Srikanth Sundarrajan

At #OpenSourceIndia2025, Aakash Bist, Member of Technical Staff at Pure Storage, is tackling one of the biggest challenges in deploying large language models — running out of GPU memory. As inputs grow longer and concurrent workloads increase, inference can slow down or even fail. Aakash is showcasing how an open-source-first approach, combined with Pure Storage’s FlashBlade, helps offload memory to fast, shared storage, keeping AI workloads efficient and stable. The session highlights how this architecture enables: ◾️ Smoother large-scale LLM inference ◾️ Cost-effective resource utilization ◾️ Improved scalability across AI workloads A highly practical session demonstrating how open source and intelligent storage design together can push the boundaries of AI performance. #OpenSourceIndia #PureStorage #AI #LLM #Inference #OpenSource #StorageInnovation #AIInfrastructure #Scalability

4

EDB Day Session — Part 2: PostgreSQL 18 / EPAS 18 — Features Are the Headline. Upgradeability Is the Story. In the same EDB Day session (Dec 8, 2025), one theme kept surfacing whenever we spoke about PostgreSQL 18 and EPAS 18. The room wasn’t asking for a “release note tour.” They were asking for something much more enterprise: “How do we take advantage of new capability… without turning upgrades into events?” Because in mature environments, the question is never just what’s new. It’s how safely can we adopt it—repeatedly. The most valuable lens I saw in the discussion was this: A major release only creates value if it’s reachable. That means: - Version currency becomes a strategy (not a heroic project every few years) - Patching becomes routine (not a mini-migration) - Monitoring becomes actionable (less noise, fewer false alerts, clearer next steps) - Change windows become workload-aware (time blocks aligned to real usage patterns, not calendar tradition) Here’s my strong belief: The best database platform isn’t the one with the most features. It’s the one where adopting features doesn’t increase fear. In the next posts, I’ll connect this to: - PGD 6.x (resilience and distributed reality), - Aurora migration thinking (what changes when you move off AWS-managed constraints), - and EDB PG AI Platform (guided operations that make upgrades and patching feel uneventful). One question for platform leaders: What’s your biggest blocker to staying current—testing, tooling, change windows, or organizational muscle memory? #EDB #PostgreSQL #EPAS #PostgreSQL18 #PlatformEngineering #DatabaseReliability #SRE #Observability #Modernization #EnterpriseIT #CloudMigration #ReleaseEngineering #DevOps

7

Excited to share the fifteenth blog of our journey to PGConf India "Terabyte-Scale PostgreSQL Logical Replication: The Challenges nobody warns us about" In this blog, Venkat Akhil Pavuluri examines the real-world challenges of operating PostgreSQL logical replication at terabyte scale. He dives into critical pressure points such as replication slot WAL accumulation, multi-day initial snapshots, schema drift, TOAST behavior, and the lack of built-in observability — highlighting how these factors can quietly destabilize large-scale deployments. Through hard-earned operational insights, the article offers a pragmatic perspective on what it truly takes to run logical replication reliably in high-volume, multi-tenant environments. Read the blog here: https://lnkd.in/gXwPyyxn Prepared. Postgres-Powered. PGConf India-Bound. See you at #PGConfIndia #PostgreSQL #BangaloreTech #PostgresCommunity #OpenSource #TechEvents #OpenSourceDB

6

Is there a function in your application that could benefit from a proof of functional correctness? Our customers have reported 60%-75% less defects in code developed with formal methods using #SPARK. Johannes Kanig explains how he iteratively improved the proof level for a character counting function using #SPARK in the AdaCore blog below raising it from the absence of runtime errors (aka SPARK Silver) to functional correctness (aka SPARK Gold). Developing code in SPARK and especially proving correctness takes more time, but the reduction in defects saves time and money in the long run.

9

Looking for a Kubernetes-native, serverless platform to run scalable, reliable event-driven applications, including inference on streaming data workloads? 👇 Hear directly from 𝐍𝐮𝐦𝐚𝐟𝐥𝐨𝐰 PM Sriharsha Yayi, Stuff Product Manager at Intuit and Krithika Vijayakumar, Senior Software Engineer at Intuit 🔥 𝐑𝐞𝐠𝐢𝐬𝐭𝐞𝐫 𝐟𝐨𝐫 𝐟𝐫𝐞𝐞 𝐟𝐨𝐫 𝐭𝐡𝐢𝐬 𝐚𝐧𝐝 𝐦𝐚𝐧𝐲 𝐦𝐨𝐫𝐞 𝐠𝐫𝐞𝐚𝐭 𝐬𝐞𝐬𝐬𝐢𝐨𝐧𝐬 𝐚𝐭 https://lnkd.in/eJkr7wqS 📆 𝐓𝐮𝐞𝐬𝐝𝐚𝐲, 𝐒𝐞𝐩𝐭𝐞𝐦𝐛𝐞𝐫 𝟐𝟑𝐫𝐝, @ 𝟒 𝐩𝐦 𝐂𝐄𝐓 | 𝟏𝟎 𝐚𝐦 𝐄𝐓 | 𝟕 𝐚𝐦 𝐏𝐓 Running inference on streaming data is complex. Teams must connect to event sources, process streams, and scale inference with minimal overhead. Numaflow, a Kubernetes-native open source platform, simplifies this by allowing teams to do stream processing and inference in the same pipelines natively on Kubernetes. It integrates with Kafka, Pulsar, SQS, and more to support diverse event sources and scales to handle high-throughput workloads. Numaflow (created by the creators of Argo Project) is currently in use by large enterprises. This talk shows how Numaflow enables ML engineers to achieve scalable stream processing and inference on streaming data without the infrastructure burden. #Numaflow #Kubernetes #InferenceStreaming #EDA #DevSecOps #KubeCrash #PlatformEngineering 

30

1 Comment

We ran Rama against the TPC-C benchmark with great results: - 40% less AWS cost than CockroachDB for the same throughput - Much tighter latency distribution than CockroachDB for each transaction (3x variance compared to 400x for writes) https://lnkd.in/gujinSnk

24

Watch back Vishesh Jindal's #CloudStackCollab session "Orchestrating GPU workloads with CloudStack"! This session dives into the technical design and implementation of native GPU orchestration in Apache CloudStack 4.21 on KVM, including device discovery, capability classification, and inventory synchronization via the KVM agent. By watching this session back, you will learn how GPU‑backed service offerings are defined and consumed by Instances. Vishesh covers operator prerequisites & host setup (IOMMU, vendor vGPU profiles), and lifecycle operations from provisioning to teardown. A live demo walks through GPU discovery on host, offering creation, and end‑to‑end deployment. https://lnkd.in/gwJhzdXn

18

Adrian Hornsby recently posted an article on MTTR (Mean Time to Recovery), and his critique of “incident‑duration‑centered thinking” underscored how a purely engineering-focused lens can miss the bigger picture—especially when incidents ride across teams and require multiple handoffs. 👉 https://lnkd.in/eZSGgd29 🔍 MTTR is a central DORA metric—but it only scratches the surface. As systems grow more complex, it makes sense that incidents may not resolve faster, even as our resilience improves. The real question is less "How quickly did we fix it?" and more "What did we learn, and how did we come out stronger?" Adrian nicely puts it that way: "as systems continue to evolve, they also tackle more complex problems that naturally take longer to solve properly". Understanding that evolution is key to stop the clock, and start the thinking. Dave Farley—co‑author of Continuous Delivery and respected DevOps voice—has even observed that while the core DORA metrics (lead time, deployment frequency, change failure rate, MTTR) are a solid starting point, they’re best used as indicators of system health, not targets to be gamed. His caution against making metrics into goals is particularly resonant here. Adrian again: 𝘖𝘳𝘨𝘢𝘯𝘪𝘻𝘢𝘵𝘪𝘰𝘯𝘴 𝘵𝘩𝘢𝘵 𝘳𝘦𝘢𝘭𝘭𝘺 𝘤𝘢𝘳𝘦 𝘢𝘣𝘰𝘶𝘵 𝘳𝘦𝘴𝘪𝘭𝘪𝘦𝘯𝘤𝘦 𝘥𝘰𝘯'𝘵 𝘰𝘣𝘴𝘦𝘴𝘴 𝘰𝘷𝘦𝘳 𝘪𝘯𝘤𝘪𝘥𝘦𝘯𝘵 𝘥𝘶𝘳𝘢𝘵𝘪𝘰𝘯. 𝘛𝘩𝘦𝘺 𝘵𝘳𝘢𝘤𝘬 𝘩𝘰𝘸 𝘸𝘦𝘭𝘭 𝘵𝘩𝘦𝘺 𝘭𝘦𝘢𝘳𝘯 𝘧𝘳𝘰𝘮 𝘦𝘢𝘤𝘩 𝘪𝘯𝘤𝘪𝘥𝘦𝘯𝘵 𝘢𝘯𝘥 𝘩𝘰𝘸 𝘱𝘴𝘺𝘤𝘩𝘰𝘭𝘰𝘨𝘪𝘤𝘢𝘭𝘭𝘺 𝘴𝘢𝘧𝘦 𝘵𝘦𝘢𝘮𝘴 𝘧𝘦𝘦𝘭 𝘸𝘩𝘦𝘯 𝘳𝘦𝘱𝘰𝘳𝘵𝘪𝘯𝘨 𝘱𝘳𝘰𝘣𝘭𝘦𝘮𝘴: · 𝘋𝘰 𝘸𝘦 𝘥𝘦𝘵𝘦𝘤𝘵 𝘴𝘪𝘮𝘪𝘭𝘢𝘳 𝘱𝘳𝘰𝘣𝘭𝘦𝘮𝘴 𝘧𝘢𝘴𝘵𝘦𝘳 𝘦𝘢𝘤𝘩 𝘵𝘪𝘮𝘦? · 𝘈𝘳𝘦 𝘵𝘦𝘢𝘮𝘴 𝘨𝘦𝘵𝘵𝘪𝘯𝘨 𝘣𝘦𝘵𝘵𝘦𝘳 𝘢𝘵 𝘪𝘮𝘱𝘳𝘰𝘷𝘪𝘴𝘪𝘯𝘨 𝘸𝘩𝘦𝘯 𝘴𝘵𝘢𝘯𝘥𝘢𝘳𝘥 𝘱𝘳𝘰𝘤𝘦𝘥𝘶𝘳𝘦𝘴 𝘧𝘢𝘪𝘭? · 𝘞𝘩𝘢𝘵 𝘴𝘶𝘳𝘱𝘳𝘪𝘴𝘦𝘴 𝘶𝘴 𝘢𝘣𝘰𝘶𝘵 𝘰𝘶𝘳 𝘴𝘺𝘴𝘵𝘦𝘮𝘴, 𝘢𝘯𝘥 𝘩𝘰𝘸 𝘥𝘰 𝘸𝘦 𝘵𝘶𝘳𝘯 𝘵𝘩𝘰𝘴𝘦 𝘴𝘶𝘳𝘱𝘳𝘪𝘴𝘦𝘴 𝘪𝘯𝘵𝘰 𝘬𝘯𝘰𝘸𝘭𝘦𝘥𝘨𝘦? · 𝘈𝘳𝘦 𝘸𝘦 𝘳𝘦𝘥𝘶𝘤𝘪𝘯𝘨 𝘤𝘰𝘰𝘳𝘥𝘪𝘯𝘢𝘵𝘪𝘰𝘯 𝘰𝘷𝘦𝘳𝘩𝘦𝘢𝘥 𝘵𝘩𝘳𝘰𝘶𝘨𝘩 𝘣𝘦𝘵𝘵𝘦𝘳 𝘢𝘳𝘤𝘩𝘪𝘵𝘦𝘤𝘵𝘶𝘳𝘦 𝘢𝘯𝘥 𝘤𝘭𝘦𝘢𝘳𝘦𝘳 𝘰𝘸𝘯𝘦𝘳𝘴𝘩𝘪𝘱? · 𝘞𝘩𝘪𝘤𝘩 𝘳𝘦𝘤𝘰𝘷𝘦𝘳𝘺 𝘴𝘵𝘳𝘢𝘵𝘦𝘨𝘪𝘦𝘴 𝘢𝘤𝘵𝘶𝘢𝘭𝘭𝘺 𝘸𝘰𝘳𝘬 𝘶𝘯𝘥𝘦𝘳 𝘱𝘳𝘦𝘴𝘴𝘶𝘳𝘦? · 𝘏𝘰𝘸 𝘲𝘶𝘪𝘤𝘬𝘭𝘺 𝘥𝘰 𝘪𝘯𝘴𝘪𝘨𝘩𝘵𝘴 𝘧𝘳𝘰𝘮 𝘰𝘯𝘦 𝘪𝘯𝘤𝘪𝘥𝘦𝘯𝘵 𝘪𝘮𝘱𝘳𝘰𝘷𝘦 𝘰𝘶𝘳 𝘳𝘦𝘴𝘱𝘰𝘯𝘴𝘦 𝘵𝘰 𝘧𝘶𝘵𝘶𝘳𝘦 𝘰𝘯𝘦𝘴? 𝘎𝘰𝘪𝘯𝘨 𝘣𝘢𝘤𝘬 𝘵𝘰 𝘵𝘩𝘢𝘵 𝘰𝘳𝘪𝘨𝘪𝘯𝘢𝘭 𝘵𝘦𝘢𝘮: 𝘸𝘩𝘦𝘯 𝘭𝘦𝘢𝘥𝘦𝘳𝘴𝘩𝘪𝘱 𝘢𝘴𝘬𝘦𝘥 "𝘐𝘧 𝘺𝘰𝘶'𝘳𝘦 𝘳𝘦𝘢𝘭𝘭𝘺 𝘪𝘮𝘱𝘳𝘰𝘷𝘪𝘯𝘨 𝘵𝘩𝘪𝘯𝘨𝘴, 𝘸𝘩𝘺 𝘪𝘴𝘯'𝘵 𝘔𝘛𝘛𝘙 𝘨𝘰𝘪𝘯𝘨 𝘥𝘰𝘸𝘯?" 𝘵𝘩𝘦𝘺 𝘸𝘦𝘳𝘦 𝘢𝘴𝘬𝘪𝘯𝘨 𝘵𝘩𝘦 𝘸𝘳𝘰𝘯𝘨 𝘲𝘶𝘦𝘴𝘵𝘪𝘰𝘯. 𝘛𝘩𝘦 𝘳𝘪𝘨𝘩𝘵 𝘲𝘶𝘦𝘴𝘵𝘪𝘰𝘯 𝘸𝘢𝘴: "𝘈𝘳𝘦 𝘺𝘰𝘶 𝘣𝘶𝘪𝘭𝘥𝘪𝘯𝘨 𝘴𝘺𝘴𝘵𝘦𝘮𝘴 𝘵𝘩𝘢𝘵 𝘭𝘦𝘢𝘳𝘯, 𝘢𝘥𝘢𝘱𝘵, 𝘢𝘯𝘥 𝘨𝘦𝘵 𝘴𝘵𝘳𝘰𝘯𝘨𝘦𝘳 𝘢𝘧𝘵𝘦𝘳 𝘦𝘢𝘤𝘩 𝘧𝘢𝘪𝘭𝘶𝘳𝘦?" Me: We are, right?

7

Engineering teams are turning AI into real results, not just experiments. Last month at the ELC Annual in SF, Redis Principal Field Engineer Josh R. joined Vinay Perneti from Augment Code, Anand K Sankaran from Workday, and Forrest Thomas from Garner Health for a fireside chat on where AI is truly adding value. Hear how they drove adoption across their teams, improved workflows, and made AI deliver real results: https://lnkd.in/gDAjZsDF

23

I have been playing around with two parallel experiments on agent long-term memory: memdb-lite and memdb, each based on radically different assumptions. memdb-lite (https://lnkd.in/exrx4Z65) treats memory as transparent files. Knowledge is stored in plain Markdown, with `MEMORY.md` for curated facts and daily logs for temporal context. Before answering questions about prior context, the agent must search memory. This is enforceable because the skill makes recall explicit. Here, you can edit the memory with the editor. The agent’s knowledge is literally `git diff`-able. memdb (https://lnkd.in/eciH8CrY) approaches memory as a temporal knowledge graph. It features an event clock (immutable evidence of what happened) and a state clock (derived canonical truth about what’s true now). Agents are instructed to emit observed facts with full provenance, while a resolver produces facts. Multiple agents can write simultaneously while maintaining consistency. In this system, you cannot simply edit the memory; the truth is determined through policy. The divergence between these approaches is not about which is superior, but rather where the complexity resides. File-based memory pushes complexity to the agent runtime, requiring them to honor the recall contract. In contrast, graph-based memory pulls complexity outside, involving resolution policies and conflict detection. These approaches are used by the leading tools, but neither is definitively right. One optimizes for transparency and human editability, while the other focuses on coordination and audit trails. This is a case of AI as learning tool: it is trivial to build poc and learn by doing…

10