Derivation join collections

derivation-join-collections

Estuary derivation that merges documents from multiple source collections into a single collection, keyed on a shared join field. Fields from each source contribute, and Estuary's reduction engine merges them.

Prereq: read derivation-basics first for concepts, the stateless-vs-stateful distinction, and reduction-annotation mechanics.

Docs:

• https://docs.estuary.dev/guides/howto_join_two_collections_typescript/ — official TypeScript walkthrough
• https://docs.estuary.dev/reference/reduction-strategies/ — reduction strategy reference (critical for joins)

Canonical examples in the Flow repo (cover the three join shapes — see "Join types" below):

• join-outer.flow.yaml — full outer (reduction-based, no state)
• join-inner.flow.yaml — inner (SQLite state, both sides must match)
• join-one-sided.flow.yaml — left / right (SQLite state, one side drives emission)

When to use this over alternatives

• Enrichment: attach customer profile fields onto every order
• Denormalisation: combine dimension + fact tables into a wide analytics table
• Customer lifetime metrics: one doc per customer with counts / totals rolled up from orders
• Pre-joined views for BI tools that prefer wide tables over joins at query time

Reach for other skills when:

• Only one source collection, aggregating by key → derivation-aggregate-metrics
• Need an embedded array of matches, not merged fields → combine with append reduction (still this skill, but see Variations)
• You want arbitrary stateful logic beyond reductions (approval flows, balance checks) → derivation-stateful-logic

Join types

Stream joins aren't quite the same as SQL joins. Pick the type before writing the derivation — the pattern differs by type:

Join type	Approach	When to use
Full outer	Reduction-based (this skill's canonical example below). Both sides emit partial documents independently; reductions merge them. Documents appear in the output as soon as either side has data for a key.	Default for enrichment / denormalisation when you want every key joined regardless of match.
Inner	Stateful (join-inner.flow.yaml). Both sides write to a SQLite state table; emission is gated on lhs IS NOT NULL AND rhs IS NOT NULL so unmatched keys never appear.	When you only want output for keys present on both sides.
Left / right (one-sided)	Stateful (join-one-sided.flow.yaml). Both sides update state, but only one side triggers emission. RHS-side updates accumulate silently; LHS-side events emit the joined record.	When a specific side is the natural "driver" (emit per-customer using the latest accumulated orders, but not vice versa).

Inner and one-sided joins are stateful — they have a state table and INSERT ... ON CONFLICT DO UPDATE style lambdas, and conceptually belong to derivation-stateful-logic. They're called out here because they share the "joining two streams" problem framing.

How it works (outer-join reduction pattern — one paragraph)

For the full-outer reduction pattern, you define one transform per source collection. Each transform shuffles on the join key and emits partial documents keyed by that same key. The output collection's schema has reduce: annotations on every field — at materialisation time, Estuary combines all partial documents with the same key into a single merged row. The "join" isn't a SQL join; it's reduction over co-keyed documents.

Canonical Example — SQL (outer join via reductions)

Customers table and orders table merge into a wide customer-lifetime collection with customer contact info + aggregate order metrics.

Project layout

my-derivation/
├── flow.yaml
└── schema.yaml

schema.yaml

type: object
properties:
  customer_id:
    type: string

  # Fields from customers source — keep the latest non-empty value
  customer_name:
    type: string
    default: ""
    reduce: { strategy: maximize }
  customer_email:
    type: string
    default: ""
    reduce: { strategy: maximize }
  tier:
    type: string
    default: ""
    reduce: { strategy: maximize }

  # Fields aggregated from orders source
  lifetime_order_count:
    type: integer
    default: 0
    reduce: { strategy: sum }
  lifetime_revenue:
    type: number
    default: 0
    reduce: { strategy: sum }
  first_order_at:
    type: string
    format: date-time
    reduce: { strategy: minimize }
  last_order_at:
    type: string
    format: date-time
    reduce: { strategy: maximize }

reduce: { strategy: merge }
required: [customer_id]

Top-level reduce: merge is mandatory. maximize on string fields keeps the lexicographically greatest value — that's usually fine for dimension fields where the source updates infrequently, but if you actually need "most recently written," use lastWriteWins instead. sum on numeric fields accumulates deltas from each order.

flow.yaml

Two transforms, one per source, both shuffled on customer_id:

collections:
  acmeCo/analytics/customer-lifetime:
    schema: schema.yaml
    key: [/customer_id]
    derive:
      using:
        sqlite: {}
      transforms:
        - name: fromCustomers
          source: acmeCo/production/customers
          shuffle: { key: [/customer_id] }
          lambda: |
            SELECT
              $customer_id  AS customer_id,
              $name         AS customer_name,
              $email        AS customer_email,
              $tier         AS tier;

        - name: fromOrders
          source: acmeCo/production/orders
          shuffle: { key: [/customer_id] }
          lambda: |
            SELECT
              $customer_id  AS customer_id,
              1             AS lifetime_order_count,
              $total        AS lifetime_revenue,
              $created_at   AS first_order_at,
              $created_at   AS last_order_at;

Each transform emits only the fields it contributes; the reduction merges them field-by-field.

Canonical Example — TypeScript

TypeScript shines when you want to collect orders into an array rather than summing them. The TS variant uses a different output collection (customer-with-orders) since its schema replaces the numeric-sum fields with an orders array:

import { IDerivation, Document, SourceFromCustomers, SourceFromOrders } from 'flow/acmeCo/analytics/customer-with-orders.ts';

export class Derivation extends IDerivation {
  fromCustomers(_read: { doc: SourceFromCustomers }): Document[] {
    return [{
      customer_id: _read.doc.customer_id,
      customer_name: _read.doc.name,
    }];
  }

  fromOrders(_read: { doc: SourceFromOrders }): Document[] {
    return [{
      customer_id: _read.doc.customer_id,
      orders: [{
        order_id: _read.doc.order_id,
        total: _read.doc.total,
        created_at: _read.doc.created_at,
      }],
    }];
  }
}

Schema difference: add an orders: { type: array, reduce: { strategy: append } } field. Each fromOrders call emits a single-element array that append appends to the array already accumulated for that customer_id.

Test

Local validation — flowctl preview --fixture

Like aggregate-metrics, preview may show partially-reduced output. With one customer doc + two order docs all for c1, you'll see multiple deltas that the materialisation will merge into the final single row.

fixture.jsonl:

["acmeCo/production/customers", {"customer_id": "c1", "name": "Alice", "email": "[email protected]", "tier": "gold"}]
["acmeCo/production/orders", {"order_id": "o1", "customer_id": "c1", "total": 100.00, "created_at": "2024-01-15T10:00:00Z"}]
["acmeCo/production/orders", {"order_id": "o2", "customer_id": "c1", "total": 250.00, "created_at": "2024-03-20T14:30:00Z"}]
["acmeCo/production/customers", {"customer_id": "c2", "name": "Bob", "email": "[email protected]", "tier": "silver"}]
{"commit": true}

flowctl preview --source flow.yaml \
  --name acmeCo/analytics/customer-lifetime \
  --fixture fixture.jsonl --timeout 30s

Expected mental reduction at materialisation:

• c1: name=Alice, email=[email protected], tier=gold, lifetime_order_count=2, lifetime_revenue=350.00, first_order_at=2024-01-15T10:00:00Z, last_order_at=2024-03-20T14:30:00Z
• c2: name=Bob, email=[email protected], tier=silver (no orders → numeric defaults, no timestamps)

Spec-defined tests — tests: block (aspirational)

tests:
  acmeCo/tests/customer-lifetime:
    - ingest:
        collection: acmeCo/production/customers
        documents:
          - { customer_id: "c1", name: "Alice", email: "[email protected]", tier: "gold" }
          - { customer_id: "c2", name: "Bob",   email: "[email protected]",   tier: "silver" }
    - ingest:
        collection: acmeCo/production/orders
        documents:
          - { order_id: "o1", customer_id: "c1", total: 100.00, created_at: "2024-01-15T10:00:00Z" }
          - { order_id: "o2", customer_id: "c1", total: 250.00, created_at: "2024-03-20T14:30:00Z" }
    - verify:
        collection: acmeCo/analytics/customer-lifetime
        documents:
          - { customer_id: "c1", customer_name: "Alice", customer_email: "[email protected]", tier: "gold",   lifetime_order_count: 2, lifetime_revenue: 350.00, first_order_at: "2024-01-15T10:00:00Z", last_order_at: "2024-03-20T14:30:00Z" }
          - { customer_id: "c2", customer_name: "Bob",   customer_email: "[email protected]",   tier: "silver" }

Variations

• Many-to-many via bridge table: three transforms — e.g. products + product_categories + categories, each shuffled on the appropriate key. Each transform emits the fields it contributes; reduction merges.
• Multi-level join (orders → line items → products): chain derivations: first join orders with line items to produce a wide line-item table, then join that with products.
• Embedded array of child records: array field with reduce: append on the parent side; each child-source transform emits a single-element array.
• Dimension with slowly-changing-value (SCD): use lastWriteWins on dimension fields so each dimension update replaces prior values. For full history, track in a separate collection.
• One-sided filter: a transform can filter its source (WHERE $status = 'active') and only contribute to the joined output for matching rows.

Gotchas specific to join-collections

• This reduction-based join pattern requires standard-updates mode on the destination. The pattern works by each transform emitting a partial document and the materialization loading the existing row, reducing the incoming partial into it, and writing back the merged result. That only happens in standard-updates mode (the default on every connector). If the destination must use delta_updates: true (opt-in on Snowflake / BigQuery / Databricks for cost or latency reasons), this pattern won't work — your destination will show fragmented per-source rows instead of joined records. For delta-updates destinations, use a stateful join via derivation-stateful-logic: maintain the join table inside SQLite (or Python) and emit already-merged documents so no downstream reduction is needed. See Delta updates.
• Shuffle keys must match across transforms. Both (or all) transforms must shuffle: { key: [/customer_id] } so docs with the same join key land on the same task shard. If one transform shuffles on /id and another on /customer_id, the join never happens.
• Shuffle key can be any JSON pointer into the source — it doesn't have to be related to the source's collection key. The constraint is cross-transform: shuffle key types and arity must align so the same logical key lands on the same shard from every source. shuffle: any is wrong for joins (it breaks co-location), but no "prefix of collection key" rule exists.
• String field reduction needs care. reduce: maximize on a string keeps the lexicographically greatest value seen — not the most-recently-written. For dimension fields (name, email), lex-max is usually fine; if you actually need latest-received, use lastWriteWins.
• Orphan keys won't have aggregated fields in the derivation collection. If a customer has no orders, fields like lifetime_order_count are simply absent from the derived doc (defaults are ignored inside derivations). Set default: 0 so the materialization writes 0 instead of NULL — defaults take effect at materialization time, not in the derivation collection itself.
• --uncommitted and preview show partial docs per source. Each transform emits its own partial doc; reduction merges at materialisation. You'll see a customer doc, then separate order docs, not a single merged doc, in preview.
• Timestamp conflicts across sources. If two sources have updated_at and you want the newest, explicitly maximize — don't rely on ordering.

Related

• derivation-basics — prerequisite
• derivation-aggregate-metrics — when aggregating within a single source, same reduction machinery
• derivation-stateful-logic — when reductions aren't enough and you need custom SQLite state
• Join Two Collections Using TypeScript (docs)
• Reduction strategies reference