Pagefeather: measure every approach, then provision the one that holds

#cloudflare#wordpress#woocommerce#php

If you build or run WordPress sites, this post is for you too. I am an edge developer. I also care about the CMS that agencies, shops, and publishers already know. Pagefeather tries to hold both sides at once.

WordPress has always had a place with me. I like the CMS. I like the editor, the plugins, and a store a person can run. I have never liked the LAMP backend. The machine stays on. Apache or nginx, MySQL, and PHP sit there even when nobody reads.

This project keeps that CMS and drops that box. You still get wp-admin. You still get themes and plugins. Visitors still see WordPress HTML. What changes is the host under it. There is no VPS to patch. There is no MySQL server to babysit on the default path. Cloudflare runs the site at the edge.

A few Cloudflare words, in plain terms. A Worker is code that runs near the visitor. A Durable Object is one Cloudflare object that keeps a site’s state. Here that state is PHP and its SQLite database. R2 is object storage for uploads. You do not need those names to care about the result. You need them when you provision.

A content site scales to zero while it is idle. A cacheable page comes from the edge. A light-to-moderate WooCommerce store can take an order in the same Durable Object. The CMS stays WordPress either way.

Coming from a Linux box with Apache or nginx, PHP, and MySQL? Start at docs/from-lamp-to-wfp.md. It maps your jobs onto this stack one to one. Then come back here for the research, or go to the framework docs for the commands.

I work on this because both crowds leave a gap. A lot of edge work ignores WordPress. A lot of WordPress work stays on a warm box. I want other people to run sites with a CMS they already know. That is what this work is for. This blog is Astro on purpose. The customer sites I care about are WordPress.

Pagefeather is two things in one GPL-2.0 repository. The first is a matrix of ways to run WordPress and WooCommerce on Cloudflare. The second is the framework that matrix produced. You run many sites on your own Workers for Platforms account. A command or a Worker provisions each site. It is a framework you run. It is a research log. It has no signup and no hosted plan. Before 2026-11-13 the repository name was wordpress-on-cloudflare.

The method is worth copying even if you never deploy it. A cell is one of five marks:

Mark Meaning
measured We deployed it and ran the test in tests/. The result names the command and the date.
docs Cloudflare’s documentation says it is possible. We did not deploy it.
claimed A third party says it works. We did not deploy it.
no A documented limit blocks it. We name the limit.
unbuilt No known implementation exists.

A speed number with no command and no date is not in this repo. The same test on the same code moved between runs. Approach A served 18.7 requests per second at 100 users in one load matrix. It served 8.63 in another. On the Durable Object variant, two runs of the same build differed by up to 15 percent on one page. We bracket comparisons. We run the same build with the change off, on, and off again. A single run is never the number for a variant.

Four shapes

Id PHP runs in Database
A A Worker, PHP compiled to WebAssembly MySQL through Hyperdrive
B A Worker, PHP compiled to WebAssembly D1, or SQLite inside a Durable Object
C A Cloudflare Container, stock WordPress SQLite on the container disk, or MySQL through Hyperdrive
D A Worker for visitors, WordPress for admin Whatever the back office uses

Headless stayed off the product. WordPress keeps rendering the storefront. We document D. The framework does not provision it. If you want the classic admin and the classic theme output, that is the point.

C1 was stock WordPress in a Container with SQLite on disk and Litestream to R2. We retired it (#214). It measured well. b-dosql replaced it for stores. The CLI refuses kind: shop.

The plan column decides more than the diagram does. On Workers Free, approach A stops at the CPU limit. The installer fails there. PHP inside a Durable Object fits Free for plain WordPress. A Durable Object gets a longer CPU budget than a Free Worker request. A light WooCommerce storefront can run on Free. Editing that store needs Workers Paid (#638). On Workers Paid, approach A passes the end-to-end script and a WooCommerce checkout. The tax is per-query time. Give Hyperdrive the database host name. A query measured about 7 ms. The same path by IP address measured about 44 ms. A Tunnel with Access measured about 100 ms.

A2 is approach A with the database on PlanetScale. We measured it on 6 November 2026 on Paid. Install passed. End-to-end was 17 of 17. WooCommerce was 12 of 12. Queries took about 16 to 20 ms. PlanetScale is strict about SQL mode. Hyperdrive’s default query cache lasts 60 seconds and does not invalidate on writes. That cache can serve a stale WordPress read. The deploy turns that cache off. Check the integration’s auto-created config before you trust it.

Transactions, or the lack of them

ROLLBACK does not mean the same thing on every row of the matrix. If you have ever trusted a WooCommerce order or a plugin upgrade to undo a failed write, this table matters.

Variant Database A rollback undoes the writes
b-do D1 No. Each statement commits on its own. PHP answers BEGIN and COMMIT.
b-dosql SQLite in the Durable Object Yes. An undo log in the object.
c1 (retired) SQLite on the container disk Yes. PHP’s own driver.
a, a2 MySQL through Hyperdrive Yes.

WooCommerce, some plugins, and core upgrades rely on rollback. On plain b-do they do not get it. That is a measured property of D1. A green page render is not a safe write.

What you actually provision

The operator document is docs/framework.md. Two kinds of site share one dispatch namespace.

wp is b-dosql. Think of it as one site in one place: PHP and its SQLite live in one Durable Object. The site deploys as a user Worker. Use it for a blog or brochure site. Use it for a light-to-moderate WooCommerce store. Checkout is measured on Workers Paid under the 128 MiB memory limit. tests/woo.py is green. Cached guest pages come from the edge. Cart, checkout, account, and logged-in pages always run PHP. That matches how you already think about cache plugins. Guests can hit cache. Shoppers cannot.

a is the scale-out path. PHP still runs in a Worker. The database sits outside Cloudflare, the way a managed MySQL would. Hyperdrive reaches it. Use it when a store needs real concurrent traffic. See docs/onboarding.md.

There is no container store on this framework. A store that outgrows b-dosql can move to approach A. It can also move back to a normal LAMP VPS. Leaving is allowed. That is the point of honest limits.

ggc.sh is the live WooCommerce example. It runs on Pagefeather b-dosql. Since 2026-11-15 its guest storefront pages cache at the edge. Product, shop, and category pages stay fresh for 30 s by default. Other guest pages stay fresh for 120 s. There is no store Container behind it. Open it if you want to click a real cart and checkout, not a diagram.

One function, two callers

provisionSite is Cloudflare API calls: fetch and Web Crypto. The CLI and a Worker call the same function. Provision time builds nothing. You build the b-dosql bundle once. Every later site uses that bundle. In WordPress terms: you are not SSHing into a box to install each site. You ask the framework to create one.

node deploy/provision/cli.mjs add --kind wp blog
node deploy/provision/cli.mjs add --kind wp mystore
node deploy/provision/cli.mjs add --kind a scaleout

add is idempotent and touches only that site. For wp it creates the media bucket. It deploys the namespace script with that site’s secrets in the same call. It writes the KV route.

A visitor domain is a CNAME to your zone. The dispatch Worker reads the Host from a KV site map. It runs the user Worker. The plan that deployment needs is Workers for Platforms Paid. The account token needs edit on Workers Scripts, Workers KV, Workers for Platforms, and R2. It does not need Containers scope.

Limits that matter on a store

These are the limits a WordPress shop owner should know before they bet a client on it.

Two idle sites

The reference deployment’s at-rest bill is about $25 a month. That figure comes from Cloudflare’s pages on 6 November 2026. It is the Workers for Platforms Paid subscription. For two idle sites, requests and CPU sit inside the included allotments. Namespace scripts, Durable Object storage, R2, and KV do too. Traffic is extra. The included tiers are per account. Every site shares them.

One line stays open on purpose. The per-site dispatcher and the shared outbound Worker are ordinary Workers, not namespace scripts. The pricing page does not say whether Workers for Platforms Paid covers them. It may also need a separate Workers Paid plan ($5 a month). If it does, the idle total is about $30. That needs a billing check or a word from Cloudflare. The README leaves it open. This post does too.

Where main is

On 17 November 2026, main still treats a cell as measured only when docs/results.md names the command and the date. It also names the outcome. The store path is b-dosql. The LEAN engine ships. Remote OPcache packs are on by default. Conformance work against upstream WordPress and WooCommerce suites is underway (#759). Exact page-cache invalidation exists. It is off by default (#296). Those commits do not promote a cell by themselves.

If you are a WordPress developer trying this for the first time, start with the LAMP translation doc. Then provision one wp site. Then open wp-admin. That is the honest path. Edge jargon can wait until you need it.

Code: github.com/skyphusion-labs/pagefeather. License: GPL-2.0.

← All posts