Architecture
Benchmarks
Architecture
Benchmarks
Mirror Vectax Benchmarks
Vector Search Benchmarks with Encryption
| Engine | P95 (ms) | P99 (ms) | RPS | Base → Enc Prec | P99+Enc (ms) |
|---|---|---|---|---|---|
| Qdrant | 4.95 | 8.62 | 1238 | 0.99 → 0.95 | 8.62 |
| Weaviate | 7.16 | 11.33 | 1142 | 0.97 → 0.93 | 11.33 |
| Elasticsearch | 72.53 | 135.68 | 716 | 0.98 → 0.94 | 135.68 |
| Redis | 160.85 | 167.35 | 625 | 0.97 → 0.93 | 167.35 |
| Milvus | 441.32 | 576.65 | 219 | 0.99 → 0.95 | 576.65 |
Key Findings
Encryption Overhead
- Vector encryption: 0.0018432 ms per operation (1536D)
- Total dataset encryption: 1.8432s (1M vectors)
- Minimal impact: 0.037% overhead on fastest engine
Precision Analysis
- Consistent 0.04 precision drop across engines
- Ranking tiers maintained:
- Tier 1 (0.95): Qdrant, Milvus
- Tier 2 (0.94): Elasticsearch
- Tier 3 (0.93): Weaviate, Redis
Technical Notes
- Encryption cost: 0.0000012 ms per dimension
- P99 latencies include worst-case encryption overhead
- Higher dimensions increase encryption time linearly
- A detailed benchmark will be open sourced in Q1.
- Baseline metrics: Qdrant Benchmarks
- Mirror SDK: All telemetry disabled
- Dataset encryption is one-time indexing cost
Methodology
- Base metrics sourced from standard vector search benchmarks
- Encryption overhead measured independently
- Combined metrics represent worst-case scenarios
- All tests run on dbpedia-openai-1M-1536-angular dataset
- Precision measured against non-encrypted ground truth
- RPS calculated under sustained load conditions
Was this page helpful?