KOLMOS Hero Background

Databases
should explain
themselves.

One unified self-compressing engine behind PostgreSQL, MySQL, and MongoDB wire protocols. Bit-exact on Cloudflare R2.

Connect Standard Drivers With Zero Code Changes

PostgreSQL
MySQL
MongoDB
Apache DataFusion
Cloudflare R2
Apache Arrow
Capabilities

Storage designed for
extreme efficiency.

Every layer of KOLMOS is engineered to minimize physical byte storage while maintaining 100% bit-exact decode fidelity and full ANSI SQL query capability.

01

Explanation Ladder

Autonomous MDL formula mining & prototype clustering reduces physical bytes.

1.66×–2.12×smaller than Parquet-zstd
02

Bit-Exact Decode

Invariant #1 guarantees decode(encode(x)) == x across all data types.

100%lossless mathematical fidelity
03

WASM Covenant

Sandboxed wasmtime decoders pinned in segment footers ensure perpetual recovery.

WASMzero-import decoder sandbox
04

Fast Append WAL

Local write-buffer buffers incoming writes before ladder background compaction.

WALwrite-buffer log & DML overlay
05

Three Wire Doors

Connect standard PostgreSQL, MySQL, and MongoDB drivers.

3 Doorszero application code changes
06

Radical Economics

Store data on Cloudflare R2 ($0.015/GB/mo) with zero data egress charges.

10×–40×lower cloud storage bills
Measured on release build benchmarks (TPC-H, NYC Taxi, Clickstream, MongoDB 1M)
FastCDC Chunks·Blake3 CAS·DataFusion 54.1
The Radical Cost & Speed Advantage

Why replace AWS RDS & MongoDB Atlas with KOLMOS?

Traditional databases trap your data on expensive SSD disks and force you to build complex ETL pipelines into separate data warehouses. KOLMOS unifies self-compressing storage directly on Cloudflare R2 at 10×–16× lower storage cost.

Live Cost Calculator

Compare Your Monthly Storage Bill

Drag the slider to select your uncompressed database volume:

Database Volume:500 GB
50 GB1 TB2.5 TB5 TB
MongoDB Atlas$0.250 / GB
$125.00 / month

Requires 500 GB provisioned high-IOPS NVMe disks + egress transfer charges.

Heavy RAM footprint & disk-scaling costs
AWS RDS (gp3 EBS)$0.115 / GB
$57.50 / month

Row-based uncompressed tables + IOPS baseline limits and snapshot surcharges.

Requires manual EBS volume expansion
Save 98%
KOLMOS on R2 / S3$0.015 / GB
$1.88 / month
Footprint: 125 GB (4× Columnar Compression)

Direct serverless object storage + $0.00 Egress fees + infinite auto-scale.

You save $123.12/month ($1477/yr)
Architecture Elimination

Stop Managing 2 Databases & 1 Fragile Pipeline

The Legacy 3-Piece Stack
1. Primary App DatabasePostgres / Mongo (Writes)
↓ Kafka / Debezium / Airflow (Sync lag & outages)
2. Analytical WarehouseClickHouse / Snowflake (Scans)
  • 2 separate server bills for app transactional DB + analytical cluster
  • Data sync delays (dashboards and reports lagged by minutes or hours)
  • Heavy engineering overhead to fix broken schema changes in ETL
The KOLMOS Unified Engine
1. Connect Standard ClientsPG · MongoDB · MySQL
KOLMOS Engine (WAL + Columnar CAS)
Writes: Fast local WAL append buffer
Scans: Vectorized SIMD scan over R2 objects
  • Zero ETL: App CRUD and analytical queries execute on the exact same database
  • Drop-in: Works out-of-the-box with Prisma, Drizzle, Mongoose, Django, and psql
  • R2 Economics: Columnar data lives directly on Cloudflare R2 ($0.015/GB, zero egress)

Direct Capabilities Matrix

Comparing traditional cloud database tiers vs KOLMOS Cloud

Read technical specs
CapabilityAWS RDS / AuroraMongoDB AtlasKOLMOS Cloud
Storage Cost per GB$0.115 / GB$0.250 / GB$0.015 / GB (10x–16x Cheaper)
Data Egress FeesUp to $0.09 / GBCloud dependent$0.00 (Zero Egress on R2)
Storage FormatRow-oriented (B-Tree blocks)BSON / WiredTiger tablesColumnar KSF1 + Formula Mining (4x–5x smaller)
Analytical Queries (`COUNT`, `GROUP BY`)Slow row scans (reads all columns)High memory pipelines10×–50× Faster (Vectorized SIMD Columnar)
Drop-in CompatibilityPostgreSQL / MySQL onlyMongoDB MQL onlyMulti-Wire (Postgres + Mongo + MySQL in one)
Disk Full RiskRequires manual EBS scalingRequires cluster tier upgradeNever (Serverless infinite S3/R2 storage)
Data That Explains Itself

From raw data
to living insight.

KOLMOS doesn't just store your data — it understands it. The Explanation Ladder computes the smallest mathematical program to regenerate exact rows, reducing storage footprint on Cloudflare R2.

  • Mathematical formulas surfaced automatically on write
  • 1.66× to 2.12× storage reduction over Apache Parquet-zstd
  • WASM decoder covenant guarantees perpetual data decodability
  • Bit-perfect fidelity verified on every single read
See all architectural capabilities
Explain Report ✦
kolmos-cli explain-report
$ kolmos explain-report --table orders
Segment 200,000 rows · 12.7 MB Parquet → 7.67 MB KSF1
Patterns & formulas mined automatically (MDL):
total_price ← derived from unit_price * quantity (Rung 1)
order_id ← follows sequence step offset (Rung 1)
user_profile ← k-medoids prototype cluster delta (Rung 2)
Storage reduced 1.656× vs Parquet-zstd (Lineitem)
Data fidelity EXACT · bit-perfect restore (18/18 Goldens) ✓
WASM Covenant Pinned · blake3 wasmtime sandbox ✓
Physical Compression Results
Compression
1.656×
Throughput
248K/s
Segment Rows
200,000
💡

MDL Rule: KOLMOS selected Rung 1 formula replacement because total program + residual bytes were lower than Rung 0 zstd dictionaries.

Data Lifecycle
Ingest
Fast WAL append
Mine AST
Formulas surfaced
Compress
1.66×–2.12× reduction
Pin WASM
Covenant hash locked
Query
Zero-copy Arrow stream
The Guarantee

Every byte written is
recoverable, bit-exact,
forever.

Measured Physical Benchmarks

Storage footprint that
redefines efficiency.

Physical benchmark measurements comparing KOLMOS against uncompressed source formats and Parquet-zstd on Cloudflare R2 storage.

1.66×–2.12×
smaller than Parquet-zstd
measured on real physical datasets
194K–329K
rows ingested / second
single-core baseline throughput
100%
bit-exact data fidelity
verified across 18 golden fixtures
Live Query Latency & Storage Comparison

Live Database Connector Measurements

🍃

MongoDB Connector (Live R2)

Document BSON

1,000,000 live BSON documents synced to Cloudflare R2. KOLMOS executes vectorized ANSI SQL aggregations over content-addressed CAS chunks instead of unindexed document scans.

// Analytical Query
SELECT event_id, AVG(latency_ms) FROM mongo_events WHERE device = 'ios' GROUP BY event_id;
Query Acceleration:51.6× Faster
Query Execution Latency51.6× faster
Source Engine Scan1,240 ms
KOLMOS Vectorized Query24 ms ⚡
Physical Storage Footprint27.4× vs Raw Saved
Original Size
71.3 MB
KOLMOS KSF1
2.60 MB
Structured SQL500,000 rows

TPC-H Lineitem

Standard 16-column analytical benchmark

Compression vs Parquet1.656×
Background Ladder EffectRung 0 Baseline
Time-Series200,000 rows

NYC Taxi Trips

Real-world fare & tip log telemetry

Compression vs Parquet1.892×
Background Ladder Effect+13% (Rung 1)
JSON Logs500,000 rows

Clickstream Events

Repetitive device & user action logs

Compression vs Parquet2.117×
Background Ladder Effect+22% (Rung 2)
Platform Workflow

Three stages.
End-to-end data lifecycle.

From ~1ms write-buffer client acks to autonomous formula optimization and zero-copy Apache DataFusion SQL execution.

01Write Path

WAL Buffer & Ingest

Ingest from PostgreSQL, MySQL, or MongoDB, or write row DML. Mutations land in the local WAL write-buffer before being partitioned into FastCDC deduplicated chunks on Cloudflare R2.

Fast WAL Append
$ kolmos ingest orders.parquet --table orders
 
Schema Hash … 0x9a8f21b7e4 (blake3)
Write Buffer … WAL Log (CRC-32 framed)
Segmenter … 200,000 rows partitioned
FastCDC … 12.7 MB → 7.67 MB (1.656× reduction)
 
✓ Atomic catalog HEAD swap committed to R2
02Background

Explanation Optimization

Background worker pool asynchronously mines affine relations, sequences, and k-medoids prototypes. If Rung 1 or Rung 2 reduces the Minimum Description Length (MDL) cost, an atomic swap is committed.

Autonomous Ladder
[re_explain] analyzing orders/seg_04a9f
 
Formula AST → total_price = quantity * unit_price
Clustering → k-medoids representative prototypes
MDL Score → Rung 1 beats Rung 0 by 13%
 
Winner: Rung 1 Formula Replacement
Commit: Zero-downtime atomic segment swap
✓ Final ratio: 1.892× smaller than Parquet
03Read Path

DataFusion SQL Execution

Execute ANSI SQL queries with Apache DataFusion 54.1. Prunes irrelevant segments via header min/max statistics before streaming zero-copy Arrow memory buffers.

Vectorized SIMD
SELECT region, SUM(total_price)
FROM orders
WHERE order_date >= '2024-01-01'
GROUP BY region;
 
Plan: Header Min/Max Statistics Pruning
Skipped 8 of 10 segments → evaluated target
 
✓ 14.2 ms · WASM covenant verified · Arrow stream
Platform Stack
013 Wire Doors & Connectors
PG · MySQL · MongoDB Wire
02DataFusion 54.1 SQL
Vectorized SIMD Arrow execution
03WAL Write-Buffer
Append log & live query overlay
04Explanation Ladder
Rungs 0/1/2/4 MDL cost optimization
05WASM Decoder Covenant
Zero-import sandbox perpetual restore
06Cloudflare R2 Storage
FastCDC deduplicated CAS objects
11 9's Durability

Immutable CAS chunks replicated across Cloudflare R2 global storage.