Modelling Postgres Performance Degradation on Burstable Cloud Instances | POSETTE 2026

Chun Lin Goh presents a POSETTE 2026 talk on how PostgreSQL performance can degrade on burstable cloud instances (for example, Azure B-series and AWS T-series) due to CPU credit depletion, and how to model the saturation point using a simple simulation approach.

Overview

What problem this talk addresses

Many teams run PostgreSQL on burstable instances to reduce cost. These instances use a CPU credit model:

• When credits are available, the VM can burst above its baseline CPU performance.
• When credits are depleted, the cloud provider throttles CPU back to the base frequency.

The key risk described is not a hard crash, but throughput exhaustion:

• PostgreSQL is not aware of the external CPU throttling.
• It continues accepting connections and work at a rate that no longer matches the throttled CPU capacity.
• This can trigger a cascading failure pattern:
  • connection pools saturate
  • p99 latencies spike
  • application requests time out
  • the database appears “online” but is effectively unavailable

Core concepts covered

• How CPU credits and bursting work on burstable instance families
• A simple explanation of the token bucket model (as an intuition for credit-based bursting)
• The “hidden risk” of CPU credit depletion and why it creates non-linear performance behavior

Approaches compared

The talk contrasts different ways to avoid getting surprised by throttling:

• Overprovisioning (buying more baseline capacity than needed)
• Load testing (often expensive to run realistically)
• Simulation as a fast decision filter to estimate when the system will saturate

Simulation approach and what it’s used for

Goh introduces simulation as a way to estimate a system’s Base Performance Ceiling (the sustainable throughput once throttled), without requiring large-scale load-testing infrastructure.

Topics called out in the session outline include:

• Discrete-event simulation fundamentals
• Building a simulation library in .NET
• A demo focused on:
  • predicting CPU credit drain and failure conditions
  • identifying a safer connection pool size for the throttled state

Cloud differences and operational visibility

• A comparison of Azure vs AWS burstable instance differences (at a conceptual level)
• The role of observability and real-time insight, with Grafana referenced for monitoring/visualization

Takeaways highlighted

• Burstable instances can be cost-effective, but the CPU credit model can create sharp performance cliffs.
• Modelling the throttled steady-state helps teams right-size instances and tune connection pools before production incidents.
• Simulation can be used as a faster, more data-driven filter than relying solely on overprovisioning or expensive load tests.