<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Mahesh</title>
    <description>The latest articles on DEV Community by Mahesh (@maheshbandaru_ba8cc2).</description>
    <link>https://dev.to/maheshbandaru_ba8cc2</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3843436%2F1914e5c5-a0e3-45cf-8611-a00da78d8092.png</url>
      <title>DEV Community: Mahesh</title>
      <link>https://dev.to/maheshbandaru_ba8cc2</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/maheshbandaru_ba8cc2"/>
    <language>en</language>
    <item>
      <title>8 system design flows interviewers ask you to draw, and the step they always poke at</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Sun, 15 Nov 2026 05:49:29 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/8-system-design-flows-interviewers-ask-you-to-draw-and-the-step-they-always-poke-at-mc</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/8-system-design-flows-interviewers-ask-you-to-draw-and-the-step-they-always-poke-at-mc</guid>
      <description>&lt;p&gt;Most system design rounds I've seen go wrong in the same place. Not the big boxes. The candidate draws a load balancer, a cache and a database, the interviewer nods, and then asks one small question about a single arrow: "what happens here when this fails?" That arrow is where the round is decided.&lt;/p&gt;

&lt;p&gt;So here are eight flows that come up again and again, what to draw for each, and the one step interviewers almost always poke at. If you can walk through that step out loud, you're in good shape.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. A request through a load balancer
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Draw:&lt;/strong&gt; client, DNS, load balancer, three app servers, a shared database.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The step they poke at:&lt;/strong&gt; what the load balancer does when one server stops answering. Talk about health checks (how often, what counts as "unhealthy"), and what happens to requests already in flight on the dying server. Bonus: sticky sessions, and why they make scaling harder.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Cache-aside reads
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Draw:&lt;/strong&gt; app checks the cache, misses, reads the database, writes the result back into the cache with a TTL.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The step they poke at:&lt;/strong&gt; the write path. When the data changes, do you update the cache or delete the key? Deleting is usually safer, because two concurrent writers can leave a stale value if both update. Then expect: "what if 10,000 requests miss the same hot key at once?" That's a cache stampede. Mention request coalescing or a short lock around the refill.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. A rate limiter
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Draw:&lt;/strong&gt; requests hitting a counter per user or per API key before they reach the service.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The step they poke at:&lt;/strong&gt; the algorithm and where the counter lives. Token bucket allows short bursts; a fixed window lets someone send double the limit across a window boundary; a sliding window fixes that at more cost. With many app servers the counter has to live somewhere shared, like Redis, and the increment has to be atomic.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. A message queue between two services
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Draw:&lt;/strong&gt; producer, queue, consumer group, a dead-letter queue.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The step they poke at:&lt;/strong&gt; delivery guarantees. Most queues give you at-least-once delivery, so the consumer will sometimes see the same message twice. Say how the consumer stays correct anyway: an idempotency key, or a "processed messages" table checked before the side effect. Then: what lands in the dead-letter queue and who looks at it.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. Database read replicas
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Draw:&lt;/strong&gt; one primary taking writes, two replicas taking reads, replication arrows between them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The step they poke at:&lt;/strong&gt; replication lag. A user updates their profile, refreshes, and sees the old value because the read went to a replica a second behind. Fixes worth naming: read-your-own-writes by sending that user's reads to the primary for a short time, or tracking the write's position and waiting for a replica to catch up.&lt;/p&gt;

&lt;h2&gt;
  
  
  6. Consistent hashing for a sharded cache
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Draw:&lt;/strong&gt; a ring, nodes placed on it, keys mapped to the next node clockwise.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The step they poke at:&lt;/strong&gt; adding or removing a node. With plain &lt;code&gt;hash(key) % N&lt;/code&gt;, changing N moves almost every key. On the ring, only the keys between the new node and its neighbour move. Then they ask about uneven load, and the answer is virtual nodes: each physical server gets many positions on the ring.&lt;/p&gt;

&lt;h2&gt;
  
  
  7. A CDN in front of static assets
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Draw:&lt;/strong&gt; user, nearest edge location, origin server, cache headers on the response.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The step they poke at:&lt;/strong&gt; invalidation. You shipped a new &lt;code&gt;app.js&lt;/code&gt; and some users still get the old one. The clean answer is versioned file names (&lt;code&gt;app.3f9c.js&lt;/code&gt;) with long cache lifetimes, so a deploy changes the URL instead of waiting for caches to expire. Purging the CDN is the fallback, not the plan.&lt;/p&gt;

&lt;h2&gt;
  
  
  8. A payment that must not run twice
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Draw:&lt;/strong&gt; client, payment service, payment provider, a payments table.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The step they poke at:&lt;/strong&gt; the retry. The client times out and sends the same payment again. Without protection, the customer is charged twice. The standard answer is an idempotency key sent by the client and stored with a unique constraint, so the second attempt returns the first result instead of creating a new charge. This one shows up a lot because it ties together retries, databases and real money.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I'd practise these
&lt;/h2&gt;

&lt;p&gt;Reading about them isn't the hard part. Explaining them while someone interrupts you is. Three things that helped me:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Say each flow out loud in two minutes&lt;/strong&gt;, then again with one failure injected ("the cache node just died"). If you can't explain the failure case simply, that's the gap.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Watch the flow move before you draw it.&lt;/strong&gt; Seeing a request actually travel through the boxes makes the failure step easier to reason about than a static diagram does.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Do one timed mock&lt;/strong&gt; where someone, or something, keeps asking "and then what?"&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Disclosure: I work on &lt;a href="https://lastroundai.com" rel="noopener noreferrer"&gt;LastRound AI&lt;/a&gt;. Its &lt;a href="https://lastroundai.com/products/concept-explainers" rel="noopener noreferrer"&gt;concept explainers&lt;/a&gt; are animated, step-through versions of flows like these (238 of them across system design, DevOps, cloud, networking and coding patterns), and they're free on every plan. There's also a &lt;a href="https://lastroundai.com/use-cases/system-design-interviews" rel="noopener noreferrer"&gt;system design mock interview&lt;/a&gt; that does the "and then what?" part. Use whatever works for you; the eight flows above are the part that matters.&lt;/p&gt;

&lt;p&gt;Which step do you get pushed on most? I'm curious whether it's the same arrows for everyone.&lt;/p&gt;

</description>
      <category>systemdesign</category>
      <category>interview</category>
      <category>career</category>
      <category>beginners</category>
    </item>
    <item>
      <title>Final Round AI vs LockedIn AI vs Parakeet AI: what their refund policies actually say (Nov 2026)</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Fri, 06 Nov 2026 06:37:02 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/final-round-ai-vs-lockedin-ai-vs-parakeet-ai-what-their-refund-policies-actually-say-nov-2026-463o</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/final-round-ai-vs-lockedin-ai-vs-parakeet-ai-what-their-refund-policies-actually-say-nov-2026-463o</guid>
      <description>&lt;p&gt;Most people compare interview copilots on features. Latency, stealth, which models they use. Fair enough. But the complaints I keep reading on Trustpilot and G2 are almost never about features. They are about money: renewals people forgot about, refund requests that went nowhere, and paying for a tool that never opened during the actual interview.&lt;/p&gt;

&lt;p&gt;So I read the refund policies of three popular tools, Final Round AI, LockedIn AI and Parakeet AI, on November 6, 2026. Here is what each one actually says, in plain words, plus a short checklist to run before you pay for any of them.&lt;/p&gt;

&lt;p&gt;Disclosure up front: I work on LastRound AI, which is in the same category. I've put our own policy at the end, held to the same standard.&lt;/p&gt;

&lt;h2&gt;
  
  
  Final Round AI: 3 days, and "substantial use" decides it
&lt;/h2&gt;

&lt;p&gt;Final Round AI's policy (last updated October 8, 2026) offers a &lt;strong&gt;3-day (72-hour) money-back guarantee&lt;/strong&gt; on initial and renewal charges, for monthly, quarterly and annual plans.&lt;/p&gt;

&lt;p&gt;The catch is the definition of use. You are not eligible if the current billing period shows "substantial use", which they define as &lt;strong&gt;at least 2 interview sessions, each lasting more than 30 minutes&lt;/strong&gt;. The guarantee is also &lt;strong&gt;once per customer&lt;/strong&gt;, across all accounts and plans.&lt;/p&gt;

&lt;p&gt;Two other things worth knowing from their pricing page: there is &lt;strong&gt;no free trial&lt;/strong&gt;, and live sessions need an active Pro subscription. So the 72-hour window is effectively your trial.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What this means for you:&lt;/strong&gt; if you buy it the night before an interview and use it for two long rounds, assume that money is spent. If you want to test it, do it inside the first 72 hours with short sessions.&lt;/p&gt;

&lt;h2&gt;
  
  
  LockedIn AI: monthly and quarterly are non-refundable once started
&lt;/h2&gt;

&lt;p&gt;LockedIn AI's policy is the strictest of the three. It says its subscriptions are digital and non-returnable, and:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Monthly and quarterly:&lt;/strong&gt; non-refundable once started. Cancelling keeps access until the cycle ends, with no refunds or prorated credits.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Annual:&lt;/strong&gt; eligible for a &lt;strong&gt;partial&lt;/strong&gt; refund if cancelled within &lt;strong&gt;7 days&lt;/strong&gt; of purchase.&lt;/li&gt;
&lt;li&gt;Purchases through the Apple App Store follow Apple's refund rules instead.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;They point people to their free trial before buying. Also note the pricing page runs heavy promo pricing (a 50% off code was showing when I checked), so compare the price you will pay at renewal, not just the first charge.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What this means for you:&lt;/strong&gt; treat a monthly or quarterly LockedIn purchase as final. Use the free trial properly first.&lt;/p&gt;

&lt;h2&gt;
  
  
  Parakeet AI: 7 days if unused, 28 days for an unused renewal
&lt;/h2&gt;

&lt;p&gt;Parakeet AI has the most detailed policy of the three:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;7-day&lt;/strong&gt; refund window for accidental or unused purchases.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;28 days&lt;/strong&gt; for a subscription renewal you haven't used since it renewed. This one is genuinely useful if you forgot to cancel.&lt;/li&gt;
&lt;li&gt;"Unused" is strict: a call session, a mock interview, a resume download and so on all count as use. There are small exceptions, for example a plan session that ran under 5 minutes.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;What this means for you:&lt;/strong&gt; the renewal rule is the friendliest of the three if you forget to cancel. But one real session counts as use, so you can't try it and then ask for your money back.&lt;/p&gt;

&lt;h2&gt;
  
  
  The pattern across all three
&lt;/h2&gt;

&lt;p&gt;Put side by side, the rules point the same way:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Refund window&lt;/th&gt;
&lt;th&gt;What kills eligibility&lt;/th&gt;
&lt;th&gt;Free way to try first&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Final Round AI&lt;/td&gt;
&lt;td&gt;72 hours, once per customer&lt;/td&gt;
&lt;td&gt;2+ sessions over 30 min&lt;/td&gt;
&lt;td&gt;No free trial&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;LockedIn AI&lt;/td&gt;
&lt;td&gt;None on monthly/quarterly; partial on annual within 7 days&lt;/td&gt;
&lt;td&gt;Starting the subscription&lt;/td&gt;
&lt;td&gt;Free trial&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Parakeet AI&lt;/td&gt;
&lt;td&gt;7 days; 28 days for an unused renewal&lt;/td&gt;
&lt;td&gt;Any session, mock or download&lt;/td&gt;
&lt;td&gt;Free tier / first mock free&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;None of them refund you because an interview went badly. That's normal for usage-based AI tools, since every live minute costs them transcription and model calls. It does mean the free trial matters more than the feature list.&lt;/p&gt;

&lt;h2&gt;
  
  
  A 5-minute checklist before you pay for any interview copilot
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Find the refund page and read the "use" definition.&lt;/strong&gt; That one sentence decides whether you can get your money back.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Test on a real call setup before the real interview.&lt;/strong&gt; Use a call with a friend on Zoom, Meet or Teams with screen share on. Check audio capture, answer delay and whether the window shows up in the shared screen.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Check the renewal price, not the promo price.&lt;/strong&gt; Several tools show a discounted first period.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Put a calendar reminder 2 days before renewal.&lt;/strong&gt; Most "I was charged again" complaints are forgotten renewals.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Check your interview's rules.&lt;/strong&gt; Some companies ban AI assistance outright. A tool can't help you if using it gets your candidacy cancelled.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  Our policy, held to the same standard
&lt;/h2&gt;

&lt;p&gt;LastRound AI's refund policy (updated September 13, 2026):&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;New accounts get &lt;strong&gt;free credits&lt;/strong&gt;, so you can test on a real call before paying.&lt;/li&gt;
&lt;li&gt;Refunds are possible within &lt;strong&gt;7 days&lt;/strong&gt; of payment, &lt;strong&gt;only if nothing was used&lt;/strong&gt; since that payment (no credits and no copilot or mock minutes).&lt;/li&gt;
&lt;li&gt;Double charges, or charges after a cancellation we already processed, are refunded.&lt;/li&gt;
&lt;li&gt;You can cancel any time. Access continues to the end of the period.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So we're in the same family as Parakeet: strict about use, but with a free way to try first. If you want the full side-by-side on features and price, we keep a comparison here: &lt;a href="https://lastroundai.com/compare/final-round-ai" rel="noopener noreferrer"&gt;Final Round AI vs LastRound AI&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Whichever tool you pick, spend your free minutes on a mock call with screen share on, not on the real interview. That's the cheapest way to avoid a refund fight.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Policies quoted from each company's own refund page, read on November 6, 2026. They change often, so check the live page before you buy.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>career</category>
      <category>interview</category>
      <category>ai</category>
      <category>productivity</category>
    </item>
    <item>
      <title>The Interview Follow-Up Developers Underprepare For: “What Would You Do Differently?”</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Tue, 08 Sep 2026 03:55:21 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/the-interview-follow-up-developers-underprepare-for-what-would-you-do-differently-35ej</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/the-interview-follow-up-developers-underprepare-for-what-would-you-do-differently-35ej</guid>
      <description>&lt;p&gt;A technical interview can be going well right up until the moment the interviewer asks a quieter question:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;“What would you do differently?”&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Many candidates treat this like a trap. They either say they would change nothing, or reach for a safe line about being too detail-oriented. Neither tells an interviewer much.&lt;/p&gt;

&lt;p&gt;What they are usually looking for is judgment. Can you look back at a real decision, understand the trade-offs, and improve your approach without becoming defensive?&lt;/p&gt;

&lt;p&gt;That is a useful signal for an engineering team. Requirements move, customers surprise you, and a choice that made sense on Monday can look different after the release is live. Teams need people who can adapt thoughtfully.&lt;/p&gt;

&lt;h2&gt;
  
  
  Pick a decision with real trade-offs
&lt;/h2&gt;

&lt;p&gt;You do not need a dramatic failure story. Choose a decision where two reasonable paths were available.&lt;/p&gt;

&lt;p&gt;Maybe you shipped a small version of a feature instead of waiting for every edge case. Maybe you selected a familiar library because the deadline was tight. Maybe you focused on the main user flow and postponed an admin workflow that later caused extra support work.&lt;/p&gt;

&lt;p&gt;The best answers are specific. Instead of saying, “I would communicate better,” describe the exact gap:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;“I kept the product and engineering teams aligned, but I did not bring support into the release review early enough. After launch, they flagged a few confusing parts of the workflow that we could have caught before shipping.”&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That says more than a polished weakness ever will. It shows you understand the work itself.&lt;/p&gt;

&lt;h2&gt;
  
  
  Give the original choice its context
&lt;/h2&gt;

&lt;p&gt;Candidates often jump straight to what went wrong. That can make a thoughtful decision sound careless.&lt;/p&gt;

&lt;p&gt;Explain why the choice made sense at the time:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;“We had two weeks before a customer commitment and a small team. I limited the first release to the highest-volume workflow so we could learn from real usage rather than wait for a complete solution.”&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;You are not trying to prove that every earlier decision was correct. You are showing that you made a call using the information, time, and resources you had.&lt;/p&gt;

&lt;p&gt;That is what good work looks like in practice. Perfect data and unlimited time are not part of most projects.&lt;/p&gt;

&lt;h2&gt;
  
  
  Describe the adjustment, not a personality flaw
&lt;/h2&gt;

&lt;p&gt;Avoid vague self-criticism. “I am a perfectionist” and “I care too much” are familiar enough that they usually disappear from the interviewer’s memory immediately.&lt;/p&gt;

&lt;p&gt;Talk about the small operating habit you would change next time.&lt;/p&gt;

&lt;p&gt;For example:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;“Next time, I would schedule a 30-minute review with support or a customer-facing teammate before launch. It would not expand the scope, but it would help us spot language and workflow issues that are hard to see from inside the product team.”&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;This answer is practical. It does not pretend the original approach was a disaster. It shows how you would make the next version stronger.&lt;/p&gt;

&lt;h2&gt;
  
  
  Finish with proof the lesson stuck
&lt;/h2&gt;

&lt;p&gt;The last line is where a decent answer becomes convincing. Explain what changed in your work afterward:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;“Since then, I have included a customer-facing review in early release planning. It has improved adoption of first releases without adding much delay.”&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;That is much stronger than “I learned a lot.” It connects reflection to a visible habit.&lt;/p&gt;

&lt;h2&gt;
  
  
  A four-part framework to practise
&lt;/h2&gt;

&lt;p&gt;Before an interview, write down four short points for one real project:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Decision:&lt;/strong&gt; What did you choose?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Context:&lt;/strong&gt; Why did it make sense then?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Adjustment:&lt;/strong&gt; What would you change now?&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Lesson:&lt;/strong&gt; What do you do differently because of it?&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Do not memorise a speech. Say the four points out loud in a few different ways. The goal is to understand the story well enough to handle a follow-up such as, “Why did you not involve support earlier?” or “How would you know the extra review was worth the time?”&lt;/p&gt;

&lt;p&gt;A script usually breaks under those questions. A real example, understood clearly, does not.&lt;/p&gt;

&lt;p&gt;“What would you do differently?” is not an invitation to tear down your past work. It is a chance to show that you can learn, make decisions under constraints, and carry better habits into the next project.&lt;/p&gt;

&lt;p&gt;Try the framework with one genuine example this week. You will quickly spot where your story is vague, where the original context is missing, and where your lesson needs to become more concrete.&lt;/p&gt;

&lt;p&gt;Those are useful things to discover before the interview, not during it.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;About LastRound AI: practical tools and practice for people preparing for real interviews. Learn more at &lt;a href="https://lastroundai.com" rel="noopener noreferrer"&gt;lastroundai.com&lt;/a&gt;.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>career</category>
      <category>productivity</category>
      <category>learning</category>
    </item>
    <item>
      <title>Real-Time Conversation Assistants: A Practical Workflow for Sales, Meetings, and Interviews</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Fri, 14 Aug 2026 06:24:42 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/real-time-conversation-assistants-a-practical-workflow-for-sales-meetings-and-interviews-3kan</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/real-time-conversation-assistants-a-practical-workflow-for-sales-meetings-and-interviews-3kan</guid>
      <description>&lt;h1&gt;
  
  
  Real-Time Conversation Assistants: A Practical Workflow for Sales, Meetings, and Interviews
&lt;/h1&gt;

&lt;p&gt;The most useful assistance during a live conversation is not a transcript that arrives later. It is a small amount of relevant context at the moment a decision, question, or follow-up appears. That is true in a technical interview, a customer call, and an internal meeting.&lt;/p&gt;

&lt;p&gt;A real-time work assistant should help with three things: understanding what is happening, deciding what deserves attention, and turning the conversation into a useful next step.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Start with context, not a generic answer
&lt;/h2&gt;

&lt;p&gt;An assistant is only as useful as the context it can use. Before a conversation begins, provide the role, customer background, agenda, or project goals. During the conversation, the assistant should connect new questions to that context instead of returning a broad answer that could apply anywhere.&lt;/p&gt;

&lt;p&gt;For an interview, that might mean relating a follow-up question to the candidate's resume and the target role. For a sales call, it might mean matching an objection to the product's capabilities and the buyer's stated priorities. For a meeting, it might mean distinguishing a decision from an open question.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Keep the assistance private and lightweight
&lt;/h2&gt;

&lt;p&gt;The best live assistant should not create another participant in the call. A private desktop overlay can provide help without joining the meeting, appearing in the guest list, or interrupting the conversation. The user remains responsible for what they say; the assistant simply makes relevant information easier to access.&lt;/p&gt;

&lt;p&gt;This design also makes the workflow flexible. It can work alongside video calls, phone conversations, screen sharing, or an in-person discussion rather than depending on one meeting platform.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Separate live help from the follow-up
&lt;/h2&gt;

&lt;p&gt;Live answers and post-conversation notes serve different purposes. During the conversation, the interface should stay focused: a suggested response, a useful follow-up question, a reminder, or a quick explanation. Afterward, the same context can become a recap, action-item list, or follow-up message.&lt;/p&gt;

&lt;p&gt;Keeping these modes separate avoids flooding the user with a complete transcript when they need one clear next step.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Use assistance to improve judgment, not replace it
&lt;/h2&gt;

&lt;p&gt;A strong workflow treats AI suggestions as prompts for better thinking. The user should be able to inspect the context, adjust the answer, and ignore a suggestion that does not fit. This matters especially in interviews and sales conversations, where authenticity and accurate claims are more valuable than a perfectly polished but irrelevant response.&lt;/p&gt;

&lt;p&gt;For teams exploring this workflow, &lt;a href="https://craqly.com" rel="noopener noreferrer"&gt;Craqly&lt;/a&gt; is a real-time work assistant for sales, meetings, and interviews. It listens to the conversation, understands context, and provides answers, notes, and next steps through a private desktop copilot without joining meetings.&lt;/p&gt;

&lt;p&gt;Disclosure: I work on Craqly.&lt;/p&gt;

&lt;h2&gt;
  
  
  A simple operating checklist
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Add the relevant context before the conversation starts.&lt;/li&gt;
&lt;li&gt;Keep live suggestions short enough to use immediately.&lt;/li&gt;
&lt;li&gt;Ask for clarification when the context is ambiguous.&lt;/li&gt;
&lt;li&gt;Capture decisions and action items separately from live assistance.&lt;/li&gt;
&lt;li&gt;Review important answers before saying them aloud or sending them.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The goal is not to automate the conversation. It is to reduce the friction around remembering, thinking, and following through while the conversation is still moving.&lt;/p&gt;

</description>
      <category>career</category>
    </item>
    <item>
      <title>We built Craqly: real-time AI help during the call, not after it</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Fri, 07 Aug 2026 10:37:09 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/we-built-craqly-real-time-ai-help-duai-productivity-showdev-careerring-the-call-not-after-it-4k53</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/we-built-craqly-real-time-ai-help-duai-productivity-showdev-careerring-the-call-not-after-it-4k53</guid>
      <description>&lt;p&gt;Every meeting tool we tried solved the wrong moment. You finish a sales call, and twenty minutes later a beautiful summary lands in your inbox telling you exactly which objection you fumbled. Thanks. The call is over.&lt;/p&gt;

&lt;p&gt;So we built &lt;a href="https://craqly.com" rel="noopener noreferrer"&gt;Craqly&lt;/a&gt; to work during the conversation instead.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it does
&lt;/h2&gt;

&lt;p&gt;Craqly is a desktop app for Mac and Windows that listens to whatever call is on your screen (Zoom, Google Meet, Teams, WebEx, Slack huddles, even a podcast in a browser tab) and surfaces short, context-aware suggestions in real time. Latency is 1-2 seconds for text answers.&lt;/p&gt;

&lt;p&gt;Three things make it different from the recorder-bots:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;No bot joins the call.&lt;/strong&gt; It reads system audio directly, so there is no awkward "Craqly Notetaker has joined the meeting" moment and no participant for procurement to object to.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The overlay is invisible to screen-share.&lt;/strong&gt; The window is excluded from screen capture at the OS level (CGWindowSharingNone on macOS, WDA_EXCLUDEFROMCAPTURE on Windows), so when you share your screen the other side sees nothing.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It answers during the call.&lt;/strong&gt; Sales mode spots objections and pulls the counter from your battle cards. Interview mode anchors answers in your resume. Meeting mode captures action items with owner and deadline while they are being said.&lt;/li&gt;
&lt;/ol&gt;

&lt;h2&gt;
  
  
  The hard parts
&lt;/h2&gt;

&lt;p&gt;The honest engineering story: real-time is unforgiving. A summary bot can take 30 seconds; nobody notices. We had to get from audio to transcript to LLM response fast enough to be useful mid-sentence, which meant streaming STT with interim results, aggressive prompt assembly, and one nasty echo-dedup problem when the mic picks up the speaker output and every line shows up twice attributed to two different people.&lt;/p&gt;

&lt;p&gt;Transcription runs in 18 languages with mid-conversation switching, which sounds like a checkbox feature until a Hyderabad standup flips between English and Telugu four times in one sentence.&lt;/p&gt;

&lt;h2&gt;
  
  
  Pricing
&lt;/h2&gt;

&lt;p&gt;There is a free tier (20 minutes/month, no credit card) so you can test it on a real call. Paid plans start at $19/month.&lt;/p&gt;

&lt;p&gt;If you try it on a live call, I would genuinely like to hear where it falls short. The failure cases teach us more than the praise does.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>showdev</category>
      <category>career</category>
    </item>
    <item>
      <title>Twelve AWS networking questions, and what the interviewer is actually listening for</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Sun, 02 Aug 2026 04:30:00 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/twelve-aws-networking-questions-and-what-the-interviewer-is-actually-listening-for-3506</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/twelve-aws-networking-questions-and-what-the-interviewer-is-actually-listening-for-3506</guid>
      <description>&lt;p&gt;I have asked some version of "walk me through why this EC2 instance can't reach the internet" in about forty interviews. Maybe six people answered it well.&lt;/p&gt;

&lt;p&gt;Not because it's hard. Because most candidates answer the question they wish had been asked. They recite what a NAT gateway is. The question was diagnostic, and diagnosis is a different skill from definition.&lt;/p&gt;

&lt;p&gt;These are the twelve that come up most, with the reasoning behind each. Worth knowing before you read on: across 1,393 interview sessions configured on LastRound AI between January 2025 and July 2026, DevOps Engineering accounted for 464 of them. That's more than backend, frontend and full-stack combined, so the competition for these roles is real and the bar has moved.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Why can't you resize a VPC CIDR block after creation?
&lt;/h2&gt;

&lt;p&gt;You can't shrink it. You can add secondary CIDR blocks, which is what most people mean when they say "resize."&lt;/p&gt;

&lt;p&gt;The reasoning being tested is whether you plan address space before you need it. A /16 gives you 65,536 addresses and costs nothing extra, so the correct habit is to allocate generously and subnet conservatively. I have watched a team spend two weeks migrating workloads because someone picked a /24 for a "small internal service" that grew.&lt;/p&gt;

&lt;p&gt;Weak answer: "You just add another CIDR." True but incomplete. Strong answer names the constraint that secondary CIDRs can't overlap with existing ones or with any peered VPC, which is exactly the trap that bites later.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. What actually makes a subnet public?
&lt;/h2&gt;

&lt;p&gt;A route to an internet gateway in its route table. Nothing else.&lt;/p&gt;

&lt;p&gt;Not the name. Not the "auto-assign public IP" setting, though that matters for whether instances get an address. A subnet called "public-subnet-1a" with no IGW route is a private subnet with a misleading name, and I have debugged that exact situation twice.&lt;/p&gt;

&lt;p&gt;This question sorts people who learned AWS from a diagram from people who have configured it. The diagram shows a box labelled public. The console shows a route table.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Security group versus network ACL: when does the difference matter?
&lt;/h2&gt;

&lt;p&gt;Security groups are stateful and evaluate at the instance level. NACLs are stateless and evaluate at the subnet boundary.&lt;/p&gt;

&lt;p&gt;Stateless is the word that does the work. If you allow inbound 443 on a NACL and forget the outbound ephemeral port range, the request arrives and the response never leaves. That failure looks like a timeout, not a rejection, which is why it eats hours.&lt;/p&gt;

&lt;p&gt;Security groups also deny by default with no explicit deny rule available. NACLs process rules in numbered order and stop at the first match, so a broad deny at rule 100 makes rule 200 irrelevant. If a candidate can explain why you would ever use a NACL given security groups exist, they have usually operated something real. The honest answer is: rarely, mostly for blanket IP blocks at the subnet edge.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Walk me through why an EC2 instance in a private subnet can't reach the internet
&lt;/h2&gt;

&lt;p&gt;This is the diagnostic one. Answer it as a chain, out loud, in order.&lt;/p&gt;

&lt;p&gt;Route table has a 0.0.0.0/0 entry pointing at a NAT gateway. NAT gateway is in a public subnet, not the private one. That public subnet routes to an internet gateway. NAT gateway has an elastic IP. Security group allows the outbound traffic. NACL allows both the outbound request and the inbound ephemeral response. DNS resolution works, which is a separate failure that presents identically.&lt;/p&gt;

&lt;p&gt;Say the chain. Candidates who name one cause and stop are guessing. Candidates who walk the path are debugging, and the interviewer can hear the difference immediately.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. NAT gateway or NAT instance?
&lt;/h2&gt;

&lt;p&gt;NAT gateway for almost everything. Managed, scales automatically, no patching.&lt;/p&gt;

&lt;p&gt;The interesting part is cost, and this is where I'll disagree with common advice. People treat NAT gateway pricing as a fixed cost of doing business. It isn't. You pay hourly plus per gigabyte processed, and the per-gigabyte charge is what surprises teams pulling container images or shipping logs out through it.&lt;/p&gt;

&lt;p&gt;A candidate who mentions VPC endpoints as the fix for S3 and ECR traffic has thought about the bill. That's a genuine signal, because it means they have seen one.&lt;/p&gt;

&lt;h2&gt;
  
  
  6. Gateway endpoints versus interface endpoints
&lt;/h2&gt;

&lt;p&gt;Gateway endpoints cover S3 and DynamoDB, attach to a route table, and are free. Interface endpoints use PrivateLink, create an ENI in your subnet, cover most other services, and cost per hour plus per gigabyte.&lt;/p&gt;

&lt;p&gt;The follow-up that catches people: gateway endpoints don't work across VPC peering or from on-premises over Direct Connect, because they're route-table constructs local to the VPC. Interface endpoints do work across those paths. If someone confidently says "just use an endpoint" for a hybrid setup, that's the gap.&lt;/p&gt;

&lt;h2&gt;
  
  
  7. VPC peering or Transit Gateway?
&lt;/h2&gt;

&lt;p&gt;Peering is point-to-point and non-transitive. Three VPCs that all need to talk means three peering connections. Ten VPCs means forty-five.&lt;/p&gt;

&lt;p&gt;That n(n-1)/2 growth is the whole argument for Transit Gateway, which gives you a hub and routes between attachments. The tradeoff is cost per attachment plus per gigabyte, so for two or three VPCs peering is usually cheaper and simpler.&lt;/p&gt;

&lt;p&gt;Non-transitive is the word that matters. If A peers with B and B peers with C, A cannot reach C. Candidates repeat "non-transitive" as a memorised phrase without being able to say what it means operationally. Ask them to draw it and you find out fast.&lt;/p&gt;

&lt;h2&gt;
  
  
  8. What happens when route table entries overlap?
&lt;/h2&gt;

&lt;p&gt;Longest prefix match wins. A route for 10.0.1.0/24 beats a route for 10.0.0.0/16 for traffic to 10.0.1.5, regardless of the order they appear.&lt;/p&gt;

&lt;p&gt;People assume rules evaluate top to bottom because NACLs do. They don't. Route tables are prefix-matched, NACLs are ordered, security groups are unordered and permissive-only. Three different evaluation models in one networking stack is genuinely confusing, and saying so out loud is fine. It signals you understand the models rather than having memorised one.&lt;/p&gt;

&lt;h2&gt;
  
  
  9. Why does DNS sometimes fail inside a VPC even when routing is correct?
&lt;/h2&gt;

&lt;p&gt;Two VPC attributes control this: enableDnsSupport and enableDnsHostnames. Support enables the Route 53 Resolver at the .2 address in your CIDR. Hostnames controls whether instances get public DNS names.&lt;/p&gt;

&lt;p&gt;Turn off support and resolution breaks entirely. Leave hostnames off and private hosted zones behave in ways that look like a routing problem. Both default differently depending on whether the VPC was created by the wizard or by Terraform, which is why this fails on the second environment and not the first.&lt;/p&gt;

&lt;p&gt;That last detail is the kind of thing you only know from having hit it.&lt;/p&gt;

&lt;h2&gt;
  
  
  10. How does cross-AZ traffic affect cost and design?
&lt;/h2&gt;

&lt;p&gt;Traffic between availability zones is charged in both directions. Within an AZ it's free.&lt;/p&gt;

&lt;p&gt;The design implication is that chatty services which must live in the same failure domain sometimes belong in the same AZ, and you accept the availability tradeoff deliberately rather than by accident. Most reference architectures spread everything across three AZs and never mention the data transfer line on the bill.&lt;/p&gt;

&lt;p&gt;I don't have a clean rule for where the line sits. It depends on how chatty the service is and what your availability target actually is, and anyone offering a universal answer is selling something.&lt;/p&gt;

&lt;h2&gt;
  
  
  11. Why do EKS clusters run out of pod capacity before they run out of CPU?
&lt;/h2&gt;

&lt;p&gt;The AWS VPC CNI assigns every pod a real IP from the subnet, through ENIs attached to the node. Each instance type supports a fixed number of ENIs and a fixed number of IPs per ENI, so pod density is bounded by instance type, not just by resources.&lt;/p&gt;

&lt;p&gt;A t3.medium supports 3 ENIs at 6 IPs each, minus one for the node. That's 17 pods, no matter how much idle CPU you have.&lt;/p&gt;

&lt;p&gt;The second half of the answer is subnet sizing. A /24 gives you 251 usable addresses, and a cluster that scales to 200 pods will exhaust it. Prefix delegation raises the density ceiling but consumes address space faster. This is the question where cloud and Kubernetes knowledge meet, and it separates people who have run a cluster from people who have deployed to one.&lt;/p&gt;

&lt;h2&gt;
  
  
  12. Can you peer two VPCs with overlapping CIDR blocks?
&lt;/h2&gt;

&lt;p&gt;No. The request fails at creation.&lt;/p&gt;

&lt;p&gt;Short answer, but the follow-up is the real question: what do you do about it? Options are re-addressing one VPC, which is disruptive, or putting a NAT layer between them, or using PrivateLink to expose the specific service rather than connecting the networks at all.&lt;/p&gt;

&lt;p&gt;PrivateLink is the answer that impresses, because it reframes the problem. You usually don't need two networks joined. You need one service reachable from the other side.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to practise these
&lt;/h2&gt;

&lt;p&gt;Reading these is not the same as being able to say them. The chain in question four takes about ninety seconds out loud and most people discover they can't sequence it under mild pressure, even knowing every component.&lt;/p&gt;

&lt;p&gt;Say them standing up, timed, and record it. What rehearsal never fixed for me was the follow-up arriving mid-sentence — "and why doesn't that work across peering?" — about three seconds after I'd finished sounding confident. That specific gap is the one &lt;a href="https://lastroundai.com/products/ai-interview-copilot" rel="noopener noreferrer"&gt;Live Interview Copilot&lt;/a&gt; is built for, since it reads the question as it's asked rather than after you have already frozen. The rehearsal still does the heavy lifting, though. Nothing bails you out of not knowing the material.&lt;/p&gt;

&lt;p&gt;For the source material, the &lt;a href="https://docs.aws.amazon.com/vpc/latest/userguide/what-is-amazon-vpc.html" rel="noopener noreferrer"&gt;AWS VPC User Guide&lt;/a&gt; is the only reference worth reading properly.&lt;/p&gt;

&lt;p&gt;Which of these have you been asked? I'm curious whether the diagnostic ones show up as often elsewhere as they do in the rounds I've sat in.&lt;/p&gt;

</description>
      <category>aws</category>
      <category>devops</category>
      <category>kubernetes</category>
      <category>interview</category>
    </item>
    <item>
      <title>Nobody wants to be your practice interviewer, and that's fine</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Sat, 01 Aug 2026 10:40:15 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/nobody-wants-to-be-your-practice-interviewer-and-thats-fine-cbn</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/nobody-wants-to-be-your-practice-interviewer-and-thats-fine-cbn</guid>
      <description>&lt;p&gt;I asked four people to run a mock interview with me before a system design round. Two said yes. One rescheduled twice and then stopped replying. The fourth ran it, was extremely encouraging, and told me everything was great. I bombed the real one nine days later.&lt;/p&gt;

&lt;p&gt;The encouraging friend was the worst of the four outcomes. Being told you're fine is more damaging than getting no practice at all, because it removes the thing that would have made you prepare harder.&lt;/p&gt;

&lt;p&gt;So I started practising alone. Badly, at first.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why solo practice usually fails
&lt;/h2&gt;

&lt;p&gt;The failure mode isn't lack of discipline. It's grading.&lt;/p&gt;

&lt;p&gt;When you answer a question in your head, you grade the version you meant. When you say it out loud to another person, they grade the version that came out of your mouth. Those are different answers. The gap between them is roughly the whole difficulty of interviewing.&lt;/p&gt;

&lt;p&gt;My first month of solo prep was reading questions and thinking "yeah, I know that one." Every single time, I did know it. And every single time, I'd have taken forty seconds too long to get to the point, buried the answer in setup, or forgotten to say the one thing the interviewer was listening for.&lt;/p&gt;

&lt;p&gt;Knowing something and being able to say it under mild social pressure in under two minutes are unrelated skills. Reading questions trains the first one.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the session data says about how people actually practise
&lt;/h2&gt;

&lt;p&gt;I work on an interview tool, so I went and looked at our own numbers to check whether other people practise the way I did.&lt;/p&gt;

&lt;p&gt;Across 1,393 interview sessions configured on LastRound AI between January 2025 and July 2026, the difficulty selector broke down like this: 1,374 people chose intermediate, 16 chose expert, and 3 chose beginner.&lt;/p&gt;

&lt;p&gt;98.6% picked the middle option.&lt;/p&gt;

&lt;p&gt;Two readings of that. The charitable one is that intermediate genuinely fits most candidates. The honest one is that the difficulty selector is close to useless, and people click the default because they don't know how to calibrate themselves. I lean towards the second, and it's an awkward thing to say about your own product.&lt;/p&gt;

&lt;p&gt;The same data shows a near-even split between live sessions and mock practice: 721 live, 672 mock. I found that genuinely surprising. I expected practice to dominate, on the theory that people prepare more than they interview. They don't, or at least these ones don't.&lt;/p&gt;

&lt;p&gt;I don't know whether that holds outside our user base. 1,393 sessions from one product skews towards people who already went looking for interview tooling, which is not a random sample of job seekers.&lt;/p&gt;

&lt;h2&gt;
  
  
  The method that fixed it for me
&lt;/h2&gt;

&lt;p&gt;Four rules. Two of them are annoying and that's the point.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Say it out loud, standing up.&lt;/strong&gt; Not muttered. Full voice, as if someone's listening. You will feel ridiculous. You will also immediately notice that your answer has no shape, because you can hear yourself running out of road at the ninety-second mark.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Record it and listen back the next day.&lt;/strong&gt; Not the same day. The next day, when you've lost the memory of what you meant and can only hear what you said. This is the single highest-value thing on the list and almost nobody does it, because listening to your own recorded voice is genuinely unpleasant.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Start the timer before you start thinking.&lt;/strong&gt; Real interviews don't give you a settling-in period. If you need thirty seconds of silence to structure an answer, that silence is part of your answer and it should be in the recording.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Write the grade before you listen.&lt;/strong&gt; Predict your score, then listen, then score it properly. The gap between predicted and actual is your calibration error, and it's the number that actually improves over weeks.&lt;/p&gt;

&lt;p&gt;That fourth one came from a friend who coaches debate, and it's the one that changed things fastest. My predictions were consistently a full point higher than reality for about three weeks. Then they converged. Once they converged I could trust my own judgement mid-interview, which is what you actually need on the day.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where a tool helps and where it doesn't
&lt;/h2&gt;

&lt;p&gt;The recording-and-reviewing loop is tedious enough that I stopped doing it properly after about two weeks. That's the real problem with solo practice: it works, and you won't keep it up.&lt;/p&gt;

&lt;p&gt;What kept me honest was moving the loop into something that ran it for me. I used &lt;a href="https://lastroundai.com/products/mock-interviews" rel="noopener noreferrer"&gt;AI Mock Interview&lt;/a&gt; for the voice sessions, mostly because it removed the two steps I kept skipping: it asks the follow-up I'd never have asked myself, and it scores the answer before I get a chance to grade myself generously.&lt;/p&gt;

&lt;p&gt;What it doesn't replace is a senior person in your actual field telling you that your architecture answer was fine but you buried the tradeoff. If you have access to that person, use them. Most people asking this question don't, which is why they're asking.&lt;/p&gt;

&lt;p&gt;The other limitation: automated feedback is good at structure and pace, weaker on domain judgement. It'll catch that you rambled. It's less reliable on whether your specific caching decision was defensible for that specific system. Treat the score as a signal about delivery, not a verdict on your engineering.&lt;/p&gt;

&lt;h2&gt;
  
  
  A practice week that actually fits around a job
&lt;/h2&gt;

&lt;p&gt;Three sessions a week, forty minutes each. Not daily. Daily practice sounds committed and collapses by day five.&lt;/p&gt;

&lt;p&gt;Monday: two behavioural questions, recorded, reviewed Tuesday. Wednesday: one technical question at full length. Saturday: review both recordings back to back and write down one thing to change.&lt;/p&gt;

&lt;p&gt;One thing. Not a list. Lists don't survive contact with a real interview, and the point of the week is to change one habit at a time.&lt;/p&gt;

&lt;p&gt;If you want to sanity-check which technologies are worth the technical slot, the &lt;a href="https://survey.stackoverflow.co/" rel="noopener noreferrer"&gt;Stack Overflow Developer Survey&lt;/a&gt; is a reasonable read on what's actually in use.&lt;/p&gt;

&lt;p&gt;What I'd genuinely like to know from anyone else doing this: does the predict-then-listen thing work for you, or did you find a faster way to calibrate? Three weeks felt slow.&lt;/p&gt;

</description>
      <category>career</category>
      <category>interview</category>
      <category>beginners</category>
    </item>
    <item>
      <title>The LeetCode Patterns That Actually Show Up in Interviews (I Stopped Grinding Random Problems and Started Tracking These)</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Thu, 23 Jul 2026 03:10:47 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/the-leetcode-patterns-that-actually-show-up-in-interviews-i-stopped-grinding-random-problems-and-4610</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/the-leetcode-patterns-that-actually-show-up-in-interviews-i-stopped-grinding-random-problems-and-4610</guid>
      <description>&lt;p&gt;Somewhere around problem 200 on my LeetCode grind, I noticed I was solving the same problem with different variable names and calling it progress. A recruiter screen for a mid-level backend role asked me to find the longest substring without repeating characters, and I froze for a solid twenty seconds before my brain connected it to a sliding window problem I'd done three weeks earlier under a completely different title. I got there, but slower than someone who'd actually understood the pattern would have. That was the moment I stopped counting solved problems and started counting patterns understood, which turned out to be a much smaller and more useful number.&lt;/p&gt;

&lt;p&gt;There are maybe a dozen patterns that cover the overwhelming majority of what gets asked in a 45-minute technical screen. Here's what I actually track when I'm prepping someone for interviews now, roughly in the order they show up.&lt;/p&gt;

&lt;p&gt;Two pointers. Anything involving a sorted array, a palindrome check, or "find a pair that sums to X" is reaching for two pointers before it's reaching for a nested loop. The tell in an interview is when someone starts writing an O(n²) brute force and the interviewer asks "can we do better than that" — that's your cue the sorted-array assumption matters. A weak candidate brute-forces it and stops there. A strong one says out loud "since it's sorted, I can move pointers from both ends" before writing a line of code.&lt;/p&gt;

&lt;p&gt;Sliding window. This is the one I froze on. Any "longest/shortest substring or subarray that satisfies some condition" problem — no repeating characters, sum equals K, at most two distinct characters — is sliding window. The pattern is: expand the right pointer until the window breaks the condition, then shrink from the left until it's valid again. People who've only memorized the code for one specific sliding window problem get stuck the moment the condition changes shape, because they never internalized that the window's job is just "track the largest valid range as you scan once."&lt;/p&gt;

&lt;p&gt;Fast and slow pointers (Floyd's cycle detection). Linked list cycle problems, finding the middle of a list in one pass, detecting a repeated number in an array treated as an implicit linked list. This one's less common than the first two but it's a great filter question because there's basically no way to stumble into the right answer without having seen the trick before — either you know slow moves one step and fast moves two, or you don't.&lt;/p&gt;

&lt;p&gt;Merge intervals. Meeting room scheduling, overlapping intervals, inserting a new interval into a sorted list. The wrong answer people give is trying to handle overlaps with a bunch of if/else branches instead of the actual insight, which is: sort by start time first, then a single pass where you either extend the current interval or start a new one. I've watched candidates write forty lines of conditional logic to avoid a five-line loop because they didn't sort first.&lt;/p&gt;

&lt;p&gt;BFS/DFS on trees and graphs. Level-order traversal, number of islands, course schedule (topological sort in disguise). The distinction that separates people who've internalized this from people who've memorized a template: can they explain why BFS is the right call for "shortest path in an unweighted graph" and DFS is the right call for "does a path exist at all" — not just recite that both exist.&lt;/p&gt;

&lt;p&gt;Binary search on the answer, not just on a sorted array. This is the pattern that trips up otherwise-strong candidates the most, because it doesn't look like binary search at first. "Minimum days to make M bouquets," "capacity to ship packages within D days" — these aren't searching an array, they're binary-searching a range of possible answers and checking feasibility at each midpoint. If someone can explain "the answer space is monotonic, so I can binary search it" without prompting, that's a strong signal regardless of whether they nail the exact implementation.&lt;/p&gt;

&lt;p&gt;Backtracking. Permutations, combinations, subsets, N-Queens. The interview differentiator here isn't whether someone can write the recursive skeleton — most people who've prepped at all can. It's whether they can identify pruning opportunities out loud, because an unpruned backtracking solution that technically works but explores way more of the search space than necessary is a common way to lose points even on a "correct" answer.&lt;/p&gt;

&lt;p&gt;Dynamic programming, but specifically: can you find the recurrence. This is the one I'll push back on common advice for. Most prep guides tell you to memorize the top twenty DP problems. I think that's close to useless, because DP problems don't repeat verbatim in interviews the way two-pointer problems sort of do — what transfers is the skill of writing out the recurrence relation in plain English before touching code. "The answer at position i depends on the answer at i-1 and i-2" is a sentence, not a LeetCode problem, and once you can write that sentence for a new problem you've never seen, you've actually learned DP instead of memorized nine specific solutions.&lt;/p&gt;

&lt;p&gt;Top-K / heap problems. Kth largest element, top K frequent elements, merge K sorted lists. The pattern-recognition trigger is any phrase with "top K" or "Kth" in it — that's a heap of size K, full stop, and the only real decision left is min-heap or max-heap depending on which end you're filtering from.&lt;/p&gt;

&lt;p&gt;Union-Find (disjoint set). Number of connected components, redundant connection, accounts merge. Less common than the others but worth having ready because it's a clean, fast answer to a class of graph-connectivity problems that would otherwise take a much messier DFS-based solution, and using it unprompted signals real breadth.&lt;/p&gt;

&lt;p&gt;What actually changed my hit rate wasn't grinding more problems — it was, for each new problem, writing one sentence before coding: "this is [pattern] because [specific reason]." If I couldn't fill in the reason, I hadn't understood the problem yet, I'd just pattern-matched the surface shape of it, which falls apart the second an interviewer tweaks the constraints. That single habit did more for me than the difference between 150 problems solved and 400.&lt;/p&gt;

&lt;p&gt;One thing I'd push back on from most interview prep advice: doing untimed practice at home doesn't prepare you for the actual experience of explaining your thinking out loud while someone watches you type, which is a genuinely different skill from just solving the problem correctly on your own. I started running mock sessions with LastRound AI's live interview copilot specifically to practice narrating my pattern-recognition process under the same time pressure and social pressure as a real screen — &lt;a href="https://lastroundai.com/products/ai-interview-copilot" rel="noopener noreferrer"&gt;https://lastroundai.com/products/ai-interview-copilot&lt;/a&gt; — because silently solving a sliding window problem on your couch and explaining your reasoning to a stranger who's grading you in real time are not the same skill, and the second one is the one that actually gets you the offer.&lt;/p&gt;

&lt;p&gt;The honest caveat: none of this replaces actually understanding the underlying data structures. Pattern recognition gets you to the right approach faster, but if you don't know how a heap or a hash map actually works under the hood, you'll pattern-match correctly and then fumble the implementation, which is its own kind of frustrating to watch as an interviewer.&lt;/p&gt;

</description>
      <category>interview</category>
      <category>algorithms</category>
      <category>careertips</category>
      <category>coding</category>
    </item>
    <item>
      <title>9 Kubernetes Interview Questions That Reveal Who's Actually Run a Cluster in Production</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Fri, 17 Jul 2026 03:54:29 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/9-kubernetes-interview-questions-that-reveal-whos-actually-run-a-cluster-in-production-kp8</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/9-kubernetes-interview-questions-that-reveal-whos-actually-run-a-cluster-in-production-kp8</guid>
      <description>&lt;p&gt;I once watched a candidate with "3 years Kubernetes" on his resume describe a Deployment and a ReplicaSet as basically interchangeable terms for "the thing that keeps my pods running." Technically not wrong, exactly — a Deployment does manage a ReplicaSet under the hood — but the follow-up question is the one that actually separates people: "if you edit a ReplicaSet directly instead of the Deployment, what happens on the next rollout?" He didn't know, because he'd never had a reason to touch anything below the Deployment layer, which is fine until the cluster does something unexpected and you're the one debugging it at 11pm. Here are nine questions that come up constantly, and what they're actually testing for.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Deployment, ReplicaSet, Pod — walk me through the hierarchy, and what breaks if you edit a ReplicaSet directly?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A Deployment owns one or more ReplicaSets (one active, plus old ones kept around for rollback), and each ReplicaSet owns a set of Pods matching its selector. Edit a ReplicaSet's spec directly and the change works — until the Deployment controller reconciles state on its next pass and overwrites your edit back to whatever the Deployment's spec says, because the Deployment is the source of truth, not the ReplicaSet. I ask this because it's the fastest way to tell if someone's ever actually debugged a "why did my manual fix disappear" incident, which is a rite of passage for anyone who's spent real time in a cluster.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Liveness, readiness, and startup probes — what failure mode does each one actually protect against?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Liveness probes answer "is this container stuck and needs a restart" — fail it enough times and kubelet kills and restarts the container. Readiness probes answer "should this pod currently receive traffic" — fail it and the pod gets pulled from the Service's endpoints without being restarted, which matters enormously during startup or a slow dependency check. Startup probes exist specifically for slow-booting apps, disabling liveness/readiness checks until the app reports it's actually initialized, so a legitimately slow Spring Boot app doesn't get killed in a restart loop before it ever finishes starting. The mistake I see constantly: teams configure only a liveness probe, pointed at the same endpoint as readiness, and then wonder why a pod that's merely slow to reconnect to its database gets killed and restarted instead of just quietly pulled from rotation until it recovers.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. What actually happens to a request when you &lt;code&gt;kubectl scale&lt;/code&gt; a Deployment down, mid-traffic?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The Deployment controller deletes the excess pods, but there's a window most people don't account for: the pod gets a SIGTERM, and unless your app handles graceful shutdown, in-flight requests get dropped mid-response while the pod is still technically in the Service's endpoint list for a few hundred milliseconds to a couple seconds, depending on kube-proxy's sync interval and how fast your app stops accepting new connections. &lt;code&gt;preStop&lt;/code&gt; hooks and a reasonable &lt;code&gt;terminationGracePeriodSeconds&lt;/code&gt; exist for exactly this — sleep for a couple seconds in preStop so kube-proxy has time to remove the pod from rotation before the app actually stops accepting connections. I've seen production incidents traced entirely to a missing preStop hook during routine autoscaling, not even a deploy — just normal scale-down dropping a percentage of requests nobody had budgeted for.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Resource requests vs. limits — what's the actual failure mode when you get each one wrong?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Requests are what the scheduler uses to decide which node has room for your pod — set them too low and you get overcommitted nodes where pods fight for actual CPU/memory at runtime. Limits are hard ceilings enforced by the kubelet — hit a memory limit and the container gets OOMKilled outright; hit a CPU limit and the container gets throttled (CFS throttling), which doesn't kill anything but silently makes your app slower in a way that's brutal to diagnose because nothing crashes, nothing logs an error, it just... gets slow under load. The wrong answer treats requests and limits as roughly the same knob turned to different values. The real signal is knowing that a memory limit without a request just means your pod schedules wherever, potentially crowds out other workloads, and then gets killed anyway once it actually hits that ceiling.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. How does traffic from a Service actually reach a Pod? Trace it end to end.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;A Service gets a stable virtual IP and DNS name; kube-proxy watches the Service and its matching Endpoints (or EndpointSlices on newer clusters) and programs iptables or IPVS rules on every node to load-balance traffic across the healthy pod IPs behind it. There's no single "load balancer process" routing packets in the way people picture it — it's distributed rule-programming on every node, which is also why a stale kube-proxy or a slow endpoint sync can mean traffic keeps hitting a pod that was just marked not-ready, for a brief but real window. A candidate who says "the Service load-balances to the pods" isn't wrong, but someone who can explain that it's iptables/IPVS rules being reconciled from EndpointSlices, not a live proxy process making per-request decisions, has clearly had to debug a networking issue that "should have just worked."&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;6. ConfigMap vs. Secret — what's the actual security difference, and where does the common assumption go wrong?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Both store key-value data and both get mounted the same way, as environment variables or files — that similarity is exactly what causes the mistake. Secrets are base64-encoded, not encrypted, by default; base64 is trivially reversible, and without encryption-at-rest enabled on etcd (which is a separate, explicit cluster configuration, not automatic), a Secret sitting in etcd is functionally as exposed as a ConfigMap to anyone with etcd access. I've interviewed people who confidently said "Secrets are encrypted, ConfigMaps aren't" as if that's a default guarantee, and it just isn't unless someone configured encryption providers on the API server. RBAC restricting who can read Secret objects is doing more real protection than the base64 encoding is.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;7. Rolling update strategy — what do &lt;code&gt;maxSurge&lt;/code&gt; and &lt;code&gt;maxUnavailable&lt;/code&gt; actually control, and what breaks if you get them wrong?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;code&gt;maxUnavailable&lt;/code&gt; caps how many pods from the old ReplicaSet can be down at once during the rollout; &lt;code&gt;maxSurge&lt;/code&gt; caps how many extra pods above the desired count can exist temporarily while new ones spin up alongside old ones. Set &lt;code&gt;maxUnavailable&lt;/code&gt; too high (or &lt;code&gt;maxSurge&lt;/code&gt; to 0 with &lt;code&gt;maxUnavailable&lt;/code&gt; high) on a service without enough replica headroom, and a rolling update can genuinely drop capacity below what your traffic needs mid-rollout — not a bug, just math nobody checked. The follow-up I actually care about: what happens if your new pods pass readiness checks but are still functionally broken — bad config, wrong image tag pointed at a broken build. The rollout completes successfully by Kubernetes' definition of success, because readiness passed, while your actual service is degraded. That gap between "rollout succeeded" and "service is healthy" is why real teams pair rollouts with actual traffic-based monitoring, not just &lt;code&gt;kubectl rollout status&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;8. Horizontal Pod Autoscaler — what's the actual mechanism, and why does it sometimes thrash?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;HPA polls the metrics API (metrics-server by default, or a custom/external metrics adapter for anything beyond CPU/memory) on a fixed interval, compares current usage against the target you set, and computes a desired replica count from that ratio — it's not instant, and it's not smart about trends, just a periodic proportional calculation. Thrashing happens when a workload's actual load oscillates faster than the HPA's stabilization window accounts for — scale up, load drops because you just added capacity, scale back down, load spikes again from the reduced capacity, repeat. The fix most people don't know by name: &lt;code&gt;behavior.scaleDown.stabilizationWindowSeconds&lt;/code&gt;, which makes the HPA wait and look at the max recommendation over a rolling window before actually scaling down, specifically to dampen this exact oscillation.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;9. StatefulSet vs. Deployment — when do you genuinely need one, and what's the tell that someone's using it out of habit rather than necessity?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;StatefulSets give you stable, predictable pod names and persistent volume claims that survive pod rescheduling — pod-0 is always pod-0, with the same storage, even after a restart. That matters for genuinely stateful workloads: databases, anything doing leader election by ordinal identity, Kafka brokers. The tell I look for: someone reaching for a StatefulSet for a stateless API "just to be safe," which buys you nothing but slower, ordered rollouts (StatefulSets update pods one at a time by default, not in parallel) and stable network identity you don't need. If your pods don't care about their own name or need their own dedicated persistent volume, a Deployment is simpler and rolls out faster, full stop.&lt;/p&gt;

&lt;p&gt;None of these are hard once you've actually run a cluster past the tutorial stage and had something quietly break in a way &lt;code&gt;kubectl get pods&lt;/code&gt; didn't explain. They're hard live, reasoning through it out loud while someone's watching for whether you actually understand the mechanism or just memorized the vocabulary. I've been running through scenario questions like these with LastRound AI's AI Mock Interview before Kubernetes-heavy interviews — saying "the rollout succeeded doesn't mean the service is healthy, those are different signals" out loud, under a bit of pressure, is a different skill than nodding along when someone else says it: &lt;a href="https://lastroundai.com/products/mock-interviews" rel="noopener noreferrer"&gt;https://lastroundai.com/products/mock-interviews&lt;/a&gt;&lt;/p&gt;

</description>
      <category>kubernetes</category>
      <category>devops</category>
      <category>interview</category>
      <category>cloudnative</category>
    </item>
    <item>
      <title>8 Azure Interview Questions That Reveal Who's Actually Run a Subscription</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Tue, 14 Jul 2026 02:51:54 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/8-azure-interview-questions-that-reveal-whos-actually-run-a-subscription-13fm</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/8-azure-interview-questions-that-reveal-whos-actually-run-a-subscription-13fm</guid>
      <description>&lt;p&gt;I once sat across from a candidate who'd listed "Azure" four separate times on his resume — AZ-104, three years of "cloud infrastructure," the works. I asked him to walk me through what happens when you assign a Contributor role at the subscription level versus the resource group level, and he went quiet for a solid ten seconds before guessing "they're basically the same thing." They are not the same thing, and that gap between "I've used Azure" and "I understand how Azure's control plane is actually structured" is exactly what these questions are built to expose. Here are the eight that come up most, and what interviewers are really listening for when they ask them.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. Walk me through the management group → subscription → resource group → resource hierarchy. Where does access actually get inherited from?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Management groups sit above subscriptions and let you apply RBAC and policy across multiple subscriptions at once — useful once an org has more than a handful of subscriptions and doesn't want to configure the same policy nine times. Access assigned at any level flows down: a Contributor role granted at the management group level applies to every subscription, resource group, and resource beneath it, with no way to "block" inheritance from above (you can only add more restrictive policy, not remove an inherited role assignment). The wrong answer treats resource groups as the top of the hierarchy, which is fine until an org actually needs cross-subscription governance and nobody on the team understands why a policy applied "everywhere" broke a dev subscription nobody meant to touch.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. RBAC scope — when do you assign a role at the subscription level versus the resource group level, and why does it matter?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;The technically correct answer is "wherever the access needs to apply," but the real signal is whether someone defaults to the narrowest scope that works. Assigning Contributor at the subscription level because it's "easier" than scoping to individual resource groups is the single most common over-permissioning mistake I see in Azure environments — it means every team, every environment, every resource group under that subscription inherits access nobody asked for. I ask this because I've had to clean up exactly this: a contractor granted Contributor at the subscription level for one project who could, technically, delete production storage accounts belonging to a completely unrelated team.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;3. NSG vs Azure Firewall — what's the actual difference, and when do you need both?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Network Security Groups are stateful, layer-3/4 packet filters that live on a subnet or NIC and evaluate rules by priority — cheap, fast, and the right tool for basic allow/deny between subnets. Azure Firewall is a managed, stateful firewall-as-a-service that operates at the VNet level, supports FQDN-based filtering, threat intelligence feeds, and centralized logging across a hub-and-spoke topology. The wrong answer treats them as redundant. The real pattern in most enterprise Azure environments is NSGs for cheap subnet-level segmentation plus a centralized Azure Firewall in the hub VNet for egress control and FQDN filtering that NSGs simply can't do — NSGs don't understand DNS names, only IP ranges and ports.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Storage account redundancy — LRS, ZRS, GRS, RA-GRS. What is each one actually protecting against, and what's the tradeoff no one mentions?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;LRS replicates three copies within a single datacenter — protects against disk/rack failure, not datacenter failure. ZRS spreads those three copies across availability zones in the same region — survives a datacenter outage, not a regional one. GRS adds asynchronous replication to a paired region hundreds of miles away — survives a full regional outage, but the replication lag means a failover can lose the last few minutes of writes. RA-GRS is GRS plus a read-only endpoint in the secondary region you can query directly. The tradeoff nobody mentions in the marketing material: geo-redundancy costs roughly double LRS, and a lot of teams pay for GRS on data that would be perfectly fine on ZRS because "geo-redundant" sounds safer without anyone checking what their actual RTO/RPO requirement is.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Managed Identity vs Service Principal — aren't they the same thing?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;They solve the same underlying problem — letting an Azure resource authenticate to other Azure services without hardcoded credentials — but a Service Principal is a standalone identity you create and manage yourself, including rotating its secret or certificate, while a Managed Identity is tied directly to a resource's lifecycle and Azure rotates the credential for you automatically, with no secret ever exposed to your code. The follow-up I actually care about: system-assigned versus user-assigned managed identity. System-assigned dies with the resource it's attached to; user-assigned is a standalone Azure object you can attach to multiple resources and reuse. If someone's answer is "just use a service principal for everything," that's the answer of someone who hasn't had to explain a credential rotation incident to a security team.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;6. App Service, Azure Functions, or AKS — how do you actually decide?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;App Service is the default for a standard web app or API you don't want to think about — managed scaling, deployment slots, built-in SSL. Azure Functions is for event-driven, short-lived workloads where you're billed per execution rather than per hour of uptime — a queue trigger processing messages, a timer job, a webhook handler. AKS is for when you actually need container orchestration complexity: multiple services, custom networking, sidecar patterns, or you're already running Kubernetes elsewhere and want consistency. The mistake I see constantly is teams reaching for AKS by default because it's the "serious" option, then spending months on cluster operations for a workload that was three App Service instances and a Function the whole time.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;7. Availability Set vs Availability Zone — what failure domain does each actually protect against?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;An Availability Set spreads VMs across multiple fault domains (separate power/network) and update domains (separate maintenance windows) within a single datacenter — protects against a rack failure or planned maintenance taking down every instance at once. An Availability Zone is a physically separate datacenter within the same region, with independent power, cooling, and networking — protects against an entire datacenter going down. They are not interchangeable, and you can't mix VMs from an Availability Set across zones. The real-world implication: if your region only has one datacenter (not all Azure regions support zones), Availability Sets are your only option, and "just use zones" isn't universally available advice.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;8. Azure Policy vs RBAC — people mix these up constantly. What's the actual difference in what each one controls?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;RBAC controls who can do what — whether a user or identity has permission to create, read, update, or delete a resource. Azure Policy controls what gets created and how, regardless of who's doing it — enforcing that storage accounts must use HTTPS-only, that VMs must come from an approved image, or that resources must carry a cost-center tag, even for someone who has full Owner-level RBAC access. A candidate who says "Policy is basically RBAC for resources" hasn't actually used Policy for compliance enforcement — the whole point is that Policy applies even when RBAC would otherwise allow the action, which is exactly the governance gap RBAC alone can't close.&lt;/p&gt;

&lt;p&gt;None of these are hard to explain once you've actually broken a subscription's access model or watched a GRS failover eat five minutes of writes. They're hard live, under a bit of pressure, with someone waiting for you to think out loud. I've started running through scenario questions like these with LastRound AI's &lt;a href="https://lastroundai.com/products/mock-interviews" rel="noopener noreferrer"&gt;AI Mock Interview&lt;/a&gt; before Azure-heavy interviews — not to memorize scripted answers, but because saying "the NSG handles subnet-level filtering, but egress FQDN filtering needs Azure Firewall in the hub" out loud, on the spot, is a different skill than recognizing it on a slide.&lt;/p&gt;

</description>
      <category>azure</category>
      <category>cloud</category>
      <category>interview</category>
      <category>devops</category>
    </item>
    <item>
      <title>9 AWS Interview Questions That Separate "Used It" From "Understand It"</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Sat, 11 Jul 2026 16:38:31 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/9-aws-interview-questions-that-separate-used-it-from-understand-it-1nf</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/9-aws-interview-questions-that-separate-used-it-from-understand-it-1nf</guid>
      <description>&lt;p&gt;A candidate at a Series B fintech I was interviewing last year hit a wall on a question I didn't think was hard: "Walk me through what happens differently when a Multi-AZ RDS instance fails over versus promoting a read replica." He'd used both. He'd never had to explain the difference out loud, under pressure, to someone deciding whether to hire him. That's the gap this list is about — not whether you've clicked around the AWS console, but whether you can reason about it when someone's watching.&lt;/p&gt;

&lt;p&gt;I've been on both sides of this table for about seven years now, first as the person sweating through EC2 questions with no real production experience, later as the person asking them. These are the nine that come up constantly, roughly in the order they tend to get asked.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. How do you choose between EC2 instance families?
&lt;/h2&gt;

&lt;p&gt;Interviewers aren't looking for a spec sheet recitation. They want to know you think about workload shape before you think about price. Compute-optimized (C-family) for CPU-bound work like batch processing or gaming servers. Memory-optimized (R-family, X1) for in-memory caches and large databases. General purpose (M/T-family) as the default until you have a reason not to. Storage-optimized (I/D-family) for workloads doing heavy local I/O, like a self-managed Cassandra cluster.&lt;/p&gt;

&lt;p&gt;The wrong answer is "I just pick T3.medium for everything." The right answer names a workload and a bottleneck. A weak candidate says "M5 is a good general instance." A strong one says "if I'm seeing sustained CPU above 80% with memory sitting at 30%, I'm moving to compute-optimized, not just bumping the instance size" — because bumping size without diagnosing the bottleneck is how teams end up overpaying for headroom they don't need.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. What's the actual difference between S3 storage classes, and when does a lifecycle policy save real money?
&lt;/h2&gt;

&lt;p&gt;Standard for anything accessed regularly. Standard-IA (Infrequent Access) for things you touch less than once a month but need fast when you do. Glacier and Glacier Deep Archive for compliance data you may never touch again but legally can't delete.&lt;/p&gt;

&lt;p&gt;Here's what actually happens if you get this wrong in a live interview: candidates often forget the retrieval cost and latency tradeoff. Glacier is cheap to store, expensive and slow to retrieve — minutes to hours depending on the retrieval tier. I've seen a team set up "smart" lifecycle rules that moved logs to Glacier after 30 days, then got hit with a five-figure retrieval bill during an incident investigation because nobody accounted for needing that data fast, under pressure, all at once.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. ALB vs NLB — what's the real distinction, not just "one is Layer 7 and one is Layer 4"?
&lt;/h2&gt;

&lt;p&gt;Everyone can recite the OSI layer answer. Fewer can explain why it matters. ALB (Application Load Balancer) operates on HTTP/HTTPS, so it can route based on path, host header, or query string — useful for microservices behind one load balancer. NLB (Network Load Balancer) handles TCP/UDP at extreme scale with static IPs and sub-millisecond latency, which matters for things like gaming backends or when a client needs to whitelist a fixed IP.&lt;/p&gt;

&lt;p&gt;My honest opinion, which goes against what a lot of "AWS certification" content pushes: don't reach for NLB by default just because it's "faster." Most web applications never come close to needing NLB's throughput ceiling, and you lose ALB's content-based routing, which usually saves more architectural complexity than the latency gain is worth.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. How do Auto Scaling Groups actually decide to scale, and what's a cooldown period doing?
&lt;/h2&gt;

&lt;p&gt;An ASG scales based on a scaling policy tied to a CloudWatch metric — target tracking (keep CPU at 50%), step scaling, or simple scaling. The cooldown period is the part people forget to mention: it's a pause after a scaling action during which the ASG won't trigger another one, specifically to stop it from overreacting to a metric spike before the last scale-out has had time to actually reduce load.&lt;/p&gt;

&lt;p&gt;A weak answer stops at "it adds instances when CPU is high." A strong answer explains why an ASG without a sensible cooldown can thrash — scaling out, watching the metric dip as new instances register, scaling back in too soon, then scaling out again ten minutes later.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. EBS vs instance store — why does persistence actually matter here?
&lt;/h2&gt;

&lt;p&gt;EBS is network-attached, persists independently of the instance lifecycle, and can be snapshotted. Instance store is physically attached to the host, is faster, and is gone the moment the instance stops (not just terminates — stops). I've watched someone lose a day of processed data because they stored intermediate results on instance store and then had a routine instance stop for maintenance.&lt;/p&gt;

&lt;h2&gt;
  
  
  6. Multi-AZ RDS vs read replicas — the question that started this whole post
&lt;/h2&gt;

&lt;p&gt;Multi-AZ is for availability: a synchronous standby in a different AZ that RDS automatically fails over to if the primary dies, usually in under two minutes, and you don't manage it — RDS handles the DNS switch. A read replica is for scaling read traffic: an asynchronous copy you query directly, that you promote manually if you want it to become a new primary, and promotion isn't instant or transparent to your application.&lt;/p&gt;

&lt;p&gt;The distinction that trips people up: Multi-AZ doesn't help with read scaling (you can't query the standby), and a read replica isn't a high-availability solution by itself (replication lag means it can be behind, and promotion requires action). You often want both, for different reasons.&lt;/p&gt;

&lt;h2&gt;
  
  
  7. Public vs private subnets — what makes a subnet "public," really?
&lt;/h2&gt;

&lt;p&gt;It's not the subnet itself, it's the route table. A subnet is public because its route table has a route to an Internet Gateway. Nothing else makes it public — not IP ranges, not naming conventions. A private subnet with no such route, reaching the internet only through a NAT Gateway in a public subnet, can still send outbound traffic (for package updates, API calls) without being reachable from outside.&lt;/p&gt;

&lt;h2&gt;
  
  
  8. Reserved Instances vs Savings Plans vs Spot — what would you actually recommend to a team trying to cut their AWS bill?
&lt;/h2&gt;

&lt;p&gt;RIs commit to a specific instance family/region for 1-3 years for a discount, and they're inflexible if your architecture changes. Savings Plans are similar in discount depth but flexible across instance families and even compute types (EC2, Fargate, Lambda), which is why most teams moving off RIs prefer them now. Spot Instances are steeply discounted spare capacity that AWS can reclaim with two minutes' notice — great for stateless, interruption-tolerant work like batch jobs or CI runners, a genuinely bad idea for anything holding session state without an external store.&lt;/p&gt;

&lt;h2&gt;
  
  
  9. What actually gets monitored via CloudWatch in a production system, beyond CPU and memory?
&lt;/h2&gt;

&lt;p&gt;This is where junior and senior answers diverge hardest. Junior answers stop at CPU, memory, disk. Senior answers mention custom application metrics (queue depth, error rate by endpoint, p99 latency), composite alarms that reduce alert fatigue by requiring multiple conditions before paging someone, and the discipline of alerting on symptoms users feel (latency, error rate) rather than every resource metric that moves.&lt;/p&gt;




&lt;p&gt;None of these questions are exotic. That's the point — the bar in a real interview isn't obscure trivia, it's whether you can explain the "why" behind something you've used, calmly, to someone who's going to make a hiring decision based on how that explanation lands. I've started running through questions like these out loud before interviews rather than just re-reading docs, since the muscle you actually need is talking it through under mild pressure, not recognition. &lt;a href="https://lastroundai.com/products/mock-interviews" rel="noopener noreferrer"&gt;LastRound AI's mock interview practice&lt;/a&gt; is what I've been using to do that — it pushes back with follow-up questions the way a real interviewer would, which flat re-reading never does.&lt;/p&gt;

</description>
      <category>aws</category>
      <category>cloud</category>
      <category>interview</category>
      <category>careertips</category>
    </item>
    <item>
      <title>I could solve it at my desk. I froze the second someone watched.</title>
      <dc:creator>Mahesh</dc:creator>
      <pubDate>Sat, 20 Jun 2026 03:14:13 +0000</pubDate>
      <link>https://dev.to/maheshbandaru_ba8cc2/i-could-solve-it-at-my-desk-i-froze-the-second-someone-watched-235e</link>
      <guid>https://dev.to/maheshbandaru_ba8cc2/i-could-solve-it-at-my-desk-i-froze-the-second-someone-watched-235e</guid>
      <description>&lt;p&gt;The first time I bombed a coding round I wasn't stuck on the algorithm. I'd done the two-pointer pattern maybe 30 times that month. What got me was the silence after "walk me through your approach," and a stranger watching me think.&lt;br&gt;
If you've done this, you know the thing nobody says: the coding round mostly isn't a test of whether you can code. You can clearly code, you do it every day. It's a test of whether you can keep thinking in a structured way while observed, on a clock. That's a different skill, and almost no one practices it directly.&lt;br&gt;
So when you grind another 40 LeetCode problems alone in a quiet room, you're training the part you're already fine at, and skipping the part that's failing you. Interviewers are reading whether you restate the problem, say your assumptions out loud, and talk through a wall instead of going silent. None of that shows up when you practice by yourself.&lt;br&gt;
Here's the part I'm less sure about, so take it as opinion. The biggest-payoff change for most mid-level engineers isn't a new pattern list. It's getting 6 or 7 reps of talking through code while a thing watches and scores you, before the real thing does.&lt;br&gt;
The reason I build around this is boring and personal: I kept watching strong engineers, including me, lose rounds we should have won, for reasons that had nothing to do with skill. That's the gap LastRound AI is aimed at. During a real screen-share it surfaces a structured approach on your screen, invisible on the share, so you don't stall at "where do I even start." And the 177 step-through explainers (HTTPS, load balancing, two-pointer, RAG) let you see the system move before you have to whiteboard it.&lt;br&gt;
You don't need the tool to start, though. Tonight: pick one problem, set a 25-minute timer, solve it out loud, record yourself, then watch it back with the sound on. It's uncomfortable, which is the point. Three rounds over a week and the room stops being novel.&lt;br&gt;
So, the question I wish someone asked me before my 11th interview: when did you last solve a problem out loud, on a timer, while something watched, and then watch yourself do it? If the answer is never, that's not a coding gap.&lt;br&gt;
The live assistant and the explainers are free to start at lastroundai.com.&lt;/p&gt;

</description>
      <category>career</category>
      <category>interview</category>
      <category>beginners</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
