Airport I/O Benchmarks

This document presents some benchmarks of core operations using the Airport extension for DuckDB. The benchmarks cover scalar functions, table creation, and selection operations involving both BIGINT and 32-byte string data. Results are measured in execution time (duration_sec), CPU usage (cpu_sec), and system CPU usage (cpu_system), and include throughput in rows per second.

All tests were conducted using 40M to 100M rows, simulating high-throughput analytical workloads.

Summary Table

Operation	Rows	Mean Duration (s)	Rows/sec
Scalar Function Echo BIGINT	100M	2.06	48.56M
Scalar Function Add BIGINT	100M	3.42	29.27M
Scalar Function Echo 32-byte string	40M	1.07	37.36M
Create remote table from 32-byte string	40M	3.53	11.33M
Select from 32-byte string table	40M	1.10	36.21M
Create remote table from BIGINT	100M	12.22	8.18M
Select from remote BIGINT table	100M	2.39	41.78M

Note: Rows rounded to 2 decimal places for clarity.

Scalar Function Benchmarks

Echo BIGINT

Duration (mean): 2.06s
Rows/sec: 48.56M
This represents a nearly zero-cost server side scalar operation, a BIGINT is sent to the server and the value is echoed back.

Add BIGINT

Duration (mean): 3.42s
Rows/sec: 29.27M
Slightly more expensive than echo, as it performs arithmetic computation, two BIGINT values are sent to the server, and a BIGINT is returned.

Echo 32-byte String

Duration (mean): 1.07s
Rows/sec: 37.36M
High throughput despite operating on 32-byte strings, showing efficient memory handling.

Remote Table Creation

Create Remove Table from 32-byte String

Duration: 3.53s
Rows/sec: 11.33M
Performance reflects single threaded serialization and data transport overhead.

Create Remote Table from BIGINT

Duration: 12.22s
Rows/sec: 8.18M
Performance reflects single threaded serialization and data transport overhead.

Select Queries

Select from 32-byte String Table

Duration (mean): 1.10s
Rows/sec: 36.21M
Very fast reads from the remote table, indicating efficient decoding.

Select from BIGINT Table

Duration (mean): 2.39s
Rows/sec: 41.78M
Excellent performance, even at scale (100M rows), shows the power of Arrow Flight transport.

Observations

Scalar functions scale well, with Echo BIGINT and Echo 32-byte String both exceeding 35M Rows/sec.
Table creation is the most costly, especially for large datasets due to single threaded insert speeds.
Select queries from remote tables using Airport are fast, often exceeding 40M Rows/sec.
System CPU usage remains low, indicating good kernel-level efficiency with I/O.

Conclusion

These benchmarks show that the Airport extension for DuckDB delivers high throughput and low-latency data access via Apache Arrow Flight, particularly for read-heavy workloads.

It is especially suitable for:

Real-time analytics
High-volume batch reads
Remote function calls with low overhead

Further work may explore concurrency scaling and performance under streaming ingestion.

--- title: "Airport I/O Benchmarks" format: html: toc: true code-fold: true theme: cosmo code-tools: true --- This document presents some benchmarks of core operations using the **Airport** extension for DuckDB. The benchmarks cover scalar functions, table creation, and selection operations involving both `BIGINT` and 32-byte string data. Results are measured in execution time (`duration_sec`), CPU usage (`cpu_sec`), and system CPU usage (`cpu_system`), and include throughput in rows per second. All tests were conducted using 40M to 100M rows, simulating high-throughput analytical workloads. # Summary Table | Operation | Rows | Mean Duration (s) | Rows/sec | |------------------------------------------|-----------:|------------------:|-----------------:| | Scalar Function Echo BIGINT | 100M | 2.06 | 48.56M | | Scalar Function Add BIGINT | 100M | 3.42 | 29.27M | | Scalar Function Echo 32-byte string | 40M | 1.07 | 37.36M | | Create remote table from 32-byte string | 40M | 3.53 | 11.33M | | Select from 32-byte string table | 40M | 1.10 | 36.21M | | Create remote table from BIGINT | 100M | 12.22 | 8.18M | | Select from remote BIGINT table | 100M | 2.39 | 41.78M | _Note: Rows rounded to 2 decimal places for clarity._ # Scalar Function Benchmarks ## Echo BIGINT - **Duration (mean):** 2.06s - **Rows/sec:** 48.56M - This represents a nearly zero-cost server side scalar operation, a `BIGINT` is sent to the server and the value is echoed back. ## Add BIGINT - **Duration (mean):** 3.42s - **Rows/sec:** 29.27M - Slightly more expensive than echo, as it performs arithmetic computation, two `BIGINT` values are sent to the server, and a `BIGINT` is returned. ## Echo 32-byte String - **Duration (mean):** 1.07s - **Rows/sec:** 37.36M - High throughput despite operating on 32-byte strings, showing efficient memory handling. # Remote Table Creation ## Create Remove Table from 32-byte String - **Duration:** 3.53s - **Rows/sec:** 11.33M - Performance reflects single threaded serialization and data transport overhead. ## Create Remote Table from BIGINT - **Duration:** 12.22s - **Rows/sec:** 8.18M - Performance reflects single threaded serialization and data transport overhead. # Select Queries ## Select from 32-byte String Table - **Duration (mean):** 1.10s - **Rows/sec:** 36.21M - Very fast reads from the remote table, indicating efficient decoding. ## Select from BIGINT Table - **Duration (mean):** 2.39s - **Rows/sec:** 41.78M - Excellent performance, even at scale (100M rows), shows the power of Arrow Flight transport. # Observations - **Scalar functions** scale well, with `Echo BIGINT` and `Echo 32-byte String` both exceeding 35M Rows/sec. - **Table creation** is the most costly, especially for large datasets due to single threaded insert speeds. - **Select queries** from remote tables using Airport are fast, often exceeding 40M Rows/sec. - System CPU usage remains low, indicating good kernel-level efficiency with I/O. # Conclusion These benchmarks show that the **Airport** extension for DuckDB delivers high throughput and low-latency data access via Apache Arrow Flight, particularly for read-heavy workloads. It is especially suitable for: - Real-time analytics - High-volume batch reads - Remote function calls with low overhead Further work may explore concurrency scaling and performance under streaming ingestion.