Kolmos replaces traditional raw page files with content-addressed explanation segments. Combining FastCDC chunk deduplication, WASM decoders, and Cloudflare R2 object storage.
High-frequency incoming records flow through a crash-safe WAL buffer before background compaction synthesizes permanent KSF segments.
Incoming INSERTs and UPDATEs append to an fsynced local log with CRC-32 frames. Queries merge WAL overlay state in-memory with sub-millisecond response.
Workers group 200,000 rows, evaluating the Minimum Description Length formula across formula miners and prototype medoids to build KSF segments.
Segments are chunked and hashed via Blake3. Deduplicated blocks upload directly to Cloudflare R2 with atomic conditional-PUT commits.
Engineered for extreme data compression, durability, and open perpetual accessibility.
Segment files utilize LEB128 variable-length integer encoding, aligned columnar chunk offsets, and 64-byte cryptographic segment footers with Blake3 digests.
Content-Defined Chunking partitions segments into normalized ~64KB chunks (4MB for high-volume S3), deduplicated globally using Blake3 hashes.
Every segment footer references a sandboxed WASM decoder artifact executing in Wasmtime with strict fuel meters, ensuring data stays decipherable decades into the future.
Header Min/Max zone maps skip 90% of segments during range scans; in-memory PK Bloom filter sidecars prune 99% of segments during point lookups.
Built ground-up for object storage at $0.015/GB/month with zero egress fees, avoiding traditional cloud database EBS and provisioned IOPS bills.
High-throughput write-ahead logging with CRC-32 framing buffers row-level mutations and merges query views in real time before background segment compaction.