spotinfo

Pick a Spot machine on AWS, GCP or Azure from one command line. No credentials. Every price ships inside the binary, so --offline answers in about a tenth of a second.

The problem

Spot capacity is far cheaper than On-Demand — between 40% and 85% off in the snapshots that ship with this tool. Finding the right machine does not scale by hand:

• Each cloud publishes prices in a different place, in a different shape, on a different schedule. AWS has a JSON feed and an interruption Advisor. GCP has server-rendered HTML pages. Azure has a REST API and a separate documentation site for vCPU and memory.
• The consoles rank by price. They do not rank by "cheapest machine with at least 4 vCPU and 16 GiB that my workload can survive losing".
• A price page has no memory of yesterday, so a script that reads one is a scraper you now own.

The result is that most teams pick one instance family, hard-code it, and stop looking.

What spotinfo does

One command ranks real machines against a real requirement, and says where the number came from.

$ spotinfo recommend --cloud azure --architecture arm64 --min-vcpu 4 --min-memory-gib 16
RANK  CLOUD  REGION        MACHINE            ARCHITECTURE  vCPU  MEMORY GiB  USD/HOUR    SAVINGS  RISK          WHY
   1  azure  centralindia  Standard_D4ps_v6   arm64            4        16.0  0.017076        81%  unavailable   ARCHITECTURE_MATCH,COST_POLICY,KNOWN_POSITIVE_PRICE,RESOURCE_MINIMUMS_MET
   2  azure  centralindia  Standard_D4ps_v5   arm64            4        16.0  0.018665        81%  unavailable   ARCHITECTURE_MATCH,COST_POLICY,KNOWN_POSITIVE_PRICE,RESOURCE_MINIMUMS_MET
   3  azure  centralindia  Standard_D4pds_v5  arm64            4        16.0  0.022361        81%  unavailable   ARCHITECTURE_MATCH,COST_POLICY,KNOWN_POSITIVE_PRICE,RESOURCE_MINIMUMS_MET

Four properties make that output worth trusting:

It works with no credentials. Price snapshots ship inside the binary, and a weekly job refreshes them through a reviewed pull request. AWS and Azure additionally read live feeds when they can reach them — AWS on every run, Azure when you name one or two regions — and fall back to the snapshot when they cannot. --offline skips every price and risk request; only --with-score still reaches a cloud, because no snapshot carries a placement figure.

It says what it does not know. GCP and Azure publish no redistributable interruption data, so every candidate reports RISK: unavailable. It is never a zero and never a low bucket. A cloud that measures nothing must not outrank a cloud that measures honestly. On GCP, --live-risk fetches a per-project preemption rate for the ranked page — Google measures it differently from AWS, so it is shown and never filtered on.

It refuses questions it cannot answer, and says why. --workload web caps interruption frequency at an AWS Spot Advisor bucket boundary. Ask for it on a cloud that measures something else and the command stops before it reads a price, naming the vendor limit rather than a feature nobody built:

$ spotinfo recommend --cloud gcp --workload web --architecture x86_64 --min-vcpu 4 --min-memory-gib 16
spotinfo: gcp: unsupported capability: risk: the web workload caps interruption frequency at 5%, an AWS Spot Advisor bucket boundary, and gcp publishes no figure measured that way; workload cost applies no ceiling and answers on every cloud