I've been experimenting with AI tools for DevOps, but I'm struggling to find real value. For example, I tried using `kubectl-ai` to apply its suggested changes in a staging cluster, and it led to unexpected pod failures and configuration drift that made rollbacks a nightmare. I also attempted to auto-generate complex Helm `values.yaml` files, but the output ignored our schema and created invalid keys. Has anyone here had positive experiences with AI in DevOps, like faster rollouts or fewer errors, while still maintaining control and visibility? Please share your honest experiences and stories!
6 Answers
I tried it for migrating pipelines from Jenkins to GitLab, and it’s hit or miss. Hallucinations tend to happen due to less data on some tasks, so you really need to know what you're doing to catch mistakes. It helps me brainstorm, though!
Definitely! I think the key is iterative prompting until you find what works for your context.
I don't view AI as a replacement for automation; it's more like an accelerator for brainstorming ideas. If it starts automating changes, we risk losing junior engineers and a lot more! Right now, it’s just not reliable enough.
For me, it's all about generating basic skeletons for scripts and unit tests. AI helps me with the groundwork, but complex changes and understanding the code are still on me. I think it’s more of a tool to assist rather than replace anything.
I’m using AI mainly for troubleshooting and generating code from scratch when I need to pick up a new tool. I always double-check everything it produces, but it’s been a major time-saver.
I’m with you; AI is not ready to manage infrastructure or deploy changes without human oversight. It’s more useful as a context-aware signal combiner that enhances decision-making rather than replacing it.
Exactly! AI can generate initial tests, documentation, and even some bootstrapping for Terraform modules, but we’re still in the phase where it's augmenting our judgment.
I've found that AI is mostly useful for debugging and writing scripts. For instance, I used it to create filters for CloudWatch logs, which really sped things up. But I always double-check everything manually because I don’t trust it 100%. It helps me with jq for some pipeline processing too.
Yeah, I'm in the same boat. I think it’s great for proof of concepts where you can control the environment and analyze the results later.
Absolutely! I also rely on it to handle Regex tasks because that can get complicated.
Have you found a specific LLM that works well for your needs? I’ve had great success with Amazon Q for Terraform tasks.