Prateek Pandey

How Meesho Transformed Shopping with Redis Meesho, India’s leading e-commerce platform, revolutionized its infrastructure by migrating to Redis on GCP. This move addressed scalability challenges during high-traffic events, enabling seamless shopping for 100+ million users. Key Highlights 1. Massive Scalability: Redis powers 20x load increases, ensuring stability during major sales like Diwali. 2. Ultra-Low Latency: Real-time personalization and gamification keep users engaged, even during peak demand. 3. Cost Efficiency: Server costs reduced by 65%, operational expenses cut by 90%. 4. Innovation at Scale: Features like “Spin the Wheel” processed 100M clicks in real time. 5. Live Commerce Revolution: Redis delivers instant, reliable messaging for smooth supplier-user interactions. Redis has become the backbone of Meesho’s innovation, driving scalability, reducing costs, and enhancing user experiences in India’s booming e-commerce market.

8

My Story of How I improved single server pod efficiency to handle 150 concurrent users operating at 100% cpu usage to 900 concurrent users operating at 85% cpu usage It has been a long day and a long time since I had posted so here's some of my recent venture into servers, server-side latency and infra costs. So the story starts with a load test request for one of our apis on a Friday morning looking forward to push a new feature to prod by evening. My mind goes well that's a double edged sword when it comes to critical requests on a Friday :) Anyways It's a simple request with a small twist to it since I wanted to publish the results for both cached and non-cached behavior of the same api. I build my jmeter scripts and run it on Azure Load Runner and bam first surprise.... As the number of virtual users increased, our poor little single pod taking up the huge loads suffered and the latency gradually went up until certain breaking point where it couldn't handle the said number of concurrent users. Surprisingly the same pod could handle same rps with way less number of users! This pod had a spec of 2 cpu and 4G memory. The server could handle merely 29 users at 191ms which is < 200ms (standards) and this seemed too low.  Moreover at ~150 users the latency shot upto 48s and at ~200 users it had reached 3mins! This led me into a rabbithole trying to debug what might be issue which led me into profiling the memory and cpu usage during such loads and long story short the cpu usage was pretty high i.e., 100% but low memory. After a live profiling of such pods via VisualVm I was able to conclude the following: 1. Number of threads must be improved. 2. Threads must be efficiently used i.e., less waiting time on threads. 3. Reduce the cpu usage. The solution for the above problem was: 1. Tomcat thread and queue management for best use of available resources. 2. Server change from BIO to NIO mode (Non-Blocking IO) 3. Connection pooling 4. Increase Heap memory and set min and max to handle throttling. Now the trick is to not over-do the number of threads and tasks handled by a thread since it can increase the memory usage again trggering gc events which is bad. After fine ballet of all above mentioned fixes lo and behold a single pod could handle  900 concurrent users 1k RPS at 950ms with cpu as low as 85% and low gc activity!! #springboot #jmeter #backend #infra #costcutting

12

After taking 100+ Mock Online System Design Interviews, I saw few generic problems. - People say to use NoSQL DB for scale. Which? How? - Is missing? - People say they will use cache - Which one? Why not Memcached? This is gone... - People replace LBs with API Gateway or BFF 🤦‍♂️ - People use MongoDb most often at scale (no matter whether we need transactional systems or write-heavy systems). - Kafka is a peanut. Use it with any drink. But yes, we don't know how to solve head-of-line blocking 🤦‍♂️ - Auto-scaling is for free (even in persistent systems) 🤦‍♂️ - Redux is required even for a simple app. - Node JS is single-threaded and what not ... What will you add?

5

6 Comments

How she left her job and made a 4000 Crore Empire in the US as an Indian | The Apache Kafka Story Video description: "🚀 The Incredible Journey of Neha Narkhede: From Building Apache Kafka to Founding Confluent 🚀 Dive into the inspiring success story of Neha Narkhede, the brilliant mind behind Apache Kafka and the co-founder of Confluent, a game-changing company in the world of real-time data. This detailed business case study in Hindi unpacks her journey from humble beginnings to becoming one of the most successful women in tech with a net worth of $520 million. Timestamps: 00:00 - Introduction to Neha Narkhede's inspiring story 01:17 - Early Life and Career Beginnings 02:21 - Landing a Job at LinkedIn: A Turning Point 03:35 - Developing Apache Kafka: Revolutionizing Data Streaming 04:45 - Understanding Real-Time Data Access in the Simplest Terms 06:01 - Scaling Apache Kafka to Global Recognition 08:00 - The Birth of Confluent: Seizing New Opportunities 08:59 - How Confluent Raised Millions in Funding 09:24 - Taking Confluent Public: The IPO Success Story 10:52 - Achieving a Net Worth of $520 Million 11:35 - Key Lessons from Neha Narkhede's Journey 12:45 - Conclusion: What We Can Learn from this Trailblazer What You’ll Learn in This Video: How Apache Kafka was built and became the backbone of real-time data processing. Insights into Neha Narkhede’s career journey and entrepreneurial mindset. The simplest explanation of real-time data access and why it's critical for businesses today. How Confluent turned Apache Kafka into a billion-dollar enterprise. Lessons from Neha Narkhede on innovation, scaling startups, and creating lasting impact. 💡 Whether you're interested in an Apache Kafka tutorial, business case studies in Hindi, or learning about real-world entrepreneurship, this video has something for everyone!  " #technology #business #BusinessCaseStudy #kafka #LinkedIn https://lnkd.in/g93AHfyi

3

🗞️ Grab the latest edition of our Confluent Developer Newsletter! 🗞️ The new roundup includes helpful resources like: - Localized Hindi (हिंदी) captions for Python developers in our #ApacheKafka 101 course! - A 6-minute video on how exactly once-processing works in #ApacheFlink - Upcoming events around the world 🌎 … and much more! Read it below ⬇️

2

Be careful with type hinting when you fetch numbers from Redis! ⚠️ Numbers from Redis are strings, not integers. Watch out for type errors and remember to cast them back to int. 🙌🧠

11

1 Comment

Worth giving your 2 minutes..

6

🚨 IMP Update: In Chrome 131, there is a new setting has been introduced due to which SelectorsHub #Debugger feature stopped working. To use SelectorsHub Debugger feature from Chrome 131 onwards, you need to do small setting in chrome DevTools. Without this setting it won't work. Special thanks to Gaurav Khurana for finding this quick fix. Please follow this video for the setting change- https://lnkd.in/gX3DwZa8 #selectorshub #update #debugger

15

2 Comments

🚀 "Attention developers! Are we overlooking Memcache in favor of Redis? While Redis is powerful and feature-rich, Memcache can sometimes be the unsung hero for specific tasks. Here’s a quick comparison: 🔹 Memcache: 🌟 Strengths: Simplicity, speed, and ease of use. Ideal for straightforward key-value caching. 🌟 Weaknesses: Limited to simple data types and lacks persistence. 🌟 Use Cases: Ideal for session caching, simple query results, and rapidly changing data where persistence isn't critical. 🔹 Redis: 🌟 Strengths: Rich data structures, persistence, and support for complex operations. 🌟 Weaknesses: More complex setup and management, can be overkill for simple caching needs. 🌟 Use Cases: Perfect for tasks needing advanced data structures like sorted sets, pub/sub messaging, and persistent data storage. So don't underestimate the power of Memcache! While Redis is great, there are times when Memcache is the better fit – it's simpler, faster for certain caching tasks, and easier to operate and manage. Follow me for more insights on #systemdesign, #distributedsystem,#distributeddatabase, #dataengineering, #dataanalytics, #iot, and #industry40. Stay tuned for the next part of the Memcahe Series!

8

1 Comment

Uber migrated trillions of entries from DynamoDB to Ledger Store (LSG) saving more than $6 million annually. 🚀 🚀 Let's understand how they migrated more than 1.2 Petabytes of data in simple words. What was Uber's original architecture ? 🤔 🤔 👉 Gulfstream - Payments platform comprising of microservices. 👉 DynamoDB - Stored all the payment-related recent data (less than 12 weeks old) 👉 TerraBlob - Similar to S3 and stored cold data (more than 12 weeks old). What is Ledger Store ? It's Uber's in-house data store optimized for storing payment-related data. It is verifiably immutable and supports Tiered storage for cost management. Why LSG was chosen for data migration ? 1️⃣ Cost savings as compared to DynamoDB. 2️⃣ Shorter indexing lag. 3️⃣ Better performance due to on-premises deployment. 4️⃣ Single source of payment data for Gulfstream services. What were the key requirements in data migration ? 🔖 🔖 ✅ Data Completeness ✅ Data Correctness ✅ Data migration throughput ✅ Latency - p99 latency of LSG to be within a bound ✅ Lag - Index updation lag to be within a limit What were the challenges in data migration ? ⚒ ⚒ 🎯 Scalability - Handle the load generated by the backfill job. 🎯 Rate limiting - To ensure consistent behaviour of the backfill job. And avoid self DDOS (Distributed Denial Of Service). 🎯 Emergency stop - Stop migration in case of an outage or overload. 🎯 Fault tolerance - Handle bad records, timeout issues and implement retries in the backfill jobs. 🎯 Logging - Reduce pressure on logging infra by rate limiting the logs generated during migration. The data migration process was smooth and was performed transparently without any disruption or downtime. Over the two years, they migrated petabytes of business-critical money data. 💰 💰 Data migration might seem like a straightforward and a menial task. But Uber's case study has shown otherwise. It brings enormous business impact and has it's own system design challenges. 🔥 🔥 Also, with most of the companies developing in-house data stores/databases, do you think cloud databases like DynamoDB will fade away in future ? Let me know your thoughts in the below comments. 👇 👇 Link to Uber's blog post - https://lnkd.in/d9DsnRug #tech #databases #Uber #SystemDesign

131

2 Comments

🚀 Enhance Your Infrastructure Monitoring with Prometheus 🚀 In modern infrastructure, real-time monitoring is a game-changer, and Prometheus has become a go-to solution for tracking performance and ensuring system reliability. Why should we choose Prometheus over others? 1.     Time-Series Metrics Collection 2.     Pull-Based Model 3.     Powerful Querying with PromQL 4.     Highly Scalable and Distributed 5.     Built-In Alerting 6.     Extensive Integrations What Can You Monitor with Prometheus? Prometheus can track metrics for: ·      CPU, memory, and disk usage on virtual machines and physical servers ·      Containerized environments (like Docker, Kubernetes) for resource consumption and performance ·      Cloud-native services, including AWS, Azure, and Google Cloud metrics ·      Application-specific metrics like request latency, throughput, and error rates From cloud services to containerized applications, Prometheus gives you the visibility needed to identify performance issues, track resource usage, and react quickly to anomalies. #prometheus #monitoring #infrastructure #devops #observability #opensource #DevOpsInsider

2

1 Comment

It is not space vs time, but space vs time vs correctness. If you want to improve on space and time, be okay with giving up on correctness. Bloom filters and the entire spectrum of approximate data structures and algorithms fork from this constraint. all the trade-offs Redis takes internally give up on correctness wherever possible to save space and time. Their eviction and expiration strategy are prime examples of it. So, space vs time was never a dilemma but a trilemma. ⚡ I keep writing and sharing my practical experience and learnings every day, so if you resonate then follow along. I keep it no fluff. arpitbhayani.me/redis #AsliEngineering #Redis #SystemDesign

413

12 Comments

Today's Topic - Why is Redis so Fast? Redis consistently ranks among the fastest databases available. Several factors contribute to its remarkable speed and performance - 1. 𝐈𝐧-𝐌𝐞𝐦𝐨𝐫𝐲 𝐃𝐚𝐭𝐚 𝐒𝐭𝐨𝐫𝐚𝐠𝐞 Lightning-Fast Access - Redis primarily operates in-memory. - This means data resides in RAM, eliminating the latency of disk reads and writes. - The image clearly demonstrates the significant speed difference between RAM and storage mediums like SSDs or HDDs. Data Persistence Options - While primarily in-memory, Redis offers options for persistence (RDB snapshots, AOF logging) to safeguard data durability. 2. 𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐞𝐝 𝐃𝐚𝐭𝐚 𝐒𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞𝐬 Efficient Implementation - Redis data structures (strings, lists, hashes, sets, sorted sets) are implemented with highly optimized algorithms and data structures, tailored for specific use cases. - The image showcases how each Redis structure maps to a well-chosen underlying implementation (e.g., ziplists for small lists, skip lists for sorted sets). Minimal Overhead - These specialized structures minimize memory overhead and computational complexity, leading to faster operations. 3. 𝐒𝐢𝐧𝐠𝐥𝐞-𝐓𝐡𝐫𝐞𝐚𝐝𝐞𝐝 𝐀𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞 No Context Switching - Redis uses a single thread for command processing, eliminating the overhead of context switching between multiple threads. Event Loop - The event loop model allows Redis to handle multiple client connections concurrently without the need for multithreading, avoiding the contention and synchronization issues that can arise. 4. 𝐈/𝐎 𝐌𝐮𝐥𝐭𝐢𝐩𝐥𝐞𝐱𝐢𝐧𝐠 (𝐑𝐞𝐝𝐢𝐬 6.0 𝐎𝐧𝐰𝐚𝐫𝐝𝐬) Enhanced Scalability - I/O multiplexing (using the io_uring interface on Linux) further improves Redis' ability to handle many concurrent connections efficiently. - The image depicts how this mechanism enables a single thread to manage multiple client connections simultaneously. Networking Performance - Redis leverages various networking optimizations (e.g., TCP backlog, SO_REUSEPORT) to enhance network throughput. Minimalistic Design - Redis focuses on core features and avoids unnecessary bloat, resulting in a lean and efficient codebase. Connect/Follow - Pallavi Ghanshani

4

Zerodha's 1.5+ million PDF contract notes processing time was reduced from 8 hours to 25 minutes is truly impressive Nomad was used instead of Kubernetes as an orchestrator. Zerodha always impresses with their engineering work. Here is another story of how rearchitecting helped Zerodha improve its overall performance and reduce costs significantly, driving customer and product success I will keep the story very short and leave a link to read the entire blog. Zerodha had a challenge processing millions of contract notes that had to be sent to the customers after executing the stocks, and this had to be completed in a day; sometimes, it took 7 to 8 hours. The old stack ran on a large single vertically scaled server to complete the process, which was very time-consuming and inefficient. The new stack is horizontally scalable and elastic. Worker nodes are used to complete the job independently. AWS S3 stores and processes the files by each worker node across various stages. Redis stores the state of each job and is also used as a backend broker, distributing the jobs among the workers. Each job's execution is distributed across EC2 instances, which write to shared storage S3 Nomad is used as an orchestration framework, and, along with Terraform, it's used to provision the nomad cluster, where auto-scaling groups are leveraged, each handling different groups of servers for various jobs like generating pdf, sending emails, and signing PDFs. Nomad deploys job worker programs on the designated nodes. Rundeck is used for administration management and execution After the processing is completed, the environment is destroyed with the help of a monitoring system along with Redis #zerodha #performanceengineering #systemdesign #performance #aws

1,469

94 Comments

New github tool : go-dnsbrute I was looking for something to brute force subdomains but all the available tools currently have too many options and does lot of other things that I did not want. So, I wrote a go program to do just that :) https://lnkd.in/gaba2DDs #github #pentesting #ethicalhacking #dns #bruteforce

7

Fill “standards” in the blanks: Like #Kafka is for Streaming.. Like #S3 is for Object Storage… Like #PostgreSQL is for RDBMS… Like #Linux is for Operating systems… ……….. is for Monitoring… ……….. is for Containers… ……….. is for Cloud Deployment… ……….. is for Web APIs… ……….. is for Systems Programming… ……….. is for Game Development… ……….. is for Mobile Development… ……….. is for Unit Testing… ………. is for API Testing… ……….. is for Load Testing… ……….. is for ………… What will you add more?

9

3 Comments

For CKA Exam aspirants, There’s some good news and some bad news. The bad news is, The CKA exam is updating after November 25, 2024. New topics will be added, focusing more on real-world Kubernetes skills like the Gateway API, Helm, Kustomize, CRDs, and Operators. The good news is, These updates are great for those who want to truly master Kubernetes and become real-world CKA experts. I’ve written a blog comparing the old and new syllabus and outlining the additions. 𝗖𝗞𝗔 𝗨𝗽𝗱𝗮𝘁𝗲 𝗕𝗹𝗼𝗴: https://lnkd.in/gsqUc5TG 𝗡𝗼𝘁𝗲: Any CKA exam taken after the updated release will test on the new set of Domains and Competencies I personally wanted the syllabus to change because Kubernetes administration has evolved a lot in the real world. These updates will help engineers with the concepts needed for real-world setups. ♻️ PS: Dont forget to Repost and share this news with the community. #DevOps #Kubernetes #CKA

681

22 Comments