<?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 Shreya Singh on Medium]]></title>
        <description><![CDATA[Stories by Shreya Singh on Medium]]></description>
        <link>https://medium.com/@siyarajpoot86?source=rss-120b28168b6a------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*rHcZ8NOcqwRbsMnyWXIY-A.jpeg</url>
            <title>Stories by Shreya Singh on Medium</title>
            <link>https://medium.com/@siyarajpoot86?source=rss-120b28168b6a------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 09:00:37 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@siyarajpoot86/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[What Really Happens When You Ask ChatGPT a Question?]]></title>
            <link>https://medium.com/@siyarajpoot86/what-really-happens-when-you-ask-chatgpt-a-question-1fcba1e00141?source=rss-120b28168b6a------2</link>
            <guid isPermaLink="false">https://medium.com/p/1fcba1e00141</guid>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[google-gemini]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[Shreya Singh]]></dc:creator>
            <pubDate>Sat, 20 Jun 2026 15:36:36 GMT</pubDate>
            <atom:updated>2026-06-20T15:36:36.400Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZxjnhW1CYJ0QSr_fzOZkCg.png" /></figure><p>User Prompt<br> ↓<br>Tokenizer<br> ↓<br>Token IDs<br> ↓<br>Embedding Layer<br> ↓<br>Positional Encoding / Positional Embeddings<br> ↓<br>Transformer Blocks (Repeated Many Times)<br> ├─ Multi-Head Self-Attention<br> ├─ Add &amp; Layer Norm<br> ├─ Feed-Forward Network<br> └─ Add &amp; Layer Norm<br> ↓<br>Final Hidden Representation<br> ↓<br>Linear Layer<br> ↓<br>Softmax<br> ↓<br>Next Token Prediction<br> ↓<br>Repeat until response is complete<br> ↓<br>Final Output</p><p>Applications like <strong>ChatGPT</strong>, <strong>Gemini</strong>, and <strong>Claude</strong> are powered by <strong>large language models</strong>.</p><p>At a very high level, when you type something like:</p><blockquote><em>“Write a polite leave application email”</em></blockquote><p>the model does something like this behind the scenes:</p><ul><li>reads your input</li><li>breaks it into smaller pieces</li><li>understands the context of those pieces</li><li>predicts what should come next</li><li>keeps generating the response token by token</li></ul><p>So in a very simplified way, we can say:</p><blockquote><strong><em>An LLM is a highly advanced next-token prediction system trained on huge amounts of language data.</em></strong></blockquote><p>That sounds simple, but there is a lot happening underneath. Let’s break it down.</p><h3>What does GPT mean?</h3><p>One of the most common LLM terms we hear is <strong>GPT</strong>.</p><p><strong>GPT = Generative Pre-trained Transformer</strong></p><p>Let’s understand each word.</p><h3>Generative</h3><p>It can <strong>generate</strong> something new.</p><p>For example:</p><ul><li>text</li><li>emails</li><li>captions</li><li>summaries</li><li>code</li><li>answers</li><li>translations</li></ul><h3>Example</h3><p><strong>Prompt:</strong> <em>Write a caption for a mountain picture</em><br> <strong>Output:</strong> <em>Chasing views, breathing peace, and collecting memories.</em></p><p>The model is generating a fresh response based on patterns it has learned.</p><h3>Pre-trained</h3><p>It is already trained on a very large amount of text before you use it.</p><p>This training data can include:</p><ul><li>books</li><li>websites</li><li>articles</li><li>documentation</li><li>code</li><li>public text datasets</li></ul><p>Because of this pre-training, the model has already seen lots of language patterns like:</p><ul><li>how emails are written</li><li>how questions are answered</li><li>how summaries are structured</li><li>how coding syntax looks</li><li>how words relate to one another</li></ul><p>So when you ask it something, it is not starting from zero. It is using patterns it already learned during training.</p><h3>Transformer</h3><p>This is the deep learning architecture behind modern LLMs.</p><p>The <strong>Transformer</strong> is the core reason LLMs became so powerful. It helps the model understand <strong>context</strong>, <strong>relationships between words</strong>, and <strong>which words matter more in a sentence</strong>.</p><p>In simple words:</p><blockquote><strong><em>Transformer helps the model understand language in context instead of looking at words in isolation.</em></strong></blockquote><h3>how does an LLM actually work?</h3><p>Let’s say the user gives a simple prompt:</p><blockquote><strong><em>“Hi, how are you?”</em></strong></blockquote><p>The model does not understand this sentence like a human does.<br> Instead, it processes the input in a series of steps.</p><p>A simplified flow looks like this:</p><p><strong>Input → Tokenization → Embeddings → Positional Encoding → Transformer Layers → Next Token Prediction → Final Response</strong></p><p>Now let’s understand each part.</p><h3>Tokenization — breaking text into smaller pieces</h3><p>The first step is <strong>tokenization</strong>.</p><p>A computer does not understand raw text like humans do, so the sentence first gets broken into smaller units called <strong>tokens</strong>.</p><p>A token can be:</p><ul><li>a word</li><li>part of a word</li><li>punctuation</li><li>sometimes even a single character</li></ul><h3>Example</h3><p>Sentence:</p><blockquote><strong><em>“Hey there”</em></strong></blockquote><p>Possible tokens:</p><ul><li>Hey</li><li>there</li></ul><p>or depending on the tokenizer:</p><ul><li>He</li><li>y</li><li>there</li></ul><p>Each token is then mapped to a number called a <strong>token ID</strong>.</p><p>For example:</p><ul><li>Hey → 10</li><li>there → 36</li></ul><blockquote><em>Different models may tokenize the same sentence differently because every model has its own tokenizer and vocabulary.</em></blockquote><h3>Why tokenization is needed</h3><p>Because models work with numbers, not words.</p><p>Before an LLM can process language, it must first convert the input into a format the machine can understand. Tokenization is that first bridge between human language and machine processing.</p><h3>Embeddings — converting tokens into meaning</h3><p>Now we have token IDs, but token IDs alone do not carry meaning.<br> The next step is to convert them into <strong>embeddings</strong>.</p><p>An <strong>embedding</strong> is a vector representation of a token that captures its semantic meaning.</p><p>Think of it like this:</p><ul><li>token ID = a label or index</li><li>embedding = a richer mathematical representation of meaning</li></ul><h3>Example</h3><p>Words like:</p><ul><li><strong>dog</strong></li><li><strong>cat</strong></li><li><strong>pet</strong></li><li><strong>animal</strong></li></ul><p>may have embeddings that are closer to each other because they are semantically related.</p><p>Similarly:</p><ul><li><strong>doctor</strong> and <strong>hospital</strong></li><li><strong>teacher</strong> and <strong>school</strong></li><li><strong>king</strong> and <strong>queen</strong></li></ul><p>can also have meaningful relationships in vector space.</p><p>So instead of seeing <strong>dog</strong> as just token number 30, the model sees it as a vector containing learned information about how that word is used in language.</p><h3>Positional Encoding — understanding word order</h3><p>Now let’s take these two sentences:</p><ul><li><strong>The dog chased the cat</strong></li><li><strong>The cat chased the dog</strong></li></ul><p>Both sentences use the same words, but the meaning is completely different.</p><p>So the model must understand not only <strong>what the words are</strong>, but also <strong>where they appear in the sentence</strong>.</p><p>That’s why we need <strong>positional encoding</strong> (or positional embeddings).</p><h3>What positional encoding does</h3><p>It tells the model the position of each token in the sequence.</p><p>This helps the model understand:</p><ul><li>which word came first</li><li>which word came after it</li><li>how the order changes the meaning</li></ul><h3>Example</h3><p>Without positional information, the model might struggle to tell the difference between:</p><ul><li><strong>I love my job</strong></li><li><strong>My job loves me</strong></li></ul><p>Same words, very different meaning.</p><p>So positional encoding helps preserve <strong>sequence order</strong>, which is extremely important in language.</p><h3>Self-Attention — the heart of the Transformer</h3><p>This is one of the most important ideas in LLMs.</p><p>Self-attention helps the model understand:</p><ul><li>which words are related to each other</li><li>which words matter more in a sentence</li><li>how the meaning of one word changes depending on context</li></ul><h3>Why self-attention is needed</h3><p>Take the word <strong>bank</strong> in these two sentences:</p><h3>Example 1</h3><blockquote><strong><em>I sat by the river bank</em></strong></blockquote><h3>Example 2</h3><blockquote><strong><em>I deposited money in the bank</em></strong></blockquote><p>The word <strong>bank</strong> appears in both sentences, but it means different things.</p><p>How does the model know which meaning is correct?</p><p>It looks at the surrounding words:</p><ul><li><strong>river</strong> suggests the side of a river</li><li><strong>money</strong> and <strong>deposited</strong> suggest a financial institution</li></ul><p>This is exactly what self-attention helps with.</p><h3>What self-attention does in simple terms</h3><p>For every word/token, the model asks:</p><ul><li>Which other words should I pay attention to?</li><li>Which words help me understand this token?</li><li>How strongly are these words connected?</li></ul><h3>Example</h3><p>Sentence:</p><blockquote><strong><em>The dog was sleeping because it was tired</em></strong></blockquote><p>The model should understand that <strong>“it”</strong> refers to <strong>“dog”</strong>.</p><p>Self-attention helps the model make that connection.</p><p>This is one of the biggest reasons why transformers are so effective at understanding language context.</p><h3>Multi-Head Attention — understanding multiple relationships at once</h3><p>Language is complex. A sentence can contain:</p><ul><li>grammar</li><li>context clues</li><li>subject-object relationships</li><li>descriptive details</li><li>references like <strong>he</strong>, <strong>she</strong>, <strong>it</strong>, <strong>they</strong></li></ul><p>One attention mechanism may not be enough to capture all of that.<br> So transformers use <strong>multi-head attention</strong>.</p><h3>Simple idea</h3><p>Instead of looking at the sentence from one angle, the model looks at it from <strong>multiple angles at the same time</strong>.</p><h3>Example</h3><p>Sentence:</p><blockquote><strong><em>The brown dog sitting on the train was sleeping quietly</em></strong></blockquote><p>Different attention heads may focus on different relationships such as:</p><ul><li><strong>dog ↔ brown</strong></li><li><strong>dog ↔ sleeping</strong></li><li><strong>dog ↔ train</strong></li><li><strong>quietly ↔ sleeping</strong></li></ul><p>This helps the model build a richer understanding of sentence</p><h3>Transformer Block — where the actual processing happens</h3><p>A Transformer model is made up of multiple repeated layers called <strong>Transformer blocks</strong>.</p><p>Each block helps the model understand the input a little better than the previous one.</p><p>A simplified Transformer block contains:</p><ol><li><strong>Multi-head self-attention</strong></li><li><strong>Add &amp; Layer Normalization</strong></li><li><strong>Feed-forward neural network</strong></li><li><strong>Add &amp; Layer Normalization</strong></li></ol><p>Let’s understand them at a high level.</p><h3>a) Multi-head self-attention</h3><p>This helps the model understand relationships between words and context inside the input.</p><h3>b) Add &amp; Layer Normalization</h3><p>This helps stabilize the model during training and makes learning smoother.</p><blockquote><em>It helps the model train better and preserve useful information across layers</em></blockquote><h3>c) Feed-forward neural network</h3><p>Once the model understands the context using attention, the output goes through a small neural network called a <strong>feed-forward network</strong>.</p><p>This further refines the representation and helps the model learn more complex patterns.</p><p>So a Transformer block is basically doing:</p><ul><li>understand context</li><li>refine the understanding</li><li>pass it to the next layer</li></ul><p>And this happens <strong>many times</strong> across stacked layers.</p><h3>How LLMs are trained</h3><p>Now let’s talk about how the model learns all this in the first place.</p><p>LLMs generally work in <strong>two major phases</strong>:</p><ol><li><strong>Training</strong></li><li><strong>Inference</strong></li></ol><h3>Phase 1: Training</h3><p>Training is the learning phase.</p><p>During training, the model is shown a huge amount of text and learns to predict the next token.</p><h3>Example</h3><p>Suppose the text is:</p><blockquote><strong><em>“The sky is blue.”</em></strong></blockquote><p>The model may be trained like this:</p><ul><li>Input: <strong>The sky is</strong></li><li>Expected next token: <strong>blue</strong></li></ul><p>Another example:</p><ul><li>Input: <strong>I am going to the</strong></li><li>Possible next token: <strong>office</strong>, <strong>market</strong>, <strong>airport</strong>, etc. depending on the training text</li></ul><p>The model keeps seeing such examples again and again and learns language patterns from them.</p><h3>What happens during training?</h3><p>The model:</p><ol><li>takes input text</li><li>predicts the next token</li><li>compares the prediction with the actual correct token</li><li>calculates the error (loss)</li><li>updates its internal weights using <strong>backpropagation</strong></li><li>repeats this process over huge datasets</li></ol><p>This is how it gradually becomes better at:</p><ul><li>grammar</li><li>sentence structure</li><li>context understanding</li><li>writing style patterns</li><li>question-answering patterns</li></ul><h3>A simple training example</h3><p>Suppose the training sentence is:</p><blockquote><strong><em>“Hi, how are you?”</em></strong></blockquote><p>The model may learn it step by step like this:</p><ul><li>Input: <strong>Hi</strong></li><li>predict next token → <strong>,</strong></li><li>Input: <strong>Hi,</strong></li><li>predict next token → <strong>how</strong></li><li>Input: <strong>Hi, how</strong></li><li>predict next token → <strong>are</strong></li><li>Input: <strong>Hi, how are</strong></li><li>predict next token → <strong>you</strong></li><li>Input: <strong>Hi, how are you</strong></li><li>predict next token → <strong>?</strong></li></ul><p>This process is repeated across billions of examples during training.</p><h3>Phase 2: Inference — when we actually use the model</h3><p>Once training is complete, the model is ready for real-world use.<br> This stage is called <strong>inference</strong>.</p><p>Now when a user enters a prompt like:</p><blockquote><strong><em>“Write a polite sick leave email to my manager”</em></strong></blockquote><p>the model is no longer learning.<br> It is simply using what it has already learned during training to generate a response.</p><h3>What happens during inference?</h3><p>The model:</p><ul><li>takes your prompt</li><li>tokenizes it</li><li>converts tokens into embeddings</li><li>processes them through transformer layers</li><li>predicts the next token</li><li>adds that token to the output</li><li>repeats until the response is complete</li></ul><h3>Example</h3><p>Prompt:</p><blockquote><strong><em>“Hi, how are you?”</em></strong></blockquote><p>Possible output:</p><blockquote><strong><em>“I’m doing well, thank you! How can I help you today?”</em></strong></blockquote><p>During inference:</p><ul><li>there is <strong>no back propagation</strong></li><li>the model weights are <strong>not updated</strong></li><li>it only performs forward processing and prediction</li></ul><h3>How does the model decide the next word?</h3><p>This is where the final prediction step happens.</p><p>After processing the prompt through transformer layers, the model produces a score for every possible token in its vocabulary.</p><p>This usually happens in two stages:</p><h3>a) Linear layer</h3><p>The model converts the final hidden representation into scores for all possible next tokens.</p><h3>Example</h3><p>Prompt:</p><blockquote><strong><em>“I am”</em></strong></blockquote><p>Possible raw scores:</p><ul><li>fine → 8.4</li><li>happy → 7.9</li><li>sleepy → 4.2</li><li>banana → 0.3</li></ul><p>These are raw values, not probabilities yet.</p><h3>b) Softmax</h3><p>The <strong>softmax</strong> function converts those scores into probabilities.</p><p>Example:</p><ul><li>fine → 0.48</li><li>happy → 0.32</li><li>sleepy → 0.18</li><li>banana → 0.02</li></ul><p>Now the model can choose the next token.</p><p>So the sentence becomes:</p><blockquote><strong><em>“I am fine”</em></strong></blockquote><p>Then it predicts the next token again, and the process continues until the response is complete.</p><h3>Why does the same prompt sometimes give different answers?</h3><p>Because the model does not always have to pick only the highest-probability token.</p><p>Generation can use different strategies such as:</p><ul><li><strong>Greedy decoding</strong> → always choose the highest probability token</li><li><strong>Temperature</strong> → control randomness and creativity</li><li><strong>Top-k sampling</strong> → choose from the top K most likely options</li><li><strong>Top-p sampling</strong> → choose from a set of tokens whose probabilities cross a threshold</li></ul><h3>Example</h3><p>Prompt</p><blockquote><strong><em>“Write a caption for coffee”</em></strong></blockquote><p>Possible response 1:</p><blockquote><strong><em>Coffee first, everything else later.</em></strong></blockquote><p>Possible response 2:</p><blockquote><strong><em>A warm cup, a calm mind, and the perfect start to the day.</em></strong></blockquote><p>Both can be valid because the model may sample from multiple likely continuations.</p><h3>A full example — from prompt to response</h3><p>Let’s take a very simple prompt:</p><blockquote><strong><em>“Translate Hello to Hindi”</em></strong></blockquote><p>A simplified LLM flow would look like this:</p><h3>Step 1: Input</h3><p>User enters:</p><blockquote><em>Translate Hello to Hindi</em></blockquote><h3>Step 2: Tokenization</h3><p>The sentence is split into tokens.</p><h3>Step 3: Embeddings</h3><p>Each token is converted into vector form.</p><h3>Step 4: Positional information</h3><p>The model understands the order of tokens.</p><h3>Step 5: Self-attention</h3><p>The model understands:</p><ul><li>this is a translation request</li><li>the word to translate is <strong>Hello</strong></li><li>the target language is <strong>Hindi</strong></li></ul><h3>Step 6: Prediction</h3><p>The model predicts the output:</p><blockquote><strong><em>Namaste / नमस्ते</em></strong></blockquote><h3>Does ChatGPT use encoder-decoder architecture?</h3><p>This is a point that often confuses beginners.</p><p>The <strong>original Transformer architecture</strong> introduced two major parts:</p><ul><li><strong>Encoder</strong></li><li><strong>Decoder</strong></li></ul><p>This setup is very useful for tasks like translation.</p><p>However, many GPT-style LLMs are <strong>decoder-only models</strong>.<br> That means they mainly generate text by reading the prompt and predicting the next token one step at a time.</p><p>So the simple takeaway is:</p><ul><li><strong>Original Transformer</strong> → encoder + decoder</li><li><strong>Many modern GPT-style LLMs</strong> → mostly decoder-based for text generation</li></ul><p><strong>Summary :</strong></p><p>LLMs may look magical, but at their core they are powerful systems built to understand context and predict the next token intelligently.</p><p><strong>About the Author</strong><br>Shreya Singh is a DevOps Engineer with experience in Kubernetes, AWS, CI/CD, Terraform, AI,LLM and Cloud Infrastructure. She is passionate about cloud-native technologies and enjoys simplifying complex technical concepts through practical and beginner-friendly content.</p><p><strong>Let’s Connect</strong><br>I’d love to connect with fellow DevOps, Cloud,AI,ML,LLM and Kubernetes enthusiasts. Feel free to reach out, share your feedback, or connect with me on LinkedIn.</p><p>🔗 LinkedIn: <a href="http://www.linkedin.com/in/shreya-singh-6817a0148">www.linkedin.com/in/shreya-singh-6817a0148</a></p><p>Thank you for reading! 🚀</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1fcba1e00141" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Basics of Containers, Virtual Machines, and Physical Servers]]></title>
            <link>https://medium.com/@siyarajpoot86/basics-of-containers-virtual-machines-and-physical-servers-b5a0484c0a84?source=rss-120b28168b6a------2</link>
            <guid isPermaLink="false">https://medium.com/p/b5a0484c0a84</guid>
            <category><![CDATA[docker]]></category>
            <category><![CDATA[containerization]]></category>
            <category><![CDATA[containers]]></category>
            <category><![CDATA[docker-basic]]></category>
            <category><![CDATA[docker-container]]></category>
            <dc:creator><![CDATA[Shreya Singh]]></dc:creator>
            <pubDate>Thu, 18 Jun 2026 03:50:36 GMT</pubDate>
            <atom:updated>2026-06-18T03:52:44.399Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xK1vb5rJlMC83JdN9RQbHQ.png" /></figure><p><strong>Flow:</strong><br>Physical Server (IBM, HP, Laptop, etc.) → Hypervisor → Operating System → Applications<br><strong>What is a Hypervisor?</strong><br>A hypervisor is a virtualization layer that allows multiple virtual machines (VMs) to be created and run on a single physical server.<br>Each virtual machine has its own operating system, and on top of that operating system, applications run independently.<br><strong>Advantages of Virtual Machines</strong><br>Virtual machines are very secure because each VM has its own complete operating system.<br>Complete isolation between VMs.<br>Issues in one VM generally do not impact another VM.<br>Suitable for workloads that require strong security and isolation.</p><p><strong>Why Move Towards Containers?</strong><br>Although virtual machines provide isolation and security, they also have some drawbacks.<br><strong>Major Drawback of Virtual Machines</strong><br>Let’s take an example:<br>Physical Server<br> |<br> Hypervisor<br> |<br> — — — — — — — — — — — -<br>| VM1 | VM2 | VM3 |<br> — — — — — — — — — — — -<br>Assume the server has:<br>100 GB Memory<br>100 CPU Units</p><p><strong>Now suppose</strong>:<br>VM1 is running Application 1.<br>Application 1 only requires 15 GB memory and 10 CPU units even at full load.<br>The remaining resources stay allocated to the VM and are often underutilized.<br>If the application is running at a lower capacity, an even larger amount of CPU and memory remains unused. This leads to inefficient resource utilization.</p><p><strong>Container Architecture</strong><br>Containers can be created in two ways:<br><strong>1. Using a Virtual Machine</strong><br>Physical Server<br> |<br> VM / EC2<br> |<br> Docker Engine<br> |<br> — — — — — — — — <br>| C1 | C2 | C3 |<br> — — — — — — — — <br>2. <strong>Directly on a Physical Server</strong><br>IBM / HP Server<br> |<br>Host Operating System<br> |<br> Docker Engine<br> |<br> — — — — — — — — <br>| C1 | C2 | C3 |<br> — — — — — — — — <br>Most organizations typically use virtual machines or cloud instances (such as EC2) and then run containers on top of them.<br>Large organizations may also have their own data centers and dedicated system engineers to manage the infrastructure.</p><p><strong>What Makes Containers Different?</strong><br>Containers are lightweight in nature because they do not include a complete operating system.<br>Instead, they share the host operating system’s kernel and only package <strong>what is required to run the application.</strong><br>Container =<br>Application<br>+ Libraries<br>+ System Dependencies<br>Unlike virtual machines, containers do not need a full guest operating system.<br><strong>Benefits of Containers</strong><br>Lightweight<br>Faster startup time<br>Better resource utilization<br>Easy to ship and deploy<br>Consistent environment across development, testing, and production</p><p><strong>Snapshot and Backup Comparison</strong><br><strong>Virtual Machine</strong><br>When taking a VM snapshot, the backup size can be large because it includes the complete operating system.<br>For example:<br>1 GB<br>2 GB<br>or even larger depending on the VM size</p><p><strong>Container</strong><br>A container image usually contains only the required application code and dependencies.<br>For example:<br>100 MB<br>500 MB</p><p>This makes containers easier to store, transfer, and deploy.</p><p><strong>Why Containers Are Easy to Deploy</strong><br>Because container images are lightweight, they can be quickly moved between environments.<br>For example, if you want to deploy an application on Kubernetes, it is much easier to deploy a small Docker image than a full virtual machine.<br>Containers package everything needed to run the application, ensuring the same behavior across all environments.</p><p><strong>Docker Basics</strong><br><strong>Flow:</strong><br>Dockerfile<br> ↓<br>Docker Engine<br> ↓<br>Image<br> ↓<br>Container<br><strong>Dockerfile</strong><br>A Dockerfile contains a set of instructions that define how a container image should be built.<br><strong>Docker Engine</strong><br>Docker Engine is the runtime responsible for building images and running containers.<br>If the Docker Engine service stops, the containers managed by it will also stop running.<br><strong>Image</strong><br>An image is a read-only template that contains:<br>Application code<br>Libraries<br>Dependencies<br>Runtime configuration</p><p><strong>Container</strong><br>A container is a running instance of an image.<br><strong>Dockerfile → Build → Image → Run → Container</strong><br>The Dockerfile sends a set of instructions to the Docker Engine, which builds an image. Once the image is executed, it creates a running container.</p><p><strong>Summary</strong><br>Virtual Machines provide complete isolation but consume more resources.<br>Containers are lightweight and share the host operating system.<br>Containers package application code, libraries, and dependencies together.<br>Docker is used to build and run containers.<br>Dockerfile → Image → Container is the basic container creation workflow.<br>Containers are easier to ship, deploy, and scale compared to virtual machines.</p><p><strong>About the Author</strong><br>Shreya Singh is a DevOps Engineer with experience in Kubernetes, AWS, CI/CD, Terraform, and Cloud Infrastructure. She is passionate about cloud-native technologies and enjoys simplifying complex technical concepts through practical and beginner-friendly content.</p><p><strong>Let’s Connect</strong><br>I’d love to connect with fellow DevOps, Cloud, and Kubernetes enthusiasts. Feel free to reach out, share your feedback, or connect with me on LinkedIn.</p><p>🔗 LinkedIn: <a href="http://www.linkedin.com/in/shreya-singh-6817a0148">www.linkedin.com/in/shreya-singh-6817a0148</a></p><p>Thank you for reading! 🚀</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b5a0484c0a84" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Complete Kubernetes Architecture Flow:]]></title>
            <link>https://medium.com/@siyarajpoot86/complete-kubernetes-architecture-flow-9b95a4bb5d29?source=rss-120b28168b6a------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b95a4bb5d29</guid>
            <category><![CDATA[learn-kubernetes-basics]]></category>
            <category><![CDATA[kubernetes-simplified]]></category>
            <category><![CDATA[kubernetes101]]></category>
            <category><![CDATA[k8s-for-beginners]]></category>
            <category><![CDATA[kubernetes-architecture]]></category>
            <dc:creator><![CDATA[Shreya Singh]]></dc:creator>
            <pubDate>Wed, 17 Jun 2026 04:39:06 GMT</pubDate>
            <atom:updated>2026-06-17T05:10:10.533Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5zkgEo1AIWKxEiVKRLyEOQ.png" /><figcaption><strong>High-level Kubernetes architecture showing the interaction between the Control Plane and Worker Nodes.</strong></figcaption></figure><h4><strong>Control Plane<br>+ — — — — — — — — — — — — — — — — — — — — — — — — — +<br>| API Server |<br>| Scheduler |<br>| ETCD |<br>| Controller Manager |<br>| Cloud Controller Manager |<br>+ — — — — — — — — — — — — — — — — — — — — — — — — — +<br> |<br> |<br> v<br> Worker Node<br>+ — — — — — — — — — — — — — — — — — — — — — — — — — +<br>| Kubelet |<br>| Kube-Proxy |<br>| Container Runtime (containerd / CRI-O) |<br>| Pods |<br>+ — — — — — — — — — — — — — — — — — — — — — — — — — +</strong></h4><p>Kubernetes (K8s) is a container orchestration platform used to deploy, manage, and scale containerized applications.</p><p>Kubernetes architecture mainly consists of two major components:</p><p>Control Plane (Master Node) — Responsible for managing the cluster.<br>Data Plane (Worker Node) — Responsible for running application workloads.</p><p>Before understanding Kubernetes architecture, let’s first understand some basic Docker concepts.</p><p>When we execute a Docker container using:</p><p>docker run nginx</p><p>Docker creates and runs a container.</p><p>To run a container, Docker requires a Container Runtime. Historically, Kubernetes supported Docker through Dockershim, which acted as a bridge between Kubernetes and Docker.</p><h4><strong>A simple analogy is Java:</strong></h4><p>To run a Java application, we need a Java Runtime Environment (JRE).<br>Similarly, to run containers, we need a Container Runtime.</p><p>Today, Kubernetes supports multiple container runtimes such as:</p><p>CRI-O<br>containerd</p><p>Kubernetes uses a Container Runtime Interface (CRI), which allows it to work with different container runtimes without being tightly coupled to a specific implementation.</p><p>Kubernetes architecture is divided into:</p><h4><strong>+ — — — — — — — — — — — +<br>| Control Plane |<br>| (Master Node) |<br>+ — — — — — — — — — — — +<br> |<br> |<br>+ — — — — — — — — — — — +<br>| Worker Node |<br>| (Data Plane) |<br>+ — — — — — — — — — — — +</strong></h4><p>Let’s start with the Worker Node.</p><h4><strong>Data Plane (Worker Node)</strong></h4><p>The Data Plane is responsible for running application workloads inside Pods.</p><h4><strong>1. Kubelet</strong></h4><p>Kubelet is an agent that runs on every worker node.</p><h4><strong>Responsibilities:</strong></h4><p>Creates and manages Pods.<br>Communicates with the Control Plane.<br>Ensures containers are running as expected.<br>Reports node and pod status back to the Control Plane.</p><p>For example, if a Pod crashes, Kubelet informs the Control Plane, which can take corrective action and restart the Pod.</p><h4><strong>2. Kube-Proxy</strong></h4><p>Kube-Proxy is responsible for networking within Kubernetes.</p><h4><strong>Responsibilities:</strong></h4><p>Provides Service networking.<br>Enables communication between Pods.<br>Performs load balancing across Pods.<br>Manages network rules using Linux iptables or IPVS.</p><p>Without Kube-Proxy, services would not be able to route traffic to application Pods.</p><h4><strong>3. Container Runtime</strong></h4><p>The Container Runtime is responsible for running containers on worker nodes.</p><p>Common runtimes include:</p><p>containerd<br>CRI-O</p><p>The runtime performs tasks such as:</p><p>Pulling container images.<br>Starting containers.<br>Stopping containers.<br>Managing container lifecycle.</p><h4><strong>Control Plane (Master Node)</strong></h4><p>The Control Plane manages the entire Kubernetes cluster.</p><p>It decides where applications run, monitors cluster health, and maintains the desired state.<strong><br>1. API Server</strong></p><p>The API Server is the heart of Kubernetes.</p><p>Every request to Kubernetes goes through the API Server.</p><p>Examples<strong>:</strong></p><p>kubectl get pods<br>kubectl apply -f deployment.yaml</p><h4><strong>Responsibilities:</strong></h4><p>Exposes Kubernetes APIs.<br>Validates requests.<br>Communicates with other control plane components.<br>Acts as the entry point to the cluster.</p><p>When a Pod creation request is received, the API Server processes it and coordinates with other components.</p><h4><strong>2. Scheduler</strong></h4><p>The Scheduler is responsible for deciding where a Pod should run.</p><h4><strong>Responsibilities:</strong></h4><p>Selects the most suitable worker node.<br>Considers CPU and memory availability.<br>Considers node affinity, taints, and tolerations.<br>Optimizes workload placement.</p><p>Note: The API Server receives the request, but the Scheduler makes the actual placement decision.</p><h4><strong>3. ETCD</strong></h4><p>ETCD is the most critical component of Kubernetes.</p><p>It is a distributed key-value store that stores the entire cluster state.</p><p>Examples of data stored in ETCD:</p><p>Pod information<br>Node information<br>Deployments<br>Services<br>Secrets<br>ConfigMaps</p><h4><strong>Key points:</strong></h4><p>Source of truth for the cluster.<br>Without ETCD, the cluster state cannot be recovered.<br>ETCD backups are essential for disaster recovery.</p><h4><strong>4. Controller Manager</strong></h4><p>The Controller Manager runs various Kubernetes controllers.</p><h4><strong>Responsibilities:</strong></h4><p>Ensures desired state matches actual state.<br>Monitors cluster resources continuously.</p><p>Example:</p><p>If a Deployment requires 3 replicas and one Pod crashes:</p><p>Desired Replicas = 3<br>Current Replicas = 2</p><p>The Controller Manager detects the difference and creates a new Pod to maintain the desired state.</p><p>Common controllers include:</p><p>ReplicaSet Controller<br>Deployment Controller<br>Node Controller<br>Endpoint Controller</p><h4><strong>5. Cloud Controller Manager (CCM)</strong></h4><p>The Cloud Controller Manager integrates Kubernetes with cloud providers.</p><p>Examples:</p><p>Amazon EKS<br>Azure AKS<br>Google GKE</p><p>Flow:</p><p>User<br> ↓<br>Kubernetes<br> ↓<br>Cloud Controller Manager<br> ↓<br>AWS / Azure / GCP</p><h4><strong>Responsibilities:</strong></h4><p>Creates Load Balancers.<br>Provisions cloud storage.<br>Manages cloud networking resources.<br>Integrates cloud-specific services with Kubernetes.</p><p>For example, when a Service of type LoadBalancer is created in EKS:</p><p><strong>type: </strong>LoadBalancer</p><p>The Cloud Controller Manager communicates with AWS and automatically provisions an AWS Load Balancer.</p><p>Without CCM, Kubernetes cannot create or manage cloud-native resources.</p><h4><strong>Conclusion</strong></h4><p>Kubernetes may look complex, but once you understand the role of each component, the architecture becomes much easier to follow. I hope this beginner-friendly overview helped you build a strong foundation in Kubernetes.</p><h4><strong>About the Author</strong></h4><p>Shreya Singh is a DevOps Engineer with experience in Kubernetes, AWS, CI/CD, Terraform, and Cloud Infrastructure. She is passionate about cloud-native technologies and enjoys simplifying complex technical concepts through practical and beginner-friendly content.</p><h4><strong>Let’s Connect</strong></h4><p>I’d love to connect with fellow DevOps, Cloud, and Kubernetes enthusiasts. Feel free to reach out, share your feedback, or connect with me on LinkedIn.</p><p><strong>🔗 LinkedIn: </strong>www.linkedin.com/in/shreya-singh-6817a0148</p><p>Thank you for reading! 🚀</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9b95a4bb5d29" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>