ZebraDB
Hybrid transactional-analytical data platform for agent-native applications.
ZebraDB is the system of record for Gen4 applications. Postgres OLTP underneath, pluggable analytical engines on the side, and a wire format (Arrow) that the rest of the DOSS stack consumes natively. It exists because every enterprise app ships with an OLTP database, a warehouse, an ETL pipeline, and a sync layer — four systems doing roughly one job. ZebraDB collapses those four into one.
What a Z-Table is
One native Postgres table per customer table
Not a sharded row in a shared table. Each customer gets a dedicated relation with its own indexes, partitions, and history. Isolation is physical, not logical.
Append-only history ledger
Every write is an insert. Current-state views are projections over the ledger. Point-in-time queries and audit trails are free.
Schema-aware
Z-Tables are declared in ZSL, not managed via ad-hoc migrations. The table, its indexes, its RLS policies, its triggers — all compiled from a single source file.
Arrow-native reads
Pair with DossGrid over FlightSQL for zero-copy reads. DataFusion produces Arrow, the wire carries Arrow, the browser consumes Arrow. Same bytes end to end.
Stack
Four systems, one substrate
| Concern | Legacy stack | ZebraDB |
|---|---|---|
| Transactional writes | Postgres / Oracle / SQL Server | Postgres (Z-Table) |
| Analytical reads | Snowflake / BigQuery | DuckDB embedded + ClickHouse bridge |
| ETL / replication | Fivetran, Airbyte, custom CDC | None — reads hit live data |
| Sync layer | Debezium, webhooks, polling | None — same store |
| Multi-tenancy | Row-level logical partitioning | Physical per-tenant tables |
| History / audit | SCD-2 tables, bolt-on CDC | Native append-only ledger |
Performance targets
Horizontal sharding keeps a single tenant dense.
Target vs. legacy OLTP + OLAP + ETL + sync.
Read-your-own-writes across OLTP and OLAP reads.
10MB batch into the grid's DataBus.
In the stack
ZSL declares Z-Tables. `zsl apply` creates/alters them.
Workflows read/write Z-Tables via the same Postgres wire; CDC emits into ZFlow events.
FlightSQL bypass streams Arrow from ZebraDB straight to the grid — one hop, zero serialization.
Pre-training data flows out of Z-Tables as OCEL-formatted event logs.
