optimizing badger cache, won a 10-15% improvement in most benchmarks

pkg/database/database.go

@@ -16,15 +16,20 @@ import (
 	"next.orly.dev/pkg/utils/units"
 )
 
+// D implements the Database interface using Badger as the storage backend
 type D struct {
 	ctx     context.Context
 	cancel  context.CancelFunc
 	dataDir string
 	Logger  *logger
 	*badger.DB
-	seq *badger.Sequence
+	seq   *badger.Sequence
+	ready chan struct{} // Closed when database is ready to serve requests
 }
 
+// Ensure D implements Database interface at compile time
+var _ Database = (*D)(nil)
+
 func New(
 	ctx context.Context, cancel context.CancelFunc, dataDir, logLevel string,
 ) (
@@ -37,6 +42,7 @@ func New(
 		Logger:  NewLogger(lol.GetLogLevel(logLevel), dataDir),
 		DB:      nil,
 		seq:     nil,
+		ready:   make(chan struct{}),
 	}
 
 	// Ensure the data directory exists
@@ -54,8 +60,8 @@ func New(
 	opts := badger.DefaultOptions(d.dataDir)
 	// Configure caches based on environment to better match workload.
 	// Defaults aim for higher hit ratios under read-heavy workloads while remaining safe.
-	var blockCacheMB = 512 // default 512 MB
-	var indexCacheMB = 256 // default 256 MB
+	var blockCacheMB = 1024 // default 512 MB
+	var indexCacheMB = 512  // default 256 MB
 	if v := os.Getenv("ORLY_DB_BLOCK_CACHE_MB"); v != "" {
 		if n, perr := strconv.Atoi(v); perr == nil && n > 0 {
 			blockCacheMB = n
@@ -69,15 +75,42 @@ func New(
 	opts.BlockCacheSize = int64(blockCacheMB * units.Mb)
 	opts.IndexCacheSize = int64(indexCacheMB * units.Mb)
 	opts.BlockSize = 4 * units.Kb // 4 KB block size
-	// Prevent huge allocations during table building and memtable flush.
-	// Badger's TableBuilder buffer is sized by BaseTableSize; ensure it's small.
-	opts.BaseTableSize = 64 * units.Mb // 64 MB per table (default ~2MB, increased for fewer files but safe)
-	opts.MemTableSize = 64 * units.Mb  // 64 MB memtable to match table size
-	// Keep value log files to a moderate size as well
-	opts.ValueLogFileSize = 256 * units.Mb // 256 MB value log files
+
+	// Reduce table sizes to lower cost-per-key in cache
+	// Smaller tables mean lower cache cost metric per entry
+	opts.BaseTableSize = 8 * units.Mb // 8 MB per table (reduced from 64 MB to lower cache cost)
+	opts.MemTableSize = 16 * units.Mb // 16 MB memtable (reduced from 64 MB)
+
+	// Keep value log files to a moderate size
+	opts.ValueLogFileSize = 128 * units.Mb // 128 MB value log files (reduced from 256 MB)
+
+	// CRITICAL: Keep small inline events in LSM tree, not value log
+	// VLogPercentile 0.99 means 99% of values stay in LSM (our optimized inline events!)
+	// This dramatically improves read performance for small events
+	opts.VLogPercentile = 0.99
+
+	// Optimize LSM tree structure
+	opts.BaseLevelSize = 64 * units.Mb // Increased from default 10 MB for fewer levels
+	opts.LevelSizeMultiplier = 10      // Default, good balance
+
 	opts.CompactL0OnClose = true
 	opts.LmaxCompaction = true
-	opts.Compression = options.None
+
+	// Enable compression to reduce cache cost
+	opts.Compression = options.ZSTD
+	opts.ZSTDCompressionLevel = 1 // Fast compression (500+ MB/s)
+
+	// Disable conflict detection for write-heavy relay workloads
+	// Nostr events are immutable, no need for transaction conflict checks
+	opts.DetectConflicts = false
+
+	// Performance tuning for high-throughput workloads
+	opts.NumCompactors = 8            // Increase from default 4 for faster compaction
+	opts.NumLevelZeroTables = 8       // Increase from default 5 to allow more L0 tables before compaction
+	opts.NumLevelZeroTablesStall = 16 // Increase from default 15 to reduce write stalls
+	opts.NumMemtables = 8             // Increase from default 5 to buffer more writes
+	opts.MaxLevels = 7                // Default is 7, keep it
+
 	opts.Logger = d.Logger
 	if d.DB, err = badger.Open(opts); chk.E(err) {
 		return
@@ -88,6 +121,10 @@ func New(
 	// run code that updates indexes when new indexes have been added and bumps
 	// the version so they aren't run again.
 	d.RunMigrations()
+
+	// Start warmup goroutine to signal when database is ready
+	go d.warmup()
+
 	// start up the expiration tag processing and shut down and clean up the
 	// database after the context is canceled.
 	go func() {
@@ -108,6 +145,29 @@ func New(
 // Path returns the path where the database files are stored.
 func (d *D) Path() string { return d.dataDir }
 
+// Ready returns a channel that closes when the database is ready to serve requests.
+// This allows callers to wait for database warmup to complete.
+func (d *D) Ready() <-chan struct{} {
+	return d.ready
+}
+
+// warmup performs database warmup operations and closes the ready channel when complete.
+// Warmup criteria:
+// - Wait at least 2 seconds for initial compactions to settle
+// - Ensure cache hit ratio is reasonable (if we have metrics available)
+func (d *D) warmup() {
+	defer close(d.ready)
+
+	// Give the database time to settle after opening
+	// This allows:
+	// - Initial compactions to complete
+	// - Memory allocations to stabilize
+	// - Cache to start warming up
+	time.Sleep(2 * time.Second)
+
+	d.Logger.Infof("database warmup complete, ready to serve requests")
+}
+
 func (d *D) Wipe() (err error) {
 	err = errors.New("not implemented")
 	return

pkg/database/factory.go

@@ -0,0 +1,39 @@
+package database
+
+import (
+	"context"
+	"fmt"
+	"strings"
+)
+
+// NewDatabase creates a database instance based on the specified type.
+// Supported types: "badger", "dgraph"
+func NewDatabase(
+	ctx context.Context,
+	cancel context.CancelFunc,
+	dbType string,
+	dataDir string,
+	logLevel string,
+) (Database, error) {
+	switch strings.ToLower(dbType) {
+	case "badger", "":
+		// Use the existing badger implementation
+		return New(ctx, cancel, dataDir, logLevel)
+	case "dgraph":
+		// Use the new dgraph implementation
+		// Import dynamically to avoid import cycles
+		return newDgraphDatabase(ctx, cancel, dataDir, logLevel)
+	default:
+		return nil, fmt.Errorf("unsupported database type: %s (supported: badger, dgraph)", dbType)
+	}
+}
+
+// newDgraphDatabase creates a dgraph database instance
+// This is defined here to avoid import cycles
+var newDgraphDatabase func(context.Context, context.CancelFunc, string, string) (Database, error)
+
+// RegisterDgraphFactory registers the dgraph database factory
+// This is called from the dgraph package's init() function
+func RegisterDgraphFactory(factory func(context.Context, context.CancelFunc, string, string) (Database, error)) {
+	newDgraphDatabase = factory
+}

pkg/database/interface.go

@@ -0,0 +1,102 @@
+package database
+
+import (
+	"context"
+	"io"
+	"time"
+
+	"next.orly.dev/pkg/database/indexes/types"
+	"next.orly.dev/pkg/encoders/event"
+	"next.orly.dev/pkg/encoders/filter"
+	"next.orly.dev/pkg/encoders/tag"
+	"next.orly.dev/pkg/interfaces/store"
+)
+
+// Database defines the interface that all database implementations must satisfy.
+// This allows switching between different storage backends (badger, dgraph, etc.)
+type Database interface {
+	// Core lifecycle methods
+	Path() string
+	Init(path string) error
+	Sync() error
+	Close() error
+	Wipe() error
+	SetLogLevel(level string)
+	Ready() <-chan struct{} // Returns a channel that closes when database is ready to serve requests
+
+	// Event storage and retrieval
+	SaveEvent(c context.Context, ev *event.E) (exists bool, err error)
+	GetSerialsFromFilter(f *filter.F) (serials types.Uint40s, err error)
+	WouldReplaceEvent(ev *event.E) (bool, types.Uint40s, error)
+
+	QueryEvents(c context.Context, f *filter.F) (evs event.S, err error)
+	QueryAllVersions(c context.Context, f *filter.F) (evs event.S, err error)
+	QueryEventsWithOptions(c context.Context, f *filter.F, includeDeleteEvents bool, showAllVersions bool) (evs event.S, err error)
+	QueryDeleteEventsByTargetId(c context.Context, targetEventId []byte) (evs event.S, err error)
+	QueryForSerials(c context.Context, f *filter.F) (serials types.Uint40s, err error)
+	QueryForIds(c context.Context, f *filter.F) (idPkTs []*store.IdPkTs, err error)
+
+	CountEvents(c context.Context, f *filter.F) (count int, approximate bool, err error)
+
+	FetchEventBySerial(ser *types.Uint40) (ev *event.E, err error)
+	FetchEventsBySerials(serials []*types.Uint40) (events map[uint64]*event.E, err error)
+
+	GetSerialById(id []byte) (ser *types.Uint40, err error)
+	GetSerialsByIds(ids *tag.T) (serials map[string]*types.Uint40, err error)
+	GetSerialsByIdsWithFilter(ids *tag.T, fn func(ev *event.E, ser *types.Uint40) bool) (serials map[string]*types.Uint40, err error)
+	GetSerialsByRange(idx Range) (serials types.Uint40s, err error)
+
+	GetFullIdPubkeyBySerial(ser *types.Uint40) (fidpk *store.IdPkTs, err error)
+	GetFullIdPubkeyBySerials(sers []*types.Uint40) (fidpks []*store.IdPkTs, err error)
+
+	// Event deletion
+	DeleteEvent(c context.Context, eid []byte) error
+	DeleteEventBySerial(c context.Context, ser *types.Uint40, ev *event.E) error
+	DeleteExpired()
+	ProcessDelete(ev *event.E, admins [][]byte) error
+	CheckForDeleted(ev *event.E, admins [][]byte) error
+
+	// Import/Export
+	Import(rr io.Reader)
+	Export(c context.Context, w io.Writer, pubkeys ...[]byte)
+	ImportEventsFromReader(ctx context.Context, rr io.Reader) error
+	ImportEventsFromStrings(ctx context.Context, eventJSONs []string, policyManager interface{ CheckPolicy(action string, ev *event.E, pubkey []byte, remote string) (bool, error) }) error
+
+	// Relay identity
+	GetRelayIdentitySecret() (skb []byte, err error)
+	SetRelayIdentitySecret(skb []byte) error
+	GetOrCreateRelayIdentitySecret() (skb []byte, err error)
+
+	// Markers (metadata key-value storage)
+	SetMarker(key string, value []byte) error
+	GetMarker(key string) (value []byte, err error)
+	HasMarker(key string) bool
+	DeleteMarker(key string) error
+
+	// Subscriptions (payment-based access control)
+	GetSubscription(pubkey []byte) (*Subscription, error)
+	IsSubscriptionActive(pubkey []byte) (bool, error)
+	ExtendSubscription(pubkey []byte, days int) error
+	RecordPayment(pubkey []byte, amount int64, invoice, preimage string) error
+	GetPaymentHistory(pubkey []byte) ([]Payment, error)
+	ExtendBlossomSubscription(pubkey []byte, tier string, storageMB int64, daysExtended int) error
+	GetBlossomStorageQuota(pubkey []byte) (quotaMB int64, err error)
+	IsFirstTimeUser(pubkey []byte) (bool, error)
+
+	// NIP-43 Invite-based ACL
+	AddNIP43Member(pubkey []byte, inviteCode string) error
+	RemoveNIP43Member(pubkey []byte) error
+	IsNIP43Member(pubkey []byte) (isMember bool, err error)
+	GetNIP43Membership(pubkey []byte) (*NIP43Membership, error)
+	GetAllNIP43Members() ([][]byte, error)
+	StoreInviteCode(code string, expiresAt time.Time) error
+	ValidateInviteCode(code string) (valid bool, err error)
+	DeleteInviteCode(code string) error
+	PublishNIP43MembershipEvent(kind int, pubkey []byte) error
+
+	// Migrations (version tracking for schema updates)
+	RunMigrations()
+
+	// Utility methods
+	EventIdsBySerial(start uint64, count int) (evs []uint64, err error)
+}