The last few PostgreSQL releases have made great strides in scaling up PostgreSQL performance on single hosts. However, you can also increase performance by scaling out horizontally across multiple copies of data on different hosts. This talk digs into scale-out designs and shows benchmark results using a set of portable scale-out benchmarks. Based on the data we can draw some conclusions about performance trade-offs of the different approaches.

The talk consists of the following parts:

# Introduction. Discussion of scale-out approaches for PostgreSQL including SLONY, pgpool, and middleware replication using Sequoia. Overview of key measures of scale-out performance: read and write scaling, query latency (for proxy approaches), large vs. small transactions, and effects of datatypes like text and BLOBs.

# Introduction to the Bristlecone performance test tools. Evaluator creates mixed loads that stress system resources like CPU. Benchmark runs focused benchmarks with systematic variation of test input parameters. Each tool includes "portable benchmarks" that can be run within a few minutes of downloading, plus you can write more.

# Off to the Races. Results of running the portable benchmarks against a few of our favorite scale-out configurations. For fun we'll throw in some MySQL results.

# Summary. Discussion of the trade-offs between different database scaling approaches and when scale-out makes sense versus scale-up.

Bristlecone tools referenced in this talk are available from http://bristlecone.continuent.org.