Back to Writing

September 26, 2026

Day 5 — One Table, Two Entities, One Mistake

Four days of Bun, Zod, tRPC and Redis kept todos in memory. Today's about making them survive a restart — but DynamoDB punishes relational habits more than it rewards them. The standard advice is one table, multiple entity types, told apart by a generic PK/SK pair. My first instinct proved why that advice exists.

What I built first (wrong)

A users table. Simple userId key. A todos attribute to hold — what, exactly? One partition key value can only ever address one item, so there was no real way to store a user's profile and several distinct todos under it. It also would have failed to even create: I'd declared a userName attribute DynamoDB never actually used in any key, which throws ValidationException: Some AttributeDefinitions are not used. AttributeDefinitions isn't a schema for your item shape — it's strictly the attributes some key schema references, nothing else.

What actually works

One table. Composite key. Entity type lives in the key values, not the table shape:

User: PK = USER#u1,  SK = USER#u1
Todo: PK = USER#u1,  SK = TODO#t1   (+ GSI1PK = STATUS#pending, GSI1SK = TODO#t1)

A Query on PK = USER#u1 returns the user's profile and every one of their todos in one round trip — the actual point of single-table design: things you fetch together, live together.

Proving the two patterns that matter

listTodosForUser("u1")   → returns only u1's 2 todos, filtered by
                            SK begins_with("TODO#") inside one partition

listTodosByStatus("pending") → returns todos from BOTH u1 AND u2 —
                            a Query on GSI1, not the base table

That second one is the point of adding a GSI at all: the base table's PK is always scoped to one user, so "every pending todo across every user" has no path through it except a full Scan — excluded by design today. GSI1PK = STATUS#<status> gives that access pattern its own index instead.

The lesson underneath both mistakes

Write the access patterns down before the schema, not after. I did it backwards — schema first, patterns as an afterthought — and got exactly the one-table-per-entity design the exercise is built to catch. The fix wasn't cleverer code, it was going back to "what do I actually need to ask this table" and letting the keys follow from that.

What's next

Day 6: full code in the repo — day-05-dynamodb-single-table/.