01 / PRINCIPLE
S3 is the only truth
No Raft, no Paxos, no replica sync. Consistency is delegated to object storage; ingest and compute are stateless, so you can kill -9 any node and rebuild from S3 alone.
BteloDB — S3-first streaming lakehouse
One engine for real-time data
A streaming lakehouse written in Rust. Ingest like Kafka, query like ClickHouse, search without Elasticsearch — one engine on S3, zero data copies, no JVM.
Scroll ↓
One engine, four jobs
KAFKA · CLICKHOUSE · ELASTICSEARCH
01IngestPOST raw JSON — no schema, no driver. Durable to WAL in microseconds, monotonic offsets.
02StreamTail the live write stream by offset. Hot from memory, cold from S3 Parquet — one seamless feed.
03QueryStandard SQL over columnar Parquet, with read-your-writes across hot and cold in one view.
04SearchLIKE, regex and contains, SIMD-scanned with Bloom pruning. Elasticsearch, deleted.
05Key / ValueLinearizable single-key writes: compare-and-set, create-only, atomic counters. No Redis.
06StoreRow→column lifecycle to immutable Parquet on S3. Compute is stateless and scales to zero.
A database earns trust by never losing a byte. Speed is what you add next.
02 / PRINCIPLE
Cost, collapsed
One storage base instead of four clusters. On R2 there is zero egress; idle tenants cost almost nothing and compute scales to zero. The target is 70% less stack TCO.
03 / PRINCIPLE
Proven, not promised
100% TDD with a dual 100%-coverage gate and deterministic simulation testing. Twenty-four scenario contracts run against a mock first, then the real engine, unchanged.
≥100Kwrites / second, per node, all the way to WAL
<20msP99 hot subscribe latency on the internal network
<1sto aggregate 10M rows across four compute nodes
→$0operating cost — elastic, scales to zero when idle
Stream·Store·Query·Search·Stream·Store·Query·Search·