Cost attribution keeps finding waste that monitoring dashboards miss - how do others handle the gap?

I work on cost attribution for a large internal cloud fleet, and a pattern keeps repeating: our observability stack says everything is healthy, and the invoice says otherwise.

Recent example: a batch worker that was fully green in monitoring — no errors, no alerts, normal resource graphs - but was costing \~$4,200/month sitting mostly idle because it was provisioned for a peak workload that moved to a different pipeline two quarters ago. Nobody's dashboard is designed to surface "this thing is fine but shouldn't exist."

What I've landed on so far:

* Attribution before optimization. Tag/label hygiene and per-team allocation first, because until spend maps to an owner, "reduce costs" is nobody's job.
* Treating the invoice as a second architecture diagram. The design doc shows what the system is supposed to be; the bill shows what it actually is. Divergence between the two is where the waste lives.
* Idle detection as its own signal, separate from utilization alerts — low-but-nonzero usage is the expensive failure mode, because it never trips anything.

Where I still don't have a good answer: shared infrastructure. NAT, load balancers, control planes, observability itself - the stuff that serves everyone and therefore belongs to no one. We've tried proportional split by traffic and even split by team, and both create weird incentives.

For those doing FinOps/cost work at any scale: how do you attribute shared platform costs without either (a) burying teams in chargeback complexity or (b) making platform costs invisible again?

https://redd.it/1v1wa6m
@r_devops

Do devils use ai workflows in enterprise

I’m building one personal project, a platform where DevOps can create ai workflows without coding(drag and drop/ chatbot) , like n8n but for DevOps. But firstly I want to ask you guys who is working in the industry, some questions to get clarity what to build.
A example of the workflow can be taking data from the aws cloud watch for perticular project and if there’s any perticular event which needs attention then send email for summary of that to user.

1. what are the tasks you have already automated ??

2. Do you guys trust ai for reasoning over any infrastructure related problems ??

3. What are the tasks that can be repeated and can be done by ai ? if you have any idea.

4. What are the platforms you use for these tasks.(Ex. Aws cloudwatch, Jenkins, Prometheus, Grafana,etc)

5. Can this project be helpful in the field

Open for discussion.
Thank you!!

https://redd.it/1v1tra1
@r_devops

If you had a 300M parameter model, what would you optimize it for?

Im working on AI infrastructure and have been thinking about where small language models actually make the most sense.

Suppose you had a 300M parameter model and your goal wasnt to compete with large frontier models at everything, but instead to consistently outperform much larger models (2B–20B) on one specific use case.

What would you optimize it for?

A few ideas that came to my mind:
Code generation for a narrow domain
Structured data extraction
Document classification
Workflow or agent planning
Log analysis
Something else entirely

I’m less interested in benchmark scores and more interested in real-world workflows where a small model could genuinely be the better choice because of specialization, latency, reliability, or deployment constraints(but ofc i also want benchmark scores to be good too lol).

If you had to pick one domain where a highly specialized 300M model could become the obvious choice over much larger models, what would it be, and why?

https://redd.it/1v1wrxd
@r_devops

I need a career

Last month, I sent a professional-grade DAM (Desarrollo de Aplicaciones Multiplataforma) in Spanish. But I don't know what to do now. I think I have to study more, but I don't know if I should pursue an engineering degree or just take courses and specializations.

https://redd.it/1v2eoyn
@r_devops

How are you setting up build systems in monorepos?

Former DevOps consultants, bootstrapped an AWS premier partner and exited to NTT Data, now we advise people building services companies.

After seeing how much easier current AI tools have made it to write code we started building tools for helping our customers and day to day work. Some of the decisions we've made so far:

1. Build out the code in separate services to keep each part well-contained. It feels AI does a far better job with smaller pieces than with larger.
2. Deploy them through lambda.
3. Use a monorepo so the AI can easily see how different things work together.
4. Migrated some apps built in Lovable into our repo.
5. All the building is done with Makefiles
6. Created a registry so it is easier for AI to add new tools in a consistent manner and so we can analyze our inventory of what exists in our development.
7. Using make for building cuz Claude made that decision when we barely had anything and I'm a dinosaur at heart.

Now, I have a monorepo with multiple different languages, a registry that I want to be the single source of truth for what I have, and a build system from the 80s. And I was curious what would people recommend for the build system rather than building a whole bunch of custom scripts?



https://redd.it/1v2ppv2
@r_devops

Good observability tool recommendation - Cloud Based

I work in two companies, one very big that have a giant budget to spend on Datadog, and another small one, that don't have that very big budget.

On the small company we are looking to migrate from Prometheus + Grafana + Alert Manager to something cloud based. Mostly because we are a small team (3 people only) and we don't have time to spend maintaining the infrastructure for it to run.

Now, is there some good alternative to Datadog? Datadog is the king, but is brutally expensive.

I've tried Signoz, looked promising, but is bad. They documentation is bad, there are just general ideas without details (they assume that you know a lot from I don't know where), they support is lame (they only have a chat to "Interact with a Human", that takes days to receive an answer, and their UI is buggy. Really bad experience.

But looking there, they are the only ones having like "close" experience to Datadog. We tried New Relic in the past, but they have all metrics and log scattered over the place without correlation. Also tried BetterStack but lacks a lot of features compared to other things.

So, is there some good observability platform cloud-based that I'm not aware of, outside Datadog? I'm not completely closed to self-hosting as long as it takes low effort to maintain.




https://redd.it/1v34gf2
@r_devops

DevOps Engineer with ~1 year of experience, but I feel like I'm not learning enough. What should I focus on?

Hi everyone,

I'm a DevOps Engineer with about 1 year of experience, but I feel like I'm not getting enough real-world exposure at work.

Most of my work is around CI/CD, AWS deployments, CloudFront, Docker, CloudFormation, Linux. I haven't had much experience with Kubernetes, production incidents, scaling, networking, or designing infrastructure.

What should I learn next to become a better DevOps/SRE engineer? Any roadmap, projects, or resources you'd recommend for gaining practical experience?

Thanks!

https://redd.it/1v37qh0
@r_devops

AI and Context Switching

Found this article discussing AI and context switching -- https://medium.com/@vibhatha/context-switching-in-the-age-of-ai-b7c53433164c

Does AI really help minimizing context switching? If you have to learn new skills in managing multiple AI agents, it may add more workloads. But if you are comfortable with the setup in the environment, why should you use AI to "rock the boat"?

Just a thought.

https://redd.it/1v36rpb
@r_devops

Multi platform build

​

I am planning to have my docker images to be multi architecture and for it my plan is to have two codebuilds for arm and amd when they both succeeds event bridge will invoke a lambda and lambda will merge the images into one

But I am unable to think a way to set up event my current plan is something like when codebuilds are triggered if anyone succeeds let's say arm build then event bridge will match the source version and and if the other build also success then it will invoke lambda

Apparently there is no ability to compare source version in eventbridge and I would have to invoke lambda and lambda has to compare the source version

Is there any way I can implement this

https://redd.it/1v39ap7
@r_devops

Currently on Falco for runtime security — anyone moved to Tetragon/KubeArmor/Tracee and regretted (or loved) it?

Running Falco in our EKS clusters right now for runtime detection, paired with Kyverno for admission control and Cosign/Vault for supply chain signing/secrets. Been solid so far, but I keep seeing Tetragon and KubeArmor come up as alternatives, especially for teams who want enforcement (block, not just alert) rather than detect-and-notify.

Curious what people are actually running in production and why:

If you moved off Falco, what pushed you? Overhead, rule fatigue, lack of enforcement?
Anyone running Tetragon specifically for the eBPF enforcement piece? Is it worth the Cilium tie-in if you're not already on Cilium for CNI?
KubeArmor folks: how's LSM behavior in practice across different node kernel configs (AppArmor vs BPF-LSM)? Heard that's where it gets messy.
Tracee: anyone using this seriously in prod or is it mostly a DFIR/forensics tool for you?

Not looking to rip out Falco, just trying to figure out if there's a compelling reason to add enforcement on top, or if pairing Falco with Falco Talon covers that gap well enough.

https://redd.it/1v3aave
@r_devops

Who owns the "why we did it this way" knowledge on your team?

Small team here and I keep running into the same thing. The reasoning behind half our setup lives in two people's heads.

Not in the runbooks, not in Confluence, not in the repo. Someone new joins and rediscovers everything the hard way and when one of those two is out we just stall.

It got noticeably worse once we started leaning on Claude Code and Cursor for real work. The agents read whatever context files we give them and treat it as truth, so now stale knowledge doesn't just slow a person down, it gets confidently baked into code by something that has no idea the decision changed months ago.

So I wonder how this actually works at bigger shops. Is there anyone whose job includes keeping that stuff current or does it just live with whoever happens to remember? And has anyone here had a documentation process survive past the first busy quarter or does it always quietly die?

Not looking for tool recommendations just trying to work out whether we're unusually bad at this or whether it's the normal state of things.

https://redd.it/1v3bzt2
@r_devops

Where is AI actually adding value in your DevOps workflow and where isn't it?

Senior DevOps here. AI is being pushed into every part of the pipeline right now, and I want to cut past the hype and hear real production experience.

Two questions:

Where have you put AI into production in your DevOps process and it genuinely adds value? (e.g. CI/CD, code review, IaC generation, monitoring/alerting, incident response, log analysis, documentation)

Where did it not prove worthwhile? Think cost, alert noise, false positives, or maintenance overhead that outweighed the benefit.

Thanks.

https://redd.it/1v3a3u7
@r_devops

How do you handle production patching for EC2 instances?

how do you handle production patching for EC2 instances in your environment?

I'm interested in learning about real-world production practices.

Some questions:

* Do you use AWS Systems Manager Patch Manager, Patch Policies, or another approach?
* Do you patch EC2 instances in place, or do you replace them with new AMIs (immutable infrastructure)?
* How do you schedule maintenance windows and minimize downtime?
* How do you handle Auto Scaling Groups during patching?
* What's your rollback strategy if a patch causes issues?
* Do you test patches in dev/staging before production?
* How much of the process is automated versus manually approved?

I'd really appreciate hearing how your organization handles production patching at scale, along with any best practices or lessons learned.

https://redd.it/1v3fu2f
@r_devops

Best way to learn K8 now?

Hi all, I’m learning Kubernetes right now for work (I use Colima + kubectl) - what is the best way to do so with AI?

I’ve used docker before in previous projects and studied concepts of containers in school - im no expert but I should know enough to learn Kubernetes.

I started the Udemy + kodekloud course: Kubernetes for the absolute beginners - Hands on. I’m not sure if it’s relevant for 2026 - so far it’s asking me to manually write yaml files, which I’m sure is important, but do I really need to be able to write yaml files with vscode extensions and AI that writes it for me?

So basically, what’s the best way to learn Kubernetes in 2026? I would appreciate any courses I should complete after my current one or a roadmap.

I have around a month or so to be good enough to collaborate with our platform and aws teams.

https://redd.it/1v3gykw
@r_devops

I am a Sr System Administrator and want to switch to DevSecOps

To begin with this is my first post or i don't know what it is called on reddit,

Apologies for the mistakes in my English please ignore it,

But I really want to switch from system admin job to DevSecOps

About myself

I have 12 years of experience starting from desktop support Engineer to laptop repairing to Data center to Monitoring to Asset management to system administration, I have knowledge of both windows and Linux operating system and currently managing Windows and Linux server (QA) no prod since my senior have never let me work on prod server even though I have worked with him since 7 years, handled a US based client for 3 years working with his team and currently handling two clients one with just some basic needs regarding systems and O365 and second with Linux servers, kind of devops but not fully the second client have a deployment using jenkins ( fetch the code from got repo, call the specific server through ssh and run the deployment and build script)

Worked with a friend's company as a contractor on his client for 6 months which ended last December and since then no work on core devops, started learning k8s and currently learning it but not able to give 100% to it also leaning python scripting and bash scripting by the help or AI can read the code and understand what it does but can not write with full confidence

I know the tools below tools

Terraform

Github

Github pipeline

Jenkins ( basic free style pipeline)

Got the idea of Sonarqube while working as a contractor but it was short lived

Understanding New relic as now the client wants to setup it

AWS basic

Azure basic and ADO basic

GCP not so much

Done a course in cybersecurity as well from local institute but not able to perform handson

I honestly want to work in Devsecops

I am already 38 and according to me not earning enough (peer pressure)

Any suggestions any help or direction is appreciated

I am even ready to do an unpaid internship ( beside the job since i have responsibilities)

Call it desperation, determination or pressure but I really want to work in Devsecops domain

Please help

https://redd.it/1v3jdnh
@r_devops

At what point do you stop treating timeouts as normal?

Hey guys! we've got a couple of internal services that will occasionally throw timeout errors under load. They retry, recover, and users never seem to notice.

The annoying part is that it's been happening long enough that nobody really reacts to it anymore. Every incident review ends up with the same general response that it's just something we've always seen.

Now I'm wondering if thats just the reality of distributed systems, or if we've gotten too comfortable ignoring something that probly deserved a closer look a while ago.

How do you decide when intermittent timeouts have crossed the line from something you live with to something worth digging into?

https://redd.it/1v3gelr
@r_devops

Looking for Terraform + AWS hands-on project ideas after completing Terraform Associate and AWS SAA

Hi everyone,

I recently completed:

HashiCorp Certified: Terraform Associate (003)
AWS Certified Solutions Architect – Associate (SAA-C03)

Now I want to move beyond certifications and gain more hands-on practical experience with Terraform on AWS.

I understand the Terraform concepts and AWS services at a theoretical level, but I want to build real-world projects that would help me improve my skills and also create some strong portfolio projects for my resume.

I am looking for recommendations on projects that simulate what a Cloud Engineer / DevOps Engineer would actually work on in an organization.


My goal is not just to deploy resources but to understand how Terraform is structured and used in real enterprise environments.

Would appreciate suggestions on:

Which projects are most valuable for learning Terraform + AWS together?
Any GitHub repositories, courses, or labs you recommend?
What kind of Terraform projects stand out on a Cloud/DevOps resume?

Thanks in advance!

https://redd.it/1v454t7
@r_devops

Which DevOps tool looked amazing during the PoC but became painful after six months?

I am Little confused ?

https://redd.it/1v445k8
@r_devops

Moving a high-memory Python application from Kubernetes to a dedicated Linux server — what would be the best setup?

Hello,

At the company where I work, we have a Python application running on a Kubernetes cluster. The application is fully integrated into a CI/CD workflow. Whenever code is merged, Jenkins pulls the source code from GitLab, builds the application, and pushes the image to Harbor. Argo CD then deploys it to the Kubernetes cluster.

However, the application consumes a large amount of RAM, and I have had to increase its memory allocation several times. We have now decided to move the application to a separate Linux server.

The application runs with Uvicorn, and I plan to make it accessible through a domain name, just as it currently is in the Kubernetes environment. However, I am not sure what the best-practice architecture would be for this setup, so I would appreciate your recommendations.

I still want Jenkins to automatically pull the application from GitLab, build it, and deploy it. After deployment, the application should automatically start and run as a web service.

My main concerns are how to restore the application as quickly as possible if a problem occurs, how to handle backups, and what kind of deployment and recovery strategy would be best.

In short, could you please advise me on the best way to design and implement this setup?

https://redd.it/1v475m2
@r_devops

How are developers and operations meant to work in this situation?

Here’s our situation - a mature operations team with pipelines building container images for a third party product going into kube

Company acquired a company with a development team who had no ops but are proficient at producing dockerfiles for their products

Ops is in azure devops, developers in bitbucket

We want to leverage the developer teams skills to get the apps into ops docker images

Some suggest that ops could build base images and developers could build their apps with that as the base - but what if the base images need patching? That means ops depends on developers for that?

Another thought I had was developers keep Dockerfiles in devops, the containers build from sources (aka git clone bitbucket) and build the project all in the ops pipeline

Would love to hear how others do this…

https://redd.it/1v480u5
@r_devops