Looking at bursty, event-driven container workloads where your containers can be idle for hours or days. The "no VM to manage" promise is key, but the pricing models diverge sharply here, and it'll gut your ROI if you pick wrong.
I've benchmarked a real internal tool scenario: ~500 tasks/month, each running for 5-7 minutes. Idle time is ~95%. Here's the raw cost breakdown from last quarter's pilot.
**Assumptions per task:**
* 1 vCPU, 2 GB RAM
* Execution time: 6 minutes avg
* 500 executions/month
**Monthly Cost Estimate:**
* **Cloud Run:** $0.48 (compute) + $0.00 (provisioned instances not used). You pay per request + execution time. The 100k free requests cover it; cost is basically just the compute seconds.
* **Azure Container Instances:** ~$12.80. You pay for the entire time the container group's image is cached and ready, plus execution. Their "per-second" billing has a 1-minute minimum for each run. Storage costs for the image add a few cents.
* **Fargate (via ECS):** ~$9.10. You pay for vCPU and memory provisioned **per minute**, rounded up. The biggest hit here is the 1-minute minimum billing, even for a 6-second task. Plus, you're paying for the load balancer if you use one.
**The verdict for true burstiness:** Cloud Run wins on pure cost for this pattern because of per-second billing with no minimum and no charge for idle. ACI and Fargate penalize short-lived, infrequent bursts.
**Negotiation tip:** This is where your cloud provider's commitment discounts (like CUDs) become almost useless. Your leverage is threatening to architect around their container service for these workloads. Use that in broader platform discussions.
Key questions for your own setup:
* What's the minimum task duration? (Under 1 minute skews harder against ACI/Fargate)
* Is the image pull time significant? (Adds to "execution" time in Cloud Run/ACI, hidden cost in Fargate)
* Are you already committed to a specific cloud's ecosystem? The operational overhead of multi-cloud might negate the savings.
—hd
—hd