About
Infrastructure engineer and founder focused on large-scale distributed systems…
Activity
5K followers
-
Suman Karumuri reposted thisSuman Karumuri reposted thisFor years, sampling kept observability costs down. Agentic AI just killed it. 🧠 In this clip from the Smooth Scaling Podcast, Suman Karumuri (founder of KalDB) chats with Jose Quaresma, and explains why you can't throw away most of your traces anymore. It used to be fine to keep half your traces, or a tenth. But every trace is unique now, and LLMs hallucinate, so when a customer sees something strange you have to explain exactly what they saw and why. On top of that, million-token contexts are turning single log messages into megabytes. The volume you store jumps an order of magnitude. 📈 You can't predict which trace will explain the next strange output, so you keep all of them. 🎧 Listen to the full episode wherever you get your Podcasts.
-
Suman Karumuri reposted thisSuman Karumuri reposted this"Nobody wants to look at dashboards or alerts anymore." That's Suman Karumuri, and he's not being flippant. Suman spent 17+ years building observability at Amazon, Twitter, Pinterest, Slack, and Airbnb. He was tech lead for Zipkin, co-authored the OpenTracing spec that fed into OpenTelemetry, and is now building KalDB, the open-source log search engine that runs at petabyte scale at Slack and Airbnb. He sat down with Jose Quaresma to work through what changes when the primary user of your logs stops being human. Have a listen for: ✅ Why agents querying logs in unpredictable bursts break a traditional log stack, and why you can't sample your data anymore ✅ Why he keeps rewriting Elasticsearch instead of tuning it, and what separating compute from storage on S3 actually unlocks ✅ The recovery-task trick that keeps your freshest logs visible when volume spikes 10x, instead of going dark for 12 hours The shift underneath all of it: observability is moving from something humans watch to something agents drive, and that changes what you store, how long, and who reads it. ▸ Save this if you own a log or observability bill heading into the agentic era. Find the episode on Youtube or your favorite Podcast Player by searching for Smooth Scaling Podcast. Episode link will be in the comments.
-
Suman Karumuri posted thisOpenAI or Anthropic should buy Slack. It sounds bold, but hear me out. Everyone is racing to build the interface for AI agents. Slack already has it. An AI agent is simply another user. You can invite it to channels, DM it, control its permissions through channel membership, share context instantly by adding it to conversations, and let it naturally participate in threads and group discussions. On top of that, Slack already has: • Identity and permissions • Enterprise security • Connections to thousands of enterprise tools • Shared channels across companies • Reminders and workflows The missing pieces are surprisingly small: • A best-in-class collaborative docs experience. Quip exists, but it’s not yet on par with Notion. • A Glean-like enterprise knowledge layer that indexes all connected systems instead of treating them as point integrations. • One-click publishing of threads, channels, and docs to the web, turning internal discussions into living documentation. • A first-class task management experience with Kanban boards, planning, and execution, where humans and AI agents can collaborate on work. Put those together and Slack becomes much more than chat. It becomes the operating system for enterprise AI agents. If I were OpenAI or Anthropic, I’d seriously consider buying Slack instead of building another chat interface. The foundation is already there. Am I missing something?
-
Suman Karumuri posted thisI'll be at Berlin Buzzwords this week. Over the last decade, I've helped build and operate search, logging, metrics, and tracing systems at very large scale. One thing I've learned is that many teams assume their OpenSearch or Elasticsearch cluster is performing as expected when it often isn't. In the past year alone, I've seen teams discover: • 30-70% unnecessary infrastructure spend • Significant data loss during ingestion they didn't realize was happening • Query patterns that silently return incomplete results • Architectures that become exponentially more expensive as observability data grows If you're running OpenSearch, Elasticsearch, or large observability workloads, I'd love to compare notes. No sales pitch. Just sharing what we're seeing across different environments and learning from others who are operating at scale. Comment below or send me a message if you'll be at Berlin Buzzwords.
-
Suman Karumuri posted thisDo I need your product before I make my first dollar? That's a surprisingly useful question, especially for infrastructure companies. Many products are valuable. Far fewer are necessary. If I can build, launch, and get customers without your product, you're competing for budget and attention. If I can't, you're on the critical path. Looking back, this explains why some infrastructure products I thought would take off didn't. And why others grew much faster than expected. Product-market fit is often discovering whether you're a dependency or an optimization.
-
Suman Karumuri posted thisDatabases are Hotel California for your data. You can check in anytime you want. But you can never leave. The lock-in isn’t just the data. It’s the transaction semantics, indexing behavior, security model, audit workflows, latency assumptions, and operational behavior that applications quietly evolve around. Sometimes even making the database faster breaks the application by exposing hidden race conditions and timing assumptions. Over time, the database stops being infrastructure. It becomes part of the application itself.
-
Suman Karumuri posted thisDatadog’s moat was integration. The agent collected the data. The storage held the data. The UI became the place humans debugged production. That bundle was powerful. It also let Datadog charge a lot for storage because leaving meant replacing the whole workflow. OpenTelemetry weakened the first part of that moat: collection. AI agents may weaken the second: the UI. If humans are no longer the main users of observability, dashboards become less central. Agents need APIs, context, retrieval, and cheap access to large volumes of logs, traces, metrics, deploys, and incidents. That changes the value chain. The problem for Datadog and Splunk is not that they cannot see this shift. It is that responding means attacking their own business model. AI-native observability wants cheaper storage, open access, and agent-first workflows. Incumbents are built around expensive storage and human-first UIs. That is why this transition is so hard for them.
-
Suman Karumuri reposted thisSuman Karumuri reposted thisWe had some great lightning talks at #haystackconf this year. - Suman Karumuri - Kaldb: Search infra for agents - Peter Dixon-Moses and Andy Edmonds - More than just a click -- Increasing training supervision with attentional foraging - Mayya Sharipova - GPU accelerated vector indexing in Elasticsearch - David Tippett - YASP - Yet another search protocol - the need for a new protocol to provide an interface into websites to make them accessible to AI agents. - Doug Turnbull - Deep Research is not Agentic Search - Nik Everett - A search war story - Bertrand Rigaldies - My first foray into LLM-based Query Understanding - Eric Pugh - Where did all the unicorns go? Some other guy* spoke on Getting better code from agentic AI, but I didn't even bother taking a picture of that one. *it was me. 😂 #agentic #ai #search #lightningtalks
-
Suman Karumuri reposted thisSuman Karumuri reposted this📚 𝗪𝗵𝘆 𝗱𝗼 𝗼𝗯𝘀𝗲𝗿𝘃𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗽𝗹𝗮𝘁𝗳𝗼𝗿𝗺𝘀 𝘀𝘁𝗿𝘂𝗴𝗴𝗹𝗲 𝘄𝗵𝗲𝗻 𝗹𝗼𝗴𝘀, 𝘁𝗿𝗮𝗰𝗲𝘀, 𝗮𝗻𝗱 𝘁𝗲𝗹𝗲𝗺𝗲𝘁𝗿𝘆 𝗮𝗹𝗹 𝗯𝗲𝗵𝗮𝘃𝗲 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁𝗹𝘆 𝘂𝗻𝗱𝗲𝗿 𝘀𝗰𝗮𝗹𝗲? Kaldb: A polystore for logs and traces Talk by Suman Karumuri from KalDB at SREday Austin 2026 Suman explains why storing every type of observability data inside a single backend creates performance and cost problems that grow over time. The session examines how polystore architectures route workloads across specialized systems while preserving unified querying across logs, traces, and high-cardinality telemetry. 👉 Suman Karumuri previously led observability and reliability engineering work at Airbnb and built large-scale logging and tracing systems at Slack, Pinterest, and Twitter before founding KalDB. 𝗦𝗲𝗲 𝗮𝗹𝗹 𝗼𝘂𝗿 𝗲𝘃𝗲𝗻𝘁𝘀 𝗵𝗲𝗿𝗲: https://sreday.com/ #Observability #DistributedSystems #Tracing #SRE
Suman Karumuri commented on a post
1mo
Great point! Make it work, make it right and make it fast/cheap applies to AI too!
-
Suman Karumuri reacted on thisSuman Karumuri reacted on thisCongratulations Sameer Agarwal and the Deductive AI team on your acquisition by Elastic! You had a vision, brought it to life and and now enterprises get to benefit from it. Thanks for trusting Umesh, Eddie and I from day 0 of this journey. PS - sharing an image from the very first holiday party of Databricks in SF which I was fortunate to attend.
-
Suman Karumuri liked thisSuman Karumuri liked thisTroubleshooting LinkedIn Profile URLs on DS-160 Applying for a US visa involves navigating the complexities of the DS-160 form, and I’ve recently encountered a technical nuance regarding LinkedIn profile URLs that might save you some frustration. If your custom LinkedIn URL includes lowercase letters (e.g., [https://lnkd.in/ds5jUjSW)), you may encounter an unexpected hurdle. The DS-160 system sometimes auto-capitalizes input fields, transforming your link into LINKEDIN.COM/IN/TRKONDURI. Because LinkedIn’s pathing is case-sensitive, this conversion often triggers a "Page Not Found" error. The Solution To ensure your profile remains accessible to consular officers regardless of how the form processes your text, try this simple workaround: Instead of using the /in/ subdirectory, you can use the /? parameter, which is case-insensitive. * Original URL: [https://lnkd.in/ds5jUjSW) (Risks 404 error if capitalized) * Alternative URL: [https://lnkd.in/d9C3pv7r) (Remains functional even if auto-capitalized by the system) By using this format, you ensure that your professional profile remains reachable and error-free throughout the visa application process. Have you encountered similar technical quirks while filling out your visa applications? Share your experiences below to help fellow applicants! #VisaTips #DS160 #LinkedIn #CareerResources #USVisa #TechTips
-
Suman Karumuri liked thisSuman Karumuri liked thisCome work with me in the Search domain at Aiven. We offer managed OpenSearch (recently on the LTS version 3.6) that has been downloaded 2B (!) times globally in the upstream. Besides making the managed search product, we contribute to the open source, most recently by extending the migration assistant to support GCP cloud. We also keep thinking about those who take the first steps and want to learn with our Free Tier and very affordable Dev Tier of OpenSearch. It is exciting times for us in Search and we are hiring a Senior Software Engineer in Helsinki. We solve for a range of use cases: classic keyword search, vector search, log search & observability. This needs to work at massive scale on a $$$ and latency budget for our customers. We spoke at the recent OpenSearchCon in Prague with our client on stage. And share our learnings in the Aiven blog. https://lnkd.in/dEadTC-TSenior Software Engineer, Search at Helsinki, Uusimaa, Finland | Careers & Jobs at AivenSenior Software Engineer, Search at Helsinki, Uusimaa, Finland | Careers & Jobs at Aiven
-
Suman Karumuri liked thisSuman Karumuri liked thisDista has been honored with the "Emerging Global Brand of the Year" award at Shiprocket SHIVIR 2026. 🏆 Being recognized among the industry's leading innovators is a proud milestone for Team Dista and a reflection of our commitment to transforming cross-border commerce through innovation. This recognition celebrates organizations creating meaningful impact in the eCommerce ecosystem, and we're proud to have won this year. Every milestone like this is powered by the trust of our customers, the support of our partners, and the passion of every member of Team Dista . Thank you for believing in our vision. We're just getting started. 🚀
-
Suman Karumuri liked thisSuman Karumuri liked this"Life comes full circle" After 18 incredible years building my career in the United States, I'm excited to share that I'm returning home to Hyderabad, India — the city where my professional journey first began. Over the past 18 years, I've had the opportunity to build products, scale engineering organizations, lead exceptional teams, and learn from some of the brightest people in the industry. The journey has been challenging, rewarding, and transformative in ways I could never have imagined when I first left India. As I step away from my role as Co-Founder & CTO of Planning Hub, I do so with deep gratitude to the team and confidence in the road ahead for the company. I'm equally excited to share that I have joined Coforge as Vice President of Product Engineering. Returning to Hyderabad while taking on this new leadership opportunity feels particularly special — coming back to the place where it all started, bringing with me experiences gathered across continents, teams, startups, and enterprises. I'm looking forward to reconnecting with old friends, building new relationships, contributing to India's rapidly evolving technology ecosystem, and helping shape the next generation of products and engineering organizations. To everyone who has been part of this journey so far — thank you. Your support, mentorship, friendship, and trust have made all the difference. Here's to new beginnings and coming full circle. #Hyderabad #India #Coforge #Leadership #ProductEngineering #Technology #AI #Startups
-
Suman Karumuri liked thisSuman Karumuri liked thisAI coding tools are changing what it means to be a specialist. A backend engineer can build frontend applications. A machine learning engineer can move more easily into infrastructure. Developers who were once constrained by a handful of languages now have access to a much broader set of tools and frameworks. In the latest episode of #UnlockedPodcast, MadLabs CEO Manju Rajashekhar discusses how coding agents are changing software development, why latency and cost are inseparable in AI systems, and why the most effective agents often succeed by working with less context, not more. Full episode in the comments 🔥
-
Suman Karumuri liked thisSuman Karumuri liked this"Nobody wants to look at dashboards or alerts anymore." That's Suman Karumuri, and he's not being flippant. Suman spent 17+ years building observability at Amazon, Twitter, Pinterest, Slack, and Airbnb. He was tech lead for Zipkin, co-authored the OpenTracing spec that fed into OpenTelemetry, and is now building KalDB, the open-source log search engine that runs at petabyte scale at Slack and Airbnb. He sat down with Jose Quaresma to work through what changes when the primary user of your logs stops being human. Have a listen for: ✅ Why agents querying logs in unpredictable bursts break a traditional log stack, and why you can't sample your data anymore ✅ Why he keeps rewriting Elasticsearch instead of tuning it, and what separating compute from storage on S3 actually unlocks ✅ The recovery-task trick that keeps your freshest logs visible when volume spikes 10x, instead of going dark for 12 hours The shift underneath all of it: observability is moving from something humans watch to something agents drive, and that changes what you store, how long, and who reads it. ▸ Save this if you own a log or observability bill heading into the agentic era. Find the episode on Youtube or your favorite Podcast Player by searching for Smooth Scaling Podcast. Episode link will be in the comments.
Experience & Education
-
KalDB
View Suman’s full experience
See their title, tenure and more.
or
Publications
-
Visualizing threads, transactions and tasks
Workshop on Program Analysis for Software Tools and Engineering (PASTE)
-
Code Bubbles: Rethinking the User Interface Paradigm of Integrated Development Environments
Proceedings of 32nd International Conference on Software Engineering (ICSE 2010)
-
Code bubbles: a working set-based interface for code understanding and maintenance
Proceedings of 28th International ACM Conference on Human Factors in Computing Systems (CHI 2010)
Patents
Languages
-
English
Native or bilingual proficiency
-
Telugu
Native or bilingual proficiency
-
Hindi
Native or bilingual proficiency
-
Bengali
Elementary proficiency
Other similar profiles
Explore more posts
-
💻 Daniele Frasca
Lambda Durable Functions were developed to address the need for pausing a Lambda execution while allowing external work, such as an LLM generating a response, to continue without incurring idle VM costs or losing execution state. This challenge is not new. It’s not a “which is better” post it walks through the same real use case implemented both ways and looks at where complexity and cost actually live. https://lnkd.in/dBDdk94z #AWS #Serverless #CloudArchitecture #StepFunctions #AI
-
Paul Baier
Last night Ashish Bhatia gave a demo session to 250 (of 420 regs) at our monthly GAI Insights "Learning Lab" Zoom call on Claude Code and OpenClaw For Claude Code, his 3 key takeaways are: - Anyone can Build: Technical/Non-Technical, any one can build. If you have an idea you can bring it to life. - Mins -> Hours -> Days: AI Tools can perform long horizon tasks. - Gulf of Specification: How do you convey your intent most effectively with AI models/agents. His 3 takeways for OpenClaw are: - Orchestration and not the model: Models are fungible and intelligence is getting commoditize. - AI is becoming the Operating system: Transition from answers to actions. - Agent Manager mindset: No upper limit to resources, what tasks would you offload to your agents. Link to recording in the comments. Let's keep Learning and Building!
2 Comments
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content