nosql-expert – Design & UI skill for AI agents

NoSQL Expert Patterns (Cassandra & DynamoDB)

Overview

This skill provides professional mental models and design patterns for distributed wide-column and key-value stores (specifically Apache Cassandra and Amazon DynamoDB).

Unlike SQL (where you model data entities), or document stores (like MongoDB), these distributed systems require you to model your queries first.

When to Use

• Designing for Scale: Moving beyond simple single-node databases to distributed clusters.
• Technology Selection: Evaluating or using Cassandra, ScyllaDB, or DynamoDB.
• Performance Tuning: Troubleshooting "hot partitions" or high latency in existing NoSQL systems.
• Microservices: Implementing "database-per-service" patterns where highly optimized reads are required.

The Mental Shift: SQL vs. Distributed NoSQL

| Feature | SQL (Relational) | Distributed NoSQL (Cassandra/DynamoDB) |

| :--- | :--- | :--- |

| Data modeling | Model Entities + Relationships | Model Queries (Access Patterns) |

| Joins | CPU-intensive, at read time | Pre-computed (Denormalized) at write time |

| Storage cost | Expensive (minimize duplication) | Cheap (duplicate data for read speed) |

| Consistency | ACID (Strong) | BASE (Eventual) / Tunable |

| Scalability | Vertical (Bigger machine) | Horizontal (More nodes/shards) |

> The Golden Rule: In SQL, you design the data model to answer *any* query. In NoSQL, you design the data model to answer *specific* queries efficiently.

Core Design Patterns

1. Query-First Modeling (Access Patterns)

You typically cannot "add a query later" without migration or creating a new table/index.

Process:

1. List all Entities (User, Order, Product).

2. List all Access Patterns ("Get User by Email", "Get Orders by User sorted by Date").

3. Design Table(s) specifically to serve those patterns with a single lookup.

2. The Partition Key is King

Data is distributed across physical nodes based on the Partition Key (PK).

• Goal: Even distribution of data and traffic.
• Anti-Pattern: Using a low-cardinality PK (e.g., status="active" or gender="m") creates Hot Partitions, limiting throughput to a single node's capacity.
• Best Practice: Use high-cardinality keys (User IDs, Device IDs, Composite Keys).

3. Clustering / Sort Keys

Within a partition, data is sorted on disk by the Clustering Key (Cassandra) or Sort Key (DynamoDB).

• This allows for efficient Range Queries (e.g., WHERE user_id=X AND date > Y).
• It effectively pre-sorts your data for specific retrieval requirements.

4. Single-Table Design (Adjacency Lists)

*Primary use: DynamoDB (but concepts apply elsewhere)*

Storing multiple entity types in one table to enable pre-joined reads.

| PK (Partition) | SK (Sort) | Data Fields... |

| :--- | :--- | :--- |

| USER#123 | PROFILE | { name: "Ian", email: "..." } |

| USER#123 | ORDER#998 | { total: 50.00, status: "shipped" } |

| USER#123 | ORDER#999 | { total: 12.00, status: "pending" } |

• Query: PK="USER#123"
• Result: Fetches profile and orders under that partition key; paginate until LastEvaluatedKey is absent rather than assuming all fit one response.

5. Denormalization & Duplication

Don't be afraid to store the same data in multiple tables to serve different query patterns.

• Table A: users_by_id (PK: uuid)
• Table B: users_by_email (PK: email)

*Trade-off: You must manage data consistency across tables (often using eventual consistency or batch writes).*

Specific Guidance

Apache Cassandra / ScyllaDB

• Primary Key Structure: ((Partition Key), Clustering Columns)
• No joins; constrained aggregation: CQL supports aggregates and GROUP BY over allowed primary-key prefixes. Check bounded scan cost; precompute only where access patterns justify the consistency/repair tradeoff.
• Avoid ALLOW FILTERING: If you see this in production, your data model is wrong. It implies a full cluster scan.
• Writes are Cheap: Inserts and Updates are just appends to the LSM tree. Don't worry about write volume as much as read efficiency.
• Tombstones: Deletes are expensive markers. Avoid high-velocity delete patterns (like queues) in standard tables.

AWS DynamoDB

• GSI (Global Secondary Index): Use GSIs to create alternative views of your data (e.g., "Search Orders by Date" instead of by User).
• *Note:* GSIs are eventually consistent.
• LSI (Local Secondary Index): Sorts data differently *within* the same partition. Must be created at table creation time.
• WCU / RCU: Understand capacity modes. Single-table design helps optimize consumed capacity units.
• TTL: Expiration is asynchronous; exclude expired items in application reads where required and verify replication charges. Do not use TTL as an exact deletion-time guarantee or apply Cassandra tombstone semantics to DynamoDB.

Expert Checklist

Before finalizing your NoSQL schema:

• [ ] Access Pattern Coverage: Does every query pattern map to a specific table or index?
• [ ] Cardinality Check: Does the Partition Key have enough unique values to spread traffic evenly?
• [ ] Split Partition Risk: For any single partition (e.g., a single user's orders), will it grow indefinitely? (DynamoDB’s 10 GB item-collection cap applies when LSIs exist; other engines and configurations have different constraints. Bucket by time or shard only with a query/capacity plan.)
• [ ] Consistency Requirement: Can the application tolerate eventual consistency for this read pattern?

Common Anti-Patterns

❌ Scatter-Gather: Querying *all* partitions to find one item (Scan).

❌ Hot Keys: Putting all "Monday" data into one partition.

❌ Relational Modeling: Creating Author and Book tables and trying to join them in code. (Instead, embed Book summaries in Author, or duplicate Author info in Books).

Example

User request:

> Use @nosql-expert for this task: Expert guidance for distributed NoSQL databases (Cassandra, DynamoDB).

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.