Stretchy Search: Benchmark Methodology

We are actively benchmarking Stretchy's hybrid keyword (BM25 full-text) and semantic vector search across Cloudflare's network. Results will be published with raw data.

Evaluation Methodology

Our benchmark framework measures search performance across varying index sizes and concurrency tiers. Testing is designed to provide reproducible metrics across three key dimensions:
• Query Latency: Measuring end-to-end client round-trip times and internal edge execution duration under both warm and cold cache states.
• Concurrency & Throughput: Simulating concurrent user queries to evaluate throughput stability without queue degradation.
• Resource Efficiency: Monitoring active heap and memory allocations during sustained ingestion and query workloads.

Workloads Under Test

The benchmark test suite evaluates distinct real-world query patterns:
• Exact & Multi-Term Keyword Search: BM25 relevance scoring across standard full-text documents.
• Filtered Multi-Clause Queries: Boolean filtering combining keyword criteria with numeric ranges and categorical tags.
• Semantic Vector Matching: Vector similarity search across document embeddings.
• Real-Time Anomaly Scoring: Single-pass statistical variance calculations executed during event ingestion.

Test Environment & Hardware Baseline

Testing environments encompass both distributed serverless edge execution on Cloudflare's global network and standalone containerized instances on standard Linux virtual machines. Results will be published with raw data, specifying exact runtime versions, core allocations, and memory constraints.

Reproducibility & Raw Data

All benchmark runners and test scripts are developed openly. You can review the test harness in the stretchy-client GitHub repository. Results will be published with raw data as validation runs complete.