Backup and Data Integrity: Preserving the Web

A vector database backup is a "Pile of Numbers." A Knowledge Graph backup is a "Complex Tapestry." If you lose a single node, you don't just lose one fact; you might break dozens of "Paths" that your AI agent relies on for reasoning. In Graph RAG, Data Integrity is not just about "Existing"; it's about "Connecting."

In this lesson, we will explore the strategies for backing up your graph store. We will learn the difference between Hot vs. Cold Backups, how to perform Integrity Checks (to find "Orphaned Nodes"), and why you should always version your graph store alongside your ingestion code.

---

1. Hot vs. Cold Backups

Cold Backups (Offline):

You shut down the database and copy the physical data files.

• Pros: 100% consistent.
• Cons: Your Graph RAG bot is offline during the process.

Hot Backups (Online):

The database continues to run while a background process streams the "Transaction Logs" to a storage bucket (S3/GCS).

• Pros: Zero downtime.
• Cons: More complex to manage; requires specialized tools (e.g., neo4j-admin backup).

---

2. The "Orphaned Node" Problem (Integrity)

An orphaned node is a node that has no relationships.

• In some cases, this is fine (a new entity).
• In Graph RAG, it usually means a Broken Ingestion Pipeline.

If your graph has 1 Million nodes but only 10,000 edges, your "Integrity" is compromised. Your agent will land on nodes but find no paths to follow. You should run a Post-Ingestion Audit to find nodes with a "Degree of 0."

---

3. Transactional Safety: ACID Compliance

Most Graph Databases (Neo4j, Neptune) are ACID compliant.

• A (Atomicity): Either the entire triplet is created, or none of it is.
• Wait: This is critical. If you create the Sudeep node but the WORKS_AT relationship fails, your graph is now "Lying" by omission.

Always use Transactions in your Python code to ensure that the Node and its primary Edges are committed as a single, unbreakable unit.

graph TD
    subgraph "Backup Strategy"
    DB[(Live Graph)] -->|Online Stream| Logs[Transaction Logs]
    Logs -->|S3 Upload| S3[(Cloud Storage)]
    DB -->|Verify| IC[Integrity Checker]
    end
    
    IC -->|Report| A[Admin Alert]
    
    style DB fill:#4285F4,color:#fff
    style S3 fill:#34A853,color:#fff

---

4. Implementation: A Daily Integrity Check with Cypher

Let's write a query that identifies "Dangling Relationships"—edges that are missing a target or nodes that are completely isolated.

// 1. Find the % of isolated nodes
MATCH (n)
WHERE NOT (n)--()
RETURN count(n) as isolated_count;

// 2. Validate against a threshold
// If more than 5% of your graph is isolated, 
// your Extraction LLM might be failing to find relationships.

---

5. Summary and Exercises

Backup is the "Safety Net" for your AI's memory.

• Hot Backups are essential for production uptime.
• Integrity Checks ensure the "Graph" is actually a network, not just a list.
• ACID Transactions prevent partial, broken knowledge from entering the system.
• Versioning helps you "Roll back" if an LLM update starts generating poor-quality triplets.

Exercises

1. Integrity Audit: Why is a node named [Person] with 0 edges "Dangerous" for Graph RAG? What will the agent do when it finds that node?
2. Backup Frequency: If your Knowledge Graph updates from Slack in real-time, how often should you perform a backup? Hourly? Daily?
3. The "Undo" Button: You just ingested 1,000 documents but realized the LLM got the "Date" format wrong. How do you "Roll Back" without losing the data you ingested 5 hours ago? (Hint: Use Transaction Logs or Snapshots).

In the next lesson, we will look at how to survive a total regional failure: High Availability and Disaster Recovery.