How Can I Reduce My Azure Costs with Complex Workloads?

For getting a grip on cost reporting, I usually export cost data to a storage account using the Cost Export feature. Then, I run a YAML pipeline to process this info with a Python script to make it more digestible. Not that I’m a Python whiz, but ChatGPT has been super handy for that. Plus, I’m checking out the Azure SDK to pull metrics directly for a more comprehensive analysis. I found filtering by tags and subscription names really helpful!

In my past job, I managed to cut costs by 30% using a few strategies: First, I extracted the cost and usage data into Power BI, then identified high-cost contributors at both subscription and resource group levels. Drilling down into those costly resources really helped in understanding what drove the expenses. I also recommended measures like shutting down or scaling back unused resources and implementing auto-sleep functions. It’s crucial to set budgets once costs stabilize. I’d also suggest considering a FinOps framework since it might suit your complex setup better.

I can relate! In our company, we had over 65 subscriptions under a single tenant. The billing section of the portal allowed us to analyze where the bulk of our spending was happening by breaking it down by resource and subscription. It’s definitely a long process, but automation can help a lot. We generate monthly reports to track spending increases or decreases, and we’re developing a daily report too. We also deployed some Optimization and Governance workbooks to get insights into potential savings, particularly through Savings Plans and Reserved Instances. Those can hit your budget goals harder than you'd think!

I've dealt with similar issues, especially once ML workloads ballooned costs. We noticed many GPUs just sat idle, so we revamped our setup: we ran ML tasks as on-demand jobs instead of keeping GPU nodes on 24/7. This made a huge difference! Additionally, we optimized inference by quantizing models and moving less traffic to CPUs while queuing the rest. Introducing department-level monitoring also worked wonders since teams could actually see their spending patterns, which naturally led to reduced usage. And let's not forget about cleaning up unused resources and optimizing containers!