Consistency Levels

Overview

Consistency levels in vector databases determine how fresh the data must be when reading. This is crucial in distributed systems with replication where writes may not instantly propagate to all replicas.

Common Levels

Strong

• Reads always return latest writes
• Highest latency
• Best for critical data

Bounded Staleness

• Data may be slightly old (e.g., < 5 seconds)
• Better performance
• Still reasonably fresh

Session

• Consistent within a session
• Different sessions may see different data
• Good balance

Eventual

• Lowest latency
• Data will eventually be consistent
• Best for high-throughput reads

Use Cases

Strong: Financial transactions, critical updates Bounded: Real-time dashboards Session: User-specific applications Eventual: Analytics, batch processing

Example (Milvus)

results = collection.search(
    data=query_vectors,
    anns_field="embedding",
    param=search_params,
    limit=10,
    consistency_level="Eventually"  # Trade consistency for speed
)

Trade-offs

• Strong: Slowest but most accurate
• Eventual: Fastest but may miss recent writes

Pricing

Not applicable (configuration option).