A single unified Rust engine behind three wire protocol doors. Built on Apache DataFusion 54.1, FastCDC content-addressed storage on Cloudflare R2, and sandboxed WASM decoder covenants.
Click on any system layer below to inspect its execution mechanics, concurrency invariants, and I/O characteristics.
PostgreSQL, MySQL, and MongoDB wire protocol listeners. Accepts standard driver connections with zero client-side code modification.
How KOLMOS executes sub-millisecond queries over billions of rows without full-table scans.
Data is partitioned into 200K-row explanation segments stored across content-addressed CAS chunks.
Range queries (WHERE date >= '2026-01-01' or WHERE amount > 500) skip non-matching segments in 0ms without downloading.
Point lookups (WHERE id = 'uuid-...' or findById) probe ~240KB sidecar Bloom filters, isolating target segments instantly.
Candidate segment decodes inside fuel-metered wasmtime sandbox, streaming zero-copy Arrow memory into DataFusion.
A clean dependency DAG — no cycles, no unsafe code, no speculative abstractions.
Fundamental types (LogicalType, SegmentId, ChunkHash, TableId, RungId)
KSF1 binary on-disk layout, LEB128 varints, TLV footers & blake3 checksums
FastCDC chunking, blake3 dedup, LocalFsStore and S3Store (Cloudflare R2)
Append-only commit log, atomic HEAD swaps, lock-free MVCC-lite snapshots
MDL cost model, Explainer/Decoder traits, formula AST mining, k-medoids
Segmenter (~200K rows), ingest connectors, background workers, GC
Apache DataFusion 54.1 TableProvider, min/max & bloom pruning execution
PostgreSQL wire protocol via datafusion-postgres
kolmos binary: ingest, query, serve, verify, explain-report, stats
Sandboxed wasm32-unknown-unknown decoder for covenant verification
FaultyStore (torn writes/partial reads), SimClock, deterministic RNG
Every segment stores the blake3 hash of its exact WASM decoder artifact. Decoders run in a zero-import wasmtime sandbox with fuel metering, permanently exempt from garbage collection.
Connect your existing PostgreSQL, MySQL, or MongoDB workloads and experience 1.66×–2.12× storage reduction today.