@@ -0,0 +1,90 @@
+package rocksdb
+
+import (
+	"errors"
+	"fmt"
+	"iter"
+	"os"
+	"path/filepath"
+
+	"github.com/linxGnu/grocksdb"
+)
+
+// loadDirName is where LoadSorted builds a file before RocksDB takes it: a
+// subdirectory of the store's own directory, which RocksDB does not scan.
+const loadDirName = "loading"
+
+// LoadSorted writes entries, which must ascend by key, as one file and moves
+// it into cf. Readers see all of the entries or none, and they survive a
+// crash once LoadSorted returns. An empty sequence is an error.
+//
+// The entries bypass the WAL and the memtable, but other writes to the store
+// wait while RocksDB installs the file. The entries' key range must not
+// overlap keys cf already holds: the file then lands below every other file
+// of cf, and cf never needs compacting. An overlap fails the load.
+//
+// Each entry is written before the next is pulled, so entries may reuse
+// their buffers. The store's lifecycle lock is held throughout, so Close
+// waits for the load and entries must not call back into the store.
+func (s *Store) LoadSorted(cf string, entries iter.Seq2[[]byte, []byte]) error {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	if err := s.checkOpen(); err != nil {
+		return err
+	}
+	if s.cfg.ReadOnly {
+		return errors.New("rocksdb: LoadSorted on a read-only store")
+	}
+	if cf == "" {
+		cf = defaultCFName
+	}
+	cfh, err := s.resolveCF(cf)
+	if err != nil {
+		return err
+	}
+	if err := s.loadSorted(cf, cfh, entries); err != nil {
+		return fmt.Errorf("rocksdb: load sorted entries into %q: %w", cf, err)
+	}
+	return nil
+}
+
+func (s *Store) loadSorted(cf string, cfh *grocksdb.ColumnFamilyHandle, entries iter.Seq2[[]byte, []byte]) error {
+	dir := filepath.Join(s.cfg.Path, loadDirName)
+	if err := os.MkdirAll(dir, dirPerm); err != nil {
+		return err
+	}
+	file, err := os.CreateTemp(dir, "*.sst")
+	if err != nil {
+		return err
+	}
+	path := file.Name()
+	// Nothing is left to remove once RocksDB has moved the file.
+	defer func() { _ = os.Remove(path) }()
+	if err := file.Close(); err != nil {
+		return err
+	}
+
+	env := grocksdb.NewDefaultEnvOptions()
+	defer env.Destroy()
+	// The writer copies cf's options, so the file carries cf's block size,
+	// filter and compression.
+	w := grocksdb.NewSSTFileWriter(env, s.cfOpts[cf])
+	defer w.Destroy()
+	if err := w.Open(path); err != nil {
+		return err
+	}
+	for key, value := range entries {
+		if err := w.Put(key, value); err != nil {
+			return err
+		}
+	}
+	if err := w.Finish(); err != nil {
+		return err
+	}
+
+	opts := grocksdb.NewDefaultIngestExternalFileOptions()
+	defer opts.Destroy()
+	opts.SetMoveFiles(true)
+	opts.SetFailIfNotBottommostLevel(true)
+	return s.db.IngestExternalFileCF(cfh, []string{path}, opts)
+}
@@ -0,0 +1,242 @@
+package rocksdb
+
+import (
+	"bytes"
+	"fmt"
+	"iter"
+	"os"
+	"path/filepath"
+	"strconv"
+	"testing"
+	"time"
+
+	"github.com/linxGnu/grocksdb"
+	"github.com/stretchr/testify/require"
+)
+
+// pairs yields key/value pairs given as alternating strings.
+func pairs(kv ...string) iter.Seq2[[]byte, []byte] {
+	return func(yield func(key, value []byte) bool) {
+		for i := 0; i < len(kv); i += 2 {
+			if !yield([]byte(kv[i]), []byte(kv[i+1])) {
+				return
+			}
+		}
+	}
+}
+
+// numbered yields n ascending keys, each with value.
+func numbered(n int, value []byte) iter.Seq2[[]byte, []byte] {
+	return func(yield func(key, value []byte) bool) {
+		for i := range n {
+			if !yield(fmt.Appendf(nil, "key-%08d", i), value) {
+				return
+			}
+		}
+	}
+}
+
+func requireValue(t *testing.T, s *Store, cf, key, want string) {
+	t.Helper()
+	got, found, err := s.Get(cf, []byte(key))
+	require.NoError(t, err)
+	require.True(t, found, "key %q", key)
+	require.Equal(t, want, string(got))
+}
+
+func loadFiles(t *testing.T, s *Store) []string {
+	t.Helper()
+	files, err := filepath.Glob(filepath.Join(s.cfg.Path, loadDirName, "*"))
+	require.NoError(t, err)
+	return files
+}
+
+func uintProperty(t *testing.T, s *Store, cf, name string) uint64 {
+	t.Helper()
+	n, err := strconv.ParseUint(s.db.GetPropertyCF(name, s.cfHandles[cf]), 10, 64)
+	require.NoError(t, err)
+	return n
+}
+
+func TestStore_LoadSorted_RoundTrip(t *testing.T) {
+	const loaded, other = "loaded", "other"
+	cfg := Config{Path: t.TempDir(), ColumnFamilies: []string{loaded, other}, Logger: silentLogger()}
+	s, err := New(cfg)
+	require.NoError(t, err)
+
+	require.NoError(t, s.LoadSorted(loaded, pairs("a1", "one", "a2", "two")))
+	// A batch to another CF sits in the memtable while a second, disjoint
+	// file is loaded.
+	require.NoError(t, s.Batch(func(b *BatchWriter) error {
+		b.Put(other, []byte("k"), []byte("batched"))
+		return nil
+	}))
+	require.NoError(t, s.LoadSorted(loaded, pairs("b1", "three")))
+	require.Error(t, s.LoadSorted(loaded, pairs("a2", "x")), "a key a loaded file holds")
+	require.Empty(t, loadFiles(t, s))
+
+	check := func(s *Store) {
+		requireValue(t, s, loaded, "a1", "one")
+		requireValue(t, s, loaded, "a2", "two")
+		requireValue(t, s, loaded, "b1", "three")
+		requireValue(t, s, other, "k", "batched")
+		_, found, err := s.Get(other, []byte("a1"))
+		require.NoError(t, err)
+		require.False(t, found, "a loaded key stays in its column family")
+		last, ok, err := s.LastKey(loaded)
+		require.NoError(t, err)
+		require.True(t, ok)
+		require.Equal(t, "b1", string(last))
+	}
+	check(s)
+
+	require.NoError(t, s.Close())
+	reopened, err := New(cfg)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = reopened.Close() })
+	check(reopened)
+}
+
+func TestStore_LoadSorted_FailedLoadLeavesNothing(t *testing.T) {
+	s := openTestStore(t, nil)
+	require.NoError(t, s.Put("", []byte("kept"), []byte("v")))
+
+	require.Error(t, s.LoadSorted("", pairs("b", "1", "a", "2")), "descending keys")
+	require.Error(t, s.LoadSorted("", pairs()), "no entries")
+	require.Error(t, s.LoadSorted("", pairs("kept", "x")), "a key the column family holds")
+	require.ErrorIs(t, s.LoadSorted("not-configured", pairs("a", "1")), ErrCFNotFound)
+
+	require.Empty(t, loadFiles(t, s))
+	for _, key := range []string{"a", "b"} {
+		_, found, err := s.Get("", []byte(key))
+		require.NoError(t, err)
+		require.False(t, found, "key %q", key)
+	}
+	requireValue(t, s, "", "kept", "v")
+}
+
+func TestStore_LoadSorted_ClosedAndReadOnlyStores(t *testing.T) {
+	path := t.TempDir()
+	s, err := New(Config{Path: path, Logger: silentLogger()})
+	require.NoError(t, err)
+	require.NoError(t, s.Close())
+	require.ErrorIs(t, s.LoadSorted("", pairs("a", "1")), ErrStoreClosed)
+
+	readOnly, err := New(Config{Path: path, Logger: silentLogger(), ReadOnly: true})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = readOnly.Close() })
+	require.Error(t, readOnly.LoadSorted("", pairs("a", "1")))
+	require.NoDirExists(t, filepath.Join(path, loadDirName))
+}
+
+// Close waits for a load in flight, so the load is durable or never started.
+func TestStore_LoadSorted_CloseWaitsForLoad(t *testing.T) {
+	path := t.TempDir()
+	s, err := New(Config{Path: path, Logger: silentLogger()})
+	require.NoError(t, err)
+
+	loading, release := make(chan struct{}), make(chan struct{})
+	loadErr := make(chan error, 1)
+	go func() {
+		loadErr <- s.LoadSorted("", func(yield func(key, value []byte) bool) {
+			close(loading)
+			<-release
+			yield([]byte("a"), []byte("1"))
+		})
+	}()
+	<-loading
+	closed := make(chan struct{})
+	go func() {
+		_ = s.Close()
+		close(closed)
+	}()
+	select {
+	case <-closed:
+		t.Fatal("Close returned while a load was in flight")
+	case <-time.After(50 * time.Millisecond):
+	}
+	close(release)
+	require.NoError(t, <-loadErr)
+	<-closed
+
+	reopened, err := New(Config{Path: path, Logger: silentLogger(), MustExist: true})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = reopened.Close() })
+	requireValue(t, reopened, "", "a", "1")
+}
+
+// The file is written with the column family's own options: a compressed
+// column family gets a compressed file.
+func TestStore_LoadSorted_UsesTheColumnFamilyOptions(t *testing.T) {
+	const raw, zstd = "raw", "zstd"
+	s, err := New(Config{
+		Path:           t.TempDir(),
+		ColumnFamilies: []string{raw, zstd},
+		Logger:         silentLogger(),
+		PerCFOptions:   map[string]CFOptions{zstd: {Compression: grocksdb.ZSTDCompression}},
+	})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = s.Close() })
+
+	value := bytes.Repeat([]byte("compressible "), 100)
+	for _, cf := range []string{raw, zstd} {
+		require.NoError(t, s.LoadSorted(cf, numbered(1000, value)))
+	}
+	const size = "rocksdb.total-sst-files-size"
+	require.Less(t, 4*uintProperty(t, s, zstd, size), uintProperty(t, s, raw, size))
+}
+
+// With CacheIndexAndFilterBlocks, an open file's index and filter blocks
+// live in the block cache, not on the heap.
+func TestStore_CacheIndexAndFilterBlocks_KeepsThemOffTheHeap(t *testing.T) {
+	const cached, plain = "cached", "plain"
+	s, err := New(Config{
+		Path:           t.TempDir(),
+		ColumnFamilies: []string{cached, plain},
+		Logger:         silentLogger(),
+		Tuning:         Tuning{BlockCacheMB: 4},
+		PerCFOptions: map[string]CFOptions{
+			cached: {BloomFilterBitsPerKey: 10, CacheIndexAndFilterBlocks: true},
+			plain:  {BloomFilterBitsPerKey: 10},
+		},
+	})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = s.Close() })
+
+	for _, cf := range []string{cached, plain} {
+		require.NoError(t, s.LoadSorted(cf, numbered(20_000, []byte("v"))))
+		requireValue(t, s, cf, "key-00000007", "v")
+	}
+	const heap = "rocksdb.estimate-table-readers-mem"
+	require.Less(t, 4*uintProperty(t, s, cached, heap), uintProperty(t, s, plain, heap))
+}
+
+// Files a stopped process was still building are removed by the next open
+// that takes the DB lock, and by no other open.
+func TestNew_RemovesUnfinishedLoadFiles(t *testing.T) {
+	path := t.TempDir()
+	cfg := Config{Path: path, Logger: silentLogger()}
+	s, err := New(cfg)
+	require.NoError(t, err)
+
+	// An open that fails on the lock leaves a live store's load alone.
+	require.NoError(t, s.LoadSorted("", func(yield func(key, value []byte) bool) {
+		_, err := New(cfg)
+		require.Error(t, err)
+		yield([]byte("a"), []byte("1"))
+	}))
+	require.NoError(t, s.Close())
+
+	orphan := filepath.Join(path, loadDirName, "orphan.sst")
+	require.NoError(t, os.WriteFile(orphan, []byte("half a file"), 0o600))
+	readOnly, err := New(Config{Path: path, Logger: silentLogger(), ReadOnly: true})
+	require.NoError(t, err)
+	require.NoError(t, readOnly.Close())
+	require.FileExists(t, orphan)
+
+	reopened, err := New(Config{Path: path, Logger: silentLogger(), MustExist: true})
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = reopened.Close() })
+	require.NoFileExists(t, orphan)
+	requireValue(t, reopened, "", "a", "1")
+}
@@ -125,7 +125,7 @@ type Store struct {
 
 	db        *grocksdb.DB
 	opts      *grocksdb.Options
-	cfOpts    []*grocksdb.Options
+	cfOpts    map[string]*grocksdb.Options
 	cfHandles map[string]*grocksdb.ColumnFamilyHandle
 	ro        *grocksdb.ReadOptions
 	wo        *grocksdb.WriteOptions
@@ -748,15 +748,22 @@ func (s *Store) constructAndOpen() error {
 		return fmt.Errorf("rocksdb: open %s: %w", abs, err)
 	}
 
-	cfMap := make(map[string]*grocksdb.ColumnFamilyHandle, len(cfHandles))
-	for i, name := range cfNames {
-		cfMap[name] = cfHandles[i]
+	if !s.cfg.ReadOnly {
+		// Files LoadSorted was still building when the last process stopped.
+		// Only the holder of the DB lock may remove them.
+		if err := os.RemoveAll(filepath.Join(abs, loadDirName)); err != nil {
+			s.cfg.Logger.WithError(err).Warnf("rocksdb: removing unfinished load files in %s", abs)
+		}
 	}
 
 	s.db = db
 	s.opts = opts
-	s.cfOpts = cfOpts
-	s.cfHandles = cfMap
+	s.cfOpts = make(map[string]*grocksdb.Options, len(cfNames))
+	s.cfHandles = make(map[string]*grocksdb.ColumnFamilyHandle, len(cfNames))
+	for i, name := range cfNames {
+		s.cfOpts[name] = cfOpts[i]
+		s.cfHandles[name] = cfHandles[i]
+	}
 	s.ro = grocksdb.NewDefaultReadOptions()
 	s.wo = grocksdb.NewDefaultWriteOptions()
 
@@ -876,6 +883,9 @@ func (s *Store) applySharedTableOptions(cfNames []string, cfOpts []*grocksdb.Opt
 		if override.BlockSize > 0 {
 			bbto.SetBlockSize(override.BlockSize)
 		}
+		if override.CacheIndexAndFilterBlocks {
+			bbto.SetCacheIndexAndFilterBlocks(true)
+		}
 		o.SetBlockBasedTableFactory(bbto)
 		s.bbtos = append(s.bbtos, bbto)
 	}
 
@@ -53,6 +53,12 @@ type CFOptions struct {
 	// BloomFilterBitsPerKey installs this CF's bloom filter; 0 = none —
 	// right for a CF never probed for keys it may not hold.
 	BloomFilterBitsPerKey int
+
+	// CacheIndexAndFilterBlocks keeps each SST's index and filter blocks in
+	// the block cache instead of on the heap for as long as the file is
+	// open. For a CF that accumulates many files, it bounds their memory by
+	// the cache size.
+	CacheIndexAndFilterBlocks bool
 }
 
 // Tuning — DB-wide RocksDB knobs shared across every CF of one store. Zero