Dynamodb

DynamoDB — access-pattern-first modeling and capacity

DynamoDB is not a relational database with a different syntax. It has no joins, no ad-hoc queries, no server-side aggregation, and it punishes schema changes after launch. So the modeling discipline IS the work: enumerate every read and write the app must perform, THEN design keys so each one is a single GetItem or Query. Add an index only when the base table physically cannot serve a pattern. Decide capacity last.

Your job here is to stop the user from treating DynamoDB like SQL.

What you produce / what you refuse

You deliver three artifacts, in this order:

1. An access-pattern list — a table of every query and write, with key condition, read/write, frequency, and what serves it. This is the contract the keys must satisfy.
2. A key-design document — entity → PK/SK encoding → which table or named GSI serves each pattern.
3. A capacity + cost recommendation — on-demand vs provisioned vs reserved, and the hot-partition risk.

You refuse to: click through the AWS console, wire the IaC/CI provisioning pipeline (that is ../deployment/SKILL.md), or model the data as normalized tables you then "join" in application code. You emit DDL, IaC snippets, and example SDK calls — you do not provision the infrastructure.

The steps below run in sequence — access patterns, then keys, then indexes, then capacity. Reordering them is the single most common DynamoDB mistake.

Step 1 — Enumerate access patterns

You cannot query what your keys do not express, and there is no WHERE over arbitrary columns, so produce this table before you name a single attribute. It is also exactly what scripts/verify.sh checks: run scripts/verify.sh <path-to-key-design.{md,json}> and every row must carry a key target (pk/sk or a named gsi) and a query_type. It FAILS on any pattern served by Scan, WARNS on FilterExpression, and asserts ≤20 GSIs / ≤5 LSIs. Read-only; exits 0 on an empty or clean target.

Bad → Good. The whole game is moving from vague to executable:

• Bad: "store users and orders." (No key condition is derivable from this.)
• Good: "get all orders for a user, newest first, page size 25." (PK + SK range + sort direction fall out.)

Rule: if a pattern needs a FilterExpression over more than ~10% of the items it scans, that is a modeling smell — redesign the key, do not patch it with a filter.

Step 2 — Key design + single-table decision

Get this wrong and you pay for it forever: changing a table's key schema after launch means a full migration (new table + backfill + dual-write).

Composite-key encoding. Use prefixed, sortable strings so one partition holds an entity and its children — items read together are co-located, so one Query returns them — and so a sort-key range query slices them:

PK = USER#123
SK = USER#123                      <- the user item itself
SK = ORDER#2026-06-02T10:15#A91    <- an order, ISO-8601 timestamp sorts lexically = chronologically
SK = ORDER#2026-06-02T11:40#B07

{ "PK": "USER#123", "SK": "ORDER#2026-06-02T10:15#A91",
  "Type": "Order", "Total": 4200, "Status": "OPEN" }

One Query with PK = USER#123 AND SK begins_with("ORDER#") returns all of that user's orders, already sorted. That is a one-to-many relationship with no join.

Many-to-many uses an adjacency list: store the edge twice (or use an inverted GSI, see Step 3) so you can traverse it from either side. Worked end-to-end in references/access-patterns.md — a multi-tenant SaaS / e-commerce model (pattern table → PK/SK map → GSI map → AWS SDK v3 @aws-sdk/lib-dynamodb calls), adjacency-list many-to-many, time-series and leaderboard patterns.

Single-table vs table-per-entity — decide honestly, do not cargo-cult:

Single-table design earns its complexity when many related entities share access patterns. When entities are independent and the queries are few, separate tables are simpler and perfectly correct — say so.

Step 3 — GSIs (only when the base table can't serve it)

Three named patterns, each with a one-line why:

• Inverted index — swap PK and SK on the GSI so you can read the relationship from the other side (GSI1PK = SK_value, GSI1SK = PK_value). Serves "given an order, who owns it" / many-to-many reverse.
• Sparse index — only items that carry the GSI's key attribute are indexed. Write the attribute only on the few items you want to find ("OPEN" orders, flagged accounts) so the index is tiny and the query is "find the few" instead of "scan the many."
• Overloaded GSI — reuse generic GSI1PK/GSI1SK attributes across multiple entity types so one index serves several patterns. Keeps you under the index quota.

Rule: add a GSI only when the base table physically cannot answer the pattern. Every GSI is a standing cost.

Warnings that bite people:

• GSIs are always eventually consistent — you cannot read-your-own-write through a GSI. If a pattern needs strong consistency, it must hit the base table.
• Each write to the base table replicates to every GSI whose keys changed, each consuming its own write capacity. More indexes = multiplied write cost.
• Project only the attributes each pattern needs. Blanket ProjectionType: ALL doubles storage and write cost; use KEYS_ONLY or INCLUDE when you can.

Quotas (default, adjustable for GSIs): up to 20 GSIs and 5 LSIs per table. LSIs must be defined at table creation, share the base partition key, and impose a 10 GB item-collection cap (all items under one partition key). GSI-only tables have no such cap — a real reason to prefer GSIs over LSIs.