> ## Documentation Index
> Fetch the complete documentation index at: https://oxy.tech/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Performance

> How Airhouse performs on ClickBench, an open analytics benchmark, compared with other warehouses and engines.

We measured Airhouse with [ClickBench](https://github.com/ClickHouse/ClickBench),
an open analytics benchmark: 43 queries over a 100-million-row table, run with
ClickBench's own harness on its reference machine, an AWS c6a.4xlarge (16 vCPU,
32 GB). Airhouse ran on that single machine, with its tables as Parquet in S3
and the DuckLake catalog in Postgres on the same machine. This is a benchmark
setup, not a production deployment. All 43 queries completed, and a warm run
of the full suite took 34.3 seconds.

## Results

ClickBench scores each system by dividing its time on every query by the
fastest system's time on that query and taking the geometric mean, so **1.0
would be fastest on every query and lower is better**.

| System | Warm-run score | Warm total | Compute \$/hour | Compute \$/warm run |
| - | - | - | - | - |
| MotherDuck Jumbo | 1.31× | 14.4 s | \$4.80 | \$0.0192 |
| ClickHouse, local disk | 1.37× | 16.6 s | \$0.61 | \$0.0028 |
| DuckDB, local file | 2.56× | 26.2 s | \$0.61 | \$0.0045 |
| **Airhouse, data in S3** | **3.50×** | **34.3 s** | **\$0.61** | **\$0.0058** |
| BigQuery | 6.04× | 27.3 s | — | — |
| Databricks SQL Medium | 6.37× | 32.2 s | — | — |
| Snowflake Medium (Gen2) | 6.99× | 60.3 s | \$10.80 | \$0.1810 |

Prices are on-demand list prices on 12 October 2026, compute only. For the
self-hosted systems, \$0.61 is the hourly price of the c6a.4xlarge the
benchmark ran on, not a price for Airhouse. Snowflake is Standard edition at
\$2.00 per credit. The cost of a warm run is the warm total multiplied by the
hourly price. MotherDuck Jumbo is its published rate on 19 October 2026,
billed per second with a one-minute minimum, so a single run on its own costs
more than the table shows. MotherDuck's pay-per-query Pulse tier isn't in the
table: its own ClickBench run took 337 seconds warm, and it bills by compute
units whose rate per query MotherDuck doesn't publish. BigQuery bills per
byte scanned, which ClickBench doesn't record, and Databricks prices depend
on the plan, so neither has a comparable hourly figure.

With its data in object storage, Airhouse scores ahead of BigQuery, Databricks,
and Snowflake, and behind MotherDuck and the engines that read from local disk.
A score is an average of per-query ratios, so the ratio of two scores is not a
speed-up. At list prices, a warm run of the suite costs about \$0.006 of
compute on Airhouse's machine and \$0.18 on a Snowflake Medium warehouse.

## Why use Airhouse

Airhouse is not the fastest system in this comparison. It's built to be fast
enough for interactive and agent use: 33 of the 43 queries finished in under a
second warm, and the median query took 0.40 seconds. The systems that score
better either read from local disk instead of object storage, or are a
separate managed service with its own copy of the data. What decides the
choice is usually what comes with the warehouse:

* **It's already part of Oxygen Nexus.** Airhouse is the warehouse in the data
  layer of Oxygen Nexus. Data that Nexus ingests lands in an Airhouse database
  for each customer, and the agents and applications built on Nexus query it
  over the Postgres protocol. Using it means no second warehouse to contract,
  connect, secure, and keep in sync.
* **Your data isn't copied to another vendor.** Tables stay as Parquet files
  in your object storage, with the catalog in your Postgres, so they remain
  readable without Airhouse.
* **Built for agent access.** Each database has its own catalog, storage
  location, users, and query limits, and agents can use short-lived
  credentials instead of stored passwords.
* **Lower compute cost.** On this benchmark, a warm run used about a
  thirtieth of the compute spend of a Snowflake Medium warehouse at list price.

If a workload needs the lowest possible latency, many large joins, or mostly
first reads of large data from object storage, a different engine may suit it
better. See the notes below.

## How to read these numbers

* **Single run.** This is one run by the Airhouse team on 12 October 2026, not
  yet submitted to or reviewed by ClickBench.
* **One setting changed.** The run used a larger Parquet row-group size than
  the default. Databases provisioned today still use the smaller default until
  the new one ships.
* **Recomputed scores.** The other systems' scores are recomputed from their
  published ClickBench timings, against only the systems in this table, so
  they differ from the ClickBench website.
* **Warm runs.** The scores use warm runs, the common case for a database in
  active use. The first query after a restart reads everything from S3 and is
  Airhouse's weakest result.
* **One table, no joins.** ClickBench doesn't measure join-heavy or
  star-schema workloads.
* **Compute prices only.** The cost columns compare an EC2 machine you
  operate with a managed service's list price, which includes operations
  you'd otherwise do yourself. They exclude storage, S3 requests, the catalog
  database, data transfer, and discounts, and they aren't what you pay for
  Airhouse.

To see where these numbers come from, read the [architecture](https://oxy.tech/docs/airhouse/concepts/architecture.md),
in particular how a query reads Parquet files from object storage.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.