Build1 publisher2 min readPublished
AWS's project spending caps delete a paused project's data after 90 days without action
AWS's project spending limits pause a project at its cap and permanently delete its data after 90 days paused without action. The cap bounds what a runaway agent experiment can bill, so someone has to own recovery and keep independent backups.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- The lowest limit AWS accepts is the greater of $20 or its own conservative estimate of what the project is likely to spend.
- AWS sends warnings at 50%, 75% and 90% of the limit, and earlier if its forecast shows the project reaching the limit within 10 days.
- Google Cloud's comparable control, announced in July, caps a single service within a project, and fixed contractual commitments keep billing under it.
- On October 3, Simon Willison argued for hard budget caps by default on usage-priced services, because a budget alert can arrive while the owner sleeps and spending continues.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- exposure Anything durable that shares a project with an agent experiment sits on the same pause-and-delete path, so isolating agents in their own project limits what one runaway loop can take down with it.
- decision Before an agent deploys into a project, its owner has to decide whether a hard pause is tolerable there, since AWS's guide allows production use only where a pause is acceptable.
- constraint Accounts without a paid plan, or outside the limited rollout, cannot create this cap, so an agent deploying there gets no project-level stopping point from AWS.
AWS builds the estimate behind its minimum limit from month-to-date activity, the resources running now and the previous month's activity. AWS says the estimate is more than a strict linear extrapolation [7]. I think the floor is good design: it will not accept a number the project is already on course to exceed. A demo that picked up an expensive supporting cast can carry a high minimum. AWS's advice for anyone who wants a lower limit is to stop resources first [8].
Enforcement is blunt. The limit measures a project's monthly pre-tax charges, with credits excluded [5]. When usage reaches the limit, AWS stops the project's resources as well as pausing it, and the data is preserved at first [3][4]. Anyone with project access can see the limit. Only project owners can manage it [12].
On a $20 limit, the alerts fire at $10, $15 and $18 [1]. The alerts are early notice only. The pause comes when usage reaches the limit, and it does not wait for anyone to read the email [3]. Willison's proposal goes further than an alert because it makes unlimited spending an explicit choice [14]. He argues that many builders would take an error response over an unexpected large invoice [14]. The write-up's author concedes that some production workloads may sensibly want the reverse, because shutting them down is costly in its own right [19].
Optional controls act earlier, based on forecast spending [11]. The first of them can affect autoscaling. An application keeps serving from its existing instances while additional instances fail to launch, and the author says that change in capacity behaviour is something to understand before a traffic spike [11].
The backup advice comes from the write-up, which is a dev.to post. Its author wrote that they read the documentation because "a capped invoice and a recoverable project need different checks" [17]. The author recommends three things: run experimental agents in their own project, check which charges the cap covers, and make someone responsible for recovery and independent backups [16]. Owners of long-held AWS accounts should check the settings on the specific project they plan to use, according to the write-up [18].
What to watch
- Whether AWS takes project spending limits out of limited rollout, or turns them on by default for new projects as Willison proposed.
- Whether AWS adds a warning, export or snapshot step before a paused project's data is deleted at 90 days.
- Whether Google extends its cap from a single service to a whole project, and how committed spend is treated under it.