treedb

TreeDB

TreeDB is a high-performance, persistent key-value store optimized for the Cosmos SDK workload. It features a B+Tree index backed by a memory-mapped file plus a persistent value log for large values, with a separate journal/redo log for crash recovery.

Features

• ACID Transactions: Atomic commits using Copy-On-Write (COW) and redundant superblocks (Meta Pages).
• Snapshot Isolation: Lock-free concurrent readers using Multi-Version Concurrency Control (MVCC) and Reference Counting.
• Hybrid Storage:
  • Index: Memory-mapped B+Tree for keys and small values.
  • Value log: Append-only log for large values (contract code, blobs) to reduce write amplification and memory pressure.
  • Journal/redo log: Commit metadata for crash recovery and checkpointing.
• Crash Recovery: Automatic recovery from torn writes using strict write-ordering and checksum verification.
• Lifecycle Management: Safe page reclamation using a Graveyard and Reader Registry to protect active snapshots.

Architecture

Storage Layout

• Root dir: Options.Dir is a root directory with two sub-databases: maindb/ (main data) and dictdb/ (dictionary store).
• Pages: 4KB fixed-size blocks in maindb/index.db.
• Nodes: Slotted pages supporting variable-length keys.
• Value log: Append-only segments under maindb/wal/ storing large values with CRC checksums.
• Journal/redo log: Commit metadata stored alongside the value log under maindb/wal/.

Write Path ("The Zipper")

Writes are batched and applied using a recursive "Zipper" merge algorithm. This creates a new version of the tree path (COW) without modifying existing on-disk pages, ensuring crash safety and snapshot isolation.

Read Path

Readers acquire a Snapshot which pins the version of the tree and the relevant value-log segments. This guarantees a consistent view of the database even while writers are committing new versions.

Usage

package main

import (
	"errors"
	"fmt"
	"log"

	treedb "github.com/snissn/gomap/TreeDB"
)

func main() {
	// Open the database (recommended: cached wrapper)
	opts := treedb.Options{Dir: "./my-db-data"}
	database, err := treedb.Open(opts)
	if err != nil {
		if errors.Is(err, treedb.ErrLocked) {
			log.Fatal("database is already open in another process")
		}
		log.Fatal(err)
	}
	defer database.Close()

	// Set a key-value pair
	if err := database.Set([]byte("key1"), []byte("value1")); err != nil {
		log.Fatal(err)
	}

	// Get a value
	val, err := database.Get([]byte("key1"))
	if err != nil {
		log.Fatal(err)
	}
	fmt.Printf("Got: %s\n", val)

	// Iterate
	it, _ := database.Iterator(nil, nil)
	defer it.Close()
	for ; it.Valid(); it.Next() {
		fmt.Printf("%s = %s\n", it.Key(), it.Value())
	}
	
	// Atomic Batch
	batch := database.NewBatch()
	batch.Set([]byte("k2"), []byte("v2"))
	batch.Delete([]byte("key1"))
	batch.Write() // Atomic commit
}

Profiles (Durable / Fast / Bench)

If you want a simple, documented “bundle” of options, start with a profile and then override a few workload-specific knobs:

opts := treedb.OptionsFor(treedb.ProfileDurable, "./my-db-data")
opts.FlushThreshold = 128 << 20 // optional tuning
db, err := treedb.Open(opts)

Profiles are intended to make intent explicit:

• ProfileDurable: safest defaults (recommended).
• ProfileFast: relax durability/integrity knobs for throughput and enables index optimizations (ForcePointers, LeafPrefixCompression, IndexColumnarLeaves, IndexPackedValuePtr, IndexInternalBaseDelta).
• ProfileWALOnFast: fast ingest profile that keeps WAL on, relaxes durability checks, and enables the same index optimizations.
• ProfileBench: deterministic benchmarking profile (not production); includes ProfileFast index optimizations.

Note: WAL-off is selected via opts.Durability = treedb.DurabilityWALOffRelaxed. The value log remains enabled in cached mode, so large values can still go through wal/ even when WAL is off.

Details: docs/TREEDB_WRITE_PATHS.md.

Durability & Safety Notes

• Safe defaults keep WAL, fsync, and read checksums enabled; relax safety knobs via Options.Durability and Options.ValueLog.ReadIntegrity.
• In relaxed durability modes (DurabilityWALOnRelaxed / DurabilityWALOffRelaxed), SetSync/WriteSync are crash-consistent only (no fsync) and may not survive power loss.
• Page checksums are verified once and cached until the page is rewritten; use VerifyOnRead for paranoid always-verify behavior. Options.ValueLog.ReadIntegrity = IntegritySkipChecksums disables value-log CRC checks entirely.
• CRC checksums detect accidental corruption, not malicious tampering; use filesystem encryption/HMAC if your threat model includes adversarial disk access.
• GetUnsafe on a Snapshot and iterator Key()/Value() return short-lived views; use Get, KeyCopy, or ValueCopy for stable bytes.
• TreeDB does not provide encryption-at-rest or secure deletion; deleted data may remain on disk until compacted. Use OS/disk encryption for confidentiality.
• Value-log segments are persistent; reclaim disk via DB.ValueLogGC (deletes fully-unreferenced segments) and treedb.ValueLogRewriteOffline (offline rewrite/compaction).
• Optional guardrail: Options.ValueLog.MaxRetainedBytes emits a warning when retained value-log bytes exceed the threshold.
• Optional hard cap: Options.ValueLog.MaxRetainedBytesHard disables value-log pointers for new large values once retained bytes exceed the threshold.
• TreeDB is pre-alpha; public APIs and on-disk format may change without backward-compatibility guarantees.

Durability Matrix (Cached Mode)

Tuning (Cached Mode)

treedb.Open defaults to cached mode (memtable + journal + value log + background flush). The most important knobs:

• Options.FlushThreshold + Options.MaxQueuedMemtables (throughput vs. backlog/memory)
• Adaptive backpressure: SlowdownBacklogSeconds, StopBacklogSeconds, MaxBacklogBytes
• Cached-mode auto checkpointing: BackgroundCheckpointInterval, BackgroundCheckpointIdleDuration
• Background pruning: PruneInterval, PruneMaxPages, PruneMaxDuration
• Optional flush build parallelism: FlushBuildConcurrency
• Optional piggyback compaction toggle: DisablePiggybackCompaction
• Value-log retention guardrails: ValueLog.MaxRetainedBytes, ValueLog.MaxRetainedBytesHard
• Value-log compression mode: ValueLog.Compression (off|block|dict|auto) and ValueLog.BlockCodec (snappy|lz4)
• Index rebuild (in-place): treedb.CompactIndex() or treedb.VacuumIndexOffline(opts) (currently an alias for CompactIndex)

ValueLog.Compression defaults to auto when unset.

Details: docs/TREEDB_TUNING.md.

Exclusive Open (Process Lock)

TreeDB acquires an exclusive lock on Options.Dir. If another process has the database open, treedb.Open returns treedb.ErrLocked.

Testing

TreeDB includes a comprehensive test suite covering unit functionality, integration, fuzzing, and crash recovery.

Unit & Integration Tests

go test -v ./...

Fuzz Testing

Model-based fuzzing verifies consistency against a simple in-memory map map model.

go test -v ./db/fuzz_test.go

Crash Simulation

The verify_crash.sh script compiles a stress tool, runs it, kills it (kill -9), and verifies database integrity upon restart.

./verify_crash.sh

License

MIT