Ecto Thinking

Mental shifts for Ecto and data layer design. These insights challenge typical ORM patterns.

Context = Setting That Changes Meaning

Context isn't just a namespace—it changes what words mean. "Product" means different things in Checkout (SKU, name), Billing (SKU, cost), and Fulfillment (SKU, warehouse). Each bounded context may have its OWN Product schema/table.

Think top-down: Subdomain → Context → Entity. Not "What context does Product belong to?" but "What is a Product in this business domain?"

Cross-Context References: IDs, Not Associations

schema "cart_items" do
  field :product_id, :integer  # Reference by ID
  # NOT: belongs_to :product, Catalog.Product
end

Query through the context, not across associations. Keeps contexts independent and testable.

DDD Patterns as Pipelines

def create_product(params) do
  params
  |> Products.build()       # Factory: unstructured → domain
  |> Products.validate()    # Aggregate: enforce invariants
  |> Products.insert()      # Repository: persist
end

Use events (as data structs) to compose bounded contexts with minimal coupling.

Schema ≠ Database Table

schema/2

embedded_schema/1

Multiple Changesets per Schema

def registration_changeset(user, attrs)  # Full validation + password
def profile_changeset(user, attrs)       # Name, bio only
def admin_changeset(user, attrs)         # Role, verified_at

Different operations = different changesets.

Multi-Tenancy: Composite Foreign Keys

add :post_id, references(:posts, with: [org_id: :org_id], match: :full)

Use

prepare_query/3

org_id

Preload vs Join Trade-offs

Join preloads can use 10x more memory for has-many.

CRUD Contexts Are Fine

"If you have a CRUD bounded context, go for it. No need to add complexity."

Use generators for simple cases. Add DDD patterns only when business logic demands it.

Gotchas from Core Team

CTE Queries Don't Inherit Schema Prefix

In multi-tenant apps, CTEs don't get the parent query's prefix.

Fix: Explicitly set prefix:

%{recursive_query | prefix: "tenant"}

Parameterized Queries ≠ Prepared Statements

• Parameterized queries:

  WHERE id = $1

  — always used by Ecto
• Prepared statements: Query plan cached by name — can be disabled

pgbouncer: Use

prepare: :unnamed

pool_count vs pool_size

More pools with fewer connections = better for benchmarks. But with mixed fast/slow queries, a single larger pool gives better latency.

Rule:

pool_count

pool_size

Sandbox Mode Doesn't Work With External Processes

Cachex, separate GenServers, or anything outside the test process won't share the sandbox transaction.

Fix: Make the external service use the test process, or accept it's not in the same transaction.

Null Bytes Crash Postgres

PostgreSQL rejects null bytes even though they're valid UTF-8.

Fix: Sanitize at boundaries:

String.replace(string, "\x00", "")

preload_order for Association Sorting

has_many :comments, Comment, preload_order: [desc: :inserted_at]

Note: Doesn't work for

through

Runtime Migrations Use List API

Ecto.Migrator.run(Repo, [{0, Migration1}, {1, Migration2}], :up, opts)

Idioms

• Prefer

  Repo.insert/1

  over

  Repo.insert!/1

  —handle

  {:ok, _}

  /

  {:error, _}

  explicitly
• Use

  Repo.transact/1

  (Ecto 3.12+) for simple transactions instead of

  Ecto.Multi

Red Flags - STOP and Reconsider

• belongs_to pointing to another context's schema
• Single changeset for all operations
• Preloading has-many with join
• CTEs in multi-tenant apps without explicit prefix
• Using pgbouncer without

  prepare: :unnamed

• Testing with Cachex/GenServers assuming sandbox shares transactions
• Accepting user input without null byte sanitization

Any of these? Re-read the Gotchas section.