diff --git a/fragment.go b/fragment.go
index e6c714937..22efb072e 100644
--- a/fragment.go
+++ b/fragment.go
@@ -1247,7 +1247,6 @@ func (f *fragment) rangeGT(bitDepth uint, predicate int64, allowEquality bool) (
 
 func (f *fragment) rangeGTUnsigned(filter *Row, bitDepth uint, predicate uint64, allowEquality bool) (*Row, error) {
 	keep := NewRow()
-
 	// Filter any bits that don't match the current bit value.
 	for i := int(bitDepth - 1); i >= 0; i-- {
 		row := f.row(uint64(bsiOffsetBit + i))
@@ -1260,12 +1259,12 @@ func (f *fragment) rangeGTUnsigned(filter *Row, bitDepth uint, predicate uint64,
 			if bit == 1 {
 				return keep, nil
 			}
-			return filter.Difference(filter.Difference(row).Difference(keep)), nil
+			return filter.Difference(filter.Difference(row, keep)), nil
 		}
 
 		// If bit is set then remove all unset columns not already kept.
 		if bit == 1 {
-			filter = filter.Difference(filter.Difference(row).Difference(keep))
+			filter = filter.Difference(filter.Difference(row, keep))
 			continue
 		}
 
@@ -1333,7 +1332,7 @@ func (f *fragment) rangeBetweenUnsigned(filter *Row, bitDepth uint, predicateMin
 		// GTE predicateMin
 		// If bit is set then remove all unset columns not already kept.
 		if bit1 == 1 {
-			filter = filter.Difference(filter.Difference(row).Difference(keep1))
+			filter = filter.Difference(filter.Difference(row, keep1))
 		} else {
 			// If bit is unset then add columns with set bit to keep.
 			// Don't bother to compute this on the final iteration.
diff --git a/go.mod b/go.mod
index 819785c3b..e8f70176e 100644
--- a/go.mod
+++ b/go.mod
@@ -8,36 +8,42 @@ require (
 	github.com/StackExchange/wmi v0.0.0-20190523213315-cbe66965904d // indirect
 	github.com/boltdb/bolt v1.3.1
 	github.com/cespare/xxhash v1.1.0
+	github.com/codahale/hdrhistogram v0.0.0-20161010025455-3a0bb77429bd // indirect
 	github.com/davecgh/go-spew v1.1.1
+	github.com/go-ole/go-ole v1.2.4 // indirect
 	github.com/gogo/protobuf v1.2.0
 	github.com/golang/protobuf v1.3.2
 	github.com/google/go-cmp v0.2.0
 	github.com/gorilla/handlers v1.3.0
 	github.com/gorilla/mux v1.7.0
 	github.com/hashicorp/memberlist v0.1.3
+	github.com/inconshreveable/mousetrap v1.0.0 // indirect
 	github.com/molecula/ext v0.0.0-20200103203257-8a458a73e8c2
 	github.com/molecula/extensions v0.0.0-20191218165536-562244600fd4
 	github.com/opentracing/opentracing-go v1.1.0
 	github.com/pelletier/go-toml v1.2.0
-	github.com/pilosa/pilosa v1.4.0
 	github.com/pkg/errors v0.8.1
 	github.com/prometheus/client_golang v0.9.3
 	github.com/prometheus/client_model v0.0.0-20190129233127-fd36f4220a90
 	github.com/remyoudompheng/bigfft v0.0.0-20190728182440-6a916e37a237 // indirect
 	github.com/satori/go.uuid v1.2.0
 	github.com/shirou/gopsutil v2.18.12+incompatible
+	github.com/shirou/w32 v0.0.0-20160930032740-bb4de0191aa4 // indirect
 	github.com/spf13/cobra v0.0.3
 	github.com/spf13/pflag v1.0.3
 	github.com/spf13/viper v1.3.1
+	github.com/uber-go/atomic v1.4.0 // indirect
 	github.com/uber/jaeger-client-go v2.16.0+incompatible
 	github.com/uber/jaeger-lib v2.2.0+incompatible // indirect
-	github.com/youtube/vitess v2.1.1+incompatible // indirect
+	go.uber.org/atomic v1.4.0 // indirect
+	golang.org/x/crypto v0.0.0-20190426145343-a29dc8fdc734 // indirect
 	golang.org/x/net v0.0.0-20190424112056-4829fb13d2c6
 	golang.org/x/sync v0.0.0-20190423024810-112230192c58
+	golang.org/x/sys v0.0.0-20190429190828-d89cdac9e872 // indirect
+	golang.org/x/text v0.3.2 // indirect
 	google.golang.org/grpc v1.24.0
 	modernc.org/mathutil v1.0.0
 	modernc.org/strutil v1.0.0
-	vitess.io/vitess v2.1.1+incompatible // indirect
 )
 
 go 1.13
diff --git a/go.sum b/go.sum
index 54cb1e76e..85e0347f6 100644
--- a/go.sum
+++ b/go.sum
@@ -7,7 +7,6 @@ github.com/DataDog/datadog-go v0.0.0-20180822151419-281ae9f2d895 h1:dmc/C8bpE5Vk
 github.com/DataDog/datadog-go v0.0.0-20180822151419-281ae9f2d895/go.mod h1:LButxg5PwREeZtORoXG3tL4fMGNddJ+vMq1mwgfaqoQ=
 github.com/OneOfOne/xxhash v1.2.2 h1:KMrpdQIwFcEqXDklaen+P1axHaj9BSKzvpUUfnHldSE=
 github.com/OneOfOne/xxhash v1.2.2/go.mod h1:HSdplMjZKSmBqAxg5vPj2TmRDmfkzw+cTzAElWljhcU=
-github.com/StackExchange/wmi v0.0.0-20181212234831-e0a55b97c705/go.mod h1:3eOhrUMpNV+6aFIbp5/iudMxNCF27Vw2OZgy4xEx0Fg=
 github.com/StackExchange/wmi v0.0.0-20190523213315-cbe66965904d h1:G0m3OIz70MZUWq3EgK3CesDbo8upS2Vm9/P3FtgI+Jk=
 github.com/StackExchange/wmi v0.0.0-20190523213315-cbe66965904d/go.mod h1:3eOhrUMpNV+6aFIbp5/iudMxNCF27Vw2OZgy4xEx0Fg=
 github.com/alecthomas/template v0.0.0-20160405071501-a0175ee3bccc/go.mod h1:LOuyumcjzFXgccqObfd/Ljyb9UuFJ6TxHnclSeseNhc=
@@ -106,8 +105,6 @@ github.com/pelletier/go-toml v1.2.0 h1:T5zMGML61Wp+FlcbWjRDT7yAxhJNAiPPLOFECq181
 github.com/pelletier/go-toml v1.2.0/go.mod h1:5z9KED0ma1S8pY6P1sdut58dfprrGBbd/94hg7ilaic=
 github.com/pilosa/memberlist v0.1.4-0.20190415211605-f6512523c021 h1:ERLyN4p3KS5Fk2ADsDENm2cq0+Lx6sF1sG8uwRlySpU=
 github.com/pilosa/memberlist v0.1.4-0.20190415211605-f6512523c021/go.mod h1:ajVTdAv/9Im8oMAAj5G31PhhMCZJV2pPBoIllUwCN7I=
-github.com/pilosa/pilosa v1.4.0 h1:nqHNIK4nDslFnem3yDp9R+6TgLdlkY9WdJD88Z83T8U=
-github.com/pilosa/pilosa v1.4.0/go.mod h1:NSTtTprtb5MSgCs4mcNqeQ2JdIMpInOi4DEImxGJeTs=
 github.com/pkg/errors v0.8.0/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
 github.com/pkg/errors v0.8.1 h1:iURUrRGxPUNPdy5/HRSm+Yj6okJ6UtLINN0Q9M4+h3I=
 github.com/pkg/errors v0.8.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
@@ -126,7 +123,6 @@ github.com/prometheus/procfs v0.0.0-20181005140218-185b4288413d/go.mod h1:c3At6R
 github.com/prometheus/procfs v0.0.0-20190507164030-5867b95ac084 h1:sofwID9zm4tzrgykg80hfFph1mryUeLRsUfoocVVmRY=
 github.com/prometheus/procfs v0.0.0-20190507164030-5867b95ac084/go.mod h1:TjEm7ze935MbeOT/UhFTIMYKhuLP4wbCsTZCD3I8kEA=
 github.com/prometheus/tsdb v0.7.1/go.mod h1:qhTCs0VvXwvX/y3TZrWD7rabWM+ijKTux40TwIPHuXU=
-github.com/remyoudompheng/bigfft v0.0.0-20190321074620-2f0d2b0e0001/go.mod h1:qqbHyh8v60DhA7CoWK5oRCqLrMHRGoxYCSS9EjAz6Eo=
 github.com/remyoudompheng/bigfft v0.0.0-20190728182440-6a916e37a237 h1:HQagqIiBmr8YXawX/le3+O26N+vPPC1PtjaF3mwnook=
 github.com/remyoudompheng/bigfft v0.0.0-20190728182440-6a916e37a237/go.mod h1:qqbHyh8v60DhA7CoWK5oRCqLrMHRGoxYCSS9EjAz6Eo=
 github.com/satori/go.uuid v1.2.0 h1:0uYX9dsZ2yD7q2RtLRtPSdGDWzjeM3TbMJP9utgA0ww=
@@ -159,13 +155,10 @@ github.com/uber-go/atomic v1.4.0 h1:yOuPqEq4ovnhEjpHmfFwsqBXDYbQeT6Nb0bwD6XnD5o=
 github.com/uber-go/atomic v1.4.0/go.mod h1:/Ct5t2lcmbJ4OSe/waGBoaVvVqtO0bmtfVNex1PFV8g=
 github.com/uber/jaeger-client-go v2.16.0+incompatible h1:Q2Pp6v3QYiocMxomCaJuwQGFt7E53bPYqEgug/AoBtY=
 github.com/uber/jaeger-client-go v2.16.0+incompatible/go.mod h1:WVhlPFC8FDjOFMMWRy2pZqQJSXxYSwNYOkTr/Z6d3Kk=
-github.com/uber/jaeger-lib v2.0.0+incompatible/go.mod h1:ComeNDZlWwrWnDv8aPp0Ba6+uUTzImX/AauajbLI56U=
 github.com/uber/jaeger-lib v2.2.0+incompatible h1:MxZXOiR2JuoANZ3J6DE/U0kSFv/eJ/GfSYVCjK7dyaw=
 github.com/uber/jaeger-lib v2.2.0+incompatible/go.mod h1:ComeNDZlWwrWnDv8aPp0Ba6+uUTzImX/AauajbLI56U=
 github.com/ugorji/go/codec v0.0.0-20181204163529-d75b2dcb6bc8/go.mod h1:VFNgLljTbGfSG7qAOspJ7OScBnGdDN/yBr0sguwnwf0=
 github.com/xordataexchange/crypt v0.0.3-0.20170626215501-b2862e3d0a77/go.mod h1:aYKd//L2LvnjZzWKhF00oedf4jCCReLcmhLdhm1A27Q=
-github.com/youtube/vitess v2.1.1+incompatible h1:SE+P7DNX/jw5RHFs5CHRhZQjq402EJFCD33JhzQMdDw=
-github.com/youtube/vitess v2.1.1+incompatible/go.mod h1:hpMim5/30F1r+0P8GGtB29d0gWHr0IZ5unS+CG0zMx8=
 go.uber.org/atomic v1.4.0 h1:cxzIVoETapQEqDhQu3QfnvXAV4AlzcvUCxkVUFw3+EU=
 go.uber.org/atomic v1.4.0/go.mod h1:gD2HeocX3+yG+ygLZcrzQJaqmWj9AIm7n08wl/qW/PE=
 golang.org/x/crypto v0.0.0-20180904163835-0709b304e793/go.mod h1:6SG95UA2DQfeDnfUPMdvaQW0Q7yPrPDi9nlGo2tz2b4=
@@ -223,5 +216,3 @@ modernc.org/mathutil v1.0.0 h1:93vKjrJopTPrtTNpZ8XIovER7iCIH1QU7wNbOQXC60I=
 modernc.org/mathutil v1.0.0/go.mod h1:wU0vUrJsVWBZ4P6e7xtFJEhFSNsfRLJ8H458uRjg03k=
 modernc.org/strutil v1.0.0 h1:XVFtQwFVwc02Wk+0L/Z/zDDXO81r5Lhe6iMKmGX3KhE=
 modernc.org/strutil v1.0.0/go.mod h1:lstksw84oURvj9y3tn8lGvRxyRC1S2+g5uuIzNfIOBs=
-vitess.io/vitess v2.1.1+incompatible h1:nuuGHiWYWpudD3gOCLeGzol2EJ25e/u5Wer2wV1O130=
-vitess.io/vitess v2.1.1+incompatible/go.mod h1:h4qvkyNYTOC0xI+vcidSWoka0gQAZc9ZPHbkHo48gP0=
diff --git a/roaring/roaring.go b/roaring/roaring.go
index 19a671dbf..1ddc5f798 100644
--- a/roaring/roaring.go
+++ b/roaring/roaring.go
@@ -833,16 +833,18 @@ func (b *Bitmap) unionIntoTargetSingle(target *Bitmap, other *Bitmap) {
 // Bitmap 4 |___X_______________________|     |      |___X_______________________|     |      |___X_______________________|
 //              _
 func (b *Bitmap) unionInPlace(others ...*Bitmap) {
+	const staticSize = 20
 	var (
 		requiredSliceSize = len(others)
-		// To avoid having to allocate a slice everytime, if the number of bitmaps
-		// being unioned is small enough we can just use this stack-allocated array.
-		staticHandledIters = [20]handledIter{}
+		// To avoid having to allocate a slice every time, if the number of bitmaps
+		// being unioned is small enough (i.e. smaller than staticSize), we can just
+		// use this stack-allocated array.
+		staticHandledIters = [staticSize]handledIter{}
 		bitmapIters        handledIters
 		target             = b
 	)
 
-	if requiredSliceSize <= 20 {
+	if requiredSliceSize <= staticSize {
 		bitmapIters = staticHandledIters[:0]
 	} else {
 		bitmapIters = make(handledIters, 0, requiredSliceSize)
@@ -993,10 +995,16 @@ func (b *Bitmap) unionInPlace(others ...*Bitmap) {
 }
 
 // Difference returns the difference of b and other.
-func (b *Bitmap) Difference(other *Bitmap) *Bitmap {
-	output := NewBitmap()
-	output.Source = b.Source
+func (b *Bitmap) Difference(other ...*Bitmap) *Bitmap {
+	output := b.singleDifference(other[0])
+	if len(other) > 1 {
+		output.DifferenceInPlace(other[1:]...)
+	}
+	return output
+}
 
+func (b *Bitmap) singleDifference(other *Bitmap) *Bitmap {
+	output := NewBitmap()
 	iiter, _ := b.Containers.Iterator(0)
 	jiter, _ := other.Containers.Iterator(0)
 	i, j := iiter.Next(), jiter.Next()
@@ -2668,6 +2676,7 @@ func (c *Container) arrayRemove(v uint16) (*Container, bool) {
 	}
 	// removing the last item? we can just return the empty container.
 	if c.N() == 1 {
+		c.n = 0
 		return nil, true
 	}
 	c = c.Thaw()
@@ -2684,6 +2693,7 @@ func (c *Container) bitmapRemove(v uint16) (*Container, bool) {
 	}
 	// removing the last item? we can just return the empty container.
 	if c.N() == 1 {
+		c.n = 0
 		return nil, true
 	}
 	c = c.Thaw()
@@ -2709,6 +2719,7 @@ func (c *Container) runRemove(v uint16) (*Container, bool) {
 	}
 	// removing the last item? we can just return the empty container.
 	if c.N() == 1 {
+		c.n = 0
 		return nil, true
 	}
 	c = c.Thaw()
@@ -5451,3 +5462,459 @@ type containerUnionSummaryStats struct {
 	// extremely dense containers.
 	hasMaxRange bool
 }
+
+// DifferenceInPlace returns the bitwise difference of b and others, modifying
+// b in place.
+func (b *Bitmap) DifferenceInPlace(others ...*Bitmap) {
+	bSize := b.Size()
+
+	// If b doesn't have any containers then return early.
+	if bSize == 0 {
+		return
+	}
+
+	const staticSize = 20
+	var (
+		requiredSliceSize = len(others)
+		// To avoid having to allocate a slice every time, if the number of bitmaps
+		// being differenced is small enough (i.e. smaller than staticSize), we can
+		// just use this stack-allocated array.
+		staticHandledIters  = [staticSize]handledIter{}
+		bitmapIters         handledIters
+		target              = b
+		removeContainerKeys = make([]uint64, 0, bSize)
+	)
+
+	if requiredSliceSize <= staticSize {
+		bitmapIters = staticHandledIters[:0]
+	} else {
+		bitmapIters = make(handledIters, 0, requiredSliceSize)
+	}
+
+	for _, other := range others {
+		otherIter, _ := other.Containers.Iterator(0)
+		if otherIter.Next() {
+			bitmapIters = append(bitmapIters, handledIter{
+				iter:    otherIter,
+				hasNext: true,
+			})
+		}
+	}
+
+	targetItr, _ := target.Containers.Iterator(0)
+	// Go through all the containers and remove the other bits
+	for targetItr.Next() {
+		targetKey, curContainer := targetItr.Value()
+		// Loop until every iters current value has been handled.
+		for _, iIter := range bitmapIters {
+			if !iIter.hasNext {
+				continue
+			}
+			iKey, iContainer := iIter.iter.Value()
+			for iKey < targetKey {
+				iIter.hasNext = iIter.iter.Next()
+				if iIter.hasNext {
+					iKey, iContainer = iIter.iter.Value()
+				} else {
+					break
+				}
+			}
+
+			if targetKey == iKey {
+				if curContainer.frozen() {
+					curContainer = curContainer.Clone()
+					b.Containers.Put(targetKey, curContainer)
+				}
+				curContainer.differenceInPlace(iContainer)
+				if curContainer.N() == 0 {
+					removeContainerKeys = append(removeContainerKeys, iKey)
+					break
+				}
+				iIter.hasNext = iIter.iter.Next()
+			}
+		}
+	}
+
+	for _, key := range removeContainerKeys {
+		b.Containers.Remove(key)
+
+	}
+	target.Containers.Repair()
+}
+
+func (c *Container) differenceInPlace(other *Container) {
+	if other.isArray() {
+		if c.isArray() {
+			differenceArrayArrayInPlace(c, other)
+		} else if c.isBitmap() {
+			differenceBitmapArrayInPlace(c, other)
+		} else if c.isRun() {
+			differenceRunArrayInPlace(c, other)
+		}
+	} else if other.isBitmap() {
+		if c.isArray() {
+			differenceArrayBitmapInPlace(c, other)
+		} else if c.isBitmap() {
+			differenceBitmapBitmapInPlace(c, other)
+		} else if c.isRun() {
+			differenceRunBitmapInPlace(c, other)
+		}
+	} else if other.isRun() {
+		if c.isArray() {
+			differenceArrayRunInPlace(c, other)
+		} else if c.isBitmap() {
+			differenceBitmapRunInPlace(c, other)
+		} else if c.isRun() {
+			differenceRunRunInPlace(c, other)
+		}
+	}
+}
+
+func differenceArrayArrayInPlace(c, other *Container) {
+	statsHit("differenceInPlace/ArrayArray")
+	aa, ab := c.array(), other.array()
+	na, nb := len(aa), len(ab)
+	if na == 0 || nb == 0 {
+		return
+	}
+	n := 0
+	for i, j := 0, 0; i < na; {
+		va := aa[i]
+		if j >= nb {
+			aa[n] = va
+			n++
+			i++
+			continue
+		}
+
+		vb := ab[j]
+		if va < vb {
+			aa[n] = va
+			n++
+			i++
+		} else if va > vb {
+			j++
+		} else {
+			i, j = i+1, j+1
+		}
+	}
+	aa = aa[:n]
+	c.setArray(aa)
+}
+
+func differenceArrayBitmapInPlace(c, other *Container) {
+	statsHit("differenceInPlace/ArrayBitmap")
+	aa := c.array()
+	n := 0
+	bitmap := other.bitmap()
+	if len(aa) == 0 || len(bitmap) == 0 {
+		return
+	}
+	for _, va := range aa {
+		bmidx := va / 64
+		bidx := va % 64
+		mask := uint64(1) << bidx
+		b := bitmap[bmidx]
+
+		if mask&^b > 0 {
+			aa[n] = va
+			n++
+		}
+	}
+	aa = aa[:n]
+	c.setArray(aa)
+}
+
+func differenceArrayRunInPlace(c, other *Container) {
+	statsHit("differenceInPlace/ArrayRun")
+
+	i := 0 // array index
+	j := 0 // run index
+	aa, rb := c.array(), other.runs()
+	if len(aa) == 0 || len(rb) == 0 {
+		return
+	}
+	n := 0
+
+	// handle overlap
+	for i < len(aa) {
+
+		// keep all array elements before beginning of runs
+		if aa[i] < rb[j].start {
+			aa[n] = aa[i]
+			n++
+			i++
+			continue
+		}
+
+		// if array element in run, skip it
+		if aa[i] >= rb[j].start && aa[i] <= rb[j].last {
+			i++
+			continue
+		}
+
+		// if array element larger than current run, check next run
+		if aa[i] > rb[j].last {
+			j++
+			if j == len(rb) {
+				break
+			}
+		}
+	}
+	for ; i < len(aa); i++ {
+		aa[n] = aa[i]
+		n++
+	}
+	aa = aa[:n]
+	c.setArray(aa)
+}
+
+func differenceBitmapArrayInPlace(c, other *Container) {
+	statsHit("differenceInPlace/BitmapArray")
+	bitmap := c.bitmap()
+	ab := other.array()
+	if len(bitmap) == 0 || len(ab) == 0 {
+		return
+	}
+
+	n := c.N()
+	for _, v := range ab {
+		if c.bitmapContains(v) {
+			bitmap[v/64] &^= (uint64(1) << uint(v%64))
+			n--
+		}
+	}
+	c.setN(n)
+	if n < ArrayMaxSize {
+		c.bitmapToArray() // With This Work
+	}
+}
+
+func differenceBitmapBitmapInPlace(c, other *Container) {
+	statsHit("differenceInPlace/BitmapBitmap")
+	// local variables added to prevent BCE checks in loop
+	// see https://go101.org/article/bounds-check-elimination.html
+	a := c.bitmap()
+	b := other.bitmap()
+	if len(a) == 0 || len(b) == 0 {
+		return
+	}
+
+	var (
+		ab = a[:bitmapN]
+		bb = b[:bitmapN]
+		n  int32
+	)
+
+	for i := 0; i < bitmapN; i++ {
+		ab[i] = ab[i] & (^bb[i])
+		n += int32(popcount(ab[i]))
+	}
+	c.setN(n)
+	if n < ArrayMaxSize {
+		c.bitmapToArray() // Will this work?
+	}
+}
+
+func differenceBitmapRunInPlace(c, other *Container) {
+	statsHit("differenceInPlace/BitmapRun")
+	if len(c.bitmap()) == 0 {
+		return
+	}
+	for _, run := range other.runs() {
+		c.bitmapZeroRange(uint64(run.start), uint64(run.last)+1)
+	}
+}
+
+func differenceRunArrayInPlace(c, other *Container) {
+	statsHit("differenceInPlace/RunArray")
+	ra, ab := c.runs(), other.array()
+	if len(ra) == 0 || len(ab) == 0 {
+		return
+	}
+	runs := make([]interval16, 0, len(ra))
+	bidx := 0
+	vb := ab[bidx]
+
+RUNLOOP:
+	for _, run := range ra {
+		start := run.start
+		for vb < run.start {
+			bidx++
+			if bidx >= len(ab) {
+				break
+			}
+			vb = ab[bidx]
+		}
+		for vb >= run.start && vb <= run.last {
+			if vb == start {
+				if vb == 65535 { // overflow
+					break RUNLOOP
+				}
+				start++
+				bidx++
+				if bidx >= len(ab) {
+					break
+				}
+				vb = ab[bidx]
+				continue
+			}
+			runs = append(runs, interval16{start: start, last: vb - 1})
+			if vb == 65535 { // overflow
+				break RUNLOOP
+			}
+			start = vb + 1
+			bidx++
+			if bidx >= len(ab) {
+				break
+			}
+			vb = ab[bidx]
+		}
+
+		if start <= run.last {
+			runs = append(runs, interval16{start: start, last: run.last})
+		}
+	}
+	c.setRuns(runs)
+	c.n = 0
+	for _, run := range runs {
+		c.n += int32(run.last-run.start) + 1
+	}
+	c.optimize()
+}
+
+func differenceRunBitmapInPlace(c, other *Container) {
+	statsHit("differenceInPlace/RunBitmap")
+	ra := c.runs()
+	if len(ra) == 0 || len(other.bitmap()) == 0 {
+		return
+	}
+	// If a is full, difference is the flip of b.
+	if len(ra) > 0 && ra[0].start == 0 && ra[0].last == 65535 {
+		clone := other.Clone()
+		bitmap := clone.bitmap()
+		for i, word := range other.bitmap() {
+			bitmap[i] = ^word
+		}
+		c.setTyp(containerBitmap)
+		c.setMapped(false)
+		c.setBitmap(bitmap)
+		c.setN(c.count())
+		return
+	}
+	runs := make([]interval16, 0, len(ra))
+	for _, inputRun := range ra {
+		run := inputRun
+		add := true
+		for bit := inputRun.start; bit <= inputRun.last; bit++ {
+			if other.bitmapContains(bit) {
+				if run.start == bit {
+					if bit == 65535 { //overflow
+						add = false
+					}
+
+					run.start++
+				} else if bit == run.last {
+					run.last--
+				} else {
+					run.last = bit - 1
+					if run.last >= run.start {
+						runs = append(runs, run)
+					}
+					run.start = bit + 1
+					run.last = inputRun.last
+				}
+				if run.start > run.last {
+					break
+				}
+			}
+
+			if bit == 65535 { //overflow
+				break
+			}
+		}
+		if run.start <= run.last {
+			if add {
+				runs = append(runs, run)
+			}
+		}
+	}
+
+	c.setRuns(runs)
+	c.n = 0
+	for _, run := range runs {
+		c.n += int32(run.last-run.start) + 1
+	}
+	if c.N() < ArrayMaxSize && int32(len(runs)) > c.N()/2 {
+		c.runToArray()
+	} else if len(runs) > runMaxSize {
+		c.runToBitmap()
+	}
+}
+
+func differenceRunRunInPlace(c, other *Container) {
+	statsHit("differenceInPlace/RunRun")
+
+	ra, rb := c.runs(), other.runs()
+	if len(ra) == 0 || len(rb) == 0 {
+		return
+	}
+	apos := 0 // current a-run index
+	bpos := 0 // current b-run index
+	astart := ra[apos].start
+	alast := ra[apos].last
+	bstart := rb[bpos].start
+	blast := rb[bpos].last
+	alen := len(ra)
+	blen := len(rb)
+
+	runs := make([]interval16, 0, alen+blen) // TODO allocate max then truncate? or something else
+	// cardinality upper bound: sum of number of runs
+	// each B-run could split an A-run in two, up to len(b.runs) times
+
+	for apos < alen && bpos < blen {
+		switch {
+		case alast < bstart:
+			// current A-run entirely precedes current B-run: keep full A-run, advance to next A-run
+			runs = append(runs, interval16{start: astart, last: alast})
+			apos++
+			if apos < alen {
+				astart = ra[apos].start
+				alast = ra[apos].last
+			}
+		case blast < astart:
+			// current B-run entirely precedes current A-run: advance to next B-run
+			bpos++
+			if bpos < blen {
+				bstart = rb[bpos].start
+				blast = rb[bpos].last
+			}
+		default:
+			// overlap
+			if astart < bstart {
+				runs = append(runs, interval16{start: astart, last: bstart - 1})
+			}
+			if alast > blast {
+				astart = blast + 1
+			} else {
+				apos++
+				if apos < alen {
+					astart = ra[apos].start
+					alast = ra[apos].last
+				}
+			}
+		}
+	}
+	if apos < alen {
+		runs = append(runs, interval16{start: astart, last: alast})
+		apos++
+		if apos < alen {
+			runs = append(runs, ra[apos:]...)
+		}
+	}
+	c.setRuns(runs)
+	c.n = 0
+	for _, run := range runs {
+		c.n += int32(run.last-run.start) + 1
+	}
+}
diff --git a/roaring/roaring_internal_test.go b/roaring/roaring_internal_test.go
index 0ecd73021..f8476459f 100644
--- a/roaring/roaring_internal_test.go
+++ b/roaring/roaring_internal_test.go
@@ -2710,6 +2710,15 @@ func unionInPlaceWrapper(a, b *Container) *Container {
 	return out.Containers.Get(0)
 }
 
+func differenceInPlaceWrapper(a, b *Container) *Container {
+	out := NewBitmap()
+	out.Containers.Put(0, a.Clone())
+	B := NewBitmap()
+	B.Containers.Put(0, b)
+	out.DifferenceInPlace(B)
+	return out.Containers.Get(0)
+}
+
 func TestContainerCombinations(t *testing.T) {
 
 	cts := setupContainerTests()
@@ -3283,6 +3292,117 @@ func TestContainerCombinations(t *testing.T) {
 		{flip, "outerBitsSet", "", "innerBitsSet"},
 		{flip, "oddBitsSet", "", "evenBitsSet"},
 		{flip, "evenBitsSet", "", "oddBitsSet"},
+
+		// differenceInPlace
+		{differenceInPlaceWrapper, "empty", "empty", "empty"},
+		{differenceInPlaceWrapper, "empty", "full", "empty"},
+		{differenceInPlaceWrapper, "empty", "firstBitSet", "empty"},
+		{differenceInPlaceWrapper, "empty", "lastBitSet", "empty"},
+		{differenceInPlaceWrapper, "empty", "firstBitUnset", "empty"},
+		{differenceInPlaceWrapper, "empty", "lastBitUnset", "empty"},
+		{differenceInPlaceWrapper, "empty", "innerBitsSet", "empty"},
+		{differenceInPlaceWrapper, "empty", "outerBitsSet", "empty"},
+		{differenceInPlaceWrapper, "empty", "oddBitsSet", "empty"},
+		{differenceInPlaceWrapper, "empty", "evenBitsSet", "empty"},
+		//
+		{differenceInPlaceWrapper, "full", "empty", "full"},
+		{differenceInPlaceWrapper, "full", "full", "empty"},
+		{differenceInPlaceWrapper, "full", "firstBitSet", "firstBitUnset"},
+		{differenceInPlaceWrapper, "full", "lastBitSet", "lastBitUnset"},
+		{differenceInPlaceWrapper, "full", "firstBitUnset", "firstBitSet"},
+		{differenceInPlaceWrapper, "full", "lastBitUnset", "lastBitSet"},
+		{differenceInPlaceWrapper, "full", "innerBitsSet", "outerBitsSet"},
+		{differenceInPlaceWrapper, "full", "outerBitsSet", "innerBitsSet"},
+		{differenceInPlaceWrapper, "full", "oddBitsSet", "evenBitsSet"},
+		{differenceInPlaceWrapper, "full", "evenBitsSet", "oddBitsSet"},
+		//
+		{differenceInPlaceWrapper, "firstBitSet", "empty", "firstBitSet"},
+		{differenceInPlaceWrapper, "firstBitSet", "full", "empty"},
+		{differenceInPlaceWrapper, "firstBitSet", "firstBitSet", "empty"},
+		{differenceInPlaceWrapper, "firstBitSet", "lastBitSet", "firstBitSet"},
+		{differenceInPlaceWrapper, "firstBitSet", "firstBitUnset", "firstBitSet"},
+		{differenceInPlaceWrapper, "firstBitSet", "lastBitUnset", "empty"},
+		{differenceInPlaceWrapper, "firstBitSet", "innerBitsSet", "firstBitSet"},
+		{differenceInPlaceWrapper, "firstBitSet", "outerBitsSet", "empty"},
+		{differenceInPlaceWrapper, "firstBitSet", "oddBitsSet", "firstBitSet"},
+		{differenceInPlaceWrapper, "firstBitSet", "evenBitsSet", "empty"},
+		//
+		{differenceInPlaceWrapper, "lastBitSet", "empty", "lastBitSet"},
+		{differenceInPlaceWrapper, "lastBitSet", "full", "empty"},
+		{differenceInPlaceWrapper, "lastBitSet", "firstBitSet", "lastBitSet"},
+		{differenceInPlaceWrapper, "lastBitSet", "lastBitSet", "empty"},
+		{differenceInPlaceWrapper, "lastBitSet", "firstBitUnset", "empty"},
+		{differenceInPlaceWrapper, "lastBitSet", "lastBitUnset", "lastBitSet"},
+		{differenceInPlaceWrapper, "lastBitSet", "innerBitsSet", "lastBitSet"},
+		{differenceInPlaceWrapper, "lastBitSet", "outerBitsSet", "empty"},
+		{differenceInPlaceWrapper, "lastBitSet", "oddBitsSet", "empty"},
+		{differenceInPlaceWrapper, "lastBitSet", "evenBitsSet", "lastBitSet"},
+		//
+		{differenceInPlaceWrapper, "firstBitUnset", "empty", "firstBitUnset"},
+		{differenceInPlaceWrapper, "firstBitUnset", "full", "empty"},
+		{differenceInPlaceWrapper, "firstBitUnset", "firstBitSet", "firstBitUnset"},
+		{differenceInPlaceWrapper, "firstBitUnset", "lastBitSet", "innerBitsSet"},
+		{differenceInPlaceWrapper, "firstBitUnset", "firstBitUnset", "empty"},
+		{differenceInPlaceWrapper, "firstBitUnset", "lastBitUnset", "lastBitSet"},
+		{differenceInPlaceWrapper, "firstBitUnset", "innerBitsSet", "lastBitSet"},
+		{differenceInPlaceWrapper, "firstBitUnset", "outerBitsSet", "innerBitsSet"},
+		//{differenceInPlaceWrapper, "firstBitUnset", "oddBitsSet", ""},
+		{differenceInPlaceWrapper, "firstBitUnset", "evenBitsSet", "oddBitsSet"},
+		//
+		{differenceInPlaceWrapper, "lastBitUnset", "empty", "lastBitUnset"},
+		{differenceInPlaceWrapper, "lastBitUnset", "full", "empty"},
+		{differenceInPlaceWrapper, "lastBitUnset", "firstBitSet", "innerBitsSet"},
+		{differenceInPlaceWrapper, "lastBitUnset", "lastBitSet", "lastBitUnset"},
+		{differenceInPlaceWrapper, "lastBitUnset", "firstBitUnset", "firstBitSet"},
+		{differenceInPlaceWrapper, "lastBitUnset", "lastBitUnset", "empty"},
+		{differenceInPlaceWrapper, "lastBitUnset", "innerBitsSet", "firstBitSet"},
+		{differenceInPlaceWrapper, "lastBitUnset", "outerBitsSet", "innerBitsSet"},
+		{differenceInPlaceWrapper, "lastBitUnset", "oddBitsSet", "evenBitsSet"},
+		//{differenceInPlaceWrapper, "lastBitUnset", "evenBitsSet", ""},
+		//
+		{differenceInPlaceWrapper, "innerBitsSet", "empty", "innerBitsSet"},
+		{differenceInPlaceWrapper, "innerBitsSet", "full", "empty"},
+		{differenceInPlaceWrapper, "innerBitsSet", "firstBitSet", "innerBitsSet"},
+		{differenceInPlaceWrapper, "innerBitsSet", "lastBitSet", "innerBitsSet"},
+		{differenceInPlaceWrapper, "innerBitsSet", "firstBitUnset", "empty"},
+		{differenceInPlaceWrapper, "innerBitsSet", "lastBitUnset", "empty"},
+		{differenceInPlaceWrapper, "innerBitsSet", "innerBitsSet", "empty"},
+		{differenceInPlaceWrapper, "innerBitsSet", "outerBitsSet", "innerBitsSet"},
+		//{differenceInPlaceWrapper, "innerBitsSet", "oddBitsSet", ""},
+		//{differenceInPlaceWrapper, "innerBitsSet", "evenBitsSet", ""},
+		//
+		{differenceInPlaceWrapper, "outerBitsSet", "empty", "outerBitsSet"},
+		{differenceInPlaceWrapper, "outerBitsSet", "full", "empty"},
+		{differenceInPlaceWrapper, "outerBitsSet", "firstBitSet", "lastBitSet"},
+		{differenceInPlaceWrapper, "outerBitsSet", "lastBitSet", "firstBitSet"},
+		{differenceInPlaceWrapper, "outerBitsSet", "firstBitUnset", "firstBitSet"},
+		{differenceInPlaceWrapper, "outerBitsSet", "lastBitUnset", "lastBitSet"},
+		{differenceInPlaceWrapper, "outerBitsSet", "innerBitsSet", "outerBitsSet"},
+		{differenceInPlaceWrapper, "outerBitsSet", "outerBitsSet", "empty"},
+		{differenceInPlaceWrapper, "outerBitsSet", "oddBitsSet", "firstBitSet"},
+		{differenceInPlaceWrapper, "outerBitsSet", "evenBitsSet", "lastBitSet"},
+		//
+		{differenceInPlaceWrapper, "oddBitsSet", "empty", "oddBitsSet"},
+		{differenceInPlaceWrapper, "oddBitsSet", "full", "empty"},
+		{differenceInPlaceWrapper, "oddBitsSet", "firstBitSet", "oddBitsSet"},
+		//{differenceInPlaceWrapper, "oddBitsSet", "lastBitSet", ""},
+		{differenceInPlaceWrapper, "oddBitsSet", "firstBitUnset", "empty"},
+		{differenceInPlaceWrapper, "oddBitsSet", "lastBitUnset", "lastBitSet"},
+		{differenceInPlaceWrapper, "oddBitsSet", "innerBitsSet", "lastBitSet"},
+		//{differenceInPlaceWrapper, "oddBitsSet", "outerBitsSet", ""},
+		{differenceInPlaceWrapper, "oddBitsSet", "oddBitsSet", "empty"},
+		{differenceInPlaceWrapper, "oddBitsSet", "evenBitsSet", "oddBitsSet"},
+		//
+		{differenceInPlaceWrapper, "evenBitsSet", "empty", "evenBitsSet"},
+		{differenceInPlaceWrapper, "evenBitsSet", "full", "empty"},
+		//{differenceInPlaceWrapper, "evenBitsSet", "firstBitSet", ""},
+		{differenceInPlaceWrapper, "evenBitsSet", "lastBitSet", "evenBitsSet"},
+		{differenceInPlaceWrapper, "evenBitsSet", "firstBitUnset", "firstBitSet"},
+		{differenceInPlaceWrapper, "evenBitsSet", "lastBitUnset", "empty"},
+		{differenceInPlaceWrapper, "evenBitsSet", "innerBitsSet", "firstBitSet"},
+		//{differenceInPlaceWrapper, "evenBitsSet", "outerBitsSet", ""},
+		{differenceInPlaceWrapper, "evenBitsSet", "oddBitsSet", "evenBitsSet"},
+		{differenceInPlaceWrapper, "evenBitsSet", "evenBitsSet", "empty"},
 	}
 	for _, testOp := range testOps {
 		for _, x := range containerTypes {
@@ -3896,3 +4016,12 @@ func TestBitmapAny(t *testing.T) {
 		t.Error("shouldn't be any left")
 	}
 }
+
+func TestDifferenceInPlace_N(t *testing.T) {
+	a := doContainer(containerRun, runFull())
+	b := doContainer(containerBitmap, bitmapFull())
+	r := differenceInPlaceWrapper(a, b)
+	if r.N() != 0 {
+		t.Error("expected difference of containers to have n=0")
+	}
+}
diff --git a/roaring/roaring_test.go b/roaring/roaring_test.go
index 836fd8de3..54ea5704a 100644
--- a/roaring/roaring_test.go
+++ b/roaring/roaring_test.go
@@ -1739,3 +1739,57 @@ func BenchmarkUnionBulk(b *testing.B) {
 			UnionInPlace(data.a1, data.a2, data.b, data.r1, data.r2)
 	}
 }
+
+func TestBitmap_DifferenceInPlace(t *testing.T) {
+	// array
+	arraybm := roaring.NewSliceBitmap()
+	for i := uint64(0); i < 1024; i += 4 {
+		_, _ = arraybm.Add((1 << 16) + i)
+	}
+	arraybm.Optimize()
+
+	// bitmap
+	bitmapbm := roaring.NewSliceBitmap()
+	for i := uint64(0); i < 16384; i += 2 {
+		_, _ = bitmapbm.Add((2 << 16) + i)
+	}
+	bitmapbm.Optimize()
+
+	// small run
+	smallrunbm := roaring.NewSliceBitmap()
+	for i := uint64(0); i < 1024; i++ {
+		_, _ = smallrunbm.Add((3 << 16) + i)
+	}
+	smallrunbm.Optimize()
+
+	// large run
+	largerunbm := roaring.NewSliceBitmap()
+	for i := uint64(0); i < 65535; i++ {
+		_, _ = largerunbm.Add((4 << 16) + i)
+	}
+	largerunbm.Optimize()
+
+	// test the difference in place
+	bm := testBM()
+	bm.DifferenceInPlace(arraybm, bitmapbm, smallrunbm, largerunbm)
+	if bm.Count() != 0 {
+		t.Fatalf("expected bitmap count to be 0, but got: %d", bm.Count())
+	}
+
+	bm = testBM()
+	bm.DifferenceInPlace(bitmapbm, smallrunbm, largerunbm)
+	if bm.Count() != 256 {
+		t.Fatalf("expected bitmap count to be 256, but got: %d", bm.Count())
+	}
+}
+
+func BenchmarkDifferencInPlace(b *testing.B) {
+	data := getBenchData(b)
+	for n := 0; n < b.N; n++ {
+		bm := roaring.NewBitmap()
+		bm.
+			UnionInPlace(data.a1, data.a2, data.b, data.r1, data.r2)
+		bm.
+			DifferenceInPlace(data.a1, data.r2, data.b, data.r1)
+	}
+}
diff --git a/row.go b/row.go
index 82a67cf8f..7c04f60e4 100644
--- a/row.go
+++ b/row.go
@@ -304,21 +304,26 @@ func (r *Row) GenericNaryOp(op ext.GenericBitmapOpBitmap, others []*Row, args ma
 }
 
 // Difference returns the diff of r and other.
-func (r *Row) Difference(other *Row) *Row {
-	var segments []rowSegment
+func (r *Row) Difference(others ...*Row) *Row {
+	var output []rowSegment
+	o := make(map[uint64][]*rowSegment)
 
-	itr := newMergeSegmentIterator(r.segments, other.segments)
-	for s0, s1 := itr.next(); s0 != nil || s1 != nil; s0, s1 = itr.next() {
-		if s0 == nil {
-			continue
-		} else if s1 == nil {
-			segments = append(segments, *s0)
-			continue
+	for x := range others {
+		for y := range others[x].segments {
+			segment := others[x].segments[y]
+			o[segment.shard] = append(o[segment.shard], &segment)
 		}
-		segments = append(segments, *s0.Difference(s1))
 	}
+	for _, segment := range r.segments {
 
-	return &Row{segments: segments}
+		dest, ok := o[segment.shard]
+		if ok {
+			output = append(output, *segment.Difference(dest...))
+		} else {
+			output = append(output, segment)
+		}
+	}
+	return &Row{segments: output}
 }
 
 // GenericUnary returns the results of a generic op on r.
@@ -576,9 +581,13 @@ func (s *rowSegment) GenericNaryOp(op ext.GenericBitmapOpBitmap, others []*rowSe
 }
 
 // Difference returns the diff of s and other.
-func (s *rowSegment) Difference(other *rowSegment) *rowSegment {
-	data := s.data.Difference(other.data)
-	data = data.Freeze()
+func (s *rowSegment) Difference(others ...*rowSegment) *rowSegment {
+	datas := make([]*roaring.Bitmap, len(others))
+	for i, other := range others {
+		datas[i] = other.data
+	}
+	data := s.data.Difference(datas...)
+	data.Freeze()
 
 	return &rowSegment{
 		data:     data,
diff --git a/row_test.go b/row_test.go
index 175faa4e7..a4740c897 100644
--- a/row_test.go
+++ b/row_test.go
@@ -138,3 +138,61 @@ func TestRow_Includes(t *testing.T) {
 		t.Fatalf("row should include %d", 2*ShardWidth)
 	}
 }
+
+func TestRow_DifferenceInPlace(t *testing.T) {
+	row0 := pilosa.NewRow(0)
+	row1 := pilosa.NewRow()
+	row2 := pilosa.NewRow(0)
+	res := row0.Difference(row1, row2)
+
+	if !row0.Includes(0) {
+		t.Fatal("row should include 0")
+	}
+	if res.Count() != 0 {
+		t.Fatal("results should be empty")
+	}
+}
+func eq(a, b []uint64) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	for i, v := range a {
+		if v != b[i] {
+			return false
+		}
+	}
+	return true
+}
+func TestRow_DifferenceInPlace2(t *testing.T) {
+	lucky := []uint64{1, 3, 7, 9, 13, 15, 21, 25, 31, 33, 37, 43, 49, 51, 63, 67, 69, 73, 75, 79, 87, 93, 99, 105, 111, 115, 127, 129, 133, 135, 141, 151, 159, 163, 169, 171, 189, 193, 195, 201, 205, 211, 219, 223, 231, 235, 237, 241, 259, 261, 267, 273, 283, 285, 289, 297}
+	src := pilosa.NewRow(lucky...)
+	row1 := pilosa.NewRow()
+	m := uint64(9)
+	for i := m; i <= lucky[len(lucky)-1]; i += m {
+		row1.SetBit(i)
+	}
+	m = uint64(3)
+	row2 := pilosa.NewRow()
+	for i := m; i <= lucky[len(lucky)-1]; i += m {
+		row2.SetBit(i)
+	}
+	row3 := pilosa.NewRow()
+	m = uint64(5)
+	for i := m; i <= lucky[len(lucky)-1]; i += m {
+		row3.SetBit(i)
+	}
+	row4 := pilosa.NewRow()
+	m = uint64(7)
+	for i := m; i <= lucky[len(lucky)-1]; i += m {
+		row4.SetBit(i)
+	}
+	res5 := src.Difference(row1, row2, row3, row4)
+	res6 := src.Difference(row4, row3, row2, row1)
+	res7 := src.Difference(row4).Difference(row3).Difference(row2).Difference(row1)
+	if !eq(res5.Columns(), res6.Columns()) {
+		t.Fatalf("results do not match: %v, %v", res5.Columns(), res6.Columns())
+	}
+	if !eq(res6.Columns(), res7.Columns()) {
+		t.Fatalf("results do not match: %v, %v", res6.Columns(), res7.Columns())
+	}
+}