diff --git a/util/checksum/checksum.go b/util/checksum/checksum.go
index 843440547100c..ef90f44e7cb00 100644
--- a/util/checksum/checksum.go
+++ b/util/checksum/checksum.go
@@ -42,11 +42,12 @@ var checksumReaderBufPool = sync.Pool{
 // | --    4B    -- | --  1020B  -- || --    4B    -- | --  1020B  -- || --    4B    -- | --   60B   -- |
 // | -- checksum -- | -- payload -- || -- checksum -- | -- payload -- || -- checksum -- | -- payload -- |
 type Writer struct {
-	err         error
-	w           io.WriteCloser
-	buf         []byte
-	payload     []byte
-	payloadUsed int
+	err                error
+	w                  io.WriteCloser
+	buf                []byte
+	payload            []byte
+	payloadUsed        int
+	flushedUserDataCnt int64
 }
 
 // NewWriter returns a new Writer which calculates and stores a CRC-32 checksum for the payload before
@@ -104,10 +105,21 @@ func (w *Writer) Flush() error {
 		w.err = err
 		return err
 	}
+	w.flushedUserDataCnt += int64(w.payloadUsed)
 	w.payloadUsed = 0
 	return nil
 }
 
+// GetCache returns the byte slice that holds the data not flushed to disk.
+func (w *Writer) GetCache() []byte {
+	return w.payload[:w.payloadUsed]
+}
+
+// GetCacheDataOffset return the user data offset in cache.
+func (w *Writer) GetCacheDataOffset() int64 {
+	return w.flushedUserDataCnt
+}
+
 // Close implements the io.Closer interface.
 func (w *Writer) Close() (err error) {
 	err = w.Flush()
diff --git a/util/checksum/checksum_test.go b/util/checksum/checksum_test.go
index b0de5b90586c9..1473903fbe080 100644
--- a/util/checksum/checksum_test.go
+++ b/util/checksum/checksum_test.go
@@ -651,3 +651,75 @@ func (s *testChecksumSuite) testTiCase3651and3652(c *check.C, encrypt bool) {
 	assertReadAt(0, make([]byte, 10200), nil, 10200, strings.Repeat("0123456789", 1020), f1)
 	assertReadAt(0, make([]byte, 10200), nil, 10200, strings.Repeat("0123456789", 1020), f2)
 }
+
+var checkFlushedData = func(c *check.C, f io.ReaderAt, off int64, readBufLen int, assertN int, assertErr interface{}, assertRes []byte) {
+	readBuf := make([]byte, readBufLen)
+	r := NewReader(f)
+	n, err := r.ReadAt(readBuf, off)
+	c.Assert(err, check.Equals, assertErr)
+	c.Assert(n, check.Equals, assertN)
+	c.Assert(bytes.Compare(readBuf, assertRes), check.Equals, 0)
+}
+
+func (s *testChecksumSuite) TestChecksumWriter(c *check.C) {
+	path := "checksum_TestChecksumWriter"
+	f, err := os.Create(path)
+	c.Assert(err, check.IsNil)
+	defer func() {
+		err = f.Close()
+		c.Assert(err, check.IsNil)
+		err = os.Remove(path)
+		c.Assert(err, check.IsNil)
+	}()
+
+	buf := bytes.NewBuffer(nil)
+	testData := "0123456789"
+	for i := 0; i < 100; i++ {
+		buf.WriteString(testData)
+	}
+
+	// Write 1000 bytes and flush.
+	w := NewWriter(f)
+	n, err := w.Write(buf.Bytes())
+	c.Assert(err, check.IsNil)
+	c.Assert(n, check.Equals, 1000)
+
+	err = w.Flush()
+	c.Assert(err, check.IsNil)
+	checkFlushedData(c, f, 0, 1000, 1000, nil, buf.Bytes())
+
+	// All data flushed, so no data in cache.
+	cacheOff := w.GetCacheDataOffset()
+	c.Assert(cacheOff, check.Equals, int64(1000))
+}
+
+func (s *testChecksumSuite) TestChecksumWriterAutoFlush(c *check.C) {
+	path := "checksum_TestChecksumWriterAutoFlush"
+	f, err := os.Create(path)
+	c.Assert(err, check.IsNil)
+	defer func() {
+		err = f.Close()
+		c.Assert(err, check.IsNil)
+		err = os.Remove(path)
+		c.Assert(err, check.IsNil)
+	}()
+
+	w := NewWriter(f)
+
+	buf := bytes.NewBuffer(nil)
+	testData := "0123456789"
+	for i := 0; i < 102; i++ {
+		buf.WriteString(testData)
+	}
+	n, err := w.Write(buf.Bytes())
+	c.Assert(err, check.IsNil)
+	c.Assert(n, check.Equals, len(buf.Bytes()))
+
+	// This write will trigger flush.
+	n, err = w.Write([]byte("0"))
+	c.Assert(err, check.IsNil)
+	c.Assert(n, check.Equals, 1)
+	checkFlushedData(c, f, 0, 1020, 1020, nil, buf.Bytes())
+	cacheOff := w.GetCacheDataOffset()
+	c.Assert(cacheOff, check.Equals, int64(len(buf.Bytes())))
+}
diff --git a/util/chunk/disk.go b/util/chunk/disk.go
index c7962c9aa9e9d..ef269213e9d0d 100644
--- a/util/chunk/disk.go
+++ b/util/chunk/disk.go
@@ -46,6 +46,9 @@ type ListInDisk struct {
 	diskTracker   *disk.Tracker // track disk usage.
 	numRowsInDisk int
 
+	checksumWriter *checksum.Writer
+	cipherWriter   *encrypt.Writer
+
 	// ctrCipher stores the key and nonce using by aes encrypt io layer
 	ctrCipher *encrypt.CtrCipher
 }
@@ -78,9 +81,11 @@ func (l *ListInDisk) initDiskFile() (err error) {
 		if err != nil {
 			return
 		}
-		underlying = encrypt.NewWriter(l.disk, l.ctrCipher)
+		l.cipherWriter = encrypt.NewWriter(l.disk, l.ctrCipher)
+		underlying = l.cipherWriter
 	}
-	l.w = checksum.NewWriter(underlying)
+	l.checksumWriter = checksum.NewWriter(underlying)
+	l.w = l.checksumWriter
 	l.bufFlushMutex = sync.RWMutex{}
 	return
 }
@@ -164,16 +169,16 @@ func (l *ListInDisk) GetChunk(chkIdx int) (*Chunk, error) {
 
 // GetRow gets a Row from the ListInDisk by RowPtr.
 func (l *ListInDisk) GetRow(ptr RowPtr) (row Row, err error) {
-	err = l.flush()
 	if err != nil {
 		return
 	}
 	off := l.offsets[ptr.ChkIdx][ptr.RowIdx]
 	var underlying io.ReaderAt = l.disk
 	if l.ctrCipher != nil {
-		underlying = encrypt.NewReader(l.disk, l.ctrCipher)
+		underlying = NewReaderWithCache(encrypt.NewReader(l.disk, l.ctrCipher), l.cipherWriter.GetCache(), l.cipherWriter.GetCacheDataOffset())
 	}
-	r := io.NewSectionReader(checksum.NewReader(underlying), off, l.offWrite-off)
+	checksumReader := NewReaderWithCache(checksum.NewReader(underlying), l.checksumWriter.GetCache(), l.checksumWriter.GetCacheDataOffset())
+	r := io.NewSectionReader(checksumReader, off, l.offWrite-off)
 	format := rowInDisk{numCol: len(l.fieldTypes)}
 	_, err = format.ReadFrom(r)
 	if err != nil {
@@ -367,3 +372,51 @@ func (format *diskFormatRow) toMutRow(fields []*types.FieldType) MutRow {
 	}
 	return MutRow{c: chk}
 }
+
+// ReaderWithCache helps to read data that has not be flushed to underlying layer.
+// By using ReaderWithCache, user can still write data into ListInDisk even after reading.
+type ReaderWithCache struct {
+	r        io.ReaderAt
+	cacheOff int64
+	cache    []byte
+}
+
+// NewReaderWithCache returns a ReaderWithCache.
+func NewReaderWithCache(r io.ReaderAt, cache []byte, cacheOff int64) *ReaderWithCache {
+	return &ReaderWithCache{
+		r:        r,
+		cacheOff: cacheOff,
+		cache:    cache,
+	}
+}
+
+// ReadAt implements the ReadAt interface.
+func (r *ReaderWithCache) ReadAt(p []byte, off int64) (readCnt int, err error) {
+	readCnt, err = r.r.ReadAt(p, off)
+	if err != io.EOF {
+		return readCnt, err
+	}
+
+	if len(p) == readCnt {
+		return readCnt, err
+	} else if len(p) < readCnt {
+		return readCnt, errors2.Trace(errors2.Errorf("cannot read more data than user requested"+
+			"(readCnt: %v, len(p): %v", readCnt, len(p)))
+	}
+
+	// When got here, user input is not filled fully, so we need read data from cache.
+	err = nil
+	p = p[readCnt:]
+	beg := off - r.cacheOff
+	if beg < 0 {
+		// This happens when only partial data of user requested resides in r.cache.
+		beg = 0
+	}
+	end := int(beg) + len(p)
+	if end > len(r.cache) {
+		err = io.EOF
+		end = len(r.cache)
+	}
+	readCnt += copy(p, r.cache[beg:end])
+	return readCnt, err
+}
diff --git a/util/chunk/disk_test.go b/util/chunk/disk_test.go
index 86461de5659c7..36750aa898244 100644
--- a/util/chunk/disk_test.go
+++ b/util/chunk/disk_test.go
@@ -14,12 +14,14 @@
 package chunk
 
 import (
+	"bytes"
 	"fmt"
 	"io"
 	"io/ioutil"
 	"math/rand"
 	"os"
 	"path/filepath"
+	"reflect"
 	"strconv"
 	"strings"
 	"testing"
@@ -30,6 +32,8 @@ import (
 	"github.com/pingcap/tidb/config"
 	"github.com/pingcap/tidb/types"
 	"github.com/pingcap/tidb/types/json"
+	"github.com/pingcap/tidb/util/checksum"
+	"github.com/pingcap/tidb/util/encrypt"
 )
 
 func initChunks(numChk, numRow int) ([]*Chunk, []*types.FieldType) {
@@ -219,6 +223,8 @@ func (s *testChunkSuite) TestListInDiskWithChecksum(c *check.C) {
 	})
 	testListInDisk(c)
 
+	testReaderWithCache(c)
+	testReaderWithCacheNoFlush(c)
 }
 
 func (s *testChunkSuite) TestListInDiskWithChecksumAndEncrypt(c *check.C) {
@@ -227,4 +233,129 @@ func (s *testChunkSuite) TestListInDiskWithChecksumAndEncrypt(c *check.C) {
 		conf.Security.SpilledFileEncryptionMethod = config.SpilledFileEncryptionMethodAES128CTR
 	})
 	testListInDisk(c)
+
+	testReaderWithCache(c)
+	testReaderWithCacheNoFlush(c)
+}
+
+// Following diagram describes the testdata we use to test:
+// 4 B: checksum of this segment.
+// 8 B: all columns' length, in the following example, we will only have one column.
+// 1012 B: data in file. because max length of each segment is 1024, so we only have 1020B for user payload.
+//
+//           Data in File                                    Data in mem cache
+// +------+------------------------------------------+ +-----------------------------+
+// |      |    1020B payload                         | |                             |
+// |4Bytes| +---------+----------------------------+ | |                             |
+// |checksum|8B collen| 1012B user data            | | |  12B remained user data     |
+// |      | +---------+----------------------------+ | |                             |
+// |      |                                          | |                             |
+// +------+------------------------------------------+ +-----------------------------+
+func testReaderWithCache(c *check.C) {
+	testData := "0123456789"
+	buf := bytes.NewBuffer(nil)
+	for i := 0; i < 102; i++ {
+		buf.WriteString(testData)
+	}
+	buf.WriteString("0123")
+
+	field := []*types.FieldType{types.NewFieldType(mysql.TypeString)}
+	chk := NewChunkWithCapacity(field, 1)
+	chk.AppendString(0, buf.String())
+	l := NewListInDisk(field)
+	err := l.Add(chk)
+	c.Assert(err, check.IsNil)
+
+	// Basic test for GetRow().
+	row, err := l.GetRow(RowPtr{0, 0})
+	c.Assert(err, check.IsNil)
+	c.Assert(row.GetDatumRow(field), check.DeepEquals, chk.GetRow(0).GetDatumRow(field))
+
+	var underlying io.ReaderAt = l.disk
+	if l.ctrCipher != nil {
+		underlying = NewReaderWithCache(encrypt.NewReader(l.disk, l.ctrCipher), l.cipherWriter.GetCache(), l.cipherWriter.GetCacheDataOffset())
+	}
+	checksumReader := NewReaderWithCache(checksum.NewReader(underlying), l.checksumWriter.GetCache(), l.checksumWriter.GetCacheDataOffset())
+
+	// Read all data.
+	data := make([]byte, 1024)
+	// Offset is 8, because we want to ignore col length.
+	readCnt, err := checksumReader.ReadAt(data, 8)
+	c.Assert(err, check.IsNil)
+	c.Assert(readCnt, check.Equals, 1024)
+	c.Assert(reflect.DeepEqual(data, buf.Bytes()), check.IsTrue)
+
+	// Only read data of mem cache.
+	data = make([]byte, 1024)
+	readCnt, err = checksumReader.ReadAt(data, 1020)
+	c.Assert(err, check.Equals, io.EOF)
+	c.Assert(readCnt, check.Equals, 12)
+	c.Assert(reflect.DeepEqual(data[:12], buf.Bytes()[1012:]), check.IsTrue)
+
+	// Read partial data of mem cache.
+	data = make([]byte, 1024)
+	readCnt, err = checksumReader.ReadAt(data, 1025)
+	c.Assert(err, check.Equals, io.EOF)
+	c.Assert(readCnt, check.Equals, 7)
+	c.Assert(reflect.DeepEqual(data[:7], buf.Bytes()[1017:]), check.IsTrue)
+
+	// Read partial data from both file and mem cache.
+	data = make([]byte, 1024)
+	readCnt, err = checksumReader.ReadAt(data, 1010)
+	c.Assert(err, check.Equals, io.EOF)
+	c.Assert(readCnt, check.Equals, 22)
+	c.Assert(reflect.DeepEqual(data[:22], buf.Bytes()[1002:]), check.IsTrue)
+
+	// Offset is too large, so no data is read.
+	data = make([]byte, 1024)
+	readCnt, err = checksumReader.ReadAt(data, 1032)
+	c.Assert(err, check.Equals, io.EOF)
+	c.Assert(readCnt, check.Equals, 0)
+	c.Assert(reflect.DeepEqual(data, make([]byte, 1024)), check.IsTrue)
+
+	// Only read 1 byte from mem cache.
+	data = make([]byte, 1024)
+	readCnt, err = checksumReader.ReadAt(data, 1031)
+	c.Assert(err, check.Equals, io.EOF)
+	c.Assert(readCnt, check.Equals, 1)
+	c.Assert(reflect.DeepEqual(data[:1], buf.Bytes()[1023:]), check.IsTrue)
+
+	// Test user requested data is small.
+	// Only request 10 bytes.
+	data = make([]byte, 10)
+	readCnt, err = checksumReader.ReadAt(data, 1010)
+	c.Assert(err, check.IsNil)
+	c.Assert(readCnt, check.Equals, 10)
+	c.Assert(reflect.DeepEqual(data, buf.Bytes()[1002:1012]), check.IsTrue)
+}
+
+// Here we test situations where size of data is small, so no data is flushed to disk.
+func testReaderWithCacheNoFlush(c *check.C) {
+	testData := "0123456789"
+
+	field := []*types.FieldType{types.NewFieldType(mysql.TypeString)}
+	chk := NewChunkWithCapacity(field, 1)
+	chk.AppendString(0, testData)
+	l := NewListInDisk(field)
+	err := l.Add(chk)
+	c.Assert(err, check.IsNil)
+
+	// Basic test for GetRow().
+	row, err := l.GetRow(RowPtr{0, 0})
+	c.Assert(err, check.IsNil)
+	c.Assert(row.GetDatumRow(field), check.DeepEquals, chk.GetRow(0).GetDatumRow(field))
+
+	var underlying io.ReaderAt = l.disk
+	if l.ctrCipher != nil {
+		underlying = NewReaderWithCache(encrypt.NewReader(l.disk, l.ctrCipher), l.cipherWriter.GetCache(), l.cipherWriter.GetCacheDataOffset())
+	}
+	checksumReader := NewReaderWithCache(checksum.NewReader(underlying), l.checksumWriter.GetCache(), l.checksumWriter.GetCacheDataOffset())
+
+	// Read all data.
+	data := make([]byte, 1024)
+	// Offset is 8, because we want to ignore col length.
+	readCnt, err := checksumReader.ReadAt(data, 8)
+	c.Assert(err, check.Equals, io.EOF)
+	c.Assert(readCnt, check.Equals, len(testData))
+	c.Assert(reflect.DeepEqual(data[:10], []byte(testData)), check.IsTrue)
 }
diff --git a/util/chunk/row_container_test.go b/util/chunk/row_container_test.go
index feed2290f38b6..a39346e34ff80 100644
--- a/util/chunk/row_container_test.go
+++ b/util/chunk/row_container_test.go
@@ -113,6 +113,28 @@ func (r *rowContainerTestSuite) TestSpillAction(c *check.C) {
 	rc.actionSpill.WaitForTest()
 	c.Assert(err, check.IsNil)
 	c.Assert(rc.AlreadySpilledSafeForTest(), check.Equals, true)
+
+	// Read
+	resChk, err := rc.GetChunk(0)
+	c.Assert(err, check.IsNil)
+	c.Assert(resChk.NumRows(), check.Equals, chk.NumRows())
+	for rowIdx := 0; rowIdx < resChk.NumRows(); rowIdx++ {
+		c.Assert(resChk.GetRow(rowIdx).GetDatumRow(fields), check.DeepEquals, chk.GetRow(rowIdx).GetDatumRow(fields))
+	}
+	// Write again
+	err = rc.Add(chk)
+	rc.actionSpill.WaitForTest()
+	c.Assert(err, check.IsNil)
+	c.Assert(rc.AlreadySpilledSafeForTest(), check.Equals, true)
+
+	// Read
+	resChk, err = rc.GetChunk(2)
+	c.Assert(err, check.IsNil)
+	c.Assert(resChk.NumRows(), check.Equals, chk.NumRows())
+	for rowIdx := 0; rowIdx < resChk.NumRows(); rowIdx++ {
+		c.Assert(resChk.GetRow(rowIdx).GetDatumRow(fields), check.DeepEquals, chk.GetRow(rowIdx).GetDatumRow(fields))
+	}
+
 	err = rc.Reset()
 	c.Assert(err, check.IsNil)
 }
diff --git a/util/encrypt/ase_layer.go b/util/encrypt/aes_layer.go
similarity index 91%
rename from util/encrypt/ase_layer.go
rename to util/encrypt/aes_layer.go
index 2bcea4373073f..a27d23da90fa6 100644
--- a/util/encrypt/ase_layer.go
+++ b/util/encrypt/aes_layer.go
@@ -71,11 +71,12 @@ func (ctr *CtrCipher) stream(counter uint64) cipher.Stream {
 
 // Writer implements an io.WriteCloser, it encrypt data using AES before writing to the underlying object.
 type Writer struct {
-	err          error
-	w            io.WriteCloser
-	n            int
-	buf          []byte
-	cipherStream cipher.Stream
+	err                error
+	w                  io.WriteCloser
+	n                  int
+	buf                []byte
+	cipherStream       cipher.Stream
+	flushedUserDataCnt int64
 }
 
 // NewWriter returns a new Writer which encrypt data using AES before writing to the underlying object.
@@ -123,6 +124,7 @@ func (w *Writer) Flush() error {
 	}
 	w.cipherStream.XORKeyStream(w.buf[:w.n], w.buf[:w.n])
 	n, err := w.w.Write(w.buf[:w.n])
+	w.flushedUserDataCnt += int64(n)
 	if n < w.n && err == nil {
 		err = io.ErrShortWrite
 	}
@@ -134,6 +136,16 @@ func (w *Writer) Flush() error {
 	return nil
 }
 
+// GetCache returns the byte slice that holds the data not flushed to disk.
+func (w *Writer) GetCache() []byte {
+	return w.buf[:w.n]
+}
+
+// GetCacheDataOffset return the user data offset in cache.
+func (w *Writer) GetCacheDataOffset() int64 {
+	return w.flushedUserDataCnt
+}
+
 // Close implements the io.Closer interface.
 func (w *Writer) Close() (err error) {
 	err = w.Flush()