Benchmarks

Reproducible benchmarks on commodity hardware. All numbers measured on a single node unless noted.

---

Hardware

Component	Spec
CPU	AMD EPYC 9654 (96 cores) / Intel Xeon Platinum 8488C
RAM	256 GB DDR5-4800 ECC
Storage	NVMe Gen4 (for WAL & Parquet HDB)
OS	Amazon Linux 2023, kernel 6.1
Compiler	Clang 19, -O3 -march=native

---

Ingestion Throughput

Scenario	Events/sec	Latency (p99)
Single-symbol tick stream	5.52M	181ns
Multi-symbol (1,000 syms)	4.8M	210ns
Kafka consumer (batch 10K)	3.2M	850μs batch
FIX 4.4 market data	1.1M	420ns parse+ingest

Lock-free MPMC ring buffer with Highway SIMD batch copy. Zero allocation on hot path.

---

Query Latency

All queries on 1M-row in-memory table, single thread.

Query	Latency
SELECT * FROM trades WHERE sym='AAPL' AND ts > now()-1h	272μs
SELECT avg(price), max(volume) FROM trades GROUP BY sym	185μs
SELECT * FROM trades ASOF JOIN quotes USING(sym, ts)	410μs
SELECT sym, ema(price, 20) FROM trades	320μs
SELECT xbar(1m, ts) as bucket, vwap(price, volume) FROM trades GROUP BY bucket	290μs
Window JOIN (±500ms)	580μs

LLVM JIT compilation. Vectorized execution with SIMD aggregation.

---

Python Zero-Copy

Operation	Latency
conn.query("SELECT * FROM trades") → NumPy array	522ns
DataFrame view (1M rows × 5 cols)	1.2μs
PyTorch tensor from query result	890ns

Direct memory-mapped view. No serialization, no copy, no Arrow conversion.

---

Comparison

	ZeptoDB	kdb+	ClickHouse	TimescaleDB	InfluxDB
Ingestion (events/sec)	5.52M	~5M	100K	50K	50K
Point query latency	272μs	~300μs	~5ms	~10ms	~15ms
ASOF JOIN	✓	✓	✗	✗	✗
SQL	Standard	q lang	✓	✓	InfluxQL
Python zero-copy	522ns	IPC (~ms)	—	—	—
License cost	Free (OSS)	$100K+/yr	Free	Free	Free

---

Reproduce

git clone https://github.com/zeptodb/zeptodb.git && cd zeptodb
mkdir -p build && cd build
cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_C_COMPILER=clang-19 -DCMAKE_CXX_COMPILER=clang++-19
ninja -j$(nproc)

# Ingestion benchmark
./bench/bench_ingestion --symbols 1 --duration 10s

# Query benchmark
./bench/bench_query --rows 1000000 --iterations 100

# Python zero-copy
python3 ../bench/bench_python_zerocopy.py

See benchmark source code for full methodology.