Query your files with SQL.
No server. No import. No extensions.

Embedded analytical database in C++20. Built-in readers for CSV, Parquet, JSON, Avro, Excel, Arrow, SQLite. 1.1×–6.6× faster than DuckDB on every benchmark we ran.

$ pip install slothdb

Get SlothDB Try it Live

LEVEL 01 — SYSTEM

whoami

# SlothDB --version

terminal~/slothdb

NAME         slothdb
VERSION      0.1.3
CATEGORY     embedded analytical database (OLAP)
LANGUAGE     C++20
LICENSE      MIT
BINARY_SIZE  8 MB, self-contained (no Java / Docker / daemon)
FORMATS      CSV · Parquet · JSON · Avro · Excel · Arrow · SQLite
APIS         CLI · Python (pip install slothdb) · C/C++ (stable ABI)
GPU          CUDA (NVIDIA) · Metal (Apple), auto at 100k rows
PLATFORMS    Linux · macOS · Windows · x86_64 + ARM64
TESTS        326 passing · 131,321 assertions
SPEED        1.1×–6.6× faster than DuckDB on 15/15 benchmarks
# end-of-profile

LEVEL 02 — INVENTORY

what ships in the binary

Every format, feature, and API — no extension downloads, no "load extension first".

◈ File readers

CSV · read_csv()
Parquet · read_parquet()
JSON / NDJSON
Excel (.xlsx) · read_xlsx()
Avro · read_avro()
Arrow IPC
SQLite · sqlite_scan()

◈ SQL surface

130+ features
Window functions · QUALIFY
Recursive CTEs
MERGE · UPSERT
Lateral joins
UNION / INTERSECT / EXCEPT

◈ Engine

Vectorized columnar (2048 values/batch)
Morsel-driven parallelism
Zero-copy VARCHAR via shared buffers
Fused WHERE + aggregate
Dict-index GROUP BY fast path
mmap file I/O

◈ APIs

CLI with linenoise (arrow keys, history, TAB)
Python: pandas DataFrame output
C/C++: stable C ABI
Structured numeric error codes

◈ Storage

Single-file .slothdb format
Atomic checkpoints
RLE · dictionary · bitpacking
Zone maps / min-max

◈ Reliability

326 passing unit tests
131,321 assertions
Bounds-checked parsing · DoS limits
UTF-8 identifiers (CJK column names)

LEVEL 03 — BUILT FOR YOUR STACK

ecosystem

SlothDB links native into your platform, language, and data format — no bindings or extensions to install separately.

◆ Platforms

🐧 Linux 🍎 macOS 🪟 Windows x86_64 ARM64 WebAssembly

◆ Clients / bindings

⌘ CLI shell 🐍 Python 🔤 C C++ Node.js Go R

◆ File formats (native)

CSV Parquet JSON / NDJSON Avro Excel .xlsx Arrow IPC SQLite

◆ Data science

🐼 pandas 📓 Jupyter NumPy Polars Apache Arrow

◆ Storage / I/O

local filesystem mmap'd reads .slothdb file GPU (CUDA) GPU (Metal) S3 · GCS · HTTP

◆ SQL surface

130+ features Window fns QUALIFY Recursive CTEs MERGE / UPSERT Lateral joins

+ = on the roadmap · current features are green-lit in v0.1.3

LEVEL 04 — TRY IT

run a query right now

Click a query — see what SlothDB would print. To run on your files, install (30s).

slothdb> ▮

↑ these outputs are real numbers from the benchmark. Install to run against your own data.

LEVEL 05 — BENCHMARKS

1M rows · warm cache · 5-run median

Sort by speedup; every bar is SlothDB winning.

15/15won

6.6×biggest

1.04×smallest

5formats

Avro · SUM

140ms vs 760ms

5.43×

CSV · COUNT(*)

33ms vs 170ms

5.08×

Avro · GROUP BY

170ms vs 800ms

4.71×

Parquet · COUNT(*)

12ms vs 34ms

2.83×

CSV · GROUP BY

100ms vs 191ms

1.91×

CSV · WHERE + GBY

107ms vs 194ms

1.81×

CSV · GBY 2-col

117ms vs 198ms

1.70×

CSV · SUM

106ms vs 177ms

1.67×

Excel · GROUP BY

2.5s vs 3.56s

1.41×

JSON · SUM

242ms vs 314ms

1.30×

Parquet · WHERE + GBY

157ms vs 198ms

1.26×

Parquet · GBY 2-col

146ms vs 173ms

1.18×

Parquet · GROUP BY

76ms vs 88ms

1.16×

JSON · GROUP BY

284ms vs 324ms

1.14×

Parquet · SUM

46ms vs 48ms

1.04×

Parquet SUM at 4% is within run-to-run noise. The Avro gap is wide partly because DuckDB reads Avro through an extension — not apples-to-apples there.

LEVEL 06 — BOSS BATTLE

SlothDB vs {DuckDB, ClickHouse, SQLite}

Honest, side-by-side. If another tool wins for your workload, we say so.

vs DuckDB — same category

	SlothDB	DuckDB
1M-row benchmark	1.04×–6.6× faster on 15/15	baseline
Built-in file formats	7 (all in binary)	3 (Excel/Avro/SQLite need extensions)
Extension ABI	Stable C ABI	Internal C++ API, breaks often
GPU offload	CUDA + Metal	CPU only
Binary size	~8 MB	~50 MB

vs ClickHouse — different category

	SlothDB	ClickHouse
Deployment	8 MB, embedded	server + Keeper + config
Cold start	< 10 ms	tens of seconds
Ops overhead	none	daemon, ports, upgrades
Cluster / distributed	no	yes

What SlothDB does not do

No distributed queries · use ClickHouse at scale
No MVCC transactions · OLTP is a poor fit
Younger codebase · open issues on GitHub when you find bugs

LEVEL 07 — FINAL BOSS

defeat the install command

You made it this far. The last boss is one line of shell.

★ INSTALL SLOTHDB v0.1.3

Pick your platform. Every method is free, MIT licensed, no signup.

$ pip install slothdb

🐍 PYTHON

pip install slothdb

🐧 LINUX / 🍎 MACOS

curl -fsSL .../install.sh | bash

🪟 WINDOWS

download slothdb.exe →

📦 DEBIAN / UBUNTU

dpkg -i slothdb_0.1.3_amd64.deb

ALL DOWNLOADS READ THE DOCS

Query your files with SQL.No server. No import. No extensions.