KaveonDB scorecard
Where KaveonDB stands against Trino for analytics and against PostgreSQL for the product's record store, rated per dimension from measured evidence.
Analytics — against Trino 483
| Dimension | Score | Evidence and gap |
|---|---|---|
| Throughput, matched corpus | 7 / 10 | 1.45× Trino on the rebuilt cluster (2.30 vs 1.58 QPS, five rounds, 2026-09-15); faster on 9 of 12 shapes, slower on high-cardinality grouping; the 1.90× gate is not met |
| Scale at 504M rows | 4 / 10 | Measured alone on the same nodes, Trino 483 answers the thirteen live corpus questions 5× faster (geomean; 1.6–21×): row-group pruning, decode throughput and string keys. Context answers hide most of it; the live path does not |
| SQL surface | 6 / 10 | CTEs, HAVING, CASE, window functions, WHERE subqueries, set operations; correlated subqueries, GROUPING SETS and approximate aggregates absent |
| Storage and formats | 5 / 10 | Parquet, Delta (no checkpoints), Iceberg readers on ADLS Gen2 and local disk; no S3, no table writes |
| Federation | 1 / 10 | No connectors; the Engine reads lake files. Out of scope by design |
| Distributed execution | 6 / 10 | Retry to an alternate worker, cancellation, restart reconciliation verified on AKS; aggregate and join spill still open |
| Memory and admission | 5 / 10 | Bounded by design; a 504M-row table found a leak on one digest, closed on the next |
| Optimizer | 4 / 10 | Pushdown, pruning, exact-statistics broadcast; no cost-based reordering or dynamic filtering |
| Clients and ecosystem | 2 / 10 | HTTP statement API and CLI only; no JDBC/ODBC or wire compatibility |
| Security | 6 / 10 | Entra identity, TLS, owner-bound records, revision conflicts; no row filters or column masks |
| Operability | 5 / 10 | Reproducible images, Helm, console, telemetry; no backup/restore or upgrade qualification |
Standing. On the workload it was built for, KaveonDB completes more exact queries per second than Trino on matched resources, with exact statistics and fail-closed semantics Trino does not have. Outside that workload — federation, SQL breadth, ecosystem, optimizer — Trino is years ahead.
Transactions — against PostgreSQL 18
| Dimension | Score | Evidence and gap |
|---|---|---|
| ACID for product records | 6 / 10 | CAS head publication, snapshot-bound reads, conflict proofs; no cross-kind transactions |
| General relational DML | 1 / 10 | Record kinds only, by design |
| Constraints and indexes | 2 / 10 | Key uniqueness and reference validation only |
| Isolation and concurrency | 4 / 10 | Snapshot reads and revision conflicts; one statement per request |
| Durability and recovery | 4 / 10 | Object-storage durability; head recovery and restore not yet qualified |
| Query over records | 5 / 10 | Bounded typed reads through the Engine |
| OLTP-shaped latency | 3 / 10 | A publication is an ADLS conditional write |
| Ecosystem | 1 / 10 | No wire protocol or drivers |
| Operational maturity | 3 / 10 | Retained snapshots and a fail-closed migration; the Engine catalog is still host-local |
Standing. KaveonDB is a typed product-record protocol on object storage, not a relational database, and does not claim to be one. Recovery, isolation and latency must reach the middle of the scale before PostgreSQL can be retired.
What may be claimed
- 26% more exact queries per second than Trino 483 on the declared corpus at matched resources — with the workload stated.
- An exact grouped aggregate over 504 million rows in about 20 seconds on three small workers — not “interactive.”
- Not 1.9× Trino, and not a category of one.
Canonical paper
The full assessment, per-dimension evidence, and the five changes that move the scores live in Where KaveonDB stands.