Impact / Economics

The Free Plan and Keeping a Site Warm

What a site costs on the free plan, what keeping it warm adds, and what the price column does not buy.

The Free Plan

The free plan covers 3.04 million views a month across the whole account, split however you like: one site at 3.04 million views, a hundred sites at 30,439 each, a thousand at 3,043 each. Past that the Workers Paid plan is $5 a month, including 10 million requests, 50 million rows written and 5 GB of storage.

Worker requests, one per view, are the meter that runs out first. Every view reaches the Worker, cached or not, so no change to rendering moves this ceiling. Only a response that never runs the Worker avoids it: a static asset, or a hostname not routed to the Worker. The 3.04 million assumes the 65 static assets beside each page do not bill as Worker requests, and a 198,000-request drive invoked no Worker (measured). The one exception is Workers Cache, which the shipped config does not enable: with it on, the same page load spends 66 requests and the account serves 1,515 views a day. Mirroring pages to R2 or publishing them as assets lifts the ceiling in the model, to 10.2 million views a month for an R2 mirror with no Cache Rule and 82.0 million with one, and to 90.5 million for the assets arm, which is not built. Each carries conditions and a lower regeneration ceiling, and The Off-Worker Arms has the scenarios and sources. At the shipping default a fill writes 1.75 rows, and the re-render ceiling is 50,916 a day, set by rows written. Renders follow saves rather than views: at five saves a day a site re-renders about 25 pages, so a thousand active sites use under half of that ceiling. For a large idle fleet the binding meter changes to storage, at 4.7 MB per site, and uploaded files are stored in each site's database against the same 5 GB.

What you pay, on your own account:

The first free cap you reach, and how much of it is used:

Against each host, one site, as the ratio of the host's bill to drupflare's:

Against each host, a thousand sites, every site asleep between requests. The 10,037.7x cell is the sleeping-fleet ceiling against the latency-matched origin at 10,000 views a site; kept warm, the same cell is 166.0x (Kept Warm):

Kept Warm

A paid plan keeps every site resident by default: the alarm re-arms every 8 s, the object never hibernates, and an editor's first request after idle answers in 156 to 371 ms rather than 2 to 4 s. The free plan leaves a site to the thermal policy, which declines any site below about 505 renders a day, and any plan can set SITE_WARM=0. One site's chain sits inside the paid allowances. A fleet's does not, because the allowances are account-wide: each warm site writes 10,896 rows and makes 10,800 object requests a day, and a thousand of them write 331 million rows a month against 50 million included. The chain is billed for its wall time, 23.3 ms a firing (n=116). derived, from the rate card and that firing, with a recomputation from the rate card alone pinned in tests/node/economics-energy.spec.ts:

What you pay with every site kept warm:

Kept warm, against each host, one site:

Kept warm, against each host, a thousand sites:

A fleet pays for keeping its sites warm, about 35 cents a site a month at 1,000 sites. At 10,000 views a site that leaves a thousand sites 30 times cheaper than the cheapest VPS, 166 times cheaper than a latency-matched origin and 119 times cheaper than the managed floor. The sleeping tables reach 1,810 times at the cheapest VPS and 10,037 times at the latency-matched one. Those are ceilings, reached by a fleet of mostly idle sites left to sleep.

What The Price Column Does Not Buy

A monthly price buys capacity from one place, and this page does not measure capacity beyond the sizing on What A Conventional Host Needs. It does not buy the latency profile: the same workload measures 100.2 ms on a VPS against 8.9 ms on drupflare once an 82 ms round trip is injected, because one serves from a region and the other from the visitor's colo. It does not buy the tail either, where the VPS p50 runs from 59 ms at 4 clients to 182 ms at 96 while drupflare holds. The latency-matched column prices the first of those; nothing here prices the second.