Google Cloud Blog - TheNote.app

Google Cloud Blog
Follow

cloud.google.com/blog is the official blog of Google Cloud. It provides news, updates, and insights on Google Cloud's products and services, as well as trends and innovations in the cloud computing industry. The blog features articles written by Google Cloud experts, engineers, and thought leaders, covering a wide range of topics such as artificial intelligence, machine learning, data analytics, security, and more. The articles often include technical tutorials, case studies, and best practices, making the blog a valuable resource for developers, IT professionals, and business leaders who use or are interested in Google Cloud. The blog is well-organized, with articles categorized by topic, product, and industry. Visitors can browse the latest articles, search for specific topics, or subscribe to the blog's RSS feed to stay up-to-date with the latest news and updates. Some of the key features of the blog include: - In-depth articles on Google Cloud products and services, such as Google Cloud Platform, Google Cloud Storage, and Google Cloud AI Platform - Technical tutorials and guides on how to use Google Cloud services - Case studies and success stories from Google Cloud customers - Insights and analysis on industry trends and innovations - News and updates on Google Cloud's partnerships and collaborations - Interviews with Google Cloud experts and thought leaders Overall, the Google Cloud blog is a valuable resource for anyone interested in cloud computing, artificial intelligence, and related technologies.

Cloud Blog cloud.google.com

RSS cloudblog.withgoogle.com

RSS Hunter • Aug 19, 2024

Thread Of Notes

Global AI routing with <1% overhead on multi-cluster GKE Inference Gateway

The demand for AI infrastructure, particularly accelerators, is extremely high, leading to shortages and fragmented compute capacity across global data centers. Modern AI workloads, with large context windows, consume accelerator memory rapidly, exacerbating utilization issues. To address this, a layered routing architecture was developed to treat globally scattered capacity as a single pool. This system features a multi-cluster GKE Inference Gateway for global traffic distribution and an LLM-d router for memory-aware scheduling. The architecture is designed to be agnostic to runtime, model, and accelerator types. Benchmarks on a large multi-region GKE deployment demonstrated near-linear throughput scaling and a 99.9% success rate. The GKE Inference Gateway added less than 1% overhead, delivering 99.5% of direct cluster call throughput. Traffic is routed based on live application signals, specifically KV-cache token utilization, enabling the system to intelligently spill traffic between regions. This memory-aware routing ensures efficient utilization by preventing accelerators from being stranded due to memory saturation. The payoff is maximized "intelligence per dollar" by achieving near-proportional performance gains from added accelerators and avoiding wasted capital. Key lessons emphasize the importance of smart load balancing, considering agentic workload memory bottlenecks, adapting to AI traffic latency, integrating routing with native serving patterns, and opting for open, portable stacks.

https://cloud.google.com/blog/products/containers-kubernetes/gpu-and-tpu-utilization-with-multi-cluster-gke-inference-gateway/ cloud.google.com

RSS Hunter • Yesterday

Scale your AI workloads faster and more efficiently with GKE Pod snapshots

Modern AI workloads face a performance and cost dilemma due to lengthy initialization times. Large language models and instant code execution require fast provisioning, often leading to infrastructure overspending. Google Kubernetes Engine (GKE) Pod snapshots offer a solution by saving and restoring running workload states. This feature drastically reduces AI inference startup times, loading large models in seconds. Cold starts, a problem for many applications, are particularly acute in AI due to significant GPU memory loading. Traditional methods involve overprovisioning or complex custom solutions for state restoration. GKE Pod snapshots enable efficient scaling of AI inference by eliminating redundant model loading for new replicas. A single initialization creates a persistent snapshot, allowing new instances to restore instantly. This accelerates scaling and reduces the need for expensive overprovisioning. For agentic workflows, Pod snapshots improve startup latency and manage idle sandboxes effectively. Snapshots quickly initialize new agent environments and allow idle sandboxes to suspend and resume instantly. This feature is adaptable for various workloads with long initialization phases. GKE Pod snapshots integrate seamlessly with Kubernetes workflows through declarative policies for snapshot management.

https://cloud.google.com/blog/products/containers-kubernetes/gke-pod-snapshots/ cloud.google.com

RSS Hunter • Yesterday

Maximizing Apache Spark availability: Mitigating compute stockouts with flexible VMs and other best practices

The increasing demand for compute power due to AI development is straining data processing and Apache Spark pipelines, leading to availability constraints. Google's Managed Service for Apache Spark offers flexible VMs to address this by allowing clusters to use a ranked list of acceptable machine families. This approach ensures pipelines remain operational even during capacity stockouts, which occur when demand exceeds available capacity for specific machine types. Flexible VMs enable multi-family blending, mixing nodes across different machine generations and types. They also support mixed storage, dynamically adapting to the host family's disk types, and provide comprehensive cluster coverage for all node types. A successful implementation requires careful ranking of preferred machine families to mitigate stockout risks without manual intervention. An example tiering strategy shows how to prioritize machine families and storage recommendations for production pipelines. For legacy workloads, a tiered strategy helps transition to newer, more available architectures. Adopting modern storage like Hyperdisk Balanced can unlock maximum availability with newer instance families. Key considerations include ensuring sufficient quotas for all specified machine types and leveraging Compute flexible Committed Use Discounts for cost savings. Performance can vary between machine generations and storage types, necessitating workload testing. Additional recommendations for improving resource availability include implementing AutoZone, using smaller machine shapes, deploying autoscaling, configuring partial cluster creation, and establishing regional fallbacks. By utilizing flexible VMs and these strategies, users can protect Spark workloads from hardware shortages and ensure critical pipelines continue running.

https://cloud.google.com/blog/products/data-analytics/maximize-apache-spark-availability-with-flexible-vms/ cloud.google.com

RSS Hunter • Yesterday

Strengthen your CI/CD pipeline with new Secure Source Manager capabilities

A secure software supply chain is vital for safe innovation and modern delivery. Software supply chain attacks have significantly increased. Google Cloud Secure Source Manager (SSM) helps address these threats by unifying source and CI/CD system management. SSM now offers two new capabilities to enhance development and CI/CD workflows. It can block unauthorized access to CI/CD systems, even if the corporate network is compromised. Additionally, a new Code Owners system provides granular control over pull request approvers per file and branch. This feature allows for specific approvers for different sections of code. It also supports branch-specific governance and nestable ownership files. Furthermore, independent approval sections can require sign-offs from multiple departments before merging. SSM also integrates with Developer Connect for secure CI/CD system and runtime connections.

https://cloud.google.com/blog/products/identity-security/strengthen-your-cicd-pipeline-with-new-secure-source-manager-capabilities/ cloud.google.com

RSS Hunter • Yesterday

Announcing Native BM25 Ranking in AlloyDB and Cloud SQL

Vector search is essential for generative AI and RAG, but it struggles with exact alphanumeric IDs. Hybrid search, a combination of semantic vector search and exact keyword full-text search, addresses this limitation. Previously, achieving BM25 ranking with AlloyDB or Cloud SQL required a separate full-text search backend, causing complexity and data silos. Now, a native BM25 index is available in preview for AlloyDB and Cloud SQL for PostgreSQL 17+, powered by the open-source pg_textsearch extension. This eliminates the need for redundant systems, providing industry-standard keyword ranking directly within the database. The native BM25 index offers C-optimized, fast scoring and eliminates data duplication and synchronization lag. AlloyDB also offers supercharged vector search with ScaNN and HNSW index types. Unlike PostgreSQL's ts_rank, BM25 provides superior relevance scoring by considering inverse document frequency and term frequency saturation. Hybrid search merges semantic understanding with precise keyword matching, ensuring that specific, relevant results are prioritized. Both AlloyDB and Cloud SQL now support native BM25 and offer methods to integrate it with vector search for a unified hybrid search experience.

https://cloud.google.com/blog/products/databases/native-bm25-search-in-alloydb-and-cloud-sql/ cloud.google.com

RSS Hunter • Sep 18

The DevFest Community Workshop Experience: Building Real Agents Together

The DevFest season began at Google Hudson Square in New York City with 80 engineers. This workshop introduced a novel approach called Workbench, prioritizing understanding core concepts over rote code memorization. Attendees focused on mental models for graph engineering, self-evolving architectures, and automated self-patching harnesses. The day involved building long-running, self-evolving multi-agent systems using Google's agentic stack. Ricky Robinett explained why prompt engineering falls short as a safety mechanism for agents. Rachel Francois and GDG Brooklyn organizers highlighted the importance of local developer communities and shared patterns. Annie Wang and Christina Lin led a lab where participants used Google's Agent Development Kit and other tools. This session demonstrated separating state from compute for efficient long-running tasks. Later, Logan Hennessy and Kartik Derasari facilitated a lab integrating BigQuery data into autonomous pipelines for better bidding strategies. This included implementing self-patching harnesses to detect and correct spend anomalies, with fast-paced quizzes adding an element of competitive engagement. The DevFest Workbench experience is touring five more cities this fall, with registration spots filling quickly.

https://cloud.google.com/blog/topics/developers-practitioners/the-devfest-community-workshop-experience-building-real-agents-together/ cloud.google.com

RSS Hunter • Sep 18

Reimagining service delivery in the agentic era with Google Public Sector

State and local governments aim to provide responsive, equitable, and accessible services but are hampered by legacy systems, siloed data, and administrative burdens. These issues create operational bottlenecks, such as manual data entry and repetitive verification processes for citizens. AI is now recognized as a top priority for state CIOs, who see it as crucial for increasing staff capacity and improving service delivery. The challenge lies in bridging the gap between old systems and modern AI-driven service delivery. Google offers an AI stack to help governments overcome integration challenges. Examples include Utah DOT using AI for property parcel mapping, saving years of manual work. The City of Hartford uses AI for real-time translation in public meetings, increasing participation and saving costs. Chattanooga uses AI to analyze crash data, improving street safety. INDOT automates compliance audits with AI, saving engineering hours. Los Angeles is embedding Gemini across departments to manage large-scale events and multilingual services. Maryland state developed a clean water app in five weeks using Gemini, automating administrative tasks. Google Public Sector partners with these organizations to leverage AI for improved government functions.

https://cloud.google.com/blog/topics/public-sector/reimagining-service-delivery-in-the-agentic-era-with-google-public-sector/ cloud.google.com

RSS Hunter • Sep 18

How to upskill enterprise AI builders by using daily micro habits

Enterprises struggle with generative AI adoption because developers face setup issues and return to their regular tasks, preventing successful implementation. Traditional enablement methods like week-long bootcamps are ineffective given developers' limited learning time. Google Cloud Consulting advocates for a four-pillar approach to AI enablement, focusing on quick, daily practice. This involves breaking down learning into short, single-skill objectives instead of lengthy training sessions. Pre-configured browser-based sandboxes eliminate setup hurdles, allowing immediate coding. Incorporating daily streaks and team challenges transforms practice into a routine workday activity. Each exercise should result in a functional component, cumulatively building a shared code library. The Advent of Agents program demonstrated this by offering daily, five-minute, browser-based exercises without setup, leading to significant developer participation and successful agent component creation. This data suggests developers will engage with learning when it fits their existing schedules, provided setup and scheduling barriers are removed. Effective AI enablement requires consistent practice and tools that integrate with regular work, not disrupt it. Google Cloud offers "Build with Gemini" workshops and the "Gemini Enterprise Agent Ready" (GEAR) program to facilitate this micro-enablement. By starting small with daily habits and ready-to-use tools, organizations can effectively build AI capabilities.

https://cloud.google.com/blog/topics/consulting/upskill-your-ai-using-daily-micro-habits/ cloud.google.com

RSS Hunter • Sep 18

Accelerating the borderless Lakehouse: Announcing preview of cross-cloud caching

The borderless Lakehouse enables unified access to an organization's complete data estate across multiple clouds. Previously, integrating distributed data required complex ETL pipelines and costly data transfers. The Lakehouse allows querying data in place by federating to various cloud catalogs and offers reduced transfer costs through Partner Cross-Cloud Interconnect. New enhancements further optimize cross-cloud query costs by minimizing data transfer.Cross-cloud caching for Lakehouse transparently accelerates queries and cuts remote transfer costs by locally caching frequently accessed data. This caching operates at a sub-file block granularity, encrypts data at rest, and maintains tenant and regional isolation. Freshness checks ensure data accuracy by verifying if remote data has changed before serving cached results. In practice, this caching significantly reduces the amount of data transferred, often to under 5% of the processed data.This efficiency leads to a lower Total Cost of Ownership, making cross-cloud analytics and AI viable at an enterprise scale. Combining this with Iceberg's compression can reduce network transfers to under 3% of the total data processed. BigQuery cross-cloud connections are also available in preview, allowing direct querying of non-Iceberg data in other clouds and accelerating workloads. These connections leverage BigQuery compute workers in Google Cloud regions, offering global availability and full BigQuery feature parity.

https://cloud.google.com/blog/products/data-analytics/borderless-lakehouse-cross-cloud-caching-and-connections/ cloud.google.com

RSS Hunter • Sep 18

Changing the game: How Google uses agentic AI to secure hundreds of millions of lines of code

Google is enhancing software security by integrating AI agents directly into its development process. This proactive approach aims to prevent vulnerabilities as code is written, rather than detecting them later. The system uses pervasive pre-submit scanning, evaluating every code change in real-time before it's committed. This contrasts with traditional, slower, one-off scans that often identify issues too late. To improve accuracy, they utilize localized threat models that leverage live codebase metadata and dependency information. Specialized triage agents then perform rapid, lightweight validation to confirm potential vulnerabilities. This two-step process ensures low latency without compromising precision. Automated bug-fix agents autonomously generate code patches based on scan results. These fixes are then submitted for human review, significantly shortening the detection-to-resolution cycle. This AI-native strategy has prevented hundreds of vulnerabilities monthly from entering Google's codebase. Organizations can adopt similar principles, separating systems, using context wisely, building robust harnesses, and automating fixes.

https://cloud.google.com/blog/topics/systems/using-ai-agents-to-secure-google-infrastructure/ cloud.google.com

RSS Hunter • Sep 18

Google named a Leader in the External Threat Intelligence Service Forrester Wave™

Google has been named a Leader in External Threat Intelligence Services by Forrester. The company received the highest scores in multiple criteria for its advanced threat intelligence capabilities. Google Threat Intelligence utilizes specialized AI agents for rapid, autonomous investigations and malware analysis. This is supported by Mandiant's incident response data and VirusTotal's crowdsourced visibility. Deep and dark web monitoring is a core strength, spotting threats before they escalate. Google's AI agents are powered by frontier models like Gemini, allowing continuous evolution and faster threat context. Their platform enables autonomous campaign attribution and complex malware analysis. Google's threat intelligence researchers provide evidence-based attribution mapped to MITRE ATT&CK. These capabilities enhance detection quality and streamline rule creation for security operations centers. Customers report significant improvements in proactive threat identification and team efficiency due to Google Threat Intelligence.

https://cloud.google.com/blog/products/identity-security/google-named-a-leader-in-the-external-threat-intelligence-service-forrester-wave/ cloud.google.com

RSS Hunter • Sep 17

The future of orchestration: Pine59’s journey to Airflow 3 on Google Cloud

Pine59 manages large data pipelines for location intelligence, producing metrics hourly to quarterly. Their Daily Foot Traffic metric processes up to 14 million locations per job, running on Google Cloud with BigQuery and Managed Service for Apache Airflow orchestrated by Airflow 3. To handle increasing data volumes and machine learning workloads, Pine59 modernized their monorepo containing hundreds of Directed Acyclic Graphs (DAGs). This transition significantly improved their MLOps capabilities, developer workflow, and pipeline speed.They stress-tested production workloads against the new Managed Airflow (Gen 3) architecture and observed immediate improvements in processing speed, task scheduling, and stability. The migration optimized MLOps by integrating a dedicated Google Kubernetes Engine cluster for model inference, separating orchestration from heavy ML execution. Airflow 3's improved developer workflow and user interface were leveraged through custom plugins like "BigQuery Auto-linkify" and "DAG Run Configuration Search" for faster debugging and troubleshooting. A compatibility shim layer streamlined operator migration between Airflow versions.Quantifiable results include dramatically reduced queue latency for DAG runs, with tasks starting almost immediately. The Daily Foot Traffic pipeline's completion time decreased by nearly 32%, from 38 minutes to under 26 minutes. Pine59 now runs all production workloads on their Gen 3 instance, establishing a faster, more resilient foundation for future data and AI pipelines. This move allowed their engineering team to focus more on value delivery than infrastructure management.

https://cloud.google.com/blog/topics/supply-chain-logistics/the-future-of-orchestration-pine59s-journey-to-airflow-3-on-google-cloud/ cloud.google.com

RSS Hunter • Sep 17

How a solo founder runs a five-continent tender platform on AlloyDB and MCP

Lucius AI operates a global tender intelligence platform on AlloyDB for PostgreSQL, consolidating relational data, audit logs, and vector embeddings. This unified approach eliminates the need for separate database systems. Migrating their semantic search to a ScaNN index dramatically improved query latency. Query times were reduced from 1.14 seconds to just 24 milliseconds, representing a 47x speedup. The company also connects an AI agent to AlloyDB using the Model Context Protocol (MCP). This integration automates crucial administrative tasks such as query analysis and data freshness checks. Strict least-privilege permissions were implemented to ensure secure database operations. Lucius AI's platform covers over 210,000 tenders across multiple continents. This architecture allows a solo founder to manage extensive operational coordination effectively. Future plans include leveraging automated vector embeddings and AlloyDB's columnar engine. Managing operations through AlloyDB and MCP enables a resilient multi-region platform managed by a single engineer.

https://cloud.google.com/blog/products/databases/solo-founder-runs-a-global-tender-platform-on-alloydb-and-mcp/ cloud.google.com

RSS Hunter • Sep 17

How Orange uses agents to make FinOps everyone's responsibility

Orange, a French telecom giant, successfully tackles cloud spend challenges through "FinOps Clean Days." These events involve engineering teams pausing their regular tasks to collaboratively optimize cloud costs, often with a competitive leaderboard and rewards. Experienced practitioners mentor newcomers, fostering learning through practical application and demonstrating tangible results to sponsors. This gamified approach has significantly boosted the engagement and satisfaction of Orange's FinOps community, achieving a high Net Promoter Score.The company's success is rooted in two core FinOps principles: shared responsibility for cloud costs and the necessity of communication for driving change. Orange emphasizes that FinOps, like cloud security, involves every stakeholder within a project. To facilitate this, they established a FinOps Community of Practice, utilizing standardized communication channels and providing actionable updates to ensure ongoing value.The concept of Clean Days arose from understanding the operational realities of agile teams, where optimization often gets sidelined by delivery priorities. By creating dedicated, collaborative, and enjoyable time for these activities, Orange overcomes this common hurdle. This strategy aligns with McKinsey's framework for change management, addressing conviction, formal mechanisms, role modeling, and talent development. Gamification and visible successes build conviction and role modeling, while peer-to-peer guidance develops skills.Extending this engagement beyond a core group of 100 people requires leveraging AI agents. These agents can handle time-intensive tasks, reduce friction for engineers, and address persistent challenges like awareness gaps, bandwidth limitations, or complexity in reporting. The strategic adoption of AI agents should begin with informing and suggesting capabilities, progressively moving towards agents that execute changes once trust is established and operational risks are managed.Building these AI solutions can be done with varying levels of coding, such as the no-code Gemini Enterprise App or the more advanced Gemini Enterprise Agent Platform for developers. Ultimately, effective Cloud FinOps is shifting from centralized reporting to decentralized action, driven by a strong cultural foundation augmented by AI agents. Orange's achievement underscores the importance of cultivating this community-driven approach as the crucial first step.

https://cloud.google.com/blog/topics/telecommunications/how-orange-uses-agents-to-make-finops-everyones-responsibility/ cloud.google.com

RSS Hunter • Sep 16

M4N VM family, now GA: Highest per-core IOPS and throughput for I/O and memory-bound workloads

Enterprise applications often face storage I/O and memory access bottlenecks, leading to over-provisioning of compute cores and increased software licensing costs. Google Cloud's new M4N machine series in Compute Engine is designed to address these challenges. These instances are purpose-built for I/O intensive, high-memory workloads, offering a significant advantage over competitors. M4N provides the highest per-core IOPS and throughput for high-memory instances. It also offers over a 20% total cost of ownership reduction for Oracle databases.The M4N instances feature higher memory ratios, up to 26:1, and can scale up to 6TB of memory. Powered by Intel's 5th Gen Xeon Scalable processors and Google Cloud's custom Titanium offload architecture, they deliver exceptional host storage performance. When paired with Hyperdisk Extreme, M4N instances achieve up to 25 GiB/s of aggregate storage performance and 1 million IOPS. This new offering complements existing memory-optimized families, specifically targeting storage and network bottlenecks.M4N is ideal for mission-critical databases, generative AI data layers, enterprise healthcare, and real-time analytics. Its high memory-to-core ratio allows for rapid data ingestion and zero-stall backup cycles. The instances also support sub-millisecond similarity searches for massive vector indexes and 400 Gbps network bandwidth. This capability is crucial for distributed model retrieval in generative AI.For enterprise healthcare and ERP systems, M4N ensures sustained I/O headroom, preventing latency spikes during peak usage. This is vital for applications like Epic Systems and SAP S/4HANA. Real-time analytics and Electronic Design Automation benefit from high memory capacity and maximum storage bandwidth for loading massive datasets. This allows for efficient checkpointing and in-memory processing.The M4N instances significantly optimize Oracle licensing costs by enabling right-sizing of infrastructure. Enterprises can reduce over-provisioning of cores while still meeting performance density requirements. This leads to a substantial TCO reduction compared to other hyperscalers. With more memory per core, larger portions of databases can reside in the system global area, reducing expensive I/O operations.Early customer experiences highlight M4N's role in transforming workloads by delivering high throughput across various instance sizes. The new compute platform demonstrates potential for scalable, resilient, and performant enterprise Oracle deployments. M4N bridges the gap between large memory footprints and extreme I/O performance, minimizing bottlenecks without additional infrastructure. This makes it a powerful solution for demanding data environments. The M4N instances are now available in select regions globally.

https://cloud.google.com/blog/products/compute/compute-engine-m4n-vms/ cloud.google.com

RSS Hunter • Sep 16

Cloud CISO Perspectives: How Google monitors AI threats and advances AI defenses

This article discusses Google's dual role as an AI developer and security company, providing unique insights into AI threats. Sandra Joyce highlights three key shifts: AI reshaping software development, expanding attack surfaces, and enhancing threat capabilities. Attackers are contaminating AI-assisted coding tools and targeting AI agents through methods like prompt injection. Google is building security directly into AI development pipelines with real-time "spellcheck for cybersecurity" features. A singular AI model for security is insufficient; Google uses a multi-model approach for comprehensive vulnerability detection. Adversaries are exploiting AI workloads by targeting GPU access and stealing valuable AI data, models, and credentials. Google advocates for a unified security graph connecting code, models, data, and identities to defend against these expanding threats. Manual incident response is being replaced by machine-speed operations to counter AI-driven attacks. Defenders possess an advantage due to their deep internal context, which can be fed into AI models for faster, more accurate security. Google actively defends the ecosystem by disabling malicious infrastructure, hardening its AI models, automating vulnerability hunting, and developing advanced defenses. The article emphasizes that securing the AI era requires AI-powered defenses and a transition to continuous, machine-speed protection.

https://cloud.google.com/blog/products/identity-security/cloud-ciso-perspectives-how-google-monitors-ai-threats-advances-ai-defenses/ cloud.google.com

RSS Hunter • Sep 16

For SeaVerse, GKE Agent Sandbox reduces infrastructure costs by 60%

SeaVerse is a gaming startup building a platform for playable AI experiences created by users from simple prompts. Users can play games, chat with AI characters, or create their own interactive apps and visual patterns. The platform prioritizes a fast and easy creative loop where ideas can be quickly refined and published without traditional coding. To achieve this, SeaVerse required robust infrastructure capable of running dynamic, multi-tenant AI workloads with strong isolation and low latency. They adopted Google Kubernetes Engine (GKE) and GKE Agent Sandbox to meet these demands. GKE provided a managed foundation for Kubernetes operations, freeing their engineers from cluster maintenance. GKE Agent Sandbox offered deep kernel-level isolation for dynamic sandbox workloads using technologies like Kata Containers and gVisor. This solution enabled SeaVerse to achieve strong security boundaries without compromising user experience speed. The platform now boasts improved observability into sandbox performance, allowing for faster issue resolution. This enhanced visibility transformed previously opaque sandbox failures into transparent processes. Furthermore, GKE Agent Sandbox provided cost flexibility, enabling SeaVerse to better match resources to workloads and reduce infrastructure costs by up to 60%. The platform now supports persistent storage for evolving creations, enhancing the creator experience. SeaVerse aims to make AI experiences immediate, expressive, and connected, with GKE and GKE Agent Sandbox providing a strong foundation for future growth and innovation.

https://cloud.google.com/blog/products/containers-kubernetes/seaverse-chooses-gke-agent-sandbox/ cloud.google.com

RSS Hunter • Sep 16

Introducing new session management tools with native, granular controls

Google Cloud has enhanced its session management to bolster security and mitigate credential theft. A default 16-hour session length is now applied to all customers who hadn't previously configured their own. This security standard has been extended globally to ensure broader protection. The session controls have evolved beyond simple administrative settings to become a granular feature integrated with Context-Aware Access. This update provides administrators with increased flexibility, better automation, and a more streamlined security workflow. Support for infrastructure as code through Terraform, gcloud CLI, and REST APIs enables programmatic management of session policies. Policies can now be targeted with precision using Google Groups, allowing for distinct session lengths for different user groups like elevated privilege users. Session controls can also be applied to specific applications such as the Google Cloud Console and gcloud CLI, preventing over-restrictive blanket policies. Administrators can now manage these policies directly within the Google Cloud Console via Access Context Manager, offering a unified experience. These changes allow for tighter reauthentication boundaries in high-risk areas without hindering developer productivity.

https://cloud.google.com/blog/products/identity-security/introducing-new-session-management-tools-with-native-granular-controls/ cloud.google.com

RSS Hunter • Sep 15

Introducing Filestore agent volumes: fully managed storage for agent workspaces

Executing data-driven tasks is crucial for enterprises, yet current agent storage solutions involve complex workarounds like shuttling state or manual disk management. These methods lead to trade-offs in latency, operational complexity, and cost as agent fleets scale. Organizations need storage that offers strict isolation, instant session resumption, pay-per-use economics, and fluid multi-agent collaboration. To address these needs, Google Cloud is introducing Filestore agent volumes, a fully managed, high-performance, elastic file storage solution. This new capability is specifically designed for scaling agentic workloads on Google Cloud, complementing dedicated compute environments like Agent Substrate on GKE. Filestore agent volumes provide instantaneous access to isolated, persistent file storage for GKE sandboxes, automating the entire volume lifecycle. Key benefits include granular isolation and enterprise guardrails, sub-second session resumption for cost-saving, smart lifecycle economics with pay-per-use pricing, and multi-agent collaboration through native RWX support. This solution unlocks various agentic workloads, such as software engineering sandboxes, collaborative multi-agent swarms, and interactive long-horizon workflows. Filestore agent volumes are currently available for non-production workloads, with production support via allowlist, offering seamless integration with GKE for scalable agent platforms.

https://cloud.google.com/blog/products/storage-data-transfer/filestore-agent-volumes/ cloud.google.com

RSS Hunter • Sep 15

Agent Substrate brings high-density, scalable, trusted infrastructure to GKE

Agent Substrate, an open-source agent execution runtime, is now available on Google Kubernetes Engine (GKE). This technology is designed for running millions of sandboxed autonomous agents with significantly higher density than standard containers. Agent Substrate boasts rapid suspend and resume operations, achieved through a native zero-trust kernel and network isolation. Leading AI teams, like Nous Research behind the top-ranked Hermes Agent, are already leveraging Agent Substrate. Scaling agent platforms presents challenges such as opaque trust boundaries and tool access friction. Traditional architectures struggle with massive bursts of agent activity and idle compute resources. Agent Substrate addresses these issues by decoupling agent execution from machine management. It offers security by default using microVMs or gVisor sandboxes, along with granular network controls. The system ensures sub-second activation of agents and high efficiency by suspending idle agents to free up compute. It runs on any Kubernetes cluster and is optimized for GKE, utilizing custom ComputeClasses and Arm-based processors for cost-effectiveness. Agent Substrate provides a scalable and reliable foundation for building production agent applications.

https://cloud.google.com/blog/products/containers-kubernetes/agent-substrate-available-on-gke/ cloud.google.com

RSS Hunter • Sep 15

Scaling Telco Autonomy: Leveraging GNNs with Distributed GraphFlow

The telecommunications industry is transitioning to autonomous network operations, driven by the increasing complexity of modern networks. Traditional manual methods and basic machine learning are insufficient for automating these sophisticated systems. AI agents offer reasoning capabilities but struggle with the massive scale of network data. To bridge this gap, Graph Neural Networks (GNNs) are being integrated with AI agents to process large-scale temporal and relational data. This combination allows for advanced diagnostics like root cause analysis and capacity planning, along with reasoning and action execution. Google Cloud's Autonomous Network Operations framework uses a three-component approach: Data, ML, and AI. The foundation is a network digital twin, a dynamic graph representation of the live network, hosted on Google Cloud's Spanner Graph for scalability and multi-model support. The ML layer utilizes Distributed Graph Flow (DGF), an open-source Python library for managing GNN lifecycles, enabling predictive insights. DGF simplifies GNN model development, allowing for applications like anomaly detection, root cause analysis, and predictive maintenance. It also facilitates "what-if" analyses by simulating network changes and their impacts. A five-step process outlines implementing root cause analysis using DGF and GNNs, involving connecting to the digital twin, training a prediction model, and deploying it for real-time inference. This GNN-powered evolution promises shorter outage times, optimized capacity, and improved customer experience.

https://cloud.google.com/blog/products/databases/run-gnns-at-scale-with-ease-introducing-distributed-graphflow/ cloud.google.com

RSS Hunter • Sep 15

Best practices for handling cloud reliability incidents

Cloud outages can significantly disrupt services, ranging from minor to widespread impacts. A structured workflow of Verify, Investigate, Report, Resolve, and Review is recommended for handling these incidents. Proactive preparation, including designing for failure and practicing response steps, is crucial. This guide focuses on foundational best practices for reliability incident handling, excluding security-specific measures. Advanced techniques involving AI agents are also emerging in Site Reliability Engineering. The initial step is preparation, encompassing design for automated responses, ensuring readily available and replicated data, developing clear playbooks with defined roles, and conducting regular training drills.The next phase is verification, where detected disruptions must be attributed to either Google, the user, or a third party. Checking Personalized Service Health and the public Cloud Service Health dashboard helps determine if Google has declared an incident. Investigating involves analyzing metrics and logs to pinpoint the cause, especially if Google hasn't declared an incident, and considering recent changes. Reporting to Google is necessary if services are failing despite green health dashboards, with appropriate priority setting and detailed case submission.Resolution involves communicating with stakeholders, failing over to secondary systems if possible, and identifying workarounds. Companies must also consider regulatory reporting requirements. After stability is restored, a blameless post-mortem analysis is essential to identify areas for improvement in playbooks, tools, and training. This review process helps refine incident response capabilities for future disruptions.

https://cloud.google.com/blog/topics/developers-practitioners/cloud-reliability-incident-handling-best-practices/ cloud.google.com

RSS Hunter • Sep 15

Agent-ready analytics: Unlocking insights with BigQuery augmented analytics

BigQuery has introduced augmented analytics Table-Valued Functions (TVFs) to automate complex data analysis. These functions combine AI, ML, and statistical methods for faster insight discovery directly within your data. Six new TVFs are available: AI.KEY_DRIVERS identifies metric change drivers, AI.CAUSAL_EFFECT quantifies action impacts, ML.CORRELATION measures metric relationships, ML.DETECT_CHANGE_POINTS finds data shifts, ML.TREND separates underlying growth from noise, and ML.SEASONALITY uncovers predictable cycles. These TVFs can be chained together for multi-step analysis. An example demonstrates using ML.DETECT_CHANGE_POINTS, AI.KEY_DRIVERS, and AI.CAUSAL_EFFECT to analyze bikeshare trip data. This chaining process first identifies significant shifts in trip volume, then pinpoints the factors driving those shifts, and finally quantifies the actual business impact. The TVFs are also integrated with Conversational Analytics, allowing natural language queries to trigger complex, multi-step data investigations. Examples illustrate how conversational prompts can leverage these TVFs for correlation and trend analysis on taxi and liquor sales data. These functions significantly enhance the ability to understand and act on data-driven insights efficiently.

https://cloud.google.com/blog/products/data-analytics/bigquery-augmented-analytics-tvfs/ cloud.google.com

RSS Hunter • Sep 14

Google is a leader in The Forrester Wave™: Public Cloud Platforms, Q3 2026

Google Cloud has been recognized as a Leader in the Forrester Wave™: Public Cloud Platforms, Q3 2026 report. The company achieved the highest score in the 'current offering' category. It also received top marks in 23 out of 30 evaluated criteria, highlighting strengths in AI development, databases, analytics, and security. This recognition underscores Google Cloud's integrated platform strategy for the agentic era. Their infrastructure has been co-designed over a decade for services like Gemini and Search. This integrated AI stack offers compute, orchestration, and networking for enhanced value. Google Cloud is well-positioned for enterprises focused on rapid innovation and AI-enabled platforms. The platform supports running autonomous agents on a secure and flexible foundation, leveraging Kubernetes. Recent enhancements include GKE Agent Sandbox for secure code execution and GKE Pod Snapshots to reduce compute costs. The Agentic Data Cloud unifies transactional and analytical data for real-time context. Google Cloud excels in database, analytics, and data governance services, offering a unified data estate. This allows agents to query live enterprise data without costly movement, fostering deterministic responsiveness.

https://cloud.google.com/blog/products/compute/forrester-wave-public-cloud-platforms-q3-2026-report/ cloud.google.com

RSS Hunter • Sep 14

Announcing Pause/Resume and NVIDIA RTX PRO 6000 Blackwell GPU support in Dataflow

Enterprises need serverless platforms for efficient AI and agentic workflows, with data preparation being a key challenge. Dataflow on Google Cloud facilitates batch and streaming pipelines for analytics and AI. Significant enhancements now address compute efficiency for long-running batch jobs and boost inference power for AI workloads. The general availability of Pause/Resume for Dataflow batch jobs allows users to restart failed jobs from their last checkpoint, saving compute and improving productivity. This feature also enables dynamic reallocation of compute resources for higher priority AI tasks. Furthermore, Dataflow now supports G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell GPUs. These new GPUs offer substantial performance improvements, including increased memory and bandwidth, for AI inference. This allows Dataflow jobs to handle models with over 70 billion parameters directly. Customers can leverage native Dataflow ML capabilities like RunInference and GPU-enabled autoscaling with these new GPUs. Together, Pause/Resume and the new GPUs optimize batch job costs and accelerate demanding AI workloads. These advancements aim to simplify AI development and deployment for customers.

https://cloud.google.com/blog/products/data-analytics/new-dataflow-features-to-enable-large-scale-ai-workloads/ cloud.google.com

RSS Hunter • Sep 14

3 Highlights from Thomas Kurian’s Keynote at the Goldman Sachs Communicopia & Technology Conference

Thomas Kurian presented Google Cloud's strategy at the Goldman Sachs Tech Conference, emphasizing their full-stack AI approach. Google Cloud offers solutions across the entire AI stack, differentiating them by performance, cost, and quality, while expanding their market and revenue. The company boasts seventeen product lines exceeding one billion dollars in revenue, with customers often surpassing their commitments by over fifty percent. Significant growth is evident in deals valued between $100 million and $1 billion, with over 300 customers holding commitments exceeding $100 million. Google Cloud's AI infrastructure, built on differentiated products, lowers costs and improves performance for AI models. Their AI servers have a two-year payback period, with TPUs offering an even faster return. The majority of their AI infrastructure total contract value comes from committed five-year agreements. Gemini Enterprise adoption is strong, providing cost-efficient, controlled business insights. Customers utilizing Google Cloud's AI products are more engaged, using 1.8 times as many products. These statements were forward-looking and subject to various risks and uncertainties.

https://cloud.google.com/blog/topics/inside-google-cloud/highlights-from-the-goldman-sachs-communicopia-and-technology-conference/ cloud.google.com

RSS Hunter • Sep 11

Introducing the Google Cloud Developer Plugin for AI Coding Agents

Agent skills enhance AI workflows, but managing them individually can be cumbersome. Plugins are introduced to solve this by bundling related skills and tools into installable units. Google Cloud now offers a new plugin for AI coding agents, empowering them with Google Cloud-specific capabilities. This plugin addresses the challenge of tool coupling by packaging related functionalities together. It ensures agents can effectively retrieve Google Cloud skills, use documentation, and interact programmatically with the platform. The plugin is built on the open Agent Plugins specification, promoting interoperability across AI coding environments. The flagship plugin, google-cloud-developer, assists agents with fundamental Google Cloud interactions like authentication and project management. It also bundles the Developer Knowledge MCP server for up-to-date documentation. The plugin streamlines processes like project onboarding and identity authentication through environment awareness and best practice reviews. Installing Google Cloud plugins is straightforward via CLI commands or marketplace additions. Users are encouraged to try the new plugin and explore further resources for guided setup and deeper insights.

https://cloud.google.com/blog/topics/developers-practitioners/introducing-the-google-cloud-developer-plugin-for-ai-coding-agents/ cloud.google.com

RSS Hunter • Sep 10

Enterprise-grade PostgreSQL with AlloyDB Omni RPM Orchestrator is generally available

The general availability of the AlloyDB Omni Red Hat RPM orchestrator now provides production-ready security and low-downtime operations for PostgreSQL workloads. This release supports four deployment modes, including standalone RPM and containerized options with Kubernetes. The RPM deployment is ideal for organizations needing cloud performance with local control, especially for workload modernization and regulated environments. It also enables edge and on-premises deployment for low-latency processing and disconnected operations. The orchestrator simplifies cluster provisioning and lifecycle management, offering flexible reference architectures. High availability is achieved through synchronous replication and automatic failure handling, while read pools allow for dynamic scaling of read workloads. Data protection is enhanced with automated backup and restore capabilities, supporting point-in-time recovery to GCS or S3-compatible storage. SELinux enforcement is provided for strict compliance, and low-downtime maintenance automates updates and scaling with rollback support. The orchestrator simplifies dynamic configuration tuning and cluster maintenance by adding or removing nodes. It offers advanced logging, custom metrics for rich observability, and integrates AlloyDB's AI features for generative AI applications.

https://cloud.google.com/blog/products/databases/alloydb-omni-rpm-orchestrator-is-generally-available/ cloud.google.com

RSS Hunter • Sep 9

Google is a Leader in the 2026 Gartner® Magic Quadrant™ for Enterprise AI Assistants

Gartner has named Google a Leader in its inaugural 2026 Magic Quadrant for Enterprise AI Assistants, recognizing its Completeness of Vision and Ability to Execute. Gemini Enterprise is highlighted as a unified platform that securely integrates AI into daily work, moving beyond basic chat to automate complex workflows with AI agents. It connects with existing tools and infrastructure, offering scalability and cost-effectiveness with built-in security and governance. Gartner praises Gemini Enterprise for its unified AI front door, open connectivity to various systems, and straightforward pricing. The platform also provides robust governance out-of-the-box and benefits from Google's full-stack AI infrastructure. Recent advancements include tailored industry solutions for legal and financial services, developer tools like Google Antigravity, and improved FinOps and cost controls. Customers like Accenture, Cleary Gottlieb, Deutsche Bank, and Williams & Connolly are already experiencing significant value and efficiency gains. Gemini Enterprise enables organizations to deploy governed, connected agents across all teams. Readers can download the full Gartner report for a comprehensive analysis of the enterprise AI assistant market.

https://cloud.google.com/blog/products/ai-machine-learning/google-is-a-leader-in-2026-gartner-magic-quadrant-for-enterprise-ai-assistants/ cloud.google.com

RSS Hunter • Sep 9

Beyond DMS: Accelerating Migrations SQL Server Logins and Users to Cloud SQL

Database Migration Service efficiently replicates data but excludes server-level objects like logins for security and compliance reasons. This deliberate omission prevents unauthorized privilege escalation by ensuring that on-premises privileges do not automatically transfer to the cloud. Migrating encrypted password hashes without oversight also violates compliance frameworks like PCI-DSS and SOC 2. The process also allows for identity modernization by removing stale credentials and shifting to cloud-native authentication. SQL Server separates logins, which authenticate connections at the server level, from users, which authorize actions within databases. A Security Identifier (SID) links server logins to database users. When databases are migrated without their corresponding logins, or with mismatched SIDs, users can become orphaned. The sp_help_revlogin script from Microsoft generates T-SQL statements to recreate logins with their original password hashes and SIDs. This script is run on the source instance to generate the migration query. This generated query is then executed on the destination Cloud SQL instance to create the necessary logins. If users become orphaned, they can be remapped to the correct server login using the ALTER USER command. For enhanced security, consider migrating to Customer-Managed Active Directory for centralized, Kerberos-based authentication.

https://cloud.google.com/blog/products/databases/how-to-replicate-sql-server-logins-and-passwords-to-cloud-sql/ cloud.google.com

RSS Hunter • Sep 9

How Airtel delivered its flawless Indian Premiere League 2026 cricket broadcasts

Airtel aimed to provide flawless live streaming for the Indian Premier League cricket season to millions of fans. They partnered with Google Cloud for the IPL 2026 season's digital delivery. The streaming infrastructure handled hundreds of petabytes of data across 74 matches, with the final match processing billions of requests. Delivering video during massive concurrency spikes required an edge architecture focused on localization and proactive monitoring. Airtel utilized Google Cloud's Media CDN and its extensive global edge network for deep local edge proximity and excellent cache efficiency. This resulted in 99.9% of tournament traffic being served locally within India, minimizing network hops. The architecture achieved superior caching efficiency with over 98% cache hit ratio, keeping origin server demands minimal. Consistent ultra-low latency, specifically p99 below 300ms, ensured fast stream start times and reduced buffering. Beyond architecture, meticulous operational preparation and proactive match-day execution were vital. Integrated operational support included pre-tournament readiness reviews and continuous Monitoring as a Service (MaaS). Dedicated match-day and weekend support ensured real-time event assistance from joint engineering teams.

https://cloud.google.com/blog/products/media-entertainment/how-airtel-delivered-its-flawless-indian-premiere-league-2026-cricket-broadcasts/ cloud.google.com

RSS Hunter • Sep 9

Spanner: Removing cumulative mutation limits for DML transactions

Google Cloud Spanner, a globally consistent relational database, has updated its transaction handling for larger, more complex operations. Previously, Spanner limited total transaction modifications to 80,000 "mutation mods." This limit, calculated as rows multiplied by columns plus indexes, could be reached by evolving applications. The new flexibility shifts this 80,000 mutation mod limit from the entire transaction to individual Data Manipulation Language (DML) statements. This means a single transaction can now contain multiple DML statements, provided each statement adheres to its own 80,000 mod limit. This allows developers to group DML statements logically by business requirements without artificial splitting. The change is backward compatible with existing Spanner client libraries, requiring no application code updates. However, developers should be mindful that larger transactions hold locks longer, potentially increasing aborts. The 80,000 mod limit applies individually to DML statements but still applies to the entire set of mutations in a single commit call using the Mutation API. Developers can monitor the total mods via mutation_count in CommitStats. Single DML statements exceeding 80,000 mods will still error, and other transaction limits like byte size remain active. Best practices include monitoring CommitStats and optimizing large operations. Spanner remains a reliable choice for scalable, always-on operational applications, now offering enhanced flexibility for developers.

https://cloud.google.com/blog/products/databases/spanner-removes-dml-mutation-limits/ cloud.google.com

RSS Hunter • Sep 9

Power agent hubs or custom harnesses with the Antigravity SDK in one toolkit

Enterprise agent adoption varies, with some opting for managed platforms and others building custom agent hubs. Developers creating their own hubs need reliable tools for predictable execution, logging, and sandboxing. The Antigravity SDK provides a runtime engine and CLI for building these hubs, offering safety policies, telemetry, and stateful persistence. It powers a multi-agent control plane that monitors and manages LLM workloads. This control plane includes the Antigravity SDK agent core for model interactions and tool execution, and an observability layer using Lifecycle Hooks for real-time telemetry. A use case involves operations engineers monitoring multiple agents, which traditionally presents challenges like fragmented logs and lack of visibility. The Antigravity SDK offers a solution with a streaming API, lifecycle hooks, a policy engine, skills management, and session state for a unified view. Behind the scenes, the runtime handles session initialization, skill resolution, concurrent stream generation, and sandboxed tool execution. Telemetry interception through lifecycle hooks allows for real-time observation and steering of agent execution. The SDK is organized into four core building blocks: modular skills, sandboxed tools, session isolation, and real-time telemetry.

https://cloud.google.com/blog/topics/developers-practitioners/power-agent-hubs-or-custom-harnesses-with-the-antigravity-sdk/ cloud.google.com

RSS Hunter • Sep 8

How KDDI built Buffmee, a faster, reliable consumer RAG app

Telecommunications company KDDI faced a challenge balancing quality and speed in their generative AI application, Buffmee. Buffmee is designed to help users grow by providing information grounded in over 100 diverse sources, ensuring reliability and preventing hallucinations. Initially, KDDI struggled with latency issues and the need to process a vast amount of proprietary content. To overcome these obstacles, they implemented an automated evaluation framework and performance optimization techniques.This systematic approach led to significant improvements, reducing total application response latency by 38% and improving Time To First Token (TTFT) by nearly 18%. KDDI utilized the Gemini Enterprise Agent Platform Evaluation Service for automated testing, replacing manual efforts with a data-driven process that improved groundedness scores by 25%. They adopted a binary evaluation system, strategic content sampling based on format and media composition, and calibrated thresholds based on product judgment.Bottlenecks were identified and resolved using BigQuery Agent Analytics and the Agent Development Kit (ADK) log analysis agent, which helped pinpoint prompt bloat and skill division issues. Large system prompts were modularized into ADK Skills to avoid attention drift and optimize response times. These techniques allowed KDDI to create a RAG system where ingested content instantly functioned, providing users with a safe and personalized way to explore media. The success of Buffmee demonstrates the importance of systematic evaluation and performance analysis for reliable generative AI applications.

https://cloud.google.com/blog/topics/customers/how-kddi-optimized-rag-performance-with-agent-development-kit/ cloud.google.com

RSS Hunter • Sep 8

Agentic analytics with the Data Agent Kit

Data practitioners face challenges with open-ended questions requiring data from diverse sources. The Data Agent Kit addresses this by enabling data workflows directly within an IDE. It utilizes the Model Context Protocol (MCP) to connect agents to various data systems. Skills, defined in markdown, teach the agent how to interact with a specific data stack. Instead of manual querying, agents execute queries and retrieve results on behalf of the user. For instance, an investigation into a dropped average order value involved querying BigQuery, Cloud SQL, and Cloud Storage. The agent discovered a new wholesale channel with a low average order value due to a promotional code. It then identified new wholesale accounts and confirmed the promotional code’s impact. The entire investigation, from data analysis to executive summary generation, occurred within a single chat session. Finally, the agent can transform ad-hoc analysis into a reproducible dbt project, including writing and correcting data quality tests.

https://cloud.google.com/blog/products/data-analytics/agentic-analytics-with-the-data-agent-kit/ cloud.google.com

RSS Hunter • Sep 8

GTIG AI Threat Tracker: From Prompting to Autonomy – The Evolution of Adversarial AI

Adversarial actors are rapidly evolving their techniques, moving from basic AI prompting to sophisticated agentic AI workflows and automation. This evolution significantly reduces response windows for defenders. In Q2 2026, threat actors executed a mass credential harvesting operation using AI in under six hours after compromising a cloud resource. Furthermore, threat actors are increasingly targeting AI assets, including proprietary models, source code, and cloud compute resources for espionage, extortion, and theft.Software supply chain risks are expanding due to the integration of AI-assisted coding tools and open-source software. Adversaries are actively targeting developers, AI coding assistants, and AI security scanners to compromise open-source software. Proprietary AI intellectual property, such as models and code, is also becoming a prime target across various sectors. The shift towards agentic AI and automation enables adversaries to autonomously manage complex attack pipelines and resolve operational errors.AI serves as a force multiplier across the entire attack lifecycle, from reconnaissance to malware obfuscation and post-exploitation activities. Threat actors are also experimenting with AI to scale information operations campaigns. To bypass access costs and circumvent security measures, adversaries are procuring illicit accounts and hijacking enterprise cloud infrastructure for unauthorized AI computations. This intensified threat landscape necessitates robust defense strategies, including proactive model safeguards, specialized threat intelligence, and targeted containment. Google is implementing these multifaceted defenses to protect its customers and infrastructure. These advancements highlight a critical need for enhanced security measures in the rapidly evolving AI landscape.

https://cloud.google.com/blog/topics/threat-intelligence/from-prompting-to-autonomy-the-evolution-of-adversarial-ai/ cloud.google.com

RSS Hunter • Sep 8

How Yahoo optimizes resources with flexible VMs in Managed Service for Apache Spark

Yahoo, a global media and technology company, faced challenges with brittle data pipelines due to fixed virtual machine (VM) configurations in their high-speed analytics workloads. When regional capacity constraints occurred, cluster provisioning in Managed Service for Apache Spark could stall, delaying critical data pipelines. To address this, Yahoo implemented flexible VMs in their Managed Service for Apache Spark clusters. This allows the system to automatically absorb resource fluctuations by defining a ranked list of acceptable VM shapes. Enabling Auto-Zone placement helps the system search for capacity across an entire region, maintaining pipeline execution without manual intervention. This optimization is part of Yahoo's broader data modernization journey, migrating on-premises big data estates to Google Cloud. Flexible configurations enhance cluster creation success by selecting from a ranked list when a preferred VM type is unavailable. They also improve regional resource utilization by searching the entire region for capacity. Configuring flexible clusters requires enabling auto-zone placement and maintaining core and memory symmetry across VM shapes. It's also crucial to align component properties like YARN and Spark resource allocations. Yahoo successfully reduced cluster provisioning failures caused by regional stockouts by 85% by implementing this flexible VM approach. This ensures continuous workload execution and prevents downstream processing delays across their massive data pipelines.

https://cloud.google.com/blog/products/data-analytics/how-yahoo-optimizes-apache-spark-with-flexible-vms/ cloud.google.com

RSS Hunter • Sep 4

Spanner migrations: Automating dual-write with Antigravity CLI for minimal disruption

Google's Finance Engineering team needed to modernize their legacy data layer with Spanner, a globally distributed and strongly consistent database. Migrating without downtime presented a significant challenge, as manually rewriting dual-write logic across numerous Data Access Objects (DAOs) was slow and error-prone. To overcome this, they developed an automated refactoring pipeline using Antigravity CLI in headless mode. This approach accelerated their migration velocity and ensured data parity in staging environments. The migration involved three phases: historical data backfill, dual-write/dual-read implementation, and automated API verification. At their scale, manual implementation of each DAO's unique requirements, like MutationConverters and rollback logic, was impractical, estimated to take months. They standardized their DAO refactoring pattern by decoupling Spanner schema translation into dedicated MutationConverter units. This created a deterministic contract for an AI coding agent to generate code reliably. Antigravity CLI's headless mode allowed for repeatable refactoring across many files without human prompts. Deterministic prompts, batch execution with automated verification, and overnight execution at scale significantly reduced migration effort. This resulted in a highly reliable data migration with sustained data fidelity and allowed engineers to focus on higher-value tasks. Key takeaways emphasize decoupling schema translation, moving from interactive chat to headless automation, and using build systems as guardrails for AI code generation.

https://cloud.google.com/blog/topics/developers-practitioners/using-antigravity-cli-to-streamline-dual-write-database-migration/ cloud.google.com

RSS Hunter • Sep 4

Not All LLM Workloads Are Equal: Benchmarking TPU Performance on Classification vs. Generation

Enterprise production deployment of LLMs reveals infrastructure's critical role in performance and cost. Not all LLM requests stress hardware equally, making standard benchmarks insufficient. This study compares Gemma 3 12B and 27B on Google Cloud TPU v6e to understand infrastructure performance at scale with distinct workloads. For decode-heavy generation, the Gemma 3 12B model scales better at high concurrency, unlike the 27B model which saturates. For prefill-heavy classification tasks, both 12B and 27B models show similar scaling, allowing larger models without a throughput penalty. Hardware saturation causes latency spikes and silent request drops, necessitating scaling based on End-to-End latency, not just CPU/memory. The architecture comprises GKE, TPU v6e, vLLM serving framework, and Gemma 3 models. Classification tasks have high input and low output, while generation tasks have low/medium input and high output. Benchmarks show the 12B model's superiority in generation scaling at high concurrency, whereas both models perform similarly in classification. To optimize, align model selection and serving configurations with specific workload input/output profiles.

https://cloud.google.com/blog/topics/developers-practitioners/not-all-llm-workloads-are-equal-benchmarking-tpu-performance-on-classification-vs-generation/ cloud.google.com

RSS Hunter • Sep 4

Announcing the Google Gen AI SDK for Kotlin 1.0: Idiomatic multiplatform access to Gemini

Google has released version 1.0 of its Kotlin Multiplatform SDK for generative AI. This SDK allows developers to easily integrate Gemini models into both JVM and Android applications. It offers idiomatic Kotlin features like Coroutines and Flow for asynchronous operations. The SDK unifies access to both Google AI Studio and Google Cloud's Gemini Enterprise Agent Platform. It supports unary and streaming text generation, as well as multi-turn chat conversations with managed context. Multimodal analysis is enhanced with Google Search grounding for fact-checking. The SDK also provides full support for Gemini's image generation models. Real-time, bidirectional interactions are possible through Gemini Live API via WebSockets. Structured tool and function calling capabilities are included for agentic workflows. The release offers a clean, multiplatform foundation for building generative AI applications.

https://cloud.google.com/blog/topics/developers-practitioners/announcing-the-google-gen-ai-sdk-for-kotlin-10-idiomatic-multiplatform-access-to-gemini/ cloud.google.com

RSS Hunter • Sep 3

Google named a Leader in 2026 Gartner® Magic Quadrant™ for SCPS

Gartner has recognized Google as a Leader in Strategic Cloud Platform Services for the ninth consecutive year, citing its Completeness of Vision. Google Cloud's dedication to customer success in building and scaling demanding workloads is highlighted. The company emphasizes accelerating customer journeys into the agentic era with a dynamic infrastructure and a unified platform for apps, data, and agents. Three key advantages are presented: a co-designed technology stack, a dynamic infrastructure for secure scaling, and digital sovereignty offering genuine choice. Google Cloud is committed to facilitating customer innovation and scale through flexibility, performance, and control. The first design principle is accelerating AI with a co-designed, open stack from silicon to applications, featuring custom hardware like TPUs and Axion processors. This approach maximizes operational efficiency without vendor lock-in, embracing open source and standards. The second principle focuses on scaling quickly and economically with a dynamic infrastructure that automates capacity management and modernizes applications. This enables organizations to securely connect data, apps, and agents for optimized global experiences. The third principle champions digital sovereignty, allowing organizations to retain data control and comply with regulations through flexible deployment options. These options include data sovereignty with boundary controls, local compliance via partner-operated clouds, and on-premises flexibility with Google Distributed Cloud.

https://cloud.google.com/blog/products/compute/google-named-a-leader-in-2026-gartner-magic-quadrant-for-scps/ cloud.google.com

RSS Hunter • Sep 3

Getting started with Mantis, our open-source bug finding-and-fixing harness

AI can now discover and exploit software vulnerabilities autonomously. To combat this, Google developed Mantis, an open-source framework that automates vulnerability discovery, triage, reproduction, and patching. Mantis aims to provide defenders with an AI advantage by enabling machine-speed security analysis. Unlike other AI tools that suffer from low true-positive rates, Mantis employs agentic techniques and sandboxed vulnerability reproduction for accuracy. It learns from past security fixes within a repository to build architectural and threat model documentation. Mantis condenses large codebases into a hierarchical security summary tree, significantly reducing token overhead. The framework distills extensive cybersecurity knowledge and is available on GitHub. Users can clone Mantis locally and then prompt a coding agent to initiate code review and vulnerability analysis. Internally at Google, this approach has successfully identified real vulnerabilities. Mantis provides sample sandboxing options and allows for custom sandbox implementations. It offers a straightforward way to begin with vulnerability discovery, filtering, and patching. The mantis-advise skill further leverages accumulated knowledge for writing secure code from the outset. To maximize AI-driven vulnerability discovery, it's recommended to provide tools with context and build cyber sandboxes with clear vulnerability acceptance criteria.

https://cloud.google.com/blog/products/identity-security/getting-started-with-the-mantis-harness-to-find-and-fix-bugs/ cloud.google.com

RSS Hunter • Sep 2

Simplify pipelines with new BigQuery identity columns

BigQuery has launched identity columns to empower customers in their data journey. This feature automatically generates sequential 64-bit integer values for unique identification within tables. Data engineers will find this a powerful, built-in mechanism for smoother and more reliable data ingestion. By offloading ID generation to BigQuery, complexity in data pipelines is reduced, allowing focus on insights. Implementing identity columns streamlines development and maintenance of data architecture. It simplifies ingestion by eliminating the need for pre-calculated unique keys in application logic or ETL tools. Boilerplate code is also reduced, leading to cleaner and more maintainable SQL. Identity columns integrate seamlessly with standard DML operations for automatic unique identifier assignment. They are flexible and adapt to existing workflows using INSERT or MERGE statements. Setting up an identity column is straightforward, with options for GENERATED ALWAYS AS IDENTITY or GENERATED BY DEFAULT AS IDENTITY. This feature underscores BigQuery's commitment to a flexible, high-performance, and standards-compliant data platform.

https://cloud.google.com/blog/products/data-analytics/bigquery-identity-columns-to-auto-generate-sequential-integers/ cloud.google.com

RSS Hunter • Sep 2

Introducing TabFM in BigQuery: Predictive analytics reimagined

Predictive analytics previously required complex custom model building and training cycles. This traditional approach involved manual feature engineering, lengthy tuning, and specialized skills, often leading to underutilization of predictive models. Google Research has introduced TabFM, a pre-trained foundation model for tabular data, now available in BigQuery. TabFM utilizes in-context learning to deliver instant, highly accurate predictions via a single SQL statement, eliminating separate training and deployment steps. This system offers zero-shot predictions, simplifying the process for users without extensive machine learning expertise. It also enhances agentic applications by providing predictive powers without managing infrastructure. TabFM demonstrates state-of-the-art accuracy, outperforming traditional models on benchmarks. The developer experience is simplified through native BigQuery integration and straightforward SQL syntax. TabFM automatically handles feature engineering tasks, and its scalability allows processing millions of rows quickly. Users can leverage AI.PREDICT for instant predictions and AI.EVALUATE to assess model performance. While TabFM excels in rapid insights and frequent retraining, traditional models remain suitable for very large datasets or when feature importance explainability is crucial. This integration makes predictive machine learning as easy as running a standard SQL query.

https://cloud.google.com/blog/products/data-analytics/tabfm-adds-predictive-ml-to-bigquery/ cloud.google.com

RSS Hunter • Sep 1

How Blackline simplifies perimeter policy intelligence with VPC Service Controls

VPC Service Controls are crucial for protecting cloud environments from data exfiltration and other threats. Google Cloud has introduced new policy intelligence capabilities to simplify VPC-SC management and troubleshooting. These new tools include a violation analyzer and a violation dashboard. Financial operations leader BlackLine uses VPC-SC to secure sensitive customer data and mitigate risks. The violation analyzer helps BlackLine administrators adapt security perimeters by identifying principals and resources involved in API request violations. Previously, resolving these violations required complex Cloud Logging queries, but the new tools provide detailed reports directly. The violation dashboard aggregates all service perimeter violations across an organization for trend identification and filtering. The violation analyzer streamlines investigations by mapping denials to specific rules without manual log analysis. These tools simplify the entire VPC-SC lifecycle, from testing new perimeters to refining policies and resolving incidents efficiently. By leveraging these advancements, organizations can spend less time on investigations and more time scaling their cloud initiatives securely.

https://cloud.google.com/blog/topics/customers/how-blackline-prevents-data-exfiltration-with-vpc-service-controls/ cloud.google.com

RSS Hunter • Sep 1

Financially Motivated Threat Actor BREEZE COMET Targets Brazil

Mandiant is tracking a financially motivated threat actor named BREEZE COMET, formerly known as UNC5669, which has been actively compromising Brazilian financial organizations since 2024. This group specializes in manipulating payment systems and banking software to conduct fraudulent transfers. BREEZE COMET employs a customized malware suite and utilizes compromised trusted websites for initial access and command and control. Their operations have expanded to include leveraging generative AI for malware development, potentially increasing their operational sophistication. The group targets entities with access to the National Financial System Network and requires mTLS credentials for authenticated fraudulent transactions. BREEZE COMET uses various methods for initial compromise, including voice phishing and exploiting compromised government websites. They also gain footholds by connecting rogue hardware devices directly into retail networks. To escalate privileges, they utilize reconnaissance tools and custom malware, specifically targeting cloud and development environments for credentials and API keys. BREEZE COMET moves laterally through networks using hijacked service accounts and specialized routing malware like COBALTSPIN. They maintain persistence through custom backdoors such as LIGHTPAINT, MILDFROST, KICKPLATE, and BOATBEAM, often disabling endpoint defenses like Windows Defender. After compromising financial applications, BREEZE COMET executes mass fraudulent transactions and then clears event logs to cover their tracks. Their evolving tactics represent a significant shift in Latin American cybercrime, moving from retail fraud to direct intrusions into core financial infrastructure. BREEZE COMET's use of AI demonstrates its impact on enhancing threat actor capabilities, speed, and scale.

https://cloud.google.com/blog/topics/threat-intelligence/financially-motivated-threat-actor-breeze-comet-targets-brazil/ cloud.google.com

RSS Hunter • Sep 1

BigQuery Graph is now GA: the knowledge foundation for the agentic era

BigQuery Graph, now generally available, integrates native graph analytics directly into the data warehouse, eliminating silos and operational overhead. This new capability allows users to perform complex graph queries alongside traditional SQL using the ISO-standard GQL. It leverages BigQuery's petabyte-scale processing, existing security measures, and integrates with BigQuery ML and AI functions. Data teams across various industries, including cybersecurity, finance, retail, and IT, are adopting BigQuery Graph. Use cases include threat and fraud detection, supply chain optimization, customer 360, knowledge graph creation, and network management.Recent enhancements include a faster and broader graph engine, offering improved GQL execution speeds and expressiveness. The new borderless Graph Lakehouse feature enables BigQuery Graph to span data residing in native BigQuery tables and open Iceberg tables across different clouds without data movement. This allows agents to traverse virtual knowledge graphs composed of data from multiple sources, like a support agent querying customer and product information across Google Cloud and AWS. Conversational analytics enables natural language interaction with graphs, reducing ambiguity and hallucination for AI agents. Furthermore, agents can now build graphs by translating datasets into nodes and edges, with capabilities for authoring and verification. The context graph feature provides an auditable memory for AI agents, capturing their reasoning and decisions as a queryable trace within BigQuery Graph. This ensures explainability and forms the basis for improving future agent actions.

https://cloud.google.com/blog/products/data-analytics/bigquery-graph-connecting-data-and-ai-at-scale/ cloud.google.com

RSS Hunter • Aug 31

From weeks to minutes: The new agentic era of data pipelines

Google Cloud is democratizing data pipeline orchestration with the new Orchestration Pipelines framework, introduced at Google Cloud NEXT ’26. This framework is made accessible through the Data Agent Kit, a free, open-source collection of data engineering tools. The kit integrates into popular IDEs and CLIs, providing a Data Engineering tab and an agentic skill for Airflow DAGs. It allows data professionals to author, deploy, and troubleshoot production-grade Airflow DAGs using natural language. The framework streamlines MLOps by decoupling orchestration logic from compute execution, using a declarative YAML DSL. This enables all data personas to bypass complex Python Airflow boilerplate. A practical example demonstrates building an MLOps architecture for proactive supply chain management. This architecture predicts transit times using BigQuery, Managed Service for Apache Spark, Gemini Enterprise Agent Platform, and dbt. The Data Agent Kit generates PySpark scripts and declarative YAML pipelines from natural language prompts. Three distinct pipelines are demonstrated: a training engine, a daily inference pipeline, and an automated evaluation and branching pipeline. Deployment is automated through CI/CD, with the Data Agent Kit generating necessary workflows. For day-two operations, the kit offers real-time monitoring and agentic troubleshooting within the IDE. It diagnoses failures, identifies root causes, and suggests inline fixes for issues. Orchestration Pipelines and the Data Agent Kit significantly reduce the time and complexity of building and maintaining MLOps architectures.

https://cloud.google.com/blog/products/data-analytics/build-data-pipelines-in-less-time-with-data-agent-kit/ cloud.google.com

RSS Hunter • Aug 31

Cloud CISO Perspectives: Tips on securing the water sector in the AI era

This Cloud CISO Perspectives for August 2026, featuring Chris Sistrunk and Stephanie Kiel, addresses critical security issues in the water sector. Threat actors are increasingly targeting critical infrastructure, especially water utilities' internet-connected programmable logic controllers. While manual overrides have historically provided a safety net, fundamental digital security is paramount in the AI era. A threat-informed, risk-managed response is recommended for water and wastewater utilities. Key actions include inventorying assets, implementing basic security hygiene, maintaining robust backups, and utilizing network segmentation and multifactor authentication. Emergency planning should integrate cyber incident responses into existing all-hazards systems, and third-party vendor access must be secured rigorously. IT and OT leaders must collaborate to build unified governance frameworks to enhance cyber-physical system resilience. AI-augmented security approaches are crucial for modernizing defenses and acting as a force multiplier for lean security teams. The "Mandiant Operational Technology (OT) Theory of 99" highlights that most compromises occur in computer workstations and servers, offering a defender's advantage through AI to protect intermediary infrastructure. Integrating AI tools strategically, with human oversight, can shift the balance in network security towards defenders. Protecting water systems is a public safety imperative, leading to increased government action and initiatives like the pilot program launched by the Office of the National Cyber Director and the State of Texas.

https://cloud.google.com/blog/products/identity-security/cloud-ciso-perspectives-tips-on-securing-water-sector-ai-era/ cloud.google.com

RSS Hunter • Aug 31

Reimagining work: How Pythian’s internal AI playbook delivers customer ROI

Pythian implemented Google Cloud's Gemini Enterprise internally to study enterprise AI ROI. They discovered many AI initiatives fail due to a tool-centric approach and a focus on minor efficiencies. Pythian developed the AI Operating Model to address these issues. This model integrates strategy, execution, and operations into a continuous loop.The framework begins with Field CTO strategy and governance to identify high-ROI use cases. Next, tooling and platform deployment establishes a secure, contextualized AI foundation. The dual Center of Excellence (COE) acts as the execution engine, with one part focused on people productivity through no-code agents and the other on process productivity via custom-coded agents.Finally, XOps ensures ongoing AI performance in production by managing model drift and prompt tuning. This comprehensive model moves beyond tool adoption to structural workflow transformation. By proving this model internally, Pythian achieved an 80% reduction in database incident resolution times and a threefold increase in user engagement. The model's success is further demonstrated through customer case studies in database operations, knowledge management, supply chain, and retail. Scaling AI effectively requires this end-to-end operating model, not just tool experimentation.

https://cloud.google.com/blog/topics/startups/how-pythians-internal-ai-playbook-delivers-customer-roi/ cloud.google.com

RSS Hunter • Aug 27