Skip to content

estuary/agent-skills / skills/derivation-stateful-logic

derivation-stateful-logic

Verified

Create an Estuary derivation with custom SQLite state (internal tables) to make per-document decisions based on accumulated history — account balances, inventory levels, approval workflows, deduplication. Derivations add complexity and cost — confirm the user wants a derivation before reaching for this. Use when user says "stateful derivation", "derivation to track balance", "transformation to approve based on state", "derive an approval workflow", "deduplication derivation", or "build a stateful derivation".

derivation-stateful-logic

Estuary derivation that maintains internal SQLite tables to record state, and whose lambdas query and update that state to decide what to emit.

Prereq: read derivation-basics first — in particular the stateless-vs-stateful distinction and the role of migrations. This skill is the "stateful via SQLite tables" path.

Docs:

When to use this over alternatives

  • Balance / account state: debit sender if funds available, reject otherwise
  • Inventory: deduct stock, mark backordered if insufficient
  • Approval workflows: accumulate approvers, emit "approved" once threshold met
  • Deduplication with history: filter events by ID seen-before (beyond what collection-key uniqueness gives)
  • Any business logic that depends on prior events for the same key

Reach for other skills when:

  • Pure aggregations (sum/count/min/max) that reductions can express → derivation-aggregate-metrics
  • Merging fields across sources → derivation-join-collections
  • Time-bounded sliding state (last 24h) → derivation-windowing (which also uses SQLite internal state, but with readDelay)
  • Stateless per-doc transforms → derivation-filter-transform

How it works (one paragraph)

You declare SQLite migrations in the derive.using.sqlite.migrations list — SQL that runs once at startup to create your state tables. Your lambda then queries and updates those tables as it processes each source document, and can SELECT a derived output row or not, depending on state. Each task shard owns its own SQLite database; events are routed to shards via a keyed shuffle, so all events for the same entity (e.g. the same account) reach the same shard and see the same state.

Canonical Example — SQL

Transfer approval system: track account balances, approve transfers with sufficient funds, reject overdrafts.

Project layout

my-derivation/
├── flow.yaml
├── schema.yaml
├── init.sql           # migration: create balances table
├── debit-sender.sql   # lambda for sender-side (decide + debit)
└── credit-recipient.sql  # lambda for recipient-side (silent credit)

schema.yaml

type: object
properties:
  transfer_id:
    type: string
  sender:
    type: string
  recipient:
    type: string
  amount:
    type: number
  outcome:
    type: string
    enum: [approved, rejected]
  reason:
    type: string
  sender_balance_after:
    type: number
required: [transfer_id, sender, outcome]

init.sql (migration)

CREATE TABLE IF NOT EXISTS balances (
  account TEXT PRIMARY KEY NOT NULL,
  balance REAL NOT NULL DEFAULT 0
);

flow.yaml

Two transforms: one evaluates transfers against the sender's balance; the second credits the recipient's account on approved transfers (reading from the derivation's own output — a self-reference).

collections:
  acmeCo/processed/transfer-outcomes:
    writeSchema: schema.yaml
    readSchema:
      allOf:
        - $ref: flow://write-schema
        - $ref: flow://inferred-schema
    key: [/transfer_id]
    projections:
      outcome:
        location: /outcome
        partition: true
    derive:
      using:
        sqlite:
          migrations:
            - init.sql
      transforms:
        - name: debitSender
          source: acmeCo/production/transfers
          shuffle: { key: [/sender] }
          lambda: debit-sender.sql

        - name: creditRecipient
          source:
            name: acmeCo/processed/transfer-outcomes   # self-reference!
            partitions:
              include:
                outcome: [approved]   # only approved transfers
          shuffle: { key: [/recipient] }
          lambda: credit-recipient.sql

Keyed shuffle is critical: /sender for debit, /recipient for credit. Each account's events go to one shard, which owns that account's balance.

debit-sender.sql

-- 1. Read current balance (default 0 if new account)
WITH current AS (
  SELECT COALESCE((SELECT balance FROM balances WHERE account = $sender), 0) AS balance
)

-- 2. Emit the outcome document
SELECT
  $transfer_id AS transfer_id,
  $sender      AS sender,
  $recipient   AS recipient,
  $amount      AS amount,
  CASE WHEN current.balance >= $amount THEN 'approved' ELSE 'rejected' END AS outcome,
  CASE WHEN current.balance >= $amount
    THEN 'Sufficient funds'
    ELSE 'Insufficient funds: balance ' || current.balance || ' < ' || $amount
  END AS reason,
  CASE WHEN current.balance >= $amount
    THEN current.balance - $amount
    ELSE current.balance
  END AS sender_balance_after
FROM current;

-- 3. Deduct from the sender's balance only if approved
INSERT INTO balances (account, balance)
SELECT $sender, -$amount
WHERE (SELECT COALESCE((SELECT balance FROM balances WHERE account = $sender), 0)) >= $amount
ON CONFLICT(account) DO UPDATE SET balance = balance - $amount;

credit-recipient.sql

-- Silently credit the recipient; this transform emits nothing.
INSERT INTO balances (account, balance)
VALUES ($recipient, $amount)
ON CONFLICT(account) DO UPDATE SET balance = balance + $amount;

-- Produce no output rows (so nothing gets published from this transform):
SELECT NULL WHERE 0;

The SELECT NULL WHERE 0 idiom is how you write a transform that updates state but emits no document. Needed because every SQL lambda must be a syntactically valid query.

Canonical Example — TypeScript

TypeScript derivations support stateful work via an in-memory state object, but the idiomatic stateful path in Estuary is SQLite. Use TypeScript for stateful work only if you have a specific reason (complex Deno-side libraries, WASM, etc.) — otherwise SQLite is more robust, survives shard splits cleanly via the recovery log, and has a clearer operational story.

Test

Local validation — flowctl preview --fixture

Stateful derivations build state within a preview run. Ingest an ordered sequence of transfers and watch the outcomes evolve. Note: flowctl preview uses an in-process runtime (no temp-data-plane subprocess), which also means shard-splitting semantics won't be exactly like production — but for single-shard stateful logic, preview is a good reflection for the primary transform.

Important caveat for self-referential derivations under --fixture: flowctl preview --fixture does NOT simulate transforms that source from the derivation's own output. The fixture file is the only input the preview reader sees, so a creditRecipient-style self-read sees no documents — the recipient side won't be credited. (Without --fixture, journal-backed preview can read historical self-output, but this requires a previously-published live collection.) For self-referential patterns under development, validate in a real data plane: ingest via HTTP or a capture and read back from the derived collection.

fixture.jsonl:

["acmeCo/production/transfers", {"transfer_id": "t1", "sender": "alice", "recipient": "bob", "amount": 50}]
["acmeCo/production/transfers", {"transfer_id": "t2", "sender": "bob", "recipient": "alice", "amount": 20}]
["acmeCo/production/transfers", {"transfer_id": "t3", "sender": "alice", "recipient": "bob", "amount": 1000}]
{"commit": true}

Starting balances are all 0. Expected outcomes:

  • t1 — alice has 0, tries to send 50 → rejected (sender_balance_after: 0)
  • t2 — bob has 0, tries to send 20 → rejected (sender_balance_after: 0)
  • t3 — alice still has 0 → rejected

To see approvals you need to seed balances first. For a richer test, add a "deposit" transform that credits accounts without checking — see the AcmeBank tutorial for the full flow.

flowctl preview --source flow.yaml \
  --name acmeCo/processed/transfer-outcomes \
  --fixture fixture.jsonl --timeout 30s

Spec-defined tests — tests: block (aspirational)

tests:
  acmeCo/tests/transfer-outcomes:
    - ingest:
        description: Three transfers with zero starting balances — all should reject.
        collection: acmeCo/production/transfers
        documents:
          - { transfer_id: "t1", sender: "alice", recipient: "bob",   amount: 50 }
          - { transfer_id: "t2", sender: "bob",   recipient: "alice", amount: 20 }
          - { transfer_id: "t3", sender: "alice", recipient: "bob",   amount: 1000 }
    - verify:
        collection: acmeCo/processed/transfer-outcomes
        documents:
          - { transfer_id: "t1", sender: "alice", outcome: "rejected", sender_balance_after: 0 }
          - { transfer_id: "t2", sender: "bob",   outcome: "rejected", sender_balance_after: 0 }
          - { transfer_id: "t3", sender: "alice", outcome: "rejected", sender_balance_after: 0 }

Variations

  • Inventory tracking: CREATE TABLE inventory (product_id TEXT PRIMARY KEY, stock INTEGER). Lambda deducts on orders, marks backordered if insufficient. Separate state-only transform for replenishments — just INSERT ... ON CONFLICT DO UPDATE, no SELECT needed.
  • Approval workflow: migration creates an approval_state table with approvers JSON array. One transform initialises requests, another adds approvers. Emit approved when json_array_length(approvers) >= required.
  • Deduplication: CREATE TABLE seen (event_id TEXT PRIMARY KEY). Lambda SELECT ... WHERE NOT EXISTS (SELECT 1 FROM seen WHERE event_id = $event_id) then INSERT OR IGNORE INTO seen. Collection key uniqueness handles simple cases; use this when source has duplicates at different key granularity.
  • Session tracking: state per user_id, open session on login, close on logout, emit session summary on close.
  • Rate limits: count events per user in state; reject once threshold hit. Combine with derivation-windowing for time-bounded rate limits.

Gotchas specific to stateful-logic

  • shuffle: any silently breaks state. With unkeyed shuffle, each event may land on a different shard and see empty state. Use shuffle: { key: [/entity_id] } wherever you read/write SQLite tables. This is the #1 cause of "my balances don't accumulate" bugs.
  • Initial activation after publish takes time. Migrations run once at first shard startup; subsequent restarts only apply new migrations. Transient errors right after publish typically resolve on their own as shards finish initializing.
  • State-only transforms can use bare DML. When a transform only updates state and emits nothing, plain INSERT, UPDATE, or DELETE statements are valid — no trailing SELECT needed. (The canonical AcmeBank creditRecipient lambda is just INSERT ... ON CONFLICT DO UPDATE.)
  • Self-referential transforms (reading from your own output) are powerful but subtle. Use partitions.include to filter which self-events the transform processes, or you can create infinite loops.
  • Migrations are append-only. You cannot change a published migration — you must add a new one. State is destroyed only on shard delete/recreate (disable/enable preserves it). For critical state, plan how you'd rebuild from source-of-truth events if a recreate is ever needed.
  • Shard splits fork the recovery log. Both new shards initially hold identical SQLite state; subsequently each processes a disjoint key range under keyed shuffle. Stale rows for the other shard's range remain in each DB but are harmless because no reads target them.
  • Don't use shuffle: any with ON CONFLICT. You might think ON CONFLICT makes it idempotent, but without keyed shuffle, concurrent updates to the same row across shards will not converge.
  • Order matters within a lambda. Query state before updating it if you want to read pre-update values. SELECT balance FROM balances ... before INSERT ... ON CONFLICT UPDATE.

Related

  • derivation-basics — prerequisite (stateless-vs-stateful, shuffle mechanics)
  • derivation-windowing — time-bounded state via readDelay
  • derivation-aggregate-metrics — when reductions can express the logic (no procedural state needed)
  • AcmeBank derivations tutorial — the canonical walk-through for this pattern
  • Internal task state
deprecation-and-migrationaddyosmani/agent-skillsManages deprecation and migration. Use when removing old systems, APIs, or features. Use when migrating users from one implementation to another. Use when migrating a database schema in production, such as renaming or dropping a column without downtime (expand/contract). Use when deciding whether to maintain or sunset existing code.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
performance-optimizationaddyosmani/agent-skillsOptimizes application performance across frontend, backend, queries, and databases. Use when performance requirements exist, when you suspect performance regressions, when Core Web Vitals or load times need improvement, when N+1 query patterns need fixing, or when profiling reveals bottlenecks.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
agent-cost-reportthedotmack/claude-memBelievable agent cost report for any period, default the last 7 full days PT, not counting today. Measured tokens from Claude Code transcripts priced at OpenRouter list prices (ESTIMATED), measured provider spend when a sanctioned source exists, note-taker cost separate, Timing-style HTML/PDF plus report.json, line-items.csv, evidence.json.Stars: 118kUpdated Claude CodeCodexVerifiedInstall
agent-cost-reportthedotmack/claude-memBelievable agent cost report for any period, default the last 7 full days PT, not counting today. Measured tokens from Claude Code transcripts priced at OpenRouter list prices (ESTIMATED), measured provider spend when a sanctioned source exists, note-taker cost separate, Timing-style HTML/PDF plus report.json, line-items.csv, evidence.json.Stars: 118kUpdated Claude CodeCodexVerifiedInstall