PolarDB PostgreSQL cluster performance test

How it works

The cluster endpoint improves read throughput in two ways:

• Read/write splitting: Write requests go to the primary node. Read requests are distributed across the primary and read-only nodes based on each node's current load, measured by the number of active (unfinished) requests.

• Transaction splitting: Read requests inside a transaction are routed to read-only nodes while maintaining read-write consistency. For details, see Transaction splitting.

Prerequisites

Before you begin, ensure that you have:

• A PolarDB PostgreSQL cluster with one primary node and at least one read-only node

• The cluster endpoint (not the primary node endpoint)

• Transaction splitting enabled on the cluster endpoint

Test environment

This test uses PolarDB PostgreSQL 16. All settings are identical across both endpoint types except the connection endpoint.

Test workload

• Tool: Sysbench

• Scenario: oltp_read_only

• Setup: Change the Sysbench connection host to the cluster endpoint. All other parameters follow the Performance testing method (OLTP).

The oltp_read_only scenario isolates read performance, which makes it suitable for measuring throughput gains from read/write splitting and transaction splitting.

Metric

TPS (Transactions Per Second): The number of transactions the database completes per second, counted as successful COMMIT operations.

Test data volume

Performance results

At 128 concurrent connections, the cluster endpoint delivers approximately 38% higher TPS than the primary node endpoint. At 256 concurrent connections, the gain increases to approximately 70%.

What's next

• Transaction splitting

• Performance testing method (OLTP)