GizmoData Blog

Hardcharger is GizmoData's new open-source server for DuckDB 2.0, built on Warp SQL, an open protocol for SQL over HTTP with Apache Arrow results. On one Graviton5 machine it ran 22 queries derived from TPC-H over 1 TB in 42.4 seconds, for about 9 cents of compute, and every feature is included: access control, SSO, mTLS, audit, instrumentation and a built-in web UI. Apache-2.0, no license keys.

adbc-replicate is our new open-source tool for copying whole schemas from Snowflake, Oracle, PostgreSQL or DuckDB into GizmoSQL. Data streams as Arrow over ADBC from end to end, masks run inside the source, and each run copies only what changed. Two customers, one on Snowflake and one on Oracle, run it every hour into GizmoData Cloud, and their AI agents query the copy over MCP all day without touching the source.

We raced six 64-vCPU GizmoSQL clusters on AWS, Google Cloud and Azure over 1 TB of TPC-H, on local NVMe and on DuckLake, with DuckDB 1.5.6 and the DuckDB 2.0 alpha. AWS Graviton5 ran the fastest lap (79 seconds for all 22 queries, $0.30), Google Cloud took four of six golds, and DuckDB 2.0 beat 1.5.6 in every lane. Twelve results with a price on every one, in 38 minutes and about $37 on GizmoData Cloud.

GizmoData Cloud is now generally available and open to everyone: fully managed GizmoSQL clusters, DuckLake catalogs and federated queries on AWS, Google Cloud and Azure, run from one control plane. Prepaid plans from $29/month, with every dollar returned as compute credit. Start free with $25 of launch credit, applied automatically, no credit card.

Nobody ran a query since Friday, yet Monday morning the GizmoSQL server is at 80% memory. The queries left; the connections didn't. Guest author Emmanuel Sibanda walks through the two flags that fix it: --max-sessions caps how many sessions can exist, and --session-idle-timeout evicts the ones that went quiet, with an idle clock that only real user SQL (including result streaming) can reset.

Scaling stateless compute is easy. Scaling independent compute nodes that all need the same mutable data is the real problem in analytics. With DuckLake holding the transactional catalog and the Parquet files in object storage, a GizmoSQL cluster can grow from one node to five and back without copying data, rebuilding metadata, or reconciling state.

Guest post from Aalborg University's Daisy research group. Seagull serves 18.7 billion AIS records from 126,400 ships as live map layers, with every vector tile computed on request from DuckDB. PostgreSQL needed a pile of pre-aggregated tables to keep up; DuckDB was fast but single-writer. GizmoSQL turned it into a centralized database that Airflow, MapLibre Martin, and the web app all share over Arrow Flight SQL.

We ran our 22-query TPC-H 1 TB benchmark on one of the first AWS Graviton5 instances: 80.5 seconds for $0.17 of compute. The same afternoon, same DuckDB, a Graviton4 r8gd.24xlarge took 89.8 seconds and an Azure Cobalt 100 E96pds_v6 took 106.8, so Graviton5 is 10% faster than Graviton4 and 25% faster than Cobalt 100, winning all 22 queries. Full methodology and every per-query timing.

We shipped an MCP server and Claude Desktop extension for GizmoSQL. Claude Desktop, Claude Code, and any Model Context Protocol client can now explore your schema and run SQL against a GizmoSQL server — read-only by default, with row caps and timeouts the server enforces, and credentials that never leave your machine.

We built pure-Go, pip-installable ADBC drivers for Oracle (no Instant Client), IBM Db2 (no CLI/JCC), and DuckDB's Quack — alongside our flagship GizmoSQL driver. Because they're all Arrow-native, moving a table from Oracle or Db2 into GizmoSQL is a few lines of Python, or a single SQL statement from DuckDB — in both directions.

GizmoSQL v1.37.0 stopped hand-converting every Arrow cell on the way into DuckDB and started streaming the whole batch through DuckDB's own arrow_scan instead — 7.2x faster bulk loads, exact decimals, and geometry columns that finally land as GEOMETRY.

The GizmoSQL ADBC driver has been rebuilt in Go with a thin, drop-in-compatible Python wrapper. One native driver now brings OAuth/SSO, automatic DDL/DML detection, and the rest of the GizmoSQL goodies to Python, Go, R, Rust, C#, C/C++, and JavaScript.

DuckDB Labs announced Quack, a native client-server protocol for DuckDB. A thank-you to the team for the kind acknowledgement — and a look at how GizmoSQL complements the growing client-server DuckDB ecosystem.

A Matrix-themed deep dive into building an AI-agent-ready architecture with GizmoSQL, DuckDB, and Claude. Simulates one million people, mutates them with automated agents, and monitors anomalies with an autonomous AI—all over Arrow Flight SQL.

A deep dive into GizmoSQL—a small server that runs DuckDB with the Arrow Flight SQL protocol wrapped around it so you can run DuckDB remotely. Originally published on HackerNoon.

Announcing the GizmoSQL Kubernetes Operator—a tool designed to reduce the operational cost of running large-scale SQL queries with reported savings of 90% compared to competing cloud data engines.

The new ADBC Flight SQL Python driver 1.10.0 brings bulk ingestion to Arrow Flight SQL servers like GizmoSQL. Load millions of rows directly from Python or DuckDB with automatic schema inference.

What happens when you give a distributed SQL engine a trillion-row dataset? GizmoEdge crushed the Coiled challenge with a 1,000-worker cluster on Azure.

GizmoSQL completed the 1 trillion row challenge! Learn how we processed massive datasets using DuckDB and Apache Arrow Flight SQL for just $0.56 on-demand.