diff --git a/db.go b/db.go
new file mode 100644
index 000000000..065205760
--- /dev/null
+++ b/db.go
@@ -0,0 +1,471 @@
+// Package tsdb implements a time series storage for float64 sample data.
+package tsdb
+
+import (
+	"encoding/binary"
+	"path/filepath"
+	"sync"
+	"time"
+
+	"github.com/fabxc/tsdb/chunks"
+	"github.com/prometheus/common/log"
+	"github.com/prometheus/common/model"
+)
+
+// DefaultOptions used for the DB.
+var DefaultOptions = &Options{
+	StalenessDelta: 5 * time.Minute,
+}
+
+// Options of the DB storage.
+type Options struct {
+	StalenessDelta time.Duration
+}
+
+// DB is a time series storage.
+type DB struct {
+	logger log.Logger
+	opts   *Options
+
+	memChunks   *memChunks
+	persistence *persistence
+	indexer     *indexer
+	stopc       chan struct{}
+}
+
+// Open or create a new DB.
+func Open(path string, l log.Logger, opts *Options) (*DB, error) {
+	if opts == nil {
+		opts = DefaultOptions
+	}
+
+	indexer, err := newMetricIndexer(filepath.Join(path, "index"), defaultIndexerQsize, defaultIndexerTimeout)
+	if err != nil {
+		return nil, err
+	}
+	persistence, err := newPersistence(filepath.Join(path, "chunks"), defaultIndexerQsize, defaultIndexerTimeout)
+	if err != nil {
+		return nil, err
+	}
+
+	mchunks := newMemChunks(l, indexer, persistence, 10, opts.StalenessDelta)
+	indexer.mc = mchunks
+	persistence.mc = mchunks
+
+	c := &DB{
+		logger:      l,
+		opts:        opts,
+		memChunks:   mchunks,
+		persistence: persistence,
+		indexer:     indexer,
+		stopc:       make(chan struct{}),
+	}
+	go c.memChunks.run(c.stopc)
+
+	return c, nil
+}
+
+// Close the storage and persist all writes.
+func (c *DB) Close() error {
+	close(c.stopc)
+	// TODO(fabxc): blocking further writes here necessary?
+	c.indexer.wait()
+	c.persistence.wait()
+
+	err0 := c.indexer.close()
+	err1 := c.persistence.close()
+	if err0 != nil {
+		return err0
+	}
+	return err1
+}
+
+// Append ingestes the samples in the scrape into the storage.
+func (c *DB) Append(scrape *Scrape) error {
+	// Sequentially add samples to in-memory chunks.
+	// TODO(fabxc): evaluate cost of making this atomic.
+	for _, s := range scrape.m {
+		if err := c.memChunks.append(s.met, scrape.ts, s.val); err != nil {
+			// TODO(fabxc): collect in multi error.
+			return err
+		}
+		// TODO(fabxc): increment ingested samples metric.
+	}
+	return nil
+}
+
+// memChunks holds the chunks that are currently being appended to.
+type memChunks struct {
+	logger         log.Logger
+	stalenessDelta time.Duration
+
+	mtx sync.RWMutex
+	// Chunks by their ID as accessed when retrieving a chunk ID from
+	// an index query.
+	chunks map[ChunkID]*chunkDesc
+	// The highest time slice chunks currently have. A new chunk can not
+	// be in a higher slice before all chunks with lower IDs have been
+	// added to the slice.
+	highTime model.Time
+
+	// Power of 2 of chunk shards.
+	num uint8
+	// Memory chunks sharded by leading bits of the chunk's metric's
+	// fingerprints. Used to quickly find chunks for new incoming samples
+	// where the metric is known but the chunk ID is not.
+	shards []*memChunksShard
+
+	indexer     *indexer
+	persistence *persistence
+}
+
+// newMemChunks returns a new memChunks sharded by n locks.
+func newMemChunks(l log.Logger, ix *indexer, p *persistence, n uint8, staleness time.Duration) *memChunks {
+	c := &memChunks{
+		logger:         l,
+		stalenessDelta: staleness,
+		num:            n,
+		chunks:         map[ChunkID]*chunkDesc{},
+		persistence:    p,
+		indexer:        ix,
+	}
+
+	if n > 63 {
+		panic("invalid shard power")
+	}
+
+	// Initialize 2^n shards.
+	for i := 0; i < 1<<n; i++ {
+		c.shards = append(c.shards, &memChunksShard{
+			descs: map[model.Fingerprint][]*chunkDesc{},
+			csize: 1024,
+		})
+	}
+	return c
+}
+
+func (mc *memChunks) run(stopc <-chan struct{}) {
+	ticker := time.NewTicker(10 * time.Minute)
+	defer ticker.Stop()
+
+	f := func() error {
+		for _, cs := range mc.shards {
+			mc.gc(cs)
+		}
+		// Wait for persistence and indexing to finish before reindexing
+		// memory chunks for the new time slice.
+		mc.persistence.wait()
+		mc.indexer.wait()
+
+		mc.mtx.Lock()
+		defer mc.mtx.Unlock()
+
+		curTimeSlice := timeSlice(model.Now())
+		// If the next time slice is in the future, we are done.
+		if curTimeSlice <= mc.highTime {
+			return nil
+		}
+
+		ids := make(ChunkIDs, 0, len(mc.chunks))
+		for id := range mc.chunks {
+			ids = append(ids, id)
+		}
+
+		if err := mc.indexer.reindexTime(ids, curTimeSlice); err != nil {
+			return err
+		}
+		mc.highTime = curTimeSlice
+		return nil
+	}
+
+	for {
+		select {
+		case <-ticker.C:
+			if err := f(); err != nil {
+				mc.logger.With("err", err).Error("memory chunk maintenance failed")
+			}
+		case <-stopc:
+			return
+		}
+	}
+}
+
+// gc writes stale and incomplete chunks to persistence and removes them
+// from the shard.
+func (mc *memChunks) gc(cs *memChunksShard) {
+	cs.RLock()
+	defer cs.RUnlock()
+
+	mint := model.Now().Add(-mc.stalenessDelta)
+
+	for fp, cdescs := range cs.descs {
+		for _, cd := range cdescs {
+			// If the last sample was added before the staleness delta, consider
+			// the chunk inactive and persist it.
+			if cd.lastSample.Timestamp.Before(mint) {
+				mc.persistence.enqueue(cd)
+				cs.del(fp, cd)
+			}
+		}
+	}
+	return
+}
+
+func (mc *memChunks) append(m model.Metric, ts model.Time, v model.SampleValue) error {
+	fp := m.FastFingerprint()
+	cs := mc.shards[fp>>(64-mc.num)]
+
+	cs.Lock()
+	defer cs.Unlock()
+
+	chkd, created := cs.get(fp, m)
+	if created {
+		mc.indexer.enqueue(chkd)
+	}
+	if err := chkd.append(ts, v); err != chunks.ErrChunkFull {
+		return err
+	}
+	// Chunk was full, remove it so a new head chunk can be created.
+	// TODO(fabxc): should we just remove them during maintenance if we set a 'persisted'
+	// flag?
+	// If we shutdown we work down the persistence queue before exiting, so we should
+	// lose no data. If we crash, the last snapshot will still have the chunk. Theoretically,
+	// deleting it here should not be a problem.
+	cs.del(fp, chkd)
+
+	mc.persistence.enqueue(chkd)
+
+	// Create a new chunk lazily and continue.
+	chkd, created = cs.get(fp, m)
+	if !created {
+		// Bug if the chunk was not newly created.
+		panic("expected newly created chunk")
+	}
+	mc.indexer.enqueue(chkd)
+
+	return chkd.append(ts, v)
+}
+
+type memChunksShard struct {
+	sync.RWMutex
+
+	// chunks holds chunk descriptors for one or more chunks
+	// with a given fingerprint.
+	descs map[model.Fingerprint][]*chunkDesc
+	csize int
+}
+
+// get returns the chunk descriptor for the given fingerprint/metric combination.
+// If none exists, a new chunk descriptor is created and true is returned.
+func (cs *memChunksShard) get(fp model.Fingerprint, m model.Metric) (*chunkDesc, bool) {
+	chks := cs.descs[fp]
+	for _, cd := range chks {
+		if cd != nil && cd.met.Equal(m) {
+			return cd, false
+		}
+	}
+	// None of the given chunks was for the metric, create a new one.
+	cd := &chunkDesc{
+		met:   m,
+		chunk: chunks.NewPlainChunk(cs.csize),
+	}
+	// Try inserting chunk in existing whole before appending.
+	for i, c := range chks {
+		if c == nil {
+			chks[i] = cd
+			return cd, true
+		}
+	}
+	cs.descs[fp] = append(chks, cd)
+	return cd, true
+}
+
+// del frees the field of the chunk descriptor for the fingerprint.
+func (cs *memChunksShard) del(fp model.Fingerprint, chkd *chunkDesc) {
+	for i, d := range cs.descs[fp] {
+		if d == chkd {
+			cs.descs[fp][i] = nil
+			return
+		}
+	}
+}
+
+// ChunkID is a unique identifier for a chunks.
+type ChunkID uint64
+
+func (id ChunkID) bytes() []byte {
+	b := make([]byte, 8)
+	binary.BigEndian.PutUint64(b, uint64(id))
+	return b
+}
+
+// ChunkIDs is a sortable list of chunk IDs.
+type ChunkIDs []ChunkID
+
+func (c ChunkIDs) Len() int           { return len(c) }
+func (c ChunkIDs) Swap(i, j int)      { c[i], c[j] = c[j], c[i] }
+func (c ChunkIDs) Less(i, j int) bool { return c[i] < c[j] }
+
+// chunkDesc wraps a plain data chunk and provides cached meta data about it.
+type chunkDesc struct {
+	id    ChunkID
+	met   model.Metric
+	chunk chunks.Chunk
+
+	// Caching fields.
+	firstTime  model.Time
+	lastSample model.SamplePair
+
+	app chunks.Appender // Current appender for the chunks.
+}
+
+func (cd *chunkDesc) append(ts model.Time, v model.SampleValue) error {
+	if cd.app == nil {
+		cd.app = cd.chunk.Appender()
+		// TODO(fabxc): set correctly once loading from snapshot is added.
+		cd.firstTime = ts
+	}
+	cd.lastSample.Timestamp = ts
+	cd.lastSample.Value = v
+
+	return cd.app.Append(ts, v)
+}
+
+// Scrape gathers samples for a single timestamp.
+type Scrape struct {
+	ts model.Time
+	m  []sample
+}
+
+type sample struct {
+	met model.Metric
+	val model.SampleValue
+}
+
+// Reset resets the scrape data and initializes it for a new scrape at
+// the given time. The underlying memory remains allocated for the next scrape.
+func (s *Scrape) Reset(ts model.Time) {
+	s.ts = ts
+	s.m = s.m[:0]
+}
+
+// Dump returns all samples that are part of the scrape.
+func (s *Scrape) Dump() []*model.Sample {
+	d := make([]*model.Sample, 0, len(s.m))
+	for _, sa := range s.m {
+		d = append(d, &model.Sample{
+			Metric:    sa.met,
+			Timestamp: s.ts,
+			Value:     sa.val,
+		})
+	}
+	return d
+}
+
+// Add adds a sample value for the given metric to the scrape.
+func (s *Scrape) Add(m model.Metric, v model.SampleValue) {
+	for ln, lv := range m {
+		if len(lv) == 0 {
+			delete(m, ln)
+		}
+	}
+	// TODO(fabxc): pre-sort added samples into the correct buckets
+	// of fingerprint shards so we only have to lock each memChunkShard once.
+	s.m = append(s.m, sample{met: m, val: v})
+}
+
+type chunkBatchProcessor struct {
+	processf func(...*chunkDesc) error
+
+	mtx    sync.RWMutex
+	logger log.Logger
+	q      []*chunkDesc
+
+	qcap    int
+	timeout time.Duration
+
+	timer   *time.Timer
+	trigger chan struct{}
+	empty   chan struct{}
+}
+
+func newChunkBatchProcessor(l log.Logger, cap int, to time.Duration) *chunkBatchProcessor {
+	if l == nil {
+		l = log.NewNopLogger()
+	}
+	p := &chunkBatchProcessor{
+		logger:  l,
+		qcap:    cap,
+		timeout: to,
+		timer:   time.NewTimer(to),
+		trigger: make(chan struct{}, 1),
+		empty:   make(chan struct{}),
+	}
+	// Start with closed channel so we don't block on wait if nothing
+	// has ever been indexed.
+	close(p.empty)
+
+	go p.run()
+	return p
+}
+
+func (p *chunkBatchProcessor) run() {
+	for {
+		// Process pending indexing batch if triggered
+		// or timeout since last indexing has passed.
+		select {
+		case <-p.trigger:
+		case <-p.timer.C:
+		}
+
+		if err := p.process(); err != nil {
+			p.logger.
+				With("err", err).With("num", len(p.q)).
+				Error("batch failed, dropping chunks descs")
+		}
+	}
+}
+
+func (p *chunkBatchProcessor) process() error {
+	// TODO(fabxc): locking the entire time will cause lock contention.
+	p.mtx.Lock()
+	defer p.mtx.Unlock()
+
+	if len(p.q) == 0 {
+		return nil
+	}
+	// Leave chunk descs behind whether successful or not.
+	defer func() {
+		p.q = p.q[:0]
+		close(p.empty)
+	}()
+
+	return p.processf(p.q...)
+}
+
+func (p *chunkBatchProcessor) enqueue(cds ...*chunkDesc) {
+	p.mtx.Lock()
+	defer p.mtx.Unlock()
+
+	if len(p.q) == 0 {
+		p.timer.Reset(p.timeout)
+		p.empty = make(chan struct{})
+	}
+
+	p.q = append(p.q, cds...)
+	if len(p.q) > p.qcap {
+		select {
+		case p.trigger <- struct{}{}:
+		default:
+			// If we cannot send a signal is already set.
+		}
+	}
+}
+
+// wait blocks until the queue becomes empty.
+func (p *chunkBatchProcessor) wait() {
+	p.mtx.RLock()
+	c := p.empty
+	p.mtx.RUnlock()
+	<-c
+}
diff --git a/db_test.go b/db_test.go
new file mode 100644
index 000000000..0436a3033
--- /dev/null
+++ b/db_test.go
@@ -0,0 +1,145 @@
+package tsdb
+
+import (
+	"fmt"
+	"io"
+	"io/ioutil"
+	"math/rand"
+	"os"
+	"testing"
+	"time"
+
+	"github.com/fabxc/tindex"
+	"github.com/prometheus/common/log"
+	"github.com/prometheus/common/model"
+	"github.com/prometheus/prometheus/storage/cinamon/chunk"
+	"github.com/prometheus/prometheus/storage/metric"
+	"github.com/stretchr/testify/require"
+)
+
+func TestE2E(t *testing.T) {
+	dir, err := ioutil.TempDir("", "cinamon_test")
+	require.NoError(t, err)
+	defer os.RemoveAll(dir)
+
+	c, err := Open(dir, log.Base(), nil)
+	require.NoError(t, err)
+
+	c.memChunks.indexer.timeout = 50 * time.Millisecond
+
+	// Set indexer size to be triggered exactly when we hit the limit.
+	// c.memChunks.indexer.qmax = 10
+
+	mets := generateMetrics(100000)
+	// var wg sync.WaitGroup
+	// for k := 0; k < len(mets)/100+1; k++ {
+	// 	wg.Add(1)
+	// 	go func(mets []model.Metric) {
+	var s Scrape
+	for i := 0; i < 2*64; i++ {
+		s.Reset(model.Time(i) * 100000)
+
+		for _, m := range mets {
+			s.Add(m, model.SampleValue(rand.Float64()))
+		}
+		require.NoError(t, c.Append(&s))
+	}
+	// 		wg.Done()
+	// 	}(mets[k*100 : (k+1)*100])
+	// }
+	// wg.Wait()
+
+	start := time.Now()
+	c.memChunks.indexer.wait()
+	fmt.Println("index wait", time.Since(start))
+
+	start = time.Now()
+	q, err := c.Querier()
+	require.NoError(t, err)
+	defer q.Close()
+
+	m1, err := metric.NewLabelMatcher(metric.Equal, "job", "somejob")
+	require.NoError(t, err)
+	m2, err := metric.NewLabelMatcher(metric.Equal, "label2", "value0")
+	require.NoError(t, err)
+	m3, err := metric.NewLabelMatcher(metric.Equal, "label4", "value0")
+	require.NoError(t, err)
+
+	it, err := q.Iterator(m1, m2, m3)
+	require.NoError(t, err)
+	res, err := tindex.ExpandIterator(it)
+	require.NoError(t, err)
+	fmt.Println("result len", len(res))
+
+	fmt.Println("querying", time.Since(start))
+}
+
+func generateMetrics(n int) (res []model.Metric) {
+	for i := 0; i < n; i++ {
+		res = append(res, model.Metric{
+			"job":    "somejob",
+			"label5": model.LabelValue(fmt.Sprintf("value%d", i%10)),
+			"label4": model.LabelValue(fmt.Sprintf("value%d", i%5)),
+			"label3": model.LabelValue(fmt.Sprintf("value%d", i%3)),
+			"label2": model.LabelValue(fmt.Sprintf("value%d", i%2)),
+			"label1": model.LabelValue(fmt.Sprintf("value%d", i)),
+		})
+	}
+	return res
+}
+
+func TestMemChunksShardGet(t *testing.T) {
+	cs := &memChunksShard{
+		descs: map[model.Fingerprint][]*chunkDesc{},
+		csize: 100,
+	}
+	cdesc1, created1 := cs.get(123, model.Metric{"x": "1"})
+	require.True(t, created1)
+	require.Equal(t, 1, len(cs.descs[123]))
+	require.Equal(t, &chunkDesc{
+		met:   model.Metric{"x": "1"},
+		chunk: chunk.NewPlainChunk(100),
+	}, cdesc1)
+
+	// Add colliding metric.
+	cdesc2, created2 := cs.get(123, model.Metric{"x": "2"})
+	require.True(t, created2)
+	require.Equal(t, 2, len(cs.descs[123]))
+	require.Equal(t, &chunkDesc{
+		met:   model.Metric{"x": "2"},
+		chunk: chunk.NewPlainChunk(100),
+	}, cdesc2)
+	// First chunk desc can still be retrieved correctly.
+	cdesc1, created1 = cs.get(123, model.Metric{"x": "1"})
+	require.False(t, created1)
+	require.Equal(t, &chunkDesc{
+		met:   model.Metric{"x": "1"},
+		chunk: chunk.NewPlainChunk(100),
+	}, cdesc1)
+}
+
+func TestChunkSeriesIterator(t *testing.T) {
+	newChunk := func(s []model.SamplePair) chunk.Chunk {
+		c := chunk.NewPlainChunk(1000)
+		app := c.Appender()
+		for _, sp := range s {
+			if err := app.Append(sp.Timestamp, sp.Value); err != nil {
+				t.Fatal(err)
+			}
+		}
+		return c
+	}
+	it := newChunkSeriesIterator(metric.Metric{}, []chunk.Chunk{
+		newChunk([]model.SamplePair{{1, 1}, {2, 2}, {3, 3}}),
+		newChunk([]model.SamplePair{{4, 4}, {5, 5}, {6, 6}}),
+		newChunk([]model.SamplePair{{7, 7}, {8, 8}, {9, 9}}),
+	})
+
+	var res []model.SamplePair
+	for sp, ok := it.Seek(0); ok; sp, ok = it.Next() {
+		fmt.Println(sp)
+		res = append(res, sp)
+	}
+	require.Equal(t, io.EOF, it.Err())
+	require.Equal(t, []model.SamplePair{{1, 1}, {2, 2}, {3, 3}, {4, 4}, {5, 5}, {6, 6}, {7, 7}, {8, 8}, {9, 9}}, res)
+}
diff --git a/index.go b/index.go
new file mode 100644
index 000000000..83506da5c
--- /dev/null
+++ b/index.go
@@ -0,0 +1,130 @@
+package tsdb
+
+import (
+	"sort"
+	"strconv"
+	"sync/atomic"
+	"time"
+
+	"github.com/fabxc/tsdb/index"
+	"github.com/prometheus/common/log"
+	"github.com/prometheus/common/model"
+)
+
+const (
+	defaultIndexerTimeout = 1 * time.Second
+	defaultIndexerQsize   = 500000
+)
+
+// indexer asynchronously indexes chunks in batches. It indexes all labels
+// of a chunk with a forward mapping and additionally indexes the chunk for
+// the time slice of its first sample.
+type indexer struct {
+	*chunkBatchProcessor
+
+	ix *index.Index
+	mc *memChunks
+}
+
+// Create batch indexer that creates new index documents
+// and indexes them by the metric fields.
+// Its post-indexing hook populates the in-memory chunk forward index.
+func newMetricIndexer(path string, qsz int, qto time.Duration) (*indexer, error) {
+	ix, err := index.Open(path, nil)
+	if err != nil {
+		return nil, err
+	}
+
+	i := &indexer{
+		ix:                  ix,
+		chunkBatchProcessor: newChunkBatchProcessor(log.Base(), qsz, qto),
+	}
+	i.chunkBatchProcessor.processf = i.index
+
+	return i, nil
+}
+
+func (ix *indexer) Querier() (*index.Querier, error) {
+	return ix.ix.Querier()
+}
+
+const (
+	timeSliceField = "__ts__"
+	timeSliceSize  = 3 * time.Hour
+)
+
+func timeSlice(t model.Time) model.Time {
+	return t - (t % model.Time(timeSliceSize/time.Millisecond))
+}
+
+func timeString(t model.Time) string {
+	return strconv.FormatInt(int64(t), 16)
+}
+
+func (ix *indexer) close() error {
+	return ix.ix.Close()
+}
+
+func (ix *indexer) index(cds ...*chunkDesc) error {
+	b, err := ix.ix.Batch()
+	if err != nil {
+		return err
+	}
+
+	ids := make([]ChunkID, len(cds))
+	for i, cd := range cds {
+		terms := make(index.Terms, 0, len(cd.met))
+		for k, v := range cd.met {
+			t := index.Term{Field: string(k), Val: string(v)}
+			terms = append(terms, t)
+		}
+		id := b.Add(terms)
+		ts := timeSlice(cd.firstTime)
+
+		// If the chunk has a higher time slice than the high one,
+		// don't index. It will be indexed when the next time slice
+		// is initiated over all memory chunks.
+		if ts <= ix.mc.highTime {
+			b.SecondaryIndex(id, index.Term{
+				Field: timeSliceField,
+				Val:   timeString(ts),
+			})
+		}
+
+		ids[i] = ChunkID(id)
+	}
+
+	if err := b.Commit(); err != nil {
+		return err
+	}
+
+	// We have to lock here already instead of post-commit as otherwise we might
+	// generate new chunk IDs, skip their indexing, and have a reindexTime being
+	// called with the chunk ID not being visible yet.
+	// TODO(fabxc): move back up
+	ix.mc.mtx.Lock()
+	defer ix.mc.mtx.Unlock()
+
+	// Make in-memory chunks visible for read.
+	for i, cd := range cds {
+		atomic.StoreUint64((*uint64)(&cd.id), uint64(ids[i]))
+		ix.mc.chunks[cd.id] = cd
+	}
+	return nil
+}
+
+// reindexTime creates an initial time slice index over all chunk IDs.
+// Any future chunks indexed for the same time slice must have higher IDs.
+func (ix *indexer) reindexTime(ids ChunkIDs, ts model.Time) error {
+	b, err := ix.ix.Batch()
+	if err != nil {
+		return err
+	}
+	sort.Sort(ids)
+	t := index.Term{Field: timeSliceField, Val: timeString(ts)}
+
+	for _, id := range ids {
+		b.SecondaryIndex(index.DocID(id), t)
+	}
+	return b.Commit()
+}
diff --git a/persist.go b/persist.go
new file mode 100644
index 000000000..56a47ed10
--- /dev/null
+++ b/persist.go
@@ -0,0 +1,163 @@
+package tsdb
+
+import (
+	"encoding/binary"
+	"os"
+	"path/filepath"
+	"time"
+
+	"github.com/boltdb/bolt"
+	"github.com/fabxc/pagebuf"
+	"github.com/prometheus/common/log"
+)
+
+type persistence struct {
+	*chunkBatchProcessor
+
+	mc     *memChunks
+	chunks *pagebuf.DB
+	index  *bolt.DB
+}
+
+func newPersistence(path string, cap int, to time.Duration) (*persistence, error) {
+	if err := os.MkdirAll(path, 0777); err != nil {
+		return nil, err
+	}
+	ix, err := bolt.Open(filepath.Join(path, "ix"), 0666, nil)
+	if err != nil {
+		return nil, err
+	}
+	if err := ix.Update(func(tx *bolt.Tx) error {
+		_, err := tx.CreateBucketIfNotExists(bktChunks)
+		return err
+	}); err != nil {
+		return nil, err
+	}
+	pb, err := pagebuf.Open(filepath.Join(path, "chunks"), 0666, nil)
+	if err != nil {
+		return nil, err
+	}
+	p := &persistence{
+		chunks:              pb,
+		index:               ix,
+		chunkBatchProcessor: newChunkBatchProcessor(log.Base(), cap, to),
+	}
+	p.chunkBatchProcessor.processf = p.persist
+
+	return p, nil
+}
+
+var bktChunks = []byte("chunks")
+
+func (p *persistence) close() error {
+	// Index must be closed first, otherwise we might deadlock.
+	err0 := p.index.Close()
+	err1 := p.chunks.Close()
+	if err0 != nil {
+		return err0
+	}
+	return err1
+}
+
+func (p *persistence) persist(cds ...*chunkDesc) error {
+	err := p.update(func(tx *persistenceTx) error {
+		bkt := tx.ix.Bucket(bktChunks)
+		for _, cd := range cds {
+			pos, err := tx.chunks.Add(cd.chunk.Data())
+			if err != nil {
+				return err
+			}
+			var buf [16]byte
+			binary.BigEndian.PutUint64(buf[:8], uint64(cd.id))
+			binary.BigEndian.PutUint64(buf[8:], pos)
+			if err := bkt.Put(buf[:8], buf[8:]); err != nil {
+				return err
+			}
+
+			tx.ids = append(tx.ids, cd.id)
+		}
+		return nil
+	})
+	return err
+}
+
+func (p *persistence) update(f func(*persistenceTx) error) error {
+	tx, err := p.begin(true)
+	if err != nil {
+		return err
+	}
+	if err := f(tx); err != nil {
+		tx.rollback()
+		return err
+	}
+	return tx.commit()
+}
+
+func (p *persistence) view(f func(*persistenceTx) error) error {
+	tx, err := p.begin(false)
+	if err != nil {
+		return err
+	}
+	if err := f(tx); err != nil {
+		tx.rollback()
+		return err
+	}
+	return tx.rollback()
+}
+
+func (p *persistence) begin(writeable bool) (*persistenceTx, error) {
+	var err error
+	tx := &persistenceTx{p: p}
+	// Index transaction is the outer one so we might end up with orphaned
+	// chunks but never with dangling pointers in the index.
+	tx.ix, err = p.index.Begin(writeable)
+	if err != nil {
+		return nil, err
+	}
+	tx.chunks, err = p.chunks.Begin(writeable)
+	if err != nil {
+		tx.ix.Rollback()
+		return nil, err
+	}
+
+	return tx, nil
+}
+
+type persistenceTx struct {
+	p      *persistence
+	ix     *bolt.Tx
+	chunks *pagebuf.Tx
+
+	ids []ChunkID
+}
+
+func (tx *persistenceTx) commit() error {
+	if err := tx.chunks.Commit(); err != nil {
+		tx.ix.Rollback()
+		return err
+	}
+	if err := tx.ix.Commit(); err != nil {
+		// TODO(fabxc): log orphaned chunks. What about overwritten ones?
+		// Should we not allows delete and add in the same tx so this cannot happen?
+		return err
+	}
+
+	// Successfully persisted chunks, clear them from the in-memory
+	// forward mapping.
+	tx.p.mc.mtx.Lock()
+	defer tx.p.mc.mtx.Unlock()
+
+	for _, id := range tx.ids {
+		delete(tx.p.mc.chunks, id)
+	}
+	return nil
+}
+
+func (tx *persistenceTx) rollback() error {
+	err0 := tx.chunks.Rollback()
+	err1 := tx.ix.Rollback()
+	if err0 != nil {
+		return err0
+	}
+	return err1
+}
diff --git a/querier.go b/querier.go
new file mode 100644
index 000000000..3313fdbd0
--- /dev/null
+++ b/querier.go
@@ -0,0 +1,248 @@
+package tsdb
+
+import (
+	"encoding/binary"
+	"errors"
+	"fmt"
+	"io"
+
+	"github.com/fabxc/tsdb/chunks"
+	"github.com/fabxc/tsdb/index"
+	"github.com/prometheus/common/model"
+	"github.com/prometheus/prometheus/storage/metric"
+)
+
+// SeriesIterator provides iteration over a time series associated with a metric.
+type SeriesIterator interface {
+	Metric() metric.Metric
+	Seek(model.Time) (model.SamplePair, bool)
+	Next() (model.SamplePair, bool)
+	Err() error
+}
+
+type chunkSeriesIterator struct {
+	m      metric.Metric
+	chunks []chunks.Chunk
+
+	err    error
+	cur    chunks.Iterator
+	curPos int
+}
+
+func newChunkSeriesIterator(m metric.Metric, chunks []chunks.Chunk) *chunkSeriesIterator {
+	return &chunkSeriesIterator{
+		m:      m,
+		chunks: chunks,
+	}
+}
+
+func (it *chunkSeriesIterator) Metric() metric.Metric {
+	return it.m
+}
+
+func (it *chunkSeriesIterator) Seek(ts model.Time) (model.SamplePair, bool) {
+	// Naively go through all chunk's first timestamps and pick the chunk
+	// containing the seeked timestamp.
+	// TODO(fabxc): this can be made smarter if it's a bottleneck.
+	for i, chk := range it.chunks {
+		cit := chk.Iterator()
+		first, ok := cit.First()
+		if !ok {
+			it.err = cit.Err()
+			return model.SamplePair{}, false
+		}
+		if first.Timestamp > ts {
+			break
+		}
+		it.cur = cit
+		it.curPos = i
+	}
+	return it.cur.Seek(ts)
+}
+
+func (it *chunkSeriesIterator) Next() (model.SamplePair, bool) {
+	sp, ok := it.cur.Next()
+	if ok {
+		return sp, true
+	}
+	if it.cur.Err() != io.EOF {
+		it.err = it.cur.Err()
+		return model.SamplePair{}, false
+	}
+	if len(it.chunks) == it.curPos+1 {
+		it.err = io.EOF
+		return model.SamplePair{}, false
+	}
+	it.curPos++
+	it.cur = it.chunks[it.curPos].Iterator()
+
+	// Return first sample of the new chunks.
+	return it.cur.Seek(0)
+}
+
+func (it *chunkSeriesIterator) Err() error {
+	return it.err
+}
+
+// Querier allows several queries over the storage with a consistent view if the data.
+type Querier struct {
+	db *DB
+	iq *index.Querier
+}
+
+// Querier returns a new Querier on the index at the current point in time.
+func (db *DB) Querier() (*Querier, error) {
+	iq, err := db.indexer.Querier()
+	if err != nil {
+		return nil, err
+	}
+	return &Querier{db: db, iq: iq}, nil
+}
+
+// Close the querier. This invalidates all previously retrieved iterators.
+func (q *Querier) Close() error {
+	return q.iq.Close()
+}
+
+// Iterator returns an iterator over all chunks that match all given
+// label matchers. The iterator is only valid until the Querier is closed.
+func (q *Querier) Iterator(matchers ...*metric.LabelMatcher) (index.Iterator, error) {
+	var its []index.Iterator
+	for _, m := range matchers {
+		var matcher index.Matcher
+		switch m.Type {
+		case metric.Equal:
+			matcher = index.NewEqualMatcher(string(m.Value))
+		case metric.RegexMatch:
+			var err error
+			matcher, err = index.NewRegexpMatcher(string(m.Value))
+			if err != nil {
+				return nil, err
+			}
+		default:
+			return nil, fmt.Errorf("matcher type %q not supported", m.Type)
+		}
+		it, err := q.iq.Search(string(m.Name), matcher)
+		if err != nil {
+			return nil, err
+		}
+		if it != nil {
+			its = append(its, it)
+		}
+	}
+	if len(its) == 0 {
+		return nil, errors.New("not found")
+	}
+	return index.Intersect(its...), nil
+}
+
+// RangeIterator returns an iterator over chunks that are present in the given time range.
+// The returned iterator is only valid until the querier is closed.
+func (q *Querier) RangeIterator(start, end model.Time) (index.Iterator, error) {
+	return nil, nil
+}
+
+// InstantIterator returns an iterator over chunks possibly containing values for
+// the given timestamp. The returned iterator is only valid until the querier is closed.
+func (q *Querier) InstantIterator(at model.Time) (index.Iterator, error) {
+	return nil, nil
+}
+
+// Series returns a list of series iterators over all chunks in the given iterator.
+// The returned series iterators are only valid until the querier is closed.
+func (q *Querier) Series(it index.Iterator) ([]SeriesIterator, error) {
+	mets := map[model.Fingerprint]metric.Metric{}
+	its := map[model.Fingerprint][]chunks.Chunk{}
+
+	id, err := it.Seek(0)
+	for ; err == nil; id, err = it.Next() {
+		terms, err := q.iq.Doc(id)
+		if err != nil {
+			return nil, err
+		}
+		met := make(model.Metric, len(terms))
+		for _, t := range terms {
+			met[model.LabelName(t.Field)] = model.LabelValue(t.Val)
+		}
+		fp := met.Fingerprint()
+
+		chunk, err := q.chunk(ChunkID(id))
+		if err != nil {
+			return nil, err
+		}
+
+		its[fp] = append(its[fp], chunk)
+		if _, ok := mets[fp]; ok {
+			continue
+		}
+		mets[fp] = metric.Metric{Metric: met, Copied: true}
+	}
+	if err != io.EOF {
+		return nil, err
+	}
+
+	res := make([]SeriesIterator, 0, len(its))
+	for fp, chks := range its {
+		res = append(res, newChunkSeriesIterator(mets[fp], chks))
+	}
+	return res, nil
+}
+
+func (q *Querier) chunk(id ChunkID) (chunks.Chunk, error) {
+	q.db.memChunks.mtx.RLock()
+	cd, ok := q.db.memChunks.chunks[id]
+	q.db.memChunks.mtx.RUnlock()
+	if ok {
+		return cd.chunk, nil
+	}
+
+	var chk chunks.Chunk
+	// TODO(fabxc): this starts a new read transaction for every
+	// chunk we have to load from persistence.
+	// Figure out what's best tradeoff between lock contention and
+	// data consistency: start transaction when instantiating the querier
+	// or lazily start transaction on first try. (Not all query operations
+	// need access to persisted chunks.)
+	err := q.db.persistence.view(func(tx *persistenceTx) error {
+		chks := tx.ix.Bucket(bktChunks)
+		ptr := chks.Get(id.bytes())
+		if ptr == nil {
+			return fmt.Errorf("chunk pointer for ID %d not found", id)
+		}
+		cdata, err := tx.chunks.Get(binary.BigEndian.Uint64(ptr))
+		if err != nil {
+			return fmt.Errorf("get chunk data for ID %d: %s", id, err)
+		}
+		chk, err = chunks.FromData(cdata)
+		return err
+	})
+	return chk, err
+}
+
+// Metrics returns the unique metrics found across all chunks in the provided iterator.
+func (q *Querier) Metrics(it index.Iterator) ([]metric.Metric, error) {
+	m := []metric.Metric{}
+	fps := map[model.Fingerprint]struct{}{}
+
+	id, err := it.Seek(0)
+	for ; err == nil; id, err = it.Next() {
+		terms, err := q.iq.Doc(id)
+		if err != nil {
+			return nil, err
+		}
+		met := make(model.Metric, len(terms))
+		for _, t := range terms {
+			met[model.LabelName(t.Field)] = model.LabelValue(t.Val)
+		}
+		fp := met.Fingerprint()
+		if _, ok := fps[fp]; ok {
+			continue
+		}
+		fps[fp] = struct{}{}
+		m = append(m, metric.Metric{Metric: met, Copied: true})
+	}
+	if err != io.EOF {
+		return nil, err
+	}
+	return m, nil
+}