Benchmarks

Results

TensorDB sled redb SQLite

Point Read Latency (lower is better)

TensorDB

276ns

sled

244ns

redb

573ns

SQLite

1.1µs

Point Write Latency (lower is better)

TensorDB

1.9µs

sled

9.2µs

redb

18.4µs

SQLite

38.6µs

Prefix Scan (100 keys, lower is better)

TensorDB

45µs

sled

38µs

redb

89µs

SQLite

156µs

Mixed Workload (80% read, 20% write)

TensorDB

2.1µs

sled

4.8µs

redb

7.2µs

SQLite

12.4µs

Benchmarks run on Linux x86_64, single-threaded, 1M pre-loaded keys. Lower is better for all measurements. See full benchmarks.

Detailed Numbers

Point Read Latency

Engine	P50	P99	Notes
TensorDB	276ns	450ns	Direct ShardReadHandle, bloom + cache
sled	244ns	380ns	B-tree, memory-mapped
redb	573ns	890ns	B-tree, copy-on-write
SQLite	1,080ns	2,100ns	B-tree, WAL mode

Point Write Latency

Engine	P50	P99	Notes
TensorDB	1.9µs	4.2µs	Fast write path, group WAL
sled	9.2µs	18µs	Write-ahead log
redb	18.4µs	35µs	Copy-on-write B-tree
SQLite	38.6µs	72µs	WAL mode, fsync per write

Prefix Scan (100 keys)

Engine	Latency	Notes
sled	38µs	B-tree range scan
TensorDB	45µs	LSM multi-level merge
redb	89µs	B-tree range scan
SQLite	156µs	B-tree + SQLite overhead

Mixed Workload (80% read, 20% write)

Engine	Avg Latency	Notes
TensorDB	2.1µs	Fast path for both reads and writes
sled	4.8µs	Memory-mapped reads
redb	7.2µs	Copy-on-write overhead
SQLite	12.4µs	Lock contention

Methodology

Platform: Linux x86_64
Dataset: 1M pre-loaded keys
Workload: Single-threaded, sequential
Measurements: 10K iterations, warm cache
SQLite: WAL mode, default settings
TensorDB: Default config (4 shards, fast write enabled)

Why TensorDB is Fast

Read Path

Direct ShardReadHandle: Bypasses actor channel (saves ~100ns)
Bloom filters: Skip SSTables with 99% accuracy
Block cache: 32MB LRU cache keeps hot data in memory
Prefix compression: Reduces data size for faster scans

Write Path

Fast write path: Atomic operations instead of channel messaging
Group WAL: Batched fdatasync reduces syscall overhead
Skip list memtable: O(log n) concurrent inserts

Running Benchmarks

# Basic benchmarks
cargo bench --bench basic

# Multi-engine comparative benchmarks
cargo bench --bench comparative

# Full overnight benchmark suite
./scripts/overnight_bench.sh