Predicting and Reducing Serverless Function Costs with a New FinOps Model

Key Takeaways

1. Despite rapid growth of Serverless, there is no effective theory for estimating total function cost; this work proposes the first cost estimation model based on FunctionGraph FinOps practice.

2. Five categories of cost‑optimization methods are identified, and Huawei Cloud launches a transparent, one‑click "User Function Cost Research Center".

Background and Terminology

Serverless charges by the millisecond, eliminating idle resource fees, but estimating total billing for a given function is difficult because multiple factors (memory size, instance concurrency, execution duration) influence cost, and traffic is often random and non‑stationary.

FinOps focuses on cloud resource management and cost optimization by linking technical, business, and financial perspectives to improve cost‑performance ratios.

Function Cost and Model

Function billing consists of resource usage (GB‑seconds) and request count. The model defines resource specification R (GB), maximum per‑instance concurrency c, average execution time μ (ms), and platform performance factor α (≈0.9‑1). The ideal per‑instance request handling capacity is: λ = (α * R) / (μ * c) Actual capacity accounts for traffic fluctuations, leading to a piecewise linear function of concurrent instances over time. The total monthly cost aggregates the resource cost and request cost across all sub‑intervals, subtracting free tier allowances.

Cost Optimization Methods

Point 1: Optimize function code to reduce execution time μ, e.g., use lightweight languages, move external service initialization out of the request path.

Point 2: Minimize code package, dependency, and image size to reduce cold‑start latency, which is billed as part of μ.

Point 3: Keep functions focused on a single purpose; split large functions into smaller ones and orchestrate with FunctionGraph flows.

Point 4: Enable per‑instance multi‑concurrency when the workload spends significant time waiting for downstream services, reducing the number of instances needed.

Point 5: Choose memory/specification wisely: larger specs increase cost but can lower μ; find the Pareto‑optimal point where marginal cost increase outweighs execution‑time gain.

Function Cost Research Center

Huawei Cloud will soon launch a Cost Analysis and Optimization Center offering offline best‑configuration tuning, online resource‑consumption awareness with dynamic spec recommendation, and predictive auto‑scaling previews, lowering the technical barrier for Serverless FinOps.

Conclusion

The paper presents the first total‑cost estimation model for Serverless functions, analyzes key cost drivers, and provides practical optimization guidance, paving the way for economical Serverless services.