动态
1332 位关注者
-
Narayan Desai 分享了此动态Very proud of the work my team has done to drive this level of reliability improvements. Offering a 4 9's SLA is seriously awesome. #bigquery #sre #googlecloudplatform https://lnkd.in/gwm_Z2MCloud data warehouse BigQuery 4 9s SLA | Google Cloud BlogCloud data warehouse BigQuery 4 9s SLA | Google Cloud Blog
-
Narayan Desai 回应了此动态Narayan Desai 回应了此动态An update on Pim ... In May of 2006 I interviewed (thirteen times, would you believe it? quite normal in that era), and landed a job at Google as an L2 in Zurich. I started in this beautiful place on August 7th 2006. For the first decade, I helped develop SRE teams: Geo, Search, Crawl/Indexing, a tiny sidequest in Borg and YouTube, and Identity. In 2015, I switched to an IC role, with my starter project Production 2020, offering a few small SRE Horizon inventions such as Automon and OMG, and helping build out Google’s unique flavor of DevOps, which has been the topic of several books. For the second decade, I devoted my time developing products and the people that worked on them, and admired our collective passion for tools and problem solving, first in SRE, then inside of Core and the wider AI, Cloud and Open Ecosystems technical community. It was dope, and I loved 99.64% of it! Time waits for no-one, and Time has called for me to turn to life’s other necessities. I have two sets of aging parents, with whom I want to enjoy some more moments. Quinn is now 19 and fun to hang out with, but matching schedules is getting increasingly difficult. I have pushed out non-work goals for many years: my piano is dusty, I have more than a few creative projects that haven’t moved much, even though I see them quietly stare at me from the corners of my blue-and-orange working room, and also that PPL(H) is not going to acquire itself. I have decided to retire from the workforce, in order to pay back some debt owed to friends and family, as well as to myself. I want to take a moment to thank you. To my managers, reports, extended team mates, technical sparring partners, organizational guides, esoteric subject matter specialists, all of whom I learned so very much from. My time with Google has been the pinnacle of my career. I wrapped up my projects, and in a homage to round numbers: my last day in office was the same as my starting day, Friday August 7th. My last day with a badge was August 31st (7306 days). I remain reachable at pim@ipng.ch, and will be staying in Brüttisellen, Switzerland with my family.
-
Narayan Desai 赞了此动态Narayan Desai 赞了此动态Thank you to Switch and to the team I’ve had the privilege of working with, especially Joshua Otremba, MBA, Matthew Fuld, and Mohit Shrivastava, P.E., CEng. I came in to work on capacity planning, forecasting, and engineering analytics and left with a much deeper, more cohesive understanding of all of the many disciplines that keep a data center running. It's not quite as convoluted as a nuclear submarine, but I'll put that in the "pros" list instead of "cons". After a short break to wrap up some personal projects, I’ll be joining Lambda as an HPC Systems Architect. I'm looking forward to the challenge and getting to turn my skills toward the state-of-the-art infrastructure powering the latest technological revolution!
-
Narayan Desai 回应了此动态Narayan Desai 回应了此动态100 books in my (cluttered) hallway, waiting for the official launch and signing next week at #SRECon!
-
Narayan Desai 回应了此动态This is big science!!! Congratulations to the Argonne Leadership Computing Facility (ALCF) visualization team: Joseph Insley, Silvio Rizzi, Victor Mateevitsi, and Janet Yonga Kim Knowles on an outstanding job bringing these incredible scientific visualizations to life! A big thank you as well to Laura Wolf, Beth Cerny, Julie Parente, and Gillian King-Cargile for helping pull everything together, and to our partners at the Electronic Visualization Laboratory (EVL): Daria Tsoupikova, Farah Kamleh, and Zainab Ahmed for their collaboration and contributions. And, of course, recognition goes to all the scientists whose groundbreaking research and data make these remarkable images possible in the first place. It’s inspiring to see what happens when great science, advanced computing, visualization, and collaboration come together. Congratulations to everyone involved!Narayan Desai 回应了此动态Starting this Thursday (Sept. 10), ALCF scientific visualizations will light up THE MART as part of Chicago’s ART on THE MART. Presented by Argonne National Laboratory and the University of Illinois Chicago, the “Where Data Becomes Light: The Beauty of Scientific Visualization” event brings together the ALCF and UIC’s Electronic Visualization Laboratory for an immersive visual experience exploring phenomena from turbulent flows to the formation of galaxies in sweeping animations projected across the building’s 2.5-acre façade. See it Thursdays through Sundays, 7:30-8 p.m., through November 15. For more information, visit: https://lnkd.in/gCdxG2uJ
出版作品
-
A Scalable Approach to Deploying and Managing Appliances
TeraGrid Conference
The use of virtualization in Grid computing has seen a lot of interest lately. However, while much effort has been expanded on developing the capabilities of Virtual Machine Monitors (VMMs) and associated tools and services relatively little has been done to investigate the requirements underlying the scalable production, deployment, and management of VM images. At the same time, the clear understanding of requirements and capabilities in this area is critical to creating progress in exploring…
The use of virtualization in Grid computing has seen a lot of interest lately. However, while much effort has been expanded on developing the capabilities of Virtual Machine Monitors (VMMs) and associated tools and services relatively little has been done to investigate the requirements underlying the scalable production, deployment, and management of VM images. At the same time, the clear understanding of requirements and capabilities in this area is critical to creating progress in exploring the applications of virtualization. In this paper, we investigate the issues and propose some of the solutions relevant to this question.
其他作者查看作品
查看Narayan的完整档案
-
浏览共同好友
-
请求引荐
-
直接联系Narayan
其他相似会员
浏览更多动态
-
Ben Sussman
3233 位关注者
I don’t post here much, but I’ll say it out loud: I’m looking to pivot to something new. After nearly two decades at Google managing infrastructure and leading their Chicago office, I moved to an excellent late-stage ‘scaling’ startup (Homebase) a couple of years ago. They have a fantastic product and people, but I’ve now decided to take my career in a different direction. 🙂 I’ve spent this month messing with agents and taking a fascinating BlueDot "AGI Strategy" course full of readings and class discussions. I’m taking my time to carefully find my next leadership role. I don’t need to remind anyone here of the massive disruption to the tech industry that’s happening, or how AI is about to disrupt the whole economy. That said, I also don’t want to be the director who is just dropping LLMs and AIs into products; I want to be the one designing infrastructure around AI, building guardrails. I’m fascinated by the field of “AI Control”, which asks: How does one design AI systems that remain safe even if some part is misaligned, compromised, or buggy? It’s a new angle on an old problem. Many of us have have spent our tech careers building traditional infrastructure stacks; we assume all the parts of the system are inherently unreliable, so we work on the challenge of guaranteeing that the system is still resilient overall. We plan for component failures, broken ACLs; in turn, we design observability, minimize blast radius, and build graceful degradation. And now as we look at stacks build around “AI agents”, all the same goals apply. There’s a world of fascinating problems we to look at: - Can we ask trusted models to monitor and ‘gate’ the actions of untrusted models? - Can we design systems around least privilege, so agents only request the ACLs they actually need? - Can we design human-approval systems that don’t make us blind to hitting the ‘y’ key over and over? We should stop treating approval as a binary condition – build trees of logic that treat actions as a risk spectrum, and teach agents to have a continuum of self-certainty. I’m reading and learning. If folks have reading recommendations, please let me know. How have y’all been exploring this space?
126
12 条评论 -
Dave Beckett
Crusoe • 1935 位关注者
Twitter Hadoop Breakfix, Four Years On. I was tech lead of the Data Platform SRE team at Twitter through 2022. Breakfix (the reactive side of fleet operations) was a critical part of the job: tens of thousands of nodes across multiple datacenters, with no single system that could tell you the truth about any of them. Four years later, I went back through the architecture to work out what actually made it work. Not the tool names, which won't mean much outside Twitter, but the shape of the system. There were 3 sources of truth reconciled by set algebra on sorted files. A deliberate Python / shell split between safety invariants and policy budgets. Per-cluster caps with asymmetric floors and ceilings doing different jobs. shuf before acting. "Do nothing" as a first-class outcome. https://lnkd.in/gw4u-svK
56
10 条评论 -
Antithesis
1万 位关注者
Many elegant theoretical proofs exist for distributed systems, but local hardware and production realities rarely play by the rules. University of Illinois Urbana-Champaign professors Ramnatthan Alagappan and Aishwarya Ganesan join us on this episode of the Bug Bash podcast to talk about bridging that gap. Their work on data infrastructure, storage interfaces, and recovery protocols has been influential at TigerBeetle, WarpStream, and many other companies. Youtube: https://lnkd.in/eFCivNgf Spotify: https://lnkd.in/eFSysKZm Apple podcasts: https://lnkd.in/eymxNt6X
21
-
Utkarsh Srivastava
Parallel Web Systems • 2582 位关注者
Proud that Parallel Web Systems the first and only non-Google web search provider to be launched as an option for grounding with Vertex AI. The accuracy boost that agents get by using our search vs others is substantial, and it has been great partnering with Google Cloud and Vertex to bring this increased accuracy to their customers.
32
-
Ilya Serov
Apple • 1036 位关注者
Three months in at Apple, and I'm more excited every day about the team, the mission, and the opportunity ahead. I joined Apple for the same reason I joined Google in 2019 and AWS before that: to build a world-class storage foundation for products that change the lives of millions. I can't believe I get to do it a third time, and I'm far from the only one. Many familiar faces from my AWS and Google Cloud days are here too, bringing years of experience building and running storage at massive scale. If you're passionate about building reliable storage at scale, we're hiring across Storage teams. Reach out, I'd love to chat! #Apple #Storage #SRE #Hiring
74
-
Todd Lipcon
Google • 2733 位关注者
Pigs are flying! Spanner can now be downloaded as a docker container/helm chart for development, integration testing, learning, etc. Includes "the whole shebang" (SQL, Graph, full text search, Paxos replication, etc) -- it's the real Spanner binary sharing 99% of the code we run in production. Check it out and send along your feedback!
238
13 条评论 -
Parag Agrawal
Parallel Web Systems Inc. • 7万 位关注者
Before agents, there was no reason to optimize web search to the extreme. Turbo is 𝐛𝐥𝐚𝐳𝐢𝐧𝐠 𝐟𝐚𝐬𝐭 enabling deep research to feel instant and voice agents searching the web without lag AND 𝐞𝐱𝐭𝐫𝐚𝐨𝐫𝐝𝐢𝐧𝐚𝐫𝐢𝐥𝐲 𝐜𝐡𝐞𝐚𝐩 allowing agents to search more and save money AND 𝐡𝐢𝐠𝐡𝐞𝐫 𝐪𝐮𝐚𝐥𝐢𝐭𝐲 than slower and more expensive alternatives.
95
7 条评论 -
Partha Ranganathan
Google • 3万 位关注者
For paper 13 in our "25 papers for 25 Years of Warehouse-Scale Computing" series, we are featuring one of my favorite papers: "The tail at scale." Horizontal scaling is awesome until it isn’t. We obsess over average latency, but _tail latency_ (the slowest 1% of requests) is what actually kills user experience. In this classic paper, Jeff Dean and Luiz André Barroso explain why. When a user request fans out to 1,000 servers and more, the chance of one of them being slow slowly starts approaching 100%. This paper didn't just define the problem; it offered clever, counter-intuitive solutions to build reliable systems out of unreliable parts. A few examples: ⚡🏁 Hedging: Send the same request to two servers; use the first answer. ⏱️⚠️ Probation: Temporarily bench slow nodes. This paper is a masterclass in pragmatism over perfection. You can't stop individual servers from hiccuping, but you can design systems that ignore hiccups. If you build high-scale systems, this paper is mandatory reading. https://lnkd.in/gag7gJqf Have you had to battle tail latency in your distributed systems? What was your biggest realization? Let me know in the comments. 👇 And given it is the day after thanksgiving, rather than an academic-style poster, today's gemini3 image for the paper is just a cool visualization! :)
240
6 条评论