Cloudflare Workflows pipeline

Cloudflare Workflows gives you durable steps on Workers via step.do. Cron schedules on the workflow binding auto-create instances — no separate cron Worker. Failed steps retry with configurable backoff.

See Running the enrich pipeline durably for the model.

Bundle setup

Your bundle runs inside the Worker (or calls out to a Node service that holds it — same bundle calls either way). For a Worker-hosted bundle, keep the database URL in secrets:

src/search.ts

import { samesake } from "@samesake/postgres";
import { llmRerank } from "@samesake/server";
import { products } from "./catalog.ts";
import { geminiEmbed, geminiGenerate } from "./gemini.ts";

export function createCatalogApp(env: Env) {
  return samesake({
    url: env.SAMESAKE_DATABASE_URL,
    collection: products,
    models: { embed: geminiEmbed, generate: geminiGenerate },
    rerank: llmRerank(geminiGenerate),
  });
}

Workflow class

src/workflows/catalog-sync.ts

import {
  WorkflowEntrypoint,
  type WorkflowEvent,
  type WorkflowStep,
} from "cloudflare:workers";
import { createCatalogApp } from "../search.ts";
import { fetchChangedRows } from "../catalog-source.ts";

type Env = {
  SAMESAKE_DATABASE_URL: string;
};

export class CatalogSync extends WorkflowEntrypoint<Env> {
  async run(event: WorkflowEvent<unknown>, step: WorkflowStep) {
    const app = createCatalogApp(this.env);

    await step.do("upsert", async () => {
      const rows = await fetchChangedRows();
      await app.enrich.upsert(rows);
    });

    await step.do(
      "enrich",
      {
        retries: { limit: 5, delay: "5 seconds", backoff: "exponential" },
      },
      () => app.enrich.enrich(),
    );
  }
}

export class CatalogMaintenance extends WorkflowEntrypoint<Env> {
  async run(_event: WorkflowEvent<unknown>, step: WorkflowStep) {
    const app = createCatalogApp(this.env);

    // Image revalidation is a caller-owned @samesake/server workflow; retry failed rows here.
    await step.do("retry-failed", () =>
      app.enrich.retryFailed({ limit: 500 }),
    );
  }
}

wrangler.toml — bindings and schedules

The schedules array on each workflow binding runs the loop on a cron — Cloudflare creates a new instance automatically.

wrangler.toml

name = "samesake-pipeline"
main = "src/index.ts"
compatibility_date = "2026-06-09"

[[workflows]]
name = "catalog-sync"
binding = "CATALOG_SYNC"
class_name = "CatalogSync"
schedules = ["*/5 * * * *"]

[[workflows]]
name = "catalog-maintenance"
binding = "CATALOG_MAINTENANCE"
class_name = "CatalogMaintenance"
schedules = ["0 * * * *"]

Webhook fast-path — create an instance from a Worker

When a product changes, create a workflow instance with that row in params, or push directly and let the scheduled sync pick it up:

src/index.ts

import { createCatalogApp } from "./search.ts";
import { CatalogSync } from "./workflows/catalog-sync.ts";

export default {
  async fetch(request: Request, env: Env): Promise<Response> {
    if (request.method === "POST" && new URL(request.url).pathname === "/webhooks/product") {
      const product = await request.json();

      // Option A: push inline, scheduled sync handles enrich
      const app = createCatalogApp(env);
      await app.enrich.upsert([product]);

      // Option B: trigger a sync workflow now
      const instance = await env.CATALOG_SYNC.create({
        params: { trigger: "webhook" },
      });

      return Response.json({ ok: true, instanceId: instance.id });
    }

    return new Response("Not found", { status: 404 });
  },
};

Use step.sleep("name", "20 seconds") between steps if you need deliberate delays — instances in a waiting state don’t count toward concurrency limits.

Note

Per-step retry config (retries: { limit, delay, backoff }) applies to that step.do only. A failure in enrich retries enrich — not upsert.

Why this fits samesake

Workers + Workflows keep everything at the edge if your app already lives there. samesake’s incrementality means each scheduled instance still runs two bundle calls over the whole collection — cheap when only dozens of rows changed. You don’t fan out per product; Cloudflare’s durability covers per-step retries.

Where to go next

• Concept and platform picker — Running the enrich pipeline durably
• Row statuses — Pipeline lifecycle
• Cloudflare Workflows docs — developers.cloudflare.com/workflows

Run it end to end

The runnable, tested version of this guide lives in examples/pipeline-cloudflare-workflows — deterministic stubs, no API key. Run bun test (Postgres-backed examples need SAMESAKE_DATABASE_URL).