Purchase Options and Spot Economics

## The purchase option is a promise about tenure, not just a price The *same* EC2 instance type, in the same Availability Zone, running the same AMI, can cost wildly different amounts depending on how you buy it. The reason is not that AWS gives volume discounts for fun — it is that each purchase option is a different bargain about **who bears the risk of capacity**. - **On-Demand** — you pay the published per-second rate and the instance is yours until *you* stop or terminate it. AWS never takes it away. You make no commitment and get no discount. - **Spot** — you run on capacity that is currently unsold. It is deeply discounted precisely because AWS reserves the right to take it back the moment an On-Demand or reserved customer wants it. That is the whole trade in one line: **Spot is cheaper because it is interruptible.** ## What "interruptible" concretely means When AWS decides to reclaim your Spot Instance, three things happen that you must design for: 1. **You get a two-minute interruption notice.** It appears in the instance metadata service and as an EventBridge event. Two minutes is short but not nothing — it is enough to stop accepting new work, flush a checkpoint, and drain. 2. **The instance then goes away** with the *interruption behavior* configured on the request. The default is terminate; stop and hibernate are available for persistent requests. Anything on an instance store volume is gone. 3. **You may not be able to get another one.** This is the failure people forget. Interruption is not the only risk — a tight pool also means a *launch* request for Spot capacity simply fails. Spot gives you no capacity guarantee going forward either. ## How the price is set — and the myth of the auction Spot began life as a genuine auction: you submitted a bid, and if your bid exceeded the current market price you ran, and your instance died the moment the price crossed your bid. That model was retired in late 2017 and **candidates still describe the old one in interviews**, which is a reliable tell. Today, AWS sets the price for each **capacity pool** — a given instance type and size, in a given Availability Zone, for a given platform — and moves it *gradually* based on long-run supply and demand for that pool. It no longer spikes to the On-Demand ceiling on a moment's notice, which is what made the old model so frightening to build on. You pay the price in effect for each second you run. You may still set a `MaxPrice` on the request. Its meaning is now narrow: if the pool's price rises above your ceiling you will not be launched, and a running instance can be interrupted for that reason. The default `MaxPrice` is the On-Demand price for that type, and leaving it at the default is usually right — setting it lower does not save you money (you pay the market price either way), it only adds a second reason to be interrupted. ```bash # The current price of one pool, most recent entries first aws ec2 describe-spot-price-history \ --instance-types m6i.large \ --product-descriptions "Linux/UNIX" \ --max-results 5 ``` ## Where Spot belongs The usable rule is: **can this workload be killed mid-flight and restarted somewhere else without a human noticing?** Good fits — CI/CD build agents, batch and ETL jobs with checkpointing, media transcoding, stateless queue consumers, big-data worker nodes, load generators, ephemeral test environments. All of them already treat a node as disposable. Bad fits — the single primary of a self-managed database, a message broker holding undelivered messages, a licence server, a long non-resumable job, anything whose restart cost or customer-visible latency exceeds what you saved. Between those poles sits the common production answer: run the fleet as a *mix*, keeping a floor of On-Demand instances for the traffic you must always serve, and letting Spot carry the elastic part above it. ## The mental model to carry into the interview Don't describe Spot as "cheap EC2". Describe it as **a different reliability contract**. You are not buying a discount; you are selling AWS an option to evict you, and the discount is the premium you collect for it. Every design decision that follows — checkpointing, diversification, an On-Demand base, graceful shutdown on the notice — is just you making sure that option is cheap for you to honour.