Flatbench

Search engine benchmark suite — compare Flatseek against Elasticsearch, tantivy, Typesense, Whoosh, ZincSearch, SQLite, and DuckDB.

Benchmarks: build speed, search latency, wildcard, range queries, and aggregations. Results saved as JSON + Markdown to ./output/.

---

Install

pip install flatbench

Requires Python ≥ 3.10, Docker (for full engine comparison).

---

Quick Start

1. Start all search engines (Docker)

flatbench make up

Starts: Flatseek API (port 8000), Elasticsearch (9200), Typesense (8108), ZincSearch (4080).

2. Generate a dataset

flatbench generate -s article -r 500000 -o ./data/article.csv

3. Run benchmark comparison

flatbench compare -e flatseek_cli,elasticsearch,tantivy,typesense,whoosh,zincsearch -s 500000

Results → output/benchmark_YYYYMMDD_HHMMSS.json + .md.

---

CLI Reference

Commands

Command	Description
flatbench generate	Generate synthetic dataset
flatbench compare	Compare multiple engines
flatbench run	Benchmark single engine
flatbench serve	Serve report viewer locally
flatbench make	Run infrastructure Makefile targets

Generate

flatbench generate --schema <schema> --rows <N> --output <path> [--format csv|jsonl]

Schemas: standard, ecommerce, logs, nested, sparse, article, adsb, campaign, devops, sosmed, blockchain

Compare

flatbench compare --engines <engines> --sizes <sizes> [options]

Options:

Flag	Description	Default
--schema	Data schema	standard
--workers, -w	Parallel index workers	1
--format	csv or jsonl	csv
--source, -S	Use existing CSV/JSONL instead of generating	—
--mode, -m	normal (disk) or tmpfs (RAM)	normal
--cache-dir, -c	Cache generated data for reuse	—
--skip-build	Skip build (use existing index)	—
--serve	After compare completes, build site and serve report	—

Engines:flatseek, flatseek_cli, elasticsearch, tantivy, typesense, whoosh, zincsearch, sqlite, duckdb

Sizes: multiple sizes supported, e.g. --sizes 1000 10000 500000

Run

flatbench run --engine <engine> --data <path> --index-dir <path> [-o output] [--iterations N]

Serve

flatbench serve [--dir ./output] [--port 8080]

Opens the report viewer in your browser automatically.

Make (Infrastructure)

flatbench make <targets...># Run Makefile targets (default: help)
flatbench make up # Start all services (docker-compose up -d)
flatbench make down # Stop services (keep volumes)
flatbench make clean # Stop and remove volumes
flatbench make status # Show service status
flatbench make logs # View logs (follow mode)
flatbench make benchmark NROWS=500000 # Run benchmark via Make

Service management:

Target	Description
up/down/clean/status/logs	Docker compose lifecycle
fs-health/fs-stats/fs-create/fs-delete	Flatseek API (port 8000)
es-health/es-stats/es-create/es-delete	Elasticsearch (port 9200)
ts-health/ts-stats/ts-create/ts-delete	Typesense (port 8108)
zs-health/zs-stats/zs-create/zs-delete	ZincSearch (port 4080)

Examples

# Generate article dataset (500K rows)
flatbench generate -s article -r 500000 -o ./data/article.csv
# Compare at single scale
flatbench compare -e flatseek_cli,elasticsearch -s 500000
# Compare at multiple scales
flatbench compare -e flatseek,tantivy -s 1000 10000 500000
# Use existing CSV (reuse generated data)
flatbench compare -e flatseek,elasticsearch -s 500000 -S ./data/article.csv
# RAM-backed index (tmpfs mode, faster builds)
flatbench compare -e flatseek,tantivy -s 500000 -m tmpfs
# Compare and auto-serve report
flatbench compare -e flatseek,tantivy -s 500000 --serve
# Run benchmark via Make
flatbench make benchmark NROWS=500000 ENGINES="flatseek_cli,elasticsearch,tantivy"

Service URLs:

Service	URL
Flatseek API	http://localhost:8000
Elasticsearch	http://localhost:9200
Typesense	http://localhost:8108
ZincSearch	http://localhost:4080
Kibana	http://localhost:5601 (dev profile)

---

Available Schemas

Schema	Fields	Description
article	8	Blog articles: id, title, content, tags, views, published_at, author
standard	12	Generic: id, name, email, phone, city, country, status, balance, created_at, updated_at, is_verified, tags
ecommerce	12	Order tracking data
logs	11	Log entries: timestamp, level, service, message, etc.
nested	6	Complex nested JSON objects
sosmed	9	Social media posts
devops	11	Infrastructure/monitoring data
adsb	10	Flight tracking data
campaign	10	Marketing campaign data
blockchain	9	Blockchain transaction data

---

Benchmark Operations

Operation	Description	Metrics
build_index	Bulk API indexing (1000 rows/batch)	duration_ms, rows/sec, index_size_mb
search	Full-text query	p50_ms, p95_ms, p99_ms, ops/sec
wildcard_search	Prefix/suffix wildcard queries	p50_ms, p95_ms, ops/sec
range_query	Numeric/date range filtering	duration_ms, hits, ops/sec
aggregate	Terms/stats aggregations	duration_ms, bucket_count, ops/sec

---

Output

Results written to ./output/ with timestamps:

output/
├── benchmark_20260501_142947.json # Full structured results
├── benchmark_20260501_142947.md # Markdown summary
└── index.json # Report manifest (for web viewer)

Report Viewer

Live:bench.flatseek.io — hosted Flatbench report viewer.

Local: Run flatbench serve --port 8080 or open report_viewer.html directly in browser.

---

Build Static Site

Build output directory as a static site (for self-hosted or Vercel deploy):

make build
# or
bash build.sh

Output → public/ directory with index.html, output/*.json, output/*.md.

Deploy to Vercel

make deploy # Deploy to production (flatbench.vercel.app)
make deploy-preview # Deploy preview build

---

Project Structure

flatbench/
├── Dockerfile # Flatseek API server container
├── docker-compose.yml # All engine containers
├── Makefile # Infrastructure + build commands
├── build.sh # Static site build script
├── report_viewer.html # Web UI for browsing results
├── pyproject.toml # flatbench package definition
├── src/flatbench/
│ ├── cli.py # CLI entry point
│ ├── benchmarks/ # Benchmark orchestration + report generation
│ ├── generators/ # Synthetic data generators (schema-aware)
│ ├── runners/ # Engine runners (HTTP API / CLI)
│ │ ├── flatseek_api.py # Flatseek HTTP API runner
│ │ ├── flatseek_cli.py # Flatseek CLI runner
│ │ ├── elasticsearch.py # Elasticsearch runner
│ │ ├── tantivy.py # tantivy (Rust) runner
│ │ ├── typesense.py # Typesense runner
│ │ ├── whoosh.py # Whoosh runner
│ │ ├── zincsearch.py # ZincSearch runner
│ │ ├── sqlite.py # SQLite FTS5 runner
│ │ └── duckdb.py # DuckDB full-text runner
│ └── output/ # Benchmark results (JSON + Markdown)

---

Adding a New Engine

fromflatbench.runnersimportBaseRunner, BenchmarkResult, register_engine@register_engine("myengine")classMyEngineRunner(BaseRunner):
name="myengine"supports_aggregate=Falsesupports_range_query=Truesupports_wildcard=Truedefbuild_index(self, data_path: str, **kwargs) ->BenchmarkResult:
# Bulk API indexing logicpassdefsearch(self, query: str, iterations: int=10, **kwargs) ->BenchmarkResult:
# Search via HTTP APIpass

Then add to --engines list: --engines flatseek,myengine,...

---

Benchmark Results (Latest: 500K rows, article schema)

Latest Full results:bench.flatseek.io

Overall Score (60% speed · 40% correctness)

Engine	Speed	Correctness	Score
Flatseek	🟢	🟢	0.878 ◀
typesense	🟢	🟢	0.832
zincsearch	🟢	🟢	0.823
elasticsearch	🟢	🟢	0.820
tantivy	🟢	🔴	0.650
whoosh	🔴	🔴	0.025

Key Takeaways

• Correctness matters: Flatseek is the only engine with zero correctness errors. Tantivy misses 99.4% of range query hits.
• Search: Tantivy fastest (0.7ms p50), but wrong. Flatseek second-fastest correct (7.9ms).
• Build: Tantivy wins (21s for 500K), but Flatseek build is reasonable (217s).
• Aggregation: Competitors (ES, tantivy) are 20–300× faster — Flatseek aggregation is a known weakness.