NVMe Postgres Battle: ClickHouse vs. PlanetScale

Decoding the NVMe Postgres Performance Gap

The ClickHouse blog post presents a compelling benchmark demonstrating superior performance of their managed Postgres offering over PlanetScale's Metal offering, despite both utilizing NVMe-backed infrastructure. The key takeaway is that the difference lies not in the hardware, but in the system-level optimizations applied by ClickHouse. Their use of 2MB huge pages, optimized WAL compression with larger max_wal_size, and careful management of backups and disk I/O are highlighted as significant contributors to their performance gains, which they quantify with impressive metrics like a 12% improvement from huge pages alone. The article is meticulously structured, providing clear benchmark commands, identical hardware configurations, and detailed results for different dataset sizes and instance types. This level of transparency and reproducibility is highly commendable and builds strong credibility for their claims.

However, a critical point to consider is the proprietary nature of some of these optimizations. While ClickHouse publishes some of their work (e.g., on huge pages, WAL-RUS), the full extent of their 'system-level optimizations' might not be entirely transparent or easily replicable by users on a standard PostgreSQL deployment. The article acknowledges this by stating, 'There may be additional implementation differences that are not externally visible through pg_settings.' This implies that users seeking to replicate these gains might face challenges if they're not using the ClickHouse managed service. Furthermore, while the TPC-B-like workload is a good proxy for OLTP, its direct applicability to every real-world application isn't guaranteed, as the authors rightly caution. Users should still perform their own workload validation. The benchmark's focus on NVMe-backed instances is also specific; performance characteristics on different storage types might vary.

Key Points

• ClickHouse Managed Postgres significantly outperforms PlanetScale Metal on NVMe-backed infrastructure for OLTP-like workloads.
• Performance differences are attributed to system-level optimizations in ClickHouse, not hardware.
• Key optimizations include 2MB huge pages, enabled WAL compression (lz4), and larger max_wal_size.
• Benchmark results show ClickHouse achieving 34-54% higher throughput and lower latency across different instance sizes and dataset scales.
• The benchmark is reproducible using PostgresBench and standard pgbench commands.
• Users are advised to validate performance with their own specific workloads.

---

📖 Source: Benchmarking NVMe-backed Managed Postgres: PlanetScale and ClickHouse