# Query Engine Design Questions

A checklist note for thinking about what kind of query engine to build.

---

## Short answer

Before building a query engine, it helps to force explicit answers to a small set of design questions.

Most architecture disagreements are really disagreements about workload, execution granularity, storage boundary, or correctness model.

---

## Core questions

### Workload

- Is the workload mostly transactional, analytical, streaming, search, or logical inference?
- Are queries mostly point lookups, scans, aggregations, recursive rules, or top-k retrieval?

### Data model

- Is the data relational, document-oriented, graph-shaped, vector-based, or rule/fact based?
- What is the engine's core internal representation?

### Unit of execution

- Does the engine run row-at-a-time, batch-at-a-time, or over fully materialized relations?
- Are blocking operators common?

### Storage boundary

- Is execution tightly coupled to one storage engine?
- Or is there a source interface with pushdown capabilities?

### Indexes

- What access patterns deserve dedicated indexes?
- Are indexes exact, approximate, ordered, inverted, or vector-oriented?

### Optimization

- Is the optimizer mostly rule-based or cost-based?
- What statistics are available?

### Distribution

- Is one machine enough?
- If not, where are exchange boundaries, partitioning choices, and failure handling defined?

### Semantics

- Is the system exact, approximate, eventually consistent, ranked, or fixpoint-based?
- Does it support recursion, inference, or witness generation?

---

## Why these questions matter

These questions determine most of the major architecture choices:

- row vs column
- iterator vs vectorized execution
- local vs distributed execution
- exact vs approximate search
- relational vs rule-based planning

If those answers are unclear, architecture discussions tend to stay vague.

---

## Practical use

This note is best used as a design checklist.

If a team can answer these questions cleanly, the likely engine shape becomes much easier to see.

If it cannot, the project is probably still mixing together several different engine ideas.

---

## Changelog

* **April 2, 2026** -- First version created.