PostgreSQL Performance Issues Often Stem from Storage, Not Database Itself

Storage Impact on PostgreSQL Performance

At POSETTE 2026, Sai Srirampur presented findings indicating that storage performance is a critical factor in PostgreSQL scalability. The presentation, titled "Postgres Isn’t Slow. Your Storage Is," explored how different storage solutions affect PostgreSQL's ability to handle large-scale workloads.

Common symptoms of performance degradation in PostgreSQL, such as slow data ingestion, unpredictable P95 read latency, autovacuum falling behind, and logical replication lag, are often attributed to the database itself. However, the analysis suggests that these issues frequently stem from underlying storage limitations.

Benchmark Methodology and Findings

To investigate the role of storage, a benchmark was conducted using a 3.3-billion-row PostgreSQL workload. This workload was run on two different storage configurations: baseline gp3 EBS and local NVMe. The comparison aimed to isolate the impact of storage on various performance metrics.

The results revealed that as the working set size exceeds available memory, PostgreSQL relies more heavily on physical I/O. Slower storage solutions lead to higher costs and queuing for these read operations, directly contributing to the observed performance bottlenecks.

Common Scaling Symptoms Linked to Storage

The presentation identified five key symptoms that emerge when PostgreSQL workloads reach hundreds of gigabytes or terabytes with high concurrency:

1. Slow ingestion: UPDATEs and UPSERTs take longer as data volume grows.

2. Inconsistent read latency: P95 latency increases significantly for some queries.

3. Autovacuum falling behind: Table bloat accumulates due to slow VACUUM processes.

4. Slower checkpoints: Checkpoints compete with workloads for I/O and CPU, causing latency spikes.

5. Logical replication lag: WAL decoding becomes slow, leading to growing replication slots.

Production Patterns for Local NVMe

The analysis concluded by proposing specific architectural patterns for deploying PostgreSQL on local NVMe in production environments. These patterns are designed to mitigate the storage-related performance issues identified in the benchmark, offering a path to improved scalability and stability for large PostgreSQL instances.