<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Anson Lee on Medium]]></title>
        <description><![CDATA[Stories by Anson Lee on Medium]]></description>
        <link>https://medium.com/@leehosanganson?source=rss-a92105dcc184------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*2-KnDTI95YHpVxRXkYDDcA.jpeg</url>
            <title>Stories by Anson Lee on Medium</title>
            <link>https://medium.com/@leehosanganson?source=rss-a92105dcc184------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 08:55:30 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@leehosanganson/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[GitHub Copilot Changed Their Pricing So I Changed Mine to Qwen3.6–35B-A3B and llama.cpp]]></title>
            <link>https://medium.com/@leehosanganson/github-copilot-changed-their-pricing-so-i-changed-mine-to-llama-cpp-db174f48e6ac?source=rss-a92105dcc184------2</link>
            <guid isPermaLink="false">https://medium.com/p/db174f48e6ac</guid>
            <category><![CDATA[open-code]]></category>
            <category><![CDATA[coding]]></category>
            <category><![CDATA[github-copilot]]></category>
            <category><![CDATA[local-llm]]></category>
            <category><![CDATA[agentic-ai]]></category>
            <dc:creator><![CDATA[Anson Lee]]></dc:creator>
            <pubDate>Fri, 01 May 2026 00:44:37 GMT</pubDate>
            <atom:updated>2026-05-03T22:25:27.893Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n-Efaj8bwwBs1wHqzjEPJg.png" /></figure><h3>The Cloud AI Pricing Trap: Why I Ditched GitHub Copilot for Local LLMs</h3><p>I used to think self-hosting an LLM stack was overkill. <a href="https://medium.com/@leehosanganson/how-i-deploy-my-own-language-model-and-why-you-should-too-4aa1f1b8d0bd">Back when I first deployed Qwen2.5–8B on my homelab</a>, I used it to process daily news and articles and it served me well. But it was also painfully small for any serious agentic coding task. The thing is, I only have 1 16GB GPU at the moment and the current market conditions halted my upgrade plan. With no available upgrade for the coming months, I heavily relied on GitHub Copilot as my main inference engine for the majority of my agentic AI coding workflows. It worked beautifully. Until it didn’t.</p><p>Now, with GitHub Copilot shifting to a usage-based pricing model on June 1, 2026, I found myself staring down a potential bill that could easily reach $200+ per month. That’s not a typo. I did some napkin math based on my usage, and the new billing structure makes my old workflow unaffordable for a hobbyist like me. So I did what any ‘rational’ engineer would do: I rebuilt my entire agentic coding pipeline around local inference, llama.cpp, unsloth, and OpenCode. And today, I’m going to walk you through exactly how I did it.</p><h3>The Math That Broke the Bank</h3><p>Here’s where things get interesting. I thought my monthly usage was somewhat reasonable: estimated at about 15 million tokens across 500 requests, averaging 30k tokens per request. Here’s my breakdown under the new pricing:</p><ul><li><strong>How the pricing works:</strong> Under the new pricing on June 1, 2026, every request counts toward your credits at a rate of 1 AI credit = $0.01 once your monthly quota runs out, which happens almost immediately since the <a href="https://docs.github.com/en/copilot/reference/copilot-billing/model-multipliers-for-annual-plans#model-multipliers-for-annual-copilot-pro-and-copilot-pro-subscribers">new multipliers</a> make the monthly quota essentially non-existent.</li><li><strong>Base premium requests:</strong> 500 requests x 30k tokens per request = ~15M tokens, using Sonnet 4.6 as a bare minimum, $50-$60/month (varies because input and output token costs differ). I wouldn’t call myself a crazy user of Copilot Pro+, given they have a 1500 premium request limit.</li><li><strong>Tool Calls &amp; Sub Agents:</strong> Each premium request will likely trigger a series of ‘normal’ request. Assuming each premium request contains on average 10 tool calls, with every single one counting towards the bill, I’m looking at roughly 500 requests × 10 tool calls × 30k tokens per tool call = ~150M total tokens. If I’m smart and configure my harness to use a cheaper model like Haiku 4.5 for sub-agents and tool calling, it still adds $150-$180/month to the bill, just for the ‘normal’ requests.</li><li><strong>Total estimate:</strong> Around a whopping $200+/month, about the same as Claude Code’s Max.</li></ul><p>I cannot go back. When I need to update a long documentation file without an agent helping me consolidate changes across the PR, the Issue and then a few other files around the codebase, it would fill me with dread. It’s not even that difficult by any means. But looping through a dozen or so files, a few webpages and making several paragraph changes accurately, both contextually and grammatically, is massively killing my productivity, especially when I’m doing this as a hobby in my spare time.</p><p>So I needed an alternative. Before getting into the technical how, let me show you my workflow looked like with Copilot and why those numbers mattered so much.</p><h3>My Agentic Coding Workflow: Before the Pricing Shift</h3><p>My workflow looks like this: I take an idea, drop it into a GitHub Project Issue, let the Copilot agent churn out a draft, then I refine and improve through more chats or just doing it myself. That’s my standard operating procedure.</p><p>The real magic happens when I connect my GitHub Copilot subscription to OpenCode, with a custom agent harness I built for my workflow. Each request I submit triggers a Planner-Generator-Evaluator flow, a pattern ‘inspired’ by <a href="https://www.anthropic.com/engineering/harness-design-long-running-apps">Anthropic’s philosophy of building harnesses for long-running applications</a>. This architecture matters because it separates planning (what to do), generation (doing the thing), and evaluation (checking correctness) into distinct steps, which dramatically improves reliability for extended coding sessions. You can see my full AI configuration setup <a href="https://github.com/leehosanganson/dotfiles/tree/main/ai/.config/ai/agents">here</a>.</p><p>This is where I do my main work. I take the draft from the GitHub Agent, start making changes, edit documentation, diagnose issues, and so on. It’s truly fantastic to speed up things you know you want to do, just vastly faster once you have your SOPs and workflows well-documented in the codebase.</p><p>That was my setup. And when the pricing wall hit, I had to rebuild it.</p><h3>From vLLM to llama.cpp with Unsloth</h3><p>The first thing that came to mind is upgrading my local LLM to something more powerful. My last deployment was Qwen2.5–8B, which served me well but was simply too small for proper agentic coding tasks. To fit something larger than an 8B would mean finding a higher parameter count with more aggressive quantisation model. This is where vLLM shows its weakness: it’s built for deployments in servers with big and heavy GPU compute to move larger models and doesn’t support most of the quantisation methods I wanted to explore. That overhead was something I could no longer bear on my single 16GB GPU.</p><p>So I turned to llama.cpp, which has excellent support for quantisation and is a first-class citizen for unsloth. The conversion was surprisingly smooth: I easily translated my vLLM manifest into a llama.cpp manifest and deployed Gemma-4–26B-E4B on my cluster. It’s an MoE (Mixture of Experts) model with great performance compared to my old deployment. But for agentic coding in OpenCode, it often got stuck in loops or couldn’t understand the tool-calling pattern properly.</p><p>That’s when I deployed <a href="https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF">Qwen3.6–35B-A3B</a>, which is technically better than Gemma 4 on paper. It works wonderfully with OpenCode, and I must say it feels eerily similar to what I was using with Sonnet 4.6.</p><p>Here’s how the hardware holds up:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*53_t3l2L4EAinhklgr_t5Q.png" /><figcaption><em>Fitting Unsloth Qwen3.6–35B-A3B UD-IQ2_M GGUF into a 16GB vRAM GPU</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zYWQgtpO2Ln26lAxVoOMSA.png" /><figcaption><em>Over 100 tool calls in a single request</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eeomGL0V090IJ6_PTGoL_g.png" /><figcaption><em>Reaching ~90 tokens per second for generation</em></figcaption></figure><p>But having the right model wasn’t enough, and I also needed the right harness. With some adjustment, I truly feel that the importance of a better harness has never been greater, and the gap between local models and proprietary models could be closed. Let me walk through what actually happened once I had both the local model and the right setup in place.</p><h3>The Agentic Experience: Local Models, Better Harnesses, and Why Both Matter</h3><p><strong>How the local model performed vs. cloud models.</strong> Small language models are not as powerful as their cloud counterparts, period, but they are good enough now to take on very small tasks or repetitive ones. If I give enough instructions on how to implement and what to avoid, and focus on architecting the solution (which is honestly the part I enjoy), I believe I can get away from a GitHub Copilot subscription for now because Qwen3.6–35B-A3B worked wonderfully with OpenCode. On top of the very usable speed, my harness breaks down my request into smaller tasks with clear instructions on what tool calls should be used. The results are being evaluated and then feed back for verdict. Every interaction is small, but complete when they are all combined.</p><p><strong>Why the architecture mattered more than just model upgrades.</strong> The current setup I’ve got is not limited to just the local LLM I’ve deployed. I connected it with LiteLLM, which is my LLM gateway to other providers like Google AI Studio (where I might investigate using Gemini 3.1 Pro as the planner and orchestrator) and OpenRouter (great for checking out free models on non-sensitive projects). This hybrid approach gives me the best of both worlds: local inference for most tasks, cloud models when I need extra power. What truly made the difference was how I structured the whole thing: the Planner-Generator-Evaluator harness, LiteLLM as my routing layer, and the discipline to know which tasks stay local and which can borrow cloud compute. That structure is what kept me from having to lock into another vendor’s pricing model altogether.</p><p><strong>The OpenCode GitHub Integration experience.</strong> GitHub’s native agentic UX is probably what I missed the most - being able to simply let the cloud agent churn out drafts for my ideas. When I set up the <a href="https://opencode.ai/docs/github/">OpenCode GitHub Integration</a>, which is practically a GitHub Action listening to events containing /oc or /opencode, the experience is not quite there compared to the native cloud Copilot agent. But I think it does the job well enough for my use case: drafting a PR from an issue and then reviewing that PR without context rot, when I&#39;m on my iPad or phone, away from the desktop.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/781/1*sJue8NEArk3yriln-dMgnw.png" /><figcaption><em>Me Triggering OpenCode to review a PR with a Local LLM hosted at home</em></figcaption></figure><h3>What’s Next: The Verdict and Beyond</h3><p>What I’ve built feels resilient. It’s affordable. And most importantly, it’s under my control. That matters more than I expected it to.</p><p>So did I actually replace GitHub Copilot? In short: <strong>yes</strong>. The cost alone tells the story. Going from a potential $200+/month cloud bill to roughly $15/month in electricity (based on the $20/month bill since the server is doing a lot more than just local LLM inference), plus hardware I already own. On the performance side, Qwen3.6–35B-A3B runs at around 90 tokens per second on my single GPU, handles hundreds of tool calls cleanly, and is genuinely workable for drafting from GitHub issues, iterating on documentation, and diagnosing problems. I didn’t lose productivity. I redirected it toward what I actually enjoy: architecting solutions rather than watching an agent fumble through them.</p><p>Should you do it? The local LLM setup truly replaced Copilot for my agentic coding use case. It wasn’t just the model swap from vLLM to llama.cpp with unsloth. It was the entire stack: the Planner-Generator-Evaluator harness, LiteLLM as a routing layer (where I have set up Google AI Studio to test Gemini 3.1 Pro, and OpenRouter for free tiers on non-sensitive projects), which gives me the freedom now. If you have the expertise, interest, and patience, I’d recommend giving it a try.</p><p>The bigger picture here goes beyond cost savings. I’m not optimistic about AI pricing trends going forward: vendors are likely going to use this opportunity to demonstrate profitability, and API rates probably will represent a price floor, not a ceiling in the coming future. Building workflows around that model feels like putting all eggs in a basket that’s actively tightening its grip. As I have to admit, I am addicted to the productivity gain working with LLMs and agentic workflows. With this setup, I have something cloud subscriptions can’t offer: resilience against arbitrary pricing changes, affordability that lets me experiment without budgeting, and control over my own toolchain. That autonomy is worth far more than $200 a month.</p><p><strong>Beyond: </strong>I will keep using it and tinkering with the configuration as part of my workflow now. Next, I’m working on fully IaC-ing my homelab with Terraform and NixOS, treating my infrastructure the same way I treat my LLM stack: versioned, reproducible, and under my own control. If you’re interested in following along, here’s my repository: <a href="https://github.com/leehosanganson/homelab">https://github.com/leehosanganson/homelab</a> and stay tuned!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=db174f48e6ac" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Deploy My Own Language Model at Home with Kubernetes and vLLM]]></title>
            <link>https://medium.com/@leehosanganson/how-i-deploy-my-own-language-model-and-why-you-should-too-4aa1f1b8d0bd?source=rss-a92105dcc184------2</link>
            <guid isPermaLink="false">https://medium.com/p/4aa1f1b8d0bd</guid>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[self-hosting]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Anson Lee]]></dc:creator>
            <pubDate>Sun, 30 Nov 2025 22:34:31 GMT</pubDate>
            <atom:updated>2026-01-20T13:01:28.607Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ogeY1WqZKB4XjKTIsRzbUw.png" /></figure><p>The world of Large Language Models (LLMs) is evolving at an astonishing pace. While many of us rely on cloud-based services for our AI needs — and frankly, for 99% of general use cases, this is likely still the most cost-efficient approach — there’s a growing movement towards self-hosting and deploying LLMs locally. I’ve recently dived deep into this world and am here to share my journey and convince you why you should consider deploying your own LLM stack.</p><h3>Why Bother with a Private Language Model?</h3><p>You can likely still depend on mainstream model providers for most of your use cases, such as drafting emails or summarising online content. Foundational models handle these tasks effectively. There are also free Language Model wrapper services for those extra niche capabilities, such as Perplexity for deep search, Nano Banana for image generation, or NotebookLM for note-taking.</p><p>So you might be asking, “Why go through the hassle when I can just use ChatGPT or Claude?” The benefits of a private LLM are compelling:</p><p><strong>Offline Resilience<br></strong>The recently increasing number of cloud service outages shows us that the internet might be more fragile than we think. The more dependent you are on AI, the more you should consider a disaster plan for when the cloud is no longer available to you. Having a Language Model that works even without an internet connection offers a layer of reliability that cloud services simply can’t match.</p><p><strong>Cost-Effective Offloading<br></strong>If you have repetitive, “low-intelligence” tasks in your current workflow, deploying an offline LLM stack, with your own infrastructure as an asset booked on your balance sheet, could prove more cost-effective in the long run. Tasks can be offloaded to a smaller, privately hosted LLM without incurring the costs associated with API calls to cloud services.</p><p><strong>Learn to Scale<br></strong>For those working in the ML/AI industry or interested in the inner workings of AI, deploying an LLM stack is an invaluable learning experience. Understanding how applications are deployed at scale, especially in environments like Kubernetes, is a highly sought-after skill in the coming years, with the increase in AI literacy, democracy, and privacy.</p><h3>My Setup: Adding a GPU to my Kubernetes cluster</h3><p>My journey began with my Kubernetes cluster hosted on Proxmox VMs. I recently added a GPU node to my Proxmox cluster specifically for AI/ML workloads. Setting up GPU passthrough correctly was important to allow my worker node to detect and fully utilise the hardware. While Proxmox recognised the GPU, I had to follow a <a href="https://forum.proxmox.com/threads/pci-gpu-passthrough-on-proxmox-ve-8-installation-and-configuration.130218/">guide</a> to enable IOMMU and ensure the node correctly detected the vendor ID.</p><p>Next, I installed <a href="https://github.com/NVIDIA/gpu-operator">Nvidia GPU Operator</a> via Helm within my Kubernetes cluster. gpu-operator simplifies much of the driver and device management on nodes and has removed most of my headaches regarding driver versions <em>(you would know if you&#39;ve ever tried to install GPU drivers on Ubuntu fully following the guide, only for Jupyter Notebook to still report no GPU device found)</em>. The only additional change I needed to make was to label the node appropriately and configure containerd socket paths to work with k3s. Also, remember to use runtimeClassName: nvidia for your GPU workloads so that they are correctly assigned to the GPU node.</p><p>Example values.yaml for gpu-operator on k3s:</p><pre>toolkit:<br>  env:<br>    - name: CONTAINERD_CONFIG<br>      value: /var/lib/rancher/k3s/agent/etc/containerd/config.toml<br>    - name: CONTAINERD_SOCKET<br>      value: /run/k3s/containerd/containerd.sock<br>    - name: CONTAINERD_RUNTIME_CLASS<br>      value: nvidia<br>    - name: CONTAINERD_SET_AS_DEFAULT<br>      value: &quot;true&quot;</pre><h3>The Stack: vLLM, LiteLLM, and OpenWebUI</h3><p>I believe an LLM stack includes three things: an inferencing engine, a model aggregator, and a front-end application. Here are the choices for my stack.</p><p><strong>vLLM: The High-Concurrency Backend<br></strong>For my LLM backend, <a href="https://github.com/vllm-project/vllm">vLLM</a> stands out. It’s designed for reliable and stable concurrent LLM API endpoints, capable of handling multiple requests simultaneously and reliably. While frameworks like Ollama and llama.cpp might offer broader model compatibility, vLLM’s ability to manage concurrency at scale is a trade-off I’m happy to make. This <a href="https://arxiv.org/abs/2511.17593">paper</a> investigates the performance of vLLM against Hugging Face TGI, and the experiments found a whopping “24x higher throughput under high-concurrency workloads”. The downsides for vLLM are that it might not support all mainstream quantisation methods and also requires more VRAM due to allocation on KV Cache for concurrency requests. This means the options for models are more limited than other inferencing engines like Ollama, given the same hardware and amount of VRAM.</p><p><strong>LiteLLM: The Simple Model Gateway</strong><br>Beyond a local open-source Language Model, I still want to experiment with and benchmark paid services like Gemini and Claude. <a href="https://github.com/BerriAI/litellm">LiteLLM</a> here acts as an LLM aggregator or gateway, providing a single endpoint and a unified set of keys to interact with various models, as long as they support the OpenAI API format. A significant benefit is its dashboard, which allows for monitoring usage across all models and setting limits to prevent unexpected costs. I can also provision unique keys for different users and use cases to improve governance and security. Today, API keys are equivalent to your wallet. Make sure you keep them safe.</p><p><strong>Open WebUI: The Most Popular User Interface</strong><br><a href="https://github.com/open-webui/open-webui">Open WebUI</a> has quickly become a popular choice for interacting with local LLMs. It integrates seamlessly with LiteLLM due to its support for third-party OpenAI API endpoints. It offers features like web search (I’ve configured Google PSE, but DuckDuckGo, SerpAPI, or SearxNG are also options), custom model creation with system prompts and knowledge bases, and the ability to import tools. One thing I found annoying with cloud services is that they limit the amount of files you upload. With Open WebUI, I am able to upload as many documents as I want, as long as they can fit into the model’s context window, for free.</p><h3>Real-World Use Cases</h3><p>After all this hassle setting up infrastructure and deploying services, here are some of my actual day-to-day use cases with local Language Models.</p><p><strong>Obsidian Integration<br></strong>My note-taking workflow in Obsidian is significantly enhanced. By configuring the <a href="https://github.com/logancyang/obsidian-copilot">Obsidian Copilot Extension</a> with my LiteLLM endpoint, I can leverage LLM capabilities directly within my notes. I even created a separate LiteLLM key for Obsidian to monitor its usage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/403/1*_DOz1VsX_WihGIR2DIKQMg.png" /><figcaption>Creating SEO title and Description without leaving Obsidian</figcaption></figure><p><strong>Automated Content Digests with n8n<br></strong>I use n8n to create an automated pipeline to digest content from various RSS feeds every day. Initially, I tried to feed the entire RSS feed content into the context window and hoped it would generate something readable using just system prompts. However, due to the lengthy context window, this ultimately resulted in unfocused responses and, more importantly, cost more money.</p><p>By introducing a ranking step with a screening agent before content generation, I can now filter out irrelevant context first and ensure that whatever goes into the context window is somewhat related to my goal. This process needs to be repeated hundreds, if not thousands, of times every day, so I decided to use a smaller, local model for this step to save money. I’ve managed to reduce the context going into the writer agent by up to 80%. Ideas on future improvements would be something related to a proper RAG system around the sourced documents. But current approach is effective enough for my simple use case.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t3UlAXj-fgD-j1hPVj6P-Q.png" /><figcaption><em>Screening Agent in n8n to rank documents with a Scoring Gate</em></figcaption></figure><h3>My Current Model Choice</h3><p>After testing various models on my 5060ti with 16GB VRAM, I’ve found <a href="https://huggingface.co/unsloth/Qwen3-8B-unsloth-bnb-4bit">unsloth/Qwen3-8B-unsloth-bnb-4bit</a> to be a good balance. I&#39;ve configured it with:</p><ul><li>max_num_seqs: 4 (parallel requests)</li><li>max_model_len: 20480 (context length)</li><li>kv-cache-memory: 7293324288 (assigned VRAM for KV cache, ~6-7GB)</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JWUv9qd_hlvio-4FGv2fTw.png" /><figcaption><em>The KV Cache details in the image might differ slightly from the text; always refer to the actual configuration.</em></figcaption></figure><h3>Conclusion</h3><p>Deploying your own LLM stack might seem unnecessary and daunting at first, but the benefits in terms of control and learning are immense. You might even save some money if you have a niche use case. Tools like vLLM, LiteLLM, and OpenWebUI, combined with Kubernetes, make this once-complex task more accessible than ever. If you’re serious about AI and want to push the boundaries of what you can do, I highly encourage you to explore building your own LLM infrastructure.</p><p>You are welcome to visit my <a href="https://github.com/leehosanganson/homelab/tree/main/kubernetes/apps/vllm">homelab repository</a> and see if it can expedite your deployment process onto your Kubernetes cluster. Follow for more!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4aa1f1b8d0bd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Journey to Kubestronaut in 2025 — Becoming a Kubestronaut as an MLE]]></title>
            <link>https://medium.com/@leehosanganson/journey-to-kubestronaut-in-2025-becoming-a-kubestronaut-as-an-mle-d8e4931f22a4?source=rss-a92105dcc184------2</link>
            <guid isPermaLink="false">https://medium.com/p/d8e4931f22a4</guid>
            <category><![CDATA[kubestronaut]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[certification]]></category>
            <dc:creator><![CDATA[Anson Lee]]></dc:creator>
            <pubDate>Thu, 02 Oct 2025 06:31:01 GMT</pubDate>
            <atom:updated>2025-10-02T06:31:01.864Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/800/1*15kv_PhZhxicp1rdBkyWMA.png" /><figcaption>Kubestronaut</figcaption></figure><h3><strong>Journey to Kubestronaut in 2025 — Becoming a Kubestronaut as an MLE</strong></h3><p>Becoming a <a href="https://www.cncf.io/training/kubestronaut/"><strong>Kubestronaut</strong></a> is more than just a certification, it’s a testament to committing significant hours to undestand Kubernetes and cloud-native technologies. My six-month journey to earning all five Kubernetes certifications has been both challenging and inspiring, proving that this path cultivates not just technical expertise but also personal and professional growth. As a Machine Learning Engineer, I would like to share with you my experience, the challenges I overcame, and the valuable lessons learned through this journey.</p><h3>What is a Kubestronaut?</h3><p>A <strong>Kubestronaut</strong> is a professional formally recognised by the <strong>Cloud Native Computing Foundation (CNCF)</strong> for holding all five core Kubernetes and cloud-native certifications:</p><ol><li><a href="https://training.linuxfoundation.org/certification/kubernetes-cloud-native-associate/"><strong>KCNA</strong></a> (Kubernetes and Cloud Native Associate)</li><li><a href="https://training.linuxfoundation.org/certification/certified-kubernetes-application-developer-ckad/"><strong>CKAD</strong></a> (Certified Kubernetes Application Developer)</li><li><a href="https://training.linuxfoundation.org/certification/certified-kubernetes-administrator-cka/"><strong>CKA</strong></a> (Certified Kubernetes Administrator)</li><li><a href="https://training.linuxfoundation.org/certification/certified-kubernetes-security-specialist/"><strong>CKS</strong></a> (Certified Kubernetes Security Specialist)</li><li><a href="https://training.linuxfoundation.org/certification/kubernetes-and-cloud-native-security-associate-kcsa/"><strong>KCSA</strong></a> (Kubernetes and Cloud Native Security Associate)</li></ol><p>These certifications collectively validate a deep understanding of Kubernetes, cloud-native principles, and essential security practices. They are tangible proof of hands-on experience, close to real-life problem-solving capabilities, and a dedicated commitment to continuous learning within the cloud-native ecosystem. This validation is increasingly critical in an era demanding resilient infrastructure for highly scalable and secure software. As a Machine Learning Engineer, becoming a Kubestronaut helps me realise the ways to leverage Cloud Native for modern AI and machine learning deployments effectively.</p><h3>Why Kubernetes and Cloud-Native Matters?</h3><p>Kubernetes has become the cornerstone of modern cloud infrastructure, driving innovation across DevOps, microservices, and automation. Engaging with this domain offers professionals significant advantages:</p><ul><li><strong>Empowers Development Freedom</strong>: As microservices, containerization, and CI/CD become standard practices, mastering these technologies grants you the freedom to choose the optimal approach for development and deployment. You’re no longer bound by vendor lock-in or cloud provider constraints, enabling solutions tailored to specific needs.</li><li><strong>Deep Technical Mastery</strong>: Navigating the vast Cloud Native ecosystem requires expertise in Kubernetes and its tools. This knowledge directly translates to securing, scaling, and optimising cloud-native applications effectively and efficiently.</li><li><strong>Accelerated Career Growth:</strong> There is significant demand for Kubernetes expertise across cloud engineering, DevOps, and Site Reliability Engineering (SRE) roles. <a href="https://www.gartner.com/en/documents/5988071">Gartner predicts</a> that more than 75% of all AI/ML deployments will be hosted in containers by 2027. Notably, ML/MLOps engineers, who bridge ML model development and production deployment, have seen a more than 5x growth in the past year.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KFqI03h5HLXVgDEMZjt51A.png" /><figcaption>Google Trend for ML Engineer over the last 5 years</figcaption></figure><p>My journey into Kubernetes began with curiosity about containerisation, evolving into a passion for applying it for my day-to-day-problem solutions. From just self-hosting a simple WebDAV server for my Zotero library, to managing a highly available Kubernetes cluster on a server rack for family media and private LLM services, hands-on experience has reinforced the value of cloud-native technologies in solving real life problems.</p><h3>My Journey with the 5 Certifications</h3><p>Earning all five certifications within six months was an intense, yet incredibly rewarding experience. It was a journey that spanned the full spectrum of Kubernetes knowledge, from foundational concepts to security practices. My preparation leveraged a mixture of structured learning with hands-on practice, and I’m grateful for the resources that made it possible.</p><p>My preparation began with <a href="https://learn.kodekloud.com/user/learning-paths/kubestronaut"><strong>KodeKloud</strong></a>, which offers a wide range of structured courses including the 5 certifications required for Kubestronauts. I focused on the areas where I needed the most improvement and practiced with their mock exams afterward.</p><p><a href="https://killercoda.com/"><strong>killercoda</strong></a> provides straightforward scenarios to practice for CKA, CKAD and CKS. They provided excellent supplementary resources to reinforce the KodeKloud materials.</p><p>I also made use of <a href="https://killer.sh/"><strong>killer.sh</strong></a> mock exams, which are also very useful to get your hands dirty with more challenging questions and a shorter time limit. Two sessions are available for free as part of your exams so make sure to make the best use of them.</p><p>For KCSA, other than the KodeKloud mock exam, I found this <a href="https://kubernetes-security-kcsa-mock.vercel.app/"><strong>KCSA mock exam app</strong></a> by <a href="https://github.com/thiago4go">Thiago S Shimada Ramos</a> to be extremely helpful for testing my understanding across different topics quickly.</p><p>For a deeper dive into my experience, check out these two articles I wrote: <a href="https://medium.com/@leehosanganson/journey-to-kubestronaut-in-2025-cka-aa5bfd4fc0c8">Journey to Kubestronaut in 2025 — CKA</a> and <a href="https://medium.com/@leehosanganson/journey-to-kubestronaut-in-2025-how-did-i-score-88-100-in-cks-ccb897735ed8">Journey to Kubestronaut in 2025 — How did I score 88/100 in CKS</a>.</p><h3>Key Lessons Learned</h3><ol><li><strong>Hands-On Practice &gt; Theory:</strong> No amount of study can replace real-world experience. I enjoyed the performance-based exams like CKAD, CKA and CKS. As the active engagement in a live Kubernetes environment and practical expertise are directly transferable to real-life experience.</li><li><strong>Time Management is Key:</strong> Understand your strength and weaknesses, prioritise things that you can control rather than wasting time to reach perfection.</li><li><strong>My Perseverance:</strong> If I bang my head hard enough onto the wall, I will eventually learn it! This new found resilience is much appreciated and essential for continuous learning and problem-solving in the fast-evolving fields of AI &amp; ML.</li></ol><h3>Final Thoughts</h3><p>Earning all five certifications has been a transformative journey. I deepened my understanding of Kubernetes and cloud-native technologies. I feel more confident than ever when managing my homelab cluster, or developing Cloud Native software in my job. This achievement demonstrates not only my technical expertise but also a proactive, problem-solving mindset crucial for designing the future of AI/ML infrastructure. For those considering this challenging yet rewarding path, remember that <strong>consistency, hands-on practice, and a growth mindset</strong> are your greatest allies.</p><p>After this challenge I will continue to<strong> </strong>create more content with my homelab and look into other certifications in the future. Follow me for more content about my personal development journey, homelab setups, and self-hosting experiences.</p><p>#Kubernetes #Kubestronaut #CNCF #Certification #CloudNative #DevOps</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d8e4931f22a4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Journey to Kubestronaut in 2025 — How did I score 88/100 in CKS]]></title>
            <link>https://medium.com/@leehosanganson/journey-to-kubestronaut-in-2025-how-did-i-score-88-100-in-cks-ccb897735ed8?source=rss-a92105dcc184------2</link>
            <guid isPermaLink="false">https://medium.com/p/ccb897735ed8</guid>
            <category><![CDATA[kubestronaut]]></category>
            <category><![CDATA[cks-exam]]></category>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[certification]]></category>
            <category><![CDATA[devops]]></category>
            <dc:creator><![CDATA[Anson Lee]]></dc:creator>
            <pubDate>Tue, 23 Sep 2025 20:01:42 GMT</pubDate>
            <atom:updated>2025-09-25T09:50:21.728Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Certified Kubernetes Security Specialist Logo" src="https://cdn-images-1.medium.com/max/1024/1*6bFgHPrG7Mbmz0dKs0Om2A.jpeg" /><figcaption>Certified Kubernetes Security Specialist (CKS)</figcaption></figure><h3><strong>Journey to Kubestronaut in 2025 — How did I score 88/100 in CKS</strong></h3><p>After tackling the challenging KCNA, CKAD, and most recently the CKA, I’ve just taken on what many consider the toughest of the Kubernetes certifications: the Certified Kubernetes Security Specialist (CKS). This exam is a critical step on my journey to becoming a Kubestronaut, a recognition from CNCF for holding all five core Kubernetes and cloud native certifications.</p><p>The CKS exam demands a deep understanding of securing the Cloud, Cluster, Container and Code. If you’re wondering how to approach this beast or are preparing for it in 2025, you’re in the right place! In this post, I’ll share my preparation strategy, a quick overview of the topics covered in September 2025, and most importantly, my personal strategy for tackling the CKS exam questions to get a score of 88/100.</p><h3>Preparation Strategy</h3><p>My preparation for the CKS exam was similar to CKA: tons of reading and watching videos, followed by a lot of mock tests. While I relied heavily on the Reddit Community for my CKAD and CKA, unfortunately, community resources for CKS are not as abundant or up-to-date.</p><h4>Key resources included</h4><p><strong>KodeKloud CKS Course:</strong> I paid for a subscription primarily for the structured learning experience, which I found more comprehensive compared to other online resources. I also found the difficulty of their mock exams to be more comparable with the actual exam.</p><p><strong>Killercoda Scenarios:</strong> I spent half a day powering through all the available scenarios. I found them helpful in identifying question types and narrowing down potential solution directions.</p><p><strong>Killer.sh Exam Simulator:</strong> Most of my success and experience come from this. My scores progressed from 41/84 on the first attempt, to 67/84 on the second, and finally 81/84 on the third, all within 2 days, which greatly boosted my confidence. Although it’s only 1 set of questions, the difficulty is much harder than the actual exam and requires significantly more effort to solve each question. Successfully completing this within the time limit (without cheating, obviously) should be a good indicator of your ability to pass the actual exam.</p><h3>Exam Day</h3><p>The CKS exam is similar to CKAD and CKA. You will get access to separate Kubernetes environment for each question and need to solve the task within that environment. While I won’t go into full detail, and if my memory serves me well, the types of questions I encountered roughly included:</p><p><strong>Admission Controller<br></strong>Setting up and managing admission controllers in kube-apiserver.</p><p><strong>ImagePolicyWebhook<br></strong>Minimal field changes to existing files and editing kube-apiserver.</p><p><strong>Docker hardening<br></strong>Tasks related to securing the Docker daemon, such as modifying TCP settings, removing users from docker group, and changing docker.sock ownership.</p><p><strong>Container image scanning<br></strong>Needing to find if the image is containing a certain package.</p><p><strong>Network Policy<br></strong>Being careful with the side effects of deny all and allow all policies.</p><p><strong>Falco Rules<br></strong>Only realised it was a Falco question after looking at the question’s documentation links. It required writing your own rule.</p><p><strong>CIS Benchmark<br></strong>Running and fixing issues according to kube-bench. This often involved only copying and pasting the right commands from suggested fixes.</p><p><strong>kube-apiserver Hardening<br></strong>Small edits in kube-apiserver manifest.</p><p><strong>Istio<br></strong>Istio was considered “free marks”. It involved very quick and simple changes to deploy sidecars.</p><p><strong>Cluster upgrade<br></strong>Only upgrading a worker node, which was much simpler than what I experienced in the CKA.</p><p><strong>Service Account tokens<br></strong>Disabling auto-mount and mounting them as a projected volume instead.</p><p>Interestingly, I did not encounter any questions on strace or Cilium, despite preparing significantly for the latter.</p><h3>My Strategy for the Exam</h3><p>CKS is considered the most difficult among the five, and it shows during my preparation. After taking multiple mock exams I developed my own little system aimed at effectively completing all questions.</p><p><strong>Initial Scan &amp; Mapping<br></strong>I started the exam by reviewing all the questions first. I quickly noted down in a terminal “to do list” with question numbers, my progress (not started, working on, done), and keywords for each question by “dry-running” my approach mentally. This helped me create a mental map and identify “quick wins” for early points in less than 10 minutes for all questions.</p><pre>1  done     cluster upgrade<br>2  n        network policy, 3 policies, 2 namespaces<br>3  n        psa, imagepolicywebhook<br>4  done     static analysis<br>5  wip      itsio<br>6  n        docker. tcp, chown<br>…</pre><p><strong>Prioritise &amp; Multi-tasking<br></strong>Time is tight for CKS. If you want to finish all the questions, you either need to be very fast or try to multitask as much as possible. My prioritisation depended on my practice strengths. I typically left network policy questions (both native Kubernetes and Cilium, though Cilium didn’t appear) for last. My go-to questions first were cluster node upgrades and static analysis, as upgrades usually takes 5+ minute to wait for and static analysis is typically easy to pick up where I left off. Next would be Istio, PSA, and CIS Benchmarks, as these often involve more scripted procedures and held fewer surprises.</p><p><strong>Screen Management<br></strong>Using a 13-inch MacBook, I kept my browser and to-do list full-screen, always keeping my plan in sight and in mind. Then, I had at most two terminals open half-screen, zoomed out, for the actual tasks I’m working on.</p><h3>Tips &amp; Handy Snippet</h3><p>I would like to also share some tips &amp; snippets I found very useful during exams.</p><p><strong>Typing</strong><br>Probably the most important thing in the exam due to all kinds of network delay and pure stress. Remember to make good use of tab to auto complete kubectl commands. Other than vim commands, use Ctrl+a and Ctrl+e to jump to start and end of line. Ctrl+r to search previous commands. Ctrl+u to clear input line. Less motion means fewer mistakes.</p><p><strong>Quick Deployment &amp; Expose</strong></p><pre>kubectl create deployment nginx --image=nginx --replicas=1<br>kubectl create service clusterip nginx --tcp=80:80</pre><p><strong>Restart </strong><strong>kube-apiserver</strong></p><pre>sudo mv /etc/kubernetes/manifests/kube-apiserver.yaml ..<br>sudo &quot;crictl ps | grep api&quot;<br>sudo mv ../kube-apiserver.yaml /etc/kubernetes/manifests/</pre><p><strong>Important Directories</strong></p><pre>/var/lib/kubelet<br>/var/lib/etcd<br>/var/lib/containerd<br><br>/etc/kubernetes<br>       /manifests<br>       /seccomp<br><br>/usr/bin/dockerd<br>/var/run/docker.sock<br>/etc/docker/daemon.json</pre><p><strong>Checksum</strong></p><pre>echo “sha filename” | sha512sum -c</pre><p><strong>Falco Targeting keyword</strong></p><pre>falco -U | grep &lt;keyword&gt;</pre><p><strong>Kube-bench Targeting keyword</strong></p><pre>kube-bench run | grep -C5 “keyword”</pre><p><strong>Update/Change kubelet Configuration</strong></p><pre>k -n kube-system edit cm kubelet-config<br>kubeadm upgrade node phase kubelet-config<br>service kubelet restart</pre><p><strong>Service Account RBAC</strong></p><pre>k create role --verb=create,delete --resources=pods,deployments,sts<br>k create rolebinding --role role --serviceaccount=namespace:saName<br>k auth can-i --as system:serviceaccount:namespace:saName …</pre><p><strong>Cilium Policy</strong></p><pre>apiVersion: “cilium.io/v2”<br>kind: CiliumNetworkPolicy<br>metadata:<br>  name: policy<br>spec:<br>  endpointSelector:<br>    matchLabels:<br>      type: messenger<br>  egressDeny:                    # deny <br>    — toEndpoints:<br>        — matchLabels:<br>            type: database<br>      icmps:                     # ping<br>        — fields:<br>            — type: 8<br>              family: IPv4<br>            — type: EchoRequest<br>              family: IPv6<br>      authentication:<br>        mode: “required”         # Enable Mutual Authentication</pre><p><strong>Network Policy To Specify a single IP</strong></p><pre>192.168.100.21/32</pre><p><strong>Write the command output to file</strong></p><pre>k exec pod &gt; /path/to/file -- &lt;command&gt;</pre><p><strong>etcd Get Secret</strong></p><pre>ETCDCTL_API=3 etcdctl \<br>  --cert /etc/kubernetes/pki/apiserver-etcd-client.crt \<br>  --key /etc/kubernetes/pki/apiserver-etcd-client.key \<br>  --cacert /etc/kubernetes/pki/etcd/ca.crt \<br>get /registry/secrets/&lt;namespace&gt;/&lt;secretName&gt;</pre><p><strong>Search Container Process</strong></p><pre>crictl inspect &lt;containerId&gt; | grep pid<br>strace -p &lt;pid&gt;</pre><p><strong>Trivy</strong></p><pre>trivy image --format &lt;format&gt; --output &lt;output&gt; &lt;imageName&gt;<br>trivy sbom --format &lt;format&gt; /path/to/file</pre><p><strong>grep Multiple Keywords</strong></p><pre>grep -E “&lt;pattren&gt;|&lt;pattern&gt;”</pre><h3>Surprises and Learnings</h3><p>The exam presented a few surprises to me, which might not apply to you.</p><p><strong>No Cilium<br></strong>I prepared extensively for Cilium questions — also due to me having trouble understanding network policy question complexities — so I was surprised that there was no Cilium question in the exam.</p><p><strong>Falco Rule Creation<br></strong>I expected to only modify existing Falco rules, but I had to improvise and write a new one from scratch as default rules didn’t capture the events.</p><p><strong>kube-apiserver Related<br></strong>These questions were easier than anticipated, especially compared to Killer.sh simulations, with volume mounts conveniently pre-configured.</p><p><strong>Docker Questions<br></strong>I hadn’t prepared much for Docker-specific questions but managed to find the necessary properties by using docker --help during the exam.</p><h3>Closing Thoughts</h3><p>CKS was a welcomed challenge, pushing my understanding of Kubernetes security and Linux fundamentals. From almost no Linux knowledge to confidently piping complex commands under pressure, I have gained a lot of confidence working with Linux in the process. I expected a much tougher time after my first mock exam in Killer.sh, on top of all the comments about CKS on the internet. However, lots of practice and completing the entire Killer.sh simulation with 20 minutes to spare rebuilt my confidence for the exam. My key takeaway, as always, is that no reading material replaces actual hands-on practice. I highly recommend trying Killer.sh a couple of days ahead of your exam to test the waters.</p><p>With the CKS behind me, I’m now setting my sights on the KCSA as my final milestone to become a Kubestronaut. Follow me for more content about my Kubestronaut journey, homelab setups, and self-hosting experiences.</p><p>#Kubernetes #CKS #CloudNative #Kubestronaut #CNCF #Certification #CloudComputing #DevOps #Security #KubernetesSecurity #Homelab #selfhost #KubernetesJourney</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ccb897735ed8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Journey to Kubestronaut in 2025 — CKA]]></title>
            <link>https://medium.com/@leehosanganson/journey-to-kubestronaut-in-2025-cka-aa5bfd4fc0c8?source=rss-a92105dcc184------2</link>
            <guid isPermaLink="false">https://medium.com/p/aa5bfd4fc0c8</guid>
            <category><![CDATA[kubernetes]]></category>
            <category><![CDATA[devops]]></category>
            <category><![CDATA[certification]]></category>
            <category><![CDATA[kubestronaut]]></category>
            <category><![CDATA[cka]]></category>
            <dc:creator><![CDATA[Anson Lee]]></dc:creator>
            <pubDate>Sat, 14 Jun 2025 22:36:37 GMT</pubDate>
            <atom:updated>2025-06-14T23:02:52.475Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*el5t0Y-6cW1_nztX7l0ZWw.png" /></figure><h3><strong>Journey to Kubestronaut in 2025 — CKA</strong></h3><p>After passing the KCNA in April and CKAD exam in May, I set my sights on the Certified Kubernetes Administrator (CKA) certification — an essential milestone on my journey to becoming a Kubestronaut, recognised by CNCF for holding all five Kubernetes and cloud native certifications.</p><h3><strong>Preparation Strategy</strong></h3><p>To prepare for the CKA, I focused on three main resources.</p><p><a href="https://learn.kodekloud.com/courses/cka-certification-course-certified-kubernetes-administrator?refererPath=%2Fuser%2Flearning-paths%2Fkubestronaut&amp;refererTitle=Kubestronaut"><strong>KodeKloud Lectures and Exam Simulations</strong></a><strong> </strong>If you already have CKAD under your belt, I recommend diving straight into Lightning Labs and Mock Exams, targeting areas you’re less familiar with.</p><p><a href="https://killer.sh/cka"><strong>Killer.sh Exam Simulator Sessions</strong></a> Registering for CKA gives you two free exam simulations. I took the first to get comfortable with the format and the second just before the real exam. While some questions differ, the problem-solving approach is similar.</p><p><a href="https://www.reddit.com/r/CKAExam/"><strong>Reddit</strong></a> Staying current with the latest exam changes and community advice was important and saved me a couple of questions. Familiarise yourself with recent question types and practice (or at least memorise) steps for common issues, even if you don’t have a cluster or simulator handy.</p><h3>Exam Day: Questions and Challenges</h3><p>CKA uses the same structure as CKAD — each question requires SSH access to a node in a cluster. I got 16 questions in total, and three of them tripped me up.</p><p>Most of them were straight forward questions, such as asking you to create a resource like HPA or PVC, adding sidecar container with logging to the same emptyDir volume and exposing a deployment via the NodePort. Some of them were simple questions but with a twist, such as migrating an Ingress resource to using Gateway while maintaining the existing TLS behaviour. These should be simple and easily translated from CKAD as they can be referred to the Kubernetes Documentation even if you did not prepare them.</p><h4>The 3 questions I failed to finish</h4><blockquote><strong>Helm Template and Installation<br></strong>Creating helm template from a Chart with a specific app version and install it with the template</blockquote><p>I added the required Helm repository and found the Chart, then used helm template --versionto generate the YAML. As someone who usually uses helm install, I didn’t realise I could simply kubectl apply -f the output. This is a neat way to snapshot manifests and will become part of my workflow.</p><blockquote><strong>CNI Installation and Pod CIDR</strong><br>Setting up the CNI by choosing between Flannel and Calico that supports Network Policy with the correct Pod CIDR</blockquote><p>I initially installed Flannel, then realised the question required Network Policy support, so I switched to Calico. However, I struggled to configure the Pod CIDR properly. The correct method is to edit the Calico installation configuration via the custom resource Installation from operator.tigera.io/v1 .</p><blockquote><strong>Cluster Troubleshooting with a non-functional </strong><strong>kube-apiserver</strong><br>Troubleshoot a cluster with non-functional kube-apiserver where only etcd, kube-controller-manager and kubelet are running</blockquote><p>I checked the kubeconfig and endpoints, then used crictl ps to confirm only etcd and kube-controller-manager were running. I tried updating the etcd URL in the kube-apiserver manifest, but ran out of time and failed to finish the problem. In hindsight, I should have checked the kubelet configuration for manifest paths right after confirming the container status from crictl.</p><p>These three questions were in the middle of the exam and cost me significant time. Unlike CKAD, I didn’t have time to review my answers, so I recommend scoring easy questions first and saving troubleshooting for the end. Use the exam’s handy commands to double-check your work during the test.</p><h3>Other Tips for the Hands-on Exam</h3><p><strong>System Preparation</strong><br>CKA uses the same proctor service (PSI) as CKAD. Ensure your machine is compatible and ready. For Mac users, know how to terminate background processes like Photos and Screen Sharing using ps aux | grep and sudo kill -9.</p><p><strong>Clean Terminal<br></strong>Start each new question by exiting the node and clearing your screen before SSHing in. This keeps your terminal history clean.</p><p><strong>Aliases and Autocomplete<br></strong>Use k as an alias for kubectl and leverage tab completion for resource names. This is a huge time-saver, especially for complex commands. For example, when getting the logs of a pod named test-pod-&lt;long-random-string&gt; in namespace testing, type k -n testing logs t, then tab will get you the rest of the command.</p><p><strong>Copy-Paste Caution<br></strong>Be patient when copying and pasting in the exam environment — delays can happen with remote connections.</p><p><strong>Screen Real Estate<br></strong>If using a laptop, resize your terminal and browser to 50% for more workspace.</p><h3>My Next Steps</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FF9h1O7bJrf97MHjbSwSqQ.png" /></figure><p>After exactly 24 hours, I got my results and am glad to say I have passed the exam with a score of 71. CKA was a challenging exam but rewarding experience. Although I might not have as much success as I did in the KCNA and CKAD exams, I have learned a lot from the lectures and my own mistakes during the exam. I am going to start playing with Gateway more in my cluster, replacing the cryptic Ingress annotations. I am now more confident working with Kubernetes control plane components and plan to redeploy my home lab with pure Kubernetes, using GitOps tools like FluxCD or ArgoCD.</p><p>Next up, I will be preparing the Certified Kubernetes Security Specialist (CKS) exam. This is widely considered the most difficult of the five, especially since I have limited experience with Kubernetes security and Linux in general. I am excited about learning how to secure the containerised applications and harden my Kubernetes environment, as one of my goals is to start hosting public-facing services on my own hardware this year.</p><p>I set out to challenge myself in April 2025 to become a Kubestronaut in 6 months. I’d love for you to join me on this journey and learn together!</p><p>#Kubernetes #CKA #CloudNative #Kubestronaut #CNCF #Certification #CloudComputing #DevOps #Homelab #selfhost #KubernetesJourney #KubestronautJourney</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aa5bfd4fc0c8" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>