GoScaler — Elastic Optimization for Serverless Cold Starts

English | 中文

Runner-up solution (2nd of 6506 teams) of the 2023 Cloud Native Programming Challenge, Track 1: elastic optimization for cold starts in Serverless scenarios.

Problem

A FaaS platform must serve bursty request traffic while minimizing both cold-start latency and idle resource cost. The scaler decides when to pre-warm containers, how many to keep, and when to reclaim them.

Approach

Built on the official contest scaffold (AliyunContainerService/scaler), the core scheduling logic is custom:

• k-means workload clustering (pkg/kmeans) — clusters historical request-rate series to recognize recurring load patterns and predict demand windows, trading warm-container inventory against cold-start risk
• Resource lifecycle management (pkg/scaler) — fine-grained warm container scheduling and reclamation (app.go / resource.go / gc.go)
• gRPC serving layer (pkg/server, cmd/scaler) implementing the contest's platform protocol, with simulator configs (sim*.yaml) for offline evaluation

Quick start

make build          # or: go build ./cmd/scaler
./run.sh            # launch against the simulator / scoring platform

Docker-based scoring is available via the provided Dockerfile.

Resources

• Detailed write-up (Chinese)