns the in-memory node, if it exists.
diff --git a/bucket_test.go b/bucket_test.go
index c506b72..dc89bcc 100644
--- a/bucket_test.go
+++ b/bucket_test.go
@@ -161,6 +161,64 @@ func TestBucket_Delete(t *testing.T) {
 	})
 }
 
+// Ensure that accessing and updating nested buckets is ok across transactions.
+func TestBucket_Nested(t *testing.T) {
+	withOpenDB(func(db *DB, path string) {
+		db.Update(func(tx *Tx) error {
+			// Create a widgets bucket.
+			b, err := tx.CreateBucket([]byte("widgets"))
+			assert.NoError(t, err)
+
+			// Create a widgets/foo bucket.
+			_, err = b.CreateBucket([]byte("foo"))
+			assert.NoError(t, err)
+
+			// Create a widgets/bar key.
+			assert.NoError(t, b.Put([]byte("bar"), []byte("0000")))
+
+			return nil
+		})
+		mustCheck(db)
+
+		// Update widgets/bar.
+		db.Update(func(tx *Tx) error {
+			var b = tx.Bucket([]byte("widgets"))
+			assert.NoError(t, b.Put([]byte("bar"), []byte("xxxx")))
+			return nil
+		})
+		mustCheck(db)
+
+		// Cause a split.
+		db.Update(func(tx *Tx) error {
+			var b = tx.Bucket([]byte("widgets"))
+			for i := 0; i < 10000; i++ {
+				assert.NoError(t, b.Put([]byte(strconv.Itoa(i)), []byte(strconv.Itoa(i))))
+			}
+			return nil
+		})
+		mustCheck(db)
+
+		// Insert into widgets/foo/baz.
+		db.Update(func(tx *Tx) error {
+			var b = tx.Bucket([]byte("widgets"))
+			assert.NoError(t, b.Bucket([]byte("foo")).Put([]byte("baz"), []byte("yyyy")))
+			return nil
+		})
+		mustCheck(db)
+
+		// Verify.
+		db.View(func(tx *Tx) error {
+			var b = tx.Bucket([]byte("widgets"))
+			assert.Equal(t, []byte("yyyy"), b.Bucket([]byte("foo")).Get([]byte("baz")))
+			assert.Equal(t, []byte("xxxx"), b.Get([]byte("bar")))
+			for i := 0; i < 10000; i++ {
+				assert.Equal(t, []byte(strconv.Itoa(i)), b.Get([]byte(strconv.Itoa(i))))
+			}
+			return nil
+		})
+	})
+}
+
 // Ensure that deleting a bucket using Delete() returns an error.
 func TestBucket_Delete_Bucket(t *testing.T) {
 	withOpenDB(func(db *DB, path string) {
@@ -550,31 +608,35 @@ func TestBucket_Stats_Large(t *testing.T) {
 	}
 
 	withOpenDB(func(db *DB, path string) {
-		db.Update(func(tx *Tx) error {
-			// Add bucket with lots of keys.
-			tx.CreateBucket([]byte("widgets"))
-			b := tx.Bucket([]byte("widgets"))
-			for i := 0; i < 100000; i++ {
-				b.Put([]byte(strconv.Itoa(i)), []byte(strconv.Itoa(i)))
-			}
-			return nil
-		})
+		var index int
+		for i := 0; i < 1000; i++ {
+			db.Update(func(tx *Tx) error {
+				// Add bucket with lots of keys.
+				b, _ := tx.CreateBucketIfNotExists([]byte("widgets"))
+				for i := 0; i < 100; i++ {
+					b.Put([]byte(strconv.Itoa(index)), []byte(strconv.Itoa(index)))
+					index++
+				}
+				return nil
+			})
+		}
 		mustCheck(db)
+
 		db.View(func(tx *Tx) error {
 			b := tx.Bucket([]byte("widgets"))
 			stats := b.Stats()
-			assert.Equal(t, stats.BranchPageN, 15)
-			assert.Equal(t, stats.BranchOverflowN, 0)
-			assert.Equal(t, stats.LeafPageN, 1281)
-			assert.Equal(t, stats.LeafOverflowN, 0)
-			assert.Equal(t, stats.KeyN, 100000)
-			assert.Equal(t, stats.Depth, 3)
+			assert.Equal(t, 19, stats.BranchPageN)
+			assert.Equal(t, 0, stats.BranchOverflowN)
+			assert.Equal(t, 1291, stats.LeafPageN)
+			assert.Equal(t, 0, stats.LeafOverflowN)
+			assert.Equal(t, 100000, stats.KeyN)
+			assert.Equal(t, 3, stats.Depth)
 			if os.Getpagesize() != 4096 {
 				// Incompatible page size
-				assert.Equal(t, stats.BranchInuse, 27289)
-				assert.Equal(t, stats.BranchAlloc, 61440)
-				assert.Equal(t, stats.LeafInuse, 2598276)
-				assert.Equal(t, stats.LeafAlloc, 5246976)
+				assert.Equal(t, 27289, stats.BranchInuse)
+				assert.Equal(t, 61440, stats.BranchAlloc)
+				assert.Equal(t, 2598276, stats.LeafInuse)
+				assert.Equal(t, 5246976, stats.LeafAlloc)
 			}
 			return nil
 		})
@@ -703,13 +765,12 @@ func TestBucket_Delete_Quick(t *testing.T) {
 				db.View(func(tx *Tx) error {
 					b := tx.Bucket([]byte("widgets"))
 					for j, exp := range items {
+						value := b.Get(exp.Key)
 						if j > i {
-							value := b.Get(exp.Key)
 							if !assert.Equal(t, exp.Value, value) {
 								t.FailNow()
 							}
 						} else {
-							value := b.Get(exp.Key)
 							if !assert.Nil(t, value) {
 								t.FailNow()
 							}
diff --git a/cursor.go b/cursor.go
index 13f2a06..520edd4 100644
--- a/cursor.go
+++ b/cursor.go
@@ -192,7 +192,7 @@ func (c *Cursor) last() {
 func (c *Cursor) search(key []byte, pgid pgid) {
 	p, n := c.bucket.pageNode(pgid)
 	if p != nil {
-		_assert((p.flags&(branchPageFlag|leafPageFlag)) != 0, "invalid page type: "+p.typ())
+		_assert((p.flags&(branchPageFlag|leafPageFlag)) != 0, "invalid page type: %d: %s", p.id, p.typ())
 	}
 	e := elemRef{page: p, node: n}
 	c.stack = append(c.stack, e)
diff --git a/node.go b/node.go
index 49de144..fd737ea 100644
--- a/node.go
+++ b/node.go
@@ -12,12 +12,20 @@ type node struct {
 	isLeaf     bool
 	unbalanced bool
 	key        []byte
-	depth      int
 	pgid       pgid
 	parent     *node
+	children   []*node
 	inodes     inodes
 }
 
+// root returns the top-level node this node is attached to.
+func (n *node) root() *node {
+	if n.parent == nil {
+		return n
+	}
+	return n.parent.root()
+}
+
 // minKeys returns the minimum number of inodes this node should have.
 func (n *node) minKeys() int {
 	if n.isLeaf {
@@ -185,12 +193,15 @@ func (n *node) write(p *page) {
 	}
 }
 
-// split divides up the node into appropriately sized nodes.
+// split breaks up a node into smaller nodes, if appropriate.
+// This should only be called from the spill() function.
 func (n *node) split(pageSize int) []*node {
+	var nodes = []*node{n}
+
 	// Ignore the split if the page doesn't have at least enough nodes for
 	// multiple pages or if the data can fit on a single page.
 	if len(n.inodes) <= (minKeysPerPage*2) || n.size() < pageSize {
-		return []*node{n}
+		return nodes
 	}
 
 	// Set fill threshold to 50%.
@@ -198,28 +209,106 @@ func (n *node) split(pageSize int) []*node {
 
 	// Group into smaller pages and target a given fill size.
 	size := pageHeaderSize
-	inodes := n.inodes
+	internalNodes := n.inodes
 	current := n
 	current.inodes = nil
-	var nodes []*node
 
-	for i, inode := range inodes {
+	// Loop over every inode and split once we reach our threshold.
+	for i, inode := range internalNodes {
 		elemSize := n.pageElementSize() + len(inode.key) + len(inode.value)
 
-		if len(current.inodes) >= minKeysPerPage && i < len(inodes)-minKeysPerPage && size+elemSize > threshold {
-			size = pageHeaderSize
+		// Split once we reach our threshold split size. However, this should
+		// only be done if we have enough keys for this node and we will have
+		// enough keys for the next node.
+		if len(current.inodes) >= minKeysPerPage && i < len(internalNodes)-minKeysPerPage && size+elemSize > threshold {
+			// If there's no parent then we need to create one.
+			if n.parent == nil {
+				n.parent = &node{bucket: n.bucket, children: []*node{n}}
+			}
+
+			// Create a new node and add it to the parent.
+			current = &node{bucket: n.bucket, isLeaf: n.isLeaf, parent: n.parent}
+			n.parent.children = append(n.parent.children, current)
 			nodes = append(nodes, current)
-			current = &node{bucket: n.bucket, isLeaf: n.isLeaf}
+
+			// Reset our running total back to zero (plus header size).
+			size = pageHeaderSize
+
+			// Update the statistics.
+			n.bucket.tx.stats.Split++
 		}
 
+		// Increase our running total of the size and append the inode.
 		size += elemSize
 		current.inodes = append(current.inodes, inode)
 	}
-	nodes = append(nodes, current)
 
 	return nodes
 }
 
+// spill writes the nodes to dirty pages and splits nodes as it goes.
+// Returns an error if dirty pages cannot be allocated.
+func (n *node) spill() error {
+	var tx = n.bucket.tx
+
+	// Spill child nodes first.
+	for _, child := range n.children {
+		if err := child.spill(); err != nil {
+			return err
+		}
+	}
+
+	// Add node's page to the freelist if it's not new.
+	if n.pgid > 0 {
+		tx.db.freelist.free(tx.id(), tx.page(n.pgid))
+	}
+
+	// Spill nodes by deepest first.
+	var nodes = n.split(tx.db.pageSize)
+	for _, node := range nodes {
+		// Allocate contiguous space for the node.
+		p, err := tx.allocate((node.size() / tx.db.pageSize) + 1)
+		if err != nil {
+			return err
+		}
+
+		// Write the node.
+		node.write(p)
+		node.pgid = p.id
+
+		// Insert into parent inodes.
+		if node.parent != nil {
+			var key = node.key
+			if key == nil {
+				key = node.inodes[0].key
+			}
+
+			node.parent.put(key, node.inodes[0].key, nil, node.pgid, 0)
+			node.key = node.inodes[0].key
+		}
+
+		// Update the statistics.
+		tx.stats.Spill++
+	}
+
+	// This is a special case where we need to write the parent if it is new
+	// and caused by a split in the root.
+	var parent = n.parent
+	if parent != nil && parent.pgid == 0 {
+		// Allocate contiguous space for the node.
+		p, err := tx.allocate((parent.size() / tx.db.pageSize) + 1)
+		if err != nil {
+			return err
+		}
+
+		// Write the new root.
+		parent.write(p)
+		parent.pgid = p.id
+	}
+
+	return nil
+}
+
 // rebalance attempts to combine the node with sibling nodes if the node fill
 // size is below a threshold or if there are not enough keys.
 func (n *node) rebalance() {
@@ -241,10 +330,11 @@ func (n *node) rebalance() {
 	if n.parent == nil {
 		// If root node is a branch and only has one node then collapse it.
 		if !n.isLeaf && len(n.inodes) == 1 {
-			// Move child's children up.
+			// Move root's child up.
 			child := n.bucket.nodes[n.inodes[0].pgid]
 			n.isLeaf = child.isLeaf
 			n.inodes = child.inodes[:]
+			n.children = child.children
 
 			// Reparent all child nodes being moved.
 			for _, inode := range n.inodes {
@@ -278,7 +368,9 @@ func (n *node) rebalance() {
 		if useNextSibling {
 			// Reparent and move node.
 			if child, ok := n.bucket.nodes[target.inodes[0].pgid]; ok {
+				child.parent.removeChild(child)
 				child.parent = n
+				child.parent.children = append(child.parent.children, child)
 			}
 			n.inodes = append(n.inodes, target.inodes[0])
 			target.inodes = target.inodes[1:]
@@ -289,7 +381,9 @@ func (n *node) rebalance() {
 		} else {
 			// Reparent and move node.
 			if child, ok := n.bucket.nodes[target.inodes[len(target.inodes)-1].pgid]; ok {
+				child.parent.removeChild(child)
 				child.parent = n
+				child.parent.children = append(child.parent.children, child)
 			}
 			n.inodes = append(n.inodes, inode{})
 			copy(n.inodes[1:], n.inodes)
@@ -309,26 +403,32 @@ func (n *node) rebalance() {
 		// Reparent all child nodes being moved.
 		for _, inode := range target.inodes {
 			if child, ok := n.bucket.nodes[inode.pgid]; ok {
+				child.parent.removeChild(child)
 				child.parent = n
+				child.parent.children = append(child.parent.children, child)
 			}
 		}
 
 		// Copy over inodes from target and remove target.
 		n.inodes = append(n.inodes, target.inodes...)
 		n.parent.del(target.key)
+		n.parent.removeChild(target)
 		delete(n.bucket.nodes, target.pgid)
 		target.free()
 	} else {
 		// Reparent all child nodes being moved.
 		for _, inode := range n.inodes {
 			if child, ok := n.bucket.nodes[inode.pgid]; ok {
+				child.parent.removeChild(child)
 				child.parent = target
+				child.parent.children = append(child.parent.children, child)
 			}
 		}
 
 		// Copy over inodes to target and remove node.
 		target.inodes = append(target.inodes, n.inodes...)
 		n.parent.del(n.key)
+		n.parent.removeChild(n)
 		n.parent.put(target.key, target.inodes[0].key, nil, target.pgid, 0)
 		delete(n.bucket.nodes, n.pgid)
 		n.free()
@@ -338,6 +438,17 @@ func (n *node) rebalance() {
 	n.parent.rebalance()
 }
 
+// removes a node from the list of in-memory children.
+// This does not affect the inodes.
+func (n *node) removeChild(target *node) {
+	for i, child := range n.children {
+		if child == target {
+			n.children = append(n.children[:i], n.children[i+1:]...)
+			return
+		}
+	}
+}
+
 // dereference causes the node to copy all its inode key/value references to heap memory.
 // This is required when the mmap is reallocated so inodes are not pointing to stale data.
 func (n *node) dereference() {
@@ -362,16 +473,10 @@ func (n *node) dereference() {
 func (n *node) free() {
 	if n.pgid != 0 {
 		n.bucket.tx.db.freelist.free(n.bucket.tx.id(), n.bucket.tx.page(n.pgid))
+		n.pgid = 0
 	}
 }
 
-// nodesByDepth sorts a list of branches by deepest first.
-type nodesByDepth []*node
-
-func (s nodesByDepth) Len() int           { return len(s) }
-func (s nodesByDepth) Swap(i, j int)      { s[i], s[j] = s[j], s[i] }
-func (s nodesByDepth) Less(i, j int) bool { return s[i].depth > s[j].depth }
-
 // inode represents an internal node inside of a node.
 // It can be used to point to elements in a page or point
 // to an element which hasn't been added to a page yet.
diff --git a/node_test.go b/node_test.go
index e58a544..f639376 100644
--- a/node_test.go
+++ b/node_test.go
@@ -85,7 +85,7 @@ func TestNode_write_LeafPage(t *testing.T) {
 // Ensure that a node can split into appropriate subgroups.
 func TestNode_split(t *testing.T) {
 	// Create a node.
-	n := &node{inodes: make(inodes, 0)}
+	n := &node{inodes: make(inodes, 0), bucket: &Bucket{tx: &Tx{}}}
 	n.put([]byte("00000001"), []byte("00000001"), []byte("0123456701234567"), 0, 0)
 	n.put([]byte("00000002"), []byte("00000002"), []byte("0123456701234567"), 0, 0)
 	n.put([]byte("00000003"), []byte("00000003"), []byte("0123456701234567"), 0, 0)
@@ -93,30 +93,30 @@ func TestNode_split(t *testing.T) {
 	n.put([]byte("00000005"), []byte("00000005"), []byte("0123456701234567"), 0, 0)
 
 	// Split between 2 & 3.
-	nodes := n.split(100)
+	n.split(100)
 
-	assert.Equal(t, len(nodes), 2)
-	assert.Equal(t, len(nodes[0].inodes), 2)
-	assert.Equal(t, len(nodes[1].inodes), 3)
+	var parent = n.parent
+	assert.Equal(t, len(parent.children), 2)
+	assert.Equal(t, len(parent.children[0].inodes), 2)
+	assert.Equal(t, len(parent.children[1].inodes), 3)
 }
 
 // Ensure that a page with the minimum number of inodes just returns a single node.
 func TestNode_split_MinKeys(t *testing.T) {
 	// Create a node.
-	n := &node{inodes: make(inodes, 0)}
+	n := &node{inodes: make(inodes, 0), bucket: &Bucket{tx: &Tx{}}}
 	n.put([]byte("00000001"), []byte("00000001"), []byte("0123456701234567"), 0, 0)
 	n.put([]byte("00000002"), []byte("00000002"), []byte("0123456701234567"), 0, 0)
 
 	// Split.
-	nodes := n.split(20)
-	assert.Equal(t, len(nodes), 1)
-	assert.Equal(t, len(nodes[0].inodes), 2)
+	n.split(20)
+	assert.Nil(t, n.parent)
 }
 
 // Ensure that a node that has keys that all fit on a page just returns one leaf.
 func TestNode_split_SinglePage(t *testing.T) {
 	// Create a node.
-	n := &node{inodes: make(inodes, 0)}
+	n := &node{inodes: make(inodes, 0), bucket: &Bucket{tx: &Tx{}}}
 	n.put([]byte("00000001"), []byte("00000001"), []byte("0123456701234567"), 0, 0)
 	n.put([]byte("00000002"), []byte("00000002"), []byte("0123456701234567"), 0, 0)
 	n.put([]byte("00000003"), []byte("00000003"), []byte("0123456701234567"), 0, 0)
@@ -124,7 +124,6 @@ func TestNode_split_SinglePage(t *testing.T) {
 	n.put([]byte("00000005"), []byte("00000005"), []byte("0123456701234567"), 0, 0)
 
 	// Split.
-	nodes := n.split(4096)
-	assert.Equal(t, len(nodes), 1)
-	assert.Equal(t, len(nodes[0].inodes), 5)
+	n.split(4096)
+	assert.Nil(t, n.parent)
 }
diff --git a/tx.go b/tx.go
index 612f493..7cdadae 100644
--- a/tx.go
+++ b/tx.go
@@ -363,8 +363,9 @@ type TxStats struct {
 	Rebalance     int           // number of node rebalances
 	RebalanceTime time.Duration // total time spent rebalancing
 
-	// Spill statistics.
-	Spill     int           // number of node spilled
+	// Split/Spill statistics.
+	Split     int           // number of nodes split
+	Spill     int           // number of nodes spilled
 	SpillTime time.Duration // total time spent spilling
 
 	// Write statistics.
@@ -380,6 +381,7 @@ func (s *TxStats) add(other *TxStats) {
 	s.NodeDeref += other.NodeDeref
 	s.Rebalance += other.Rebalance
 	s.RebalanceTime += other.RebalanceTime
+	s.Split += other.Split
 	s.Spill += other.Spill
 	s.SpillTime += other.SpillTime
 	s.Write += other.Write
@@ -398,6 +400,7 @@ func (s *TxStats) Sub(other *TxStats) TxStats {
 	diff.NodeDeref = s.NodeDeref - other.NodeDeref
 	diff.Rebalance = s.Rebalance - other.Rebalance
 	diff.RebalanceTime = s.RebalanceTime - other.RebalanceTime
+	diff.Split = s.Split - other.Split
 	diff.Spill = s.Spill - other.Spill
 	diff.SpillTime = s.SpillTime - other.SpillTime
 	diff.Write = s.Write - other.Write