콘텐츠로 이동

FAQ

Frequently Asked Questions


Data Management

Q1. Are external model (GPT-4, Claude) captions also stored in Milvus?

A1. Only stored in Feature Store.

Milvus has schema stability issues that make it difficult to manage multiple model results in an integrated manner. We set Feature Store as the Single Source of Truth (SSOT) for consistent data management.

Role Division:

  • Feature Store: Caption data version management and research/analysis repository
  • Milvus: Vector search system for VSS services (RAG, Video QA)

Q2. How is data consistency guaranteed when synchronization fails?

A2. Create a new version with a new timestamp and automatically retry.

On failure, the system doesn't overwrite existing data. Instead, it creates a new version with a new timestamp and automatically retries in the next cycle. Even if data is stored in Feature Store, it won't be searchable if not registered in Metadata Store.

Core Principles:

  • Immutable: Created data cannot be modified
  • Append-only: No overwriting, all history preserved
  • Automatic Retry: Save checkpoint and auto-recover in next cycle

Performance

Q1. Doesn't concurrent Milvus access affect existing system performance?

A1. Designed to minimize load, but requires validation through experiments.

Operation Method:

  • Initial (once): Full data sync during early morning hours
  • Every 10 minutes: Query only newly added data (e.g., 30 records)

Validation Required:

  • Performance impact when querying new data during VSS write operations
  • Latency changes during concurrent access
  • Apply rate limiting if performance degradation occurs

Architecture

Q1. Why connect to VSS instead of the serving system (AiD)?

A1. Currently focused on video understanding purposes. AiD/AiM integration is next step.

The current system is built for video understanding purposes, so we only consider integration with VSS.


Use Cases

Q1. If only evaluation is needed, isn't direct Milvus query simpler?

A1. Feature Store solves Milvus's structural limitations and supports various use cases.

Issue Direct Milvus Query Feature Store
Schema Stability Breaks on VSS update ✅ Stable
Multi-model Management Requires schema change ✅ Flexible addition
Version Management Time point unclear ✅ Point-in-Time

Solves Milvus's schema stability, scalability constraints, and lack of time-based version management. Supports model comparison, experiment reproduction, config analysis, etc., beyond simple evaluation.