derivation-flatten-array

Stateless Estuary derivation that emits many output documents from a single source document by iterating over a nested array.

Prereq: read derivation-basics first for concepts, project layout, workflow, and language choice.

Docs: https://docs.estuary.dev/guides/flatten-array/ — the official TypeScript walkthrough for this exact pattern. Read it alongside this skill.

When to use this over alternatives

• Normalising nested JSON: {order_id, line_items: [...]} → one doc per line item
• Exploding tags/categories: {article_id, tags: ["a", "b", "c"]} → one doc per tag
• Unpacking array columns from source databases for analytics
• Preparing data for systems that prefer normalised 1-row-per-item tables

If the array isn't in the source document itself (e.g., you need to join across collections to assemble it), use derivation-join-collections. If you just need to keep the array intact but change how it's materialised (e.g., MongoDB flow_document), you don't need a derivation at all.

Canonical Example — TypeScript (recommended)

Orders arrive with an embedded array of line items. We emit one document per line item, carrying order context (customer, date) onto each row.

Project layout

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

schema.yaml

The output schema describes one line item — not the array:

type: object
properties:
  order_id:
    type: string
  line_id:
    type: string
  customer_id:
    type: string
  order_date:
    type: string
    format: date-time
  product_id:
    type: string
  product_name:
    type: string
  quantity:
    type: integer
  unit_price:
    type: number
  line_total:
    type: number
required: [order_id, line_id, product_id]

flow.yaml

collections:
  acmeCo/normalized/order-line-items:
    writeSchema: schema.yaml
    readSchema:
      allOf:
        - $ref: flow://write-schema
        - $ref: flow://inferred-schema
    key: [/order_id, /line_id]
    derive:
      using:
        typescript:
          module: flatten.ts
      transforms:
        - name: flattenLineItems
          source: acmeCo/production/orders
          shuffle: any

Key fields: the collection key is composite (order_id + line_id) — the array element id keeps output rows unique. shuffle: any because flattening is stateless. The writeSchema declares what the lambda emits; readSchema lets the inferred schema widen automatically as the lambda evolves. See derivation-basics § Schema setup.

flatten.ts

Generate the stub with flowctl generate --source flow.yaml, then fill in:

import { IDerivation, Document, SourceFlattenLineItems } from 'flow/acmeCo/normalized/order-line-items.ts';

export class Derivation extends IDerivation {
  flattenLineItems(_read: { doc: SourceFlattenLineItems }): Document[] {
    const order = _read.doc;

    // Missing or empty array → emit nothing
    if (!order.line_items || order.line_items.length === 0) {
      return [];
    }

    return order.line_items.map(item => ({
      order_id: order.order_id,
      line_id: item.line_id,
      customer_id: order.customer_id,
      order_date: order.order_date,
      product_id: item.product_id,
      product_name: item.product_name,
      quantity: item.quantity,
      unit_price: item.unit_price,
      line_total: item.quantity * item.unit_price,
    }));
  }
}

Return [] to drop the document entirely; return one document per array element to flatten.

Canonical Example — SQL

SQLite can flatten via json_each, which produces one row per array element. Works well for shallow arrays of primitives or simple objects; gets verbose for deeply nested structures (TypeScript is usually clearer then).

Tags array (simple case)

collections:
  acmeCo/normalized/article-tags:
    schema:
      type: object
      properties:
        article_id: { type: string }
        tag: { type: string }
      required: [article_id, tag]
    key: [/article_id, /tag]
    derive:
      using:
        sqlite: {}
      transforms:
        - name: flattenTags
          source: acmeCo/articles
          shuffle: any
          lambda: |
            SELECT
              $article_id AS article_id,
              json_each.value AS tag
            FROM json_each($tags);

Object array (more verbose)

lambda: |
  SELECT
    $order_id AS order_id,
    json_extract(json_each.value, '$.line_id')      AS line_id,
    json_extract(json_each.value, '$.product_id')   AS product_id,
    json_extract(json_each.value, '$.product_name') AS product_name,
    CAST(json_extract(json_each.value, '$.quantity') AS INTEGER)   AS quantity,
    json_extract(json_each.value, '$.unit_price')   AS unit_price
  FROM json_each($line_items);

For anything beyond simple fields, prefer the TypeScript form.

Test

Local validation — flowctl preview --fixture (works today)

fixture.jsonl:

["acmeCo/production/orders", {"order_id": "o1", "customer_id": "c1", "order_date": "2024-06-01T10:00:00Z", "line_items": [{"line_id": "li_1", "product_id": "p_a", "product_name": "Widget", "quantity": 2, "unit_price": 10.00}, {"line_id": "li_2", "product_id": "p_b", "product_name": "Gadget", "quantity": 1, "unit_price": 25.00}, {"line_id": "li_3", "product_id": "p_c", "product_name": "Gizmo", "quantity": 3, "unit_price": 5.50}]}]
["acmeCo/production/orders", {"order_id": "o2", "customer_id": "c2", "order_date": "2024-06-02T10:00:00Z", "line_items": []}]
{"commit": true}

flowctl preview --source flow.yaml \
  --name acmeCo/normalized/order-line-items \
  --fixture fixture.jsonl --timeout 30s

Expected output (3 docs from o1, nothing from o2's empty array):

{"order_id":"o1","line_id":"li_1","customer_id":"c1","order_date":"2024-06-01T10:00:00Z","product_id":"p_a","product_name":"Widget","quantity":2,"unit_price":10.00,"line_total":20.00}
{"order_id":"o1","line_id":"li_2","customer_id":"c1","order_date":"2024-06-01T10:00:00Z","product_id":"p_b","product_name":"Gadget","quantity":1,"unit_price":25.00,"line_total":25.00}
{"order_id":"o1","line_id":"li_3","customer_id":"c1","order_date":"2024-06-01T10:00:00Z","product_id":"p_c","product_name":"Gizmo","quantity":3,"unit_price":16.50}

Spec-defined tests — tests: block

tests:
  acmeCo/tests/flatten-line-items:
    - ingest:
        collection: acmeCo/production/orders
        documents:
          - order_id: "o1"
            customer_id: "c1"
            order_date: "2024-06-01T10:00:00Z"
            line_items:
              - { line_id: "li_1", product_id: "p_a", product_name: "Widget", quantity: 2, unit_price: 10.00 }
              - { line_id: "li_2", product_id: "p_b", product_name: "Gadget", quantity: 1, unit_price: 25.00 }
              - { line_id: "li_3", product_id: "p_c", product_name: "Gizmo",  quantity: 3, unit_price: 5.50 }
          - order_id: "o2"
            customer_id: "c2"
            order_date: "2024-06-02T10:00:00Z"
            line_items: []
    - verify:
        collection: acmeCo/normalized/order-line-items
        documents:
          - { order_id: "o1", line_id: "li_1", customer_id: "c1", order_date: "2024-06-01T10:00:00Z", product_id: "p_a", product_name: "Widget", quantity: 2, unit_price: 10.00, line_total: 20.00 }
          - { order_id: "o1", line_id: "li_2", customer_id: "c1", order_date: "2024-06-01T10:00:00Z", product_id: "p_b", product_name: "Gadget", quantity: 1, unit_price: 25.00, line_total: 25.00 }
          - { order_id: "o1", line_id: "li_3", customer_id: "c1", order_date: "2024-06-01T10:00:00Z", product_id: "p_c", product_name: "Gizmo",  quantity: 3, unit_price: 5.50,  line_total: 16.50 }

Variations

• Simple string array (tags, prefs): json_each($tags) in SQL, _read.doc.tags.map(tag => ({...})) in TS.
• JSON-string-encoded array: wrap JSON.parse in try/catch for TS; json_each(json($field)) in SQL where $field is a string.
• Deeply nested arrays (categories→products): nested for...of loops in TS. SQL gets painful fast — prefer TS.
• Conditional flattening: filter array elements before emitting (_read.doc.addresses.filter(a => a.is_primary).map(...)).
• Include array index as a field: TS has .map((item, i) => ...); SQL uses json_each.key as the index.
• Partial re-emit on upstream update: if the source document's array shrinks on update, stale output documents remain. See the gotcha below.

Gotchas specific to flatten-array

• Key uniqueness is on you. The derived collection key must uniquely identify a flattened row. If the array elements have stable IDs (line_id), use them. If they don't, include the array index (item_index) in the key — otherwise multiple elements collide and reduce into one.
• Source updates don't remove stale flattened rows. If an upstream order initially had 3 line items and the array later shrinks to 2, the derivation re-runs on the source update and emits 2 new rows — but the 3rd earlier row stays in the derived collection. If you need strict 1:1 sync with the source array, model deletions explicitly (e.g., source emits full array each update and you compare; or include a deleted flag per line item).
• Empty/missing array is not an error. return [] in TS or FROM json_each(NULL) in SQL both produce zero output docs, which is the right behaviour — but surprises people expecting a per-source-doc output. The source document itself is not reflected in the derived collection.
• json_each behavior on non-array input. Only json_each(NULL) and json_each('[]') silently yield zero rows. A JSON object yields one row per member, a scalar (json_each(42)) yields a single row, and a non-JSON string raises a malformed JSON error that surfaces as a transform failure. If you hit "no output", inspect the source with flowctl preview --source flow.yaml --name <source-collection> to see what the field actually contains.

Related

• derivation-basics — prerequisite
• derivation-filter-transform — when you want zero-or-one output per input, not many
• derivation-join-collections — when the array you want to flatten is spread across multiple source collections
• How to Flatten an Array Using TypeScript (docs walkthrough)
• Derivations concept