Friday, September 18, 2026

Percona developed a custom harness to evaluate how well large language models handle specialized database administration workloads. The testing focuses on practical systems-administration tasks executed against remote systems rather than theoretical prompts. This approach aims to measure the actual utility of LLMs in operational database environments.

• Percona built a specific harness to test LLMs on real-world DBA tasks
• Evaluation targets practical systems-administration rather than abstract queries
• Tests run against remote systems to simulate actual operational conditions
• Goal is to determine LLM readiness for specialized database administration

Christophe Pettus explains the introduction of max_notify_queue_pages, a new GUC that enforces a hard limit on the LISTEN/NOTIFY queue size. Previously, this queue, stored as SLRU pages in pg_notify, relied on implicit bounds that ceased to function effectively. This new parameter provides explicit control to prevent unbounded memory consumption in PostgreSQL clusters.

• Explicitly limits LISTEN/NOTIFY queue memory via pg_notify SLRU pages
• Fixes broken self-enforcement from versions prior to PostgreSQL 17
• Prevents runaway memory usage in high-throughput notification scenarios
• Configurable via GUC for operational tuning and stability

The pgAdmin 4 development team released version 9.18, addressing four security vulnerabilities (CVE-2026-86861 through CVE-2026-86864) alongside 29 bug fixes. Key features include a toggleable Object Explorer that mimics VS Code's sidebar behavior and persists state across refreshes. The release also hardens default Content-Security-Policy settings and introduces a customizable keyboard shortcut for toggling the explorer.

• Urgently patch to mitigate four new CVEs affecting pgAdmin 4 installations.
• Adopt the new toggle_object_explorer preference for faster UI navigation.
• Review Content-Security-Policy changes to ensure no dependency breakage.
• Update to v9.18 to resolve 29 assorted bugs and stability issues.

The fourth alpha release of pgColumnar focuses on improving data layout and query skipping capabilities. It introduces Hilbert curve-based table layout to keep neighboring keys closer together compared to Z-ordering. Additionally, the release enhances star-schema joins by enabling the skipping of fact-table groups during execution.

• Hilbert curve layout offers better locality than Z-ordering for neighboring keys.
• Star-schema joins now support skipping fact-table groups for improved performance.
• This alpha release targets layout efficiency and query execution optimization.
• pgColumnar continues to evolve as a columnar access method for PostgreSQL.

PostgreSQL 19's release cycle encountered significant delays due to an aggressive pace that saw eight major features land in the five weeks preceding the code freeze. The intense development pressure led to three committers withdrawing their contributions, causing the release process to falter under its own scope. This highlights the risks associated with compressing large-scale feature integration into tight pre-freeze windows.

• PostgreSQL 19 release is delayed or unstable due to scope creep.
• Eight major features merged late, destabilizing the freeze period.
• Three committers withdrew work, indicating burnout or quality concerns.
• Late-merge risk management is critical for major version stability.

Qwen has launched the 3.8 Omni Flash model, a new entry in its multimodal lineup focused on speed and efficiency. This release targets practitioners needing rapid inference capabilities for mixed media workloads without sacrificing core performance metrics. The model joins the existing suite of tools available for deployment in cloud-native environments.

• New multimodal model optimized for inference speed and lower latency
• Suits high-throughput pipelines requiring rapid text and vision processing
• Available via Qwen's official blog for immediate technical evaluation
• Part of ongoing updates to the Qwen family for production readiness

This research introduces an approach where LLM weights are not static but generated and adapted dynamically from live data streams. Unlike traditional models with fixed parameters, this method allows for continuous adaptation without retraining the entire network. The core contribution lies in the mechanism for weight generation, enabling the model to scale effectively while maintaining relevance to current inputs.

• Models can adapt to new data without full retraining cycles.
• Weight generation replaces static parameter storage for flexibility.
• Enables continuous learning from live data streams.
• Reduces infrastructure costs associated with large static models.

VisKG-LM decouples graph encoding from language reasoning by serializing retrieved subgraphs as Relation-Labeled Paths and rendering them as images. This offline compilation allows the model to access knowledge as read-only visual memory, avoiding the need to re-encode the same static subgraph during every online inference step. The approach preserves the branching structure of the graph in a two-dimensional layout for efficient multi-choice question answering.

• Eliminates redundant graph re-encoding by compiling subgraphs offline once.
• Treats knowledge as read-only visual memory, speeding up inference loops.
• Decouples structural encoding from the language model's reasoning path.
• Uses Relation-Labeled Paths rendered as images to preserve graph topology.
• Optimizes scoring across multiple epochs and evaluation runs.

ScientistTwo is a fully autonomous multi-agent framework designed to execute problem-driven research. It takes a fundamental challenge from a human expert and independently navigates the scientific landscape to identify theoretical and empirical bottlenecks. The system establishes baselines, formulates novel hypotheses, and coordinates specialized agents to orchestrate an end-to-end discovery cycle.

• Enables fully autonomous end-to-end scientific discovery cycles without human intervention.
• Coordinates specialized agents to handle hypothesis formulation and bottleneck identification.
• Aims to expand the human knowledge frontier by venturing into unexplored scientific territory.
• Moves beyond assisted research to problem-driven autonomous AI navigation of scientific landscapes.

Amazon SageMaker AI now accepts prioritized lists of instance types for training and processing workloads, allowing the scheduler to select from multiple acceptable options rather than a single fixed type. This change eliminates the need for custom retry logic or concurrent job submissions to handle GPU contention during peak demand. The feature automates capacity discovery, reducing wait times and simplifying job configuration for workloads that are flexible across instance sizes.

• Submit a prioritized list of instance types instead of a single option to increase scheduling flexibility.
• Avoid complex custom retry logic or concurrent job submissions for GPU availability issues.
• Reduce job start times during peak periods by allowing SageMaker to pick the first available instance.
• Ideal for training and processing workloads that perform comparably across multiple instance families.

Amazon SageMaker JumpStart now hosts models from IBM, Kakao, and the OpenFold Consortium. The update introduces granite-speech-4.1-2b for multilingual AS and AST, kanana-2-30b-a3b-instruct for bilingual agentic tasks, and OpenFold3 for biomolecular structure prediction. These additions expand the available foundation model portfolio for scalable AWS deployments.

• IBM's granite-speech-4.1-2b handles multilingual ASR and translation for 6 languages.
• Kakao's kanana-2-30b-a3b-instruct targets bilingual agentic AI workflows.
• OpenFold3 supports biomolecular structure prediction via SageMaker JumpStart.
• All three models are now deployable as standard JumpStart foundation models.

This paper addresses the challenge of language-model agents performing tasks that span days or weeks, exceeding standard context windows and human attention spans. The authors propose that agents must run continually without forgetting, a capability achieved through an external harness rather than model architecture changes. The proposed solution is a hierarchical system featuring time-indexed levels, bounded file summaries, and a clocked tick mechanism for autonomous action.

• Agents need a continuous, non-forgetting harness to handle long-term tasks beyond context windows.
• Architecture uses time-scale levels where each keeps a bounded summary of the level below.
• Clocked ticks serve as the fundamental unit for autonomous agent actions.
• Cascaded intelligence manages complexity across these hierarchical levels.

This paper addresses the challenge of AI agents operating in enterprise systems where business rules are dynamic and undocumented. The authors introduce a method for continual world model discovery, allowing agents to infer causal relationships by observing outcomes of their actions on records. Evaluation uses EnterpriseWorldShift, a benchmark built on a live ServiceNow environment with nine tables and 25 hidden rules.

• Agents must learn business logic dynamically rather than relying on static schemas.
• World models are updated continuously as organizational rules change over time.
• Benchmark uses live ServiceNow data to test rule discovery accuracy.
• Focuses on causal inference from record interactions in complex enterprise workflows.

This paper challenges the assumption that RL gains in agentic language models reflect better decision-making, arguing that endpoint success conflates arrival state with execution capability. Because agents in closed loops generate their own observation sequences, SFT and RL checkpoints are evaluated from fundamentally different state distributions. The authors introduce checkpoint handoffs to isolate these effects, revealing that restricting comparisons to shared states can misleadingly flip the perceived impact of RL training.

• RL gains may reflect state selection rather than improved policy execution.
• Closed-loop agents create dependent observations, biasing standard evaluations.
• Comparing SFT vs RL on identical tasks is flawed due to divergent state paths.
• Checkpoint handoffs isolate arrival state from in-state decision quality.
• Selection bias in restricted state comparisons can invert performance signals.

LangChain 1.4.1 is a patch release addressing two specific issues in the 1.4.0 series. It resolves a bug where open Model Context Protocol (MCP) object arguments were not being preserved correctly. Additionally, the release corrects documentation for the InterruptOnConfig feature.

• Updates are required if you rely on open MCP objects to ensure arguments are passed correctly.
• Review InterruptOnConfig usage against the corrected documentation to avoid configuration errors.
• This is a minor patch release focused on stability and documentation accuracy.
• Upgrade from 1.4.0 to ensure proper behavior with MCP integrations.

A new position paper argues that the fragmentation in agentic AI stacks mirrors the pre-OS era of computing. It proposes a Foundation Model Operating System (FMOS) layer to virtualize interactions with foundation models. This abstraction aims to provide portable state, memory, and guardrails, eliminating the need for each framework to re-implement these core services.

• Current agentic frameworks embed implicit runtimes, making behavior non-portable across tools.
• FMOS would abstract FM interactions like VMs abstract hardware for standardized access.
• Centralizing governance and state management could resolve brittle control planes.
• Protocols like MCP and A2A handle connectivity but not the underlying runtime abstraction.

Cilium is gaining traction on GitHub as an eBPF-based solution for Kubernetes networking and security. It replaces kube-proxy with efficient hash tables for distributed load balancing and supports L3-L7 identity-based policies. The project enables flat Layer 3 networks across clusters using native routing or overlay modes.

• Replaces kube-proxy with eBPF hash tables for near-limiting scale and performance.
• Enforces L3-L7 network policies decoupled from IP addresses via identity-based security.
• Provides native routing or overlay modes for spanning multiple clusters.
• Offers integrated observability and security without traditional overlay complexity.

AWS Batch introduces three new APIs—CancelJobs, TerminateJobs, and TerminateServiceJobs—that allow engineers to cancel or terminate up to 50 jobs in a single call. This update reduces operational overhead for large-scale workloads by consolidating job management actions. Additionally, ListJobs and ListServiceJobs now include specific fields to track cancellation and termination states for better lifecycle visibility.

• Use CancelJobs for jobs in SUBMITTED, PENDING, or RUNNABLE states.
• Use TerminateJobs or TerminateServiceJobs to stop jobs in any state, including RUNNING.
• Each API call supports up to 50 job IDs, including array job components.
• ListJobs now returns isCancelled and isTerminated fields for state tracking.
• ListServiceJobs now returns isTerminated to simplify service job monitoring.

Amazon ECS now offers a unified deployment view for Managed Daemons directly in the AWS Management Console. This feature consolidates progress tracking, failure diagnostics, and lifecycle timelines into a single interface. Engineers can monitor rollout status, review completed deployments, and identify blockers without aggregating data from multiple sources.

• Eliminates the need to piece together deployment status from disparate logs or APIs.
• Provides a timeline with timestamps and total duration for each deployment step.
• Tracks instance states including completed, in-progress, remaining, and draining.
• Visualizes rollback paths immediately when a deployment is interrupted.

This paper proposes a framework to automatically build business semantic layers from noisy, heterogeneous application logs. The method uses a two-stage abstraction process: first identifying high-level business features via LLM inference with domain knowledge, then deriving fine-grained business nodes. This automation aims to eliminate the manual effort currently required to reconcile data discrepancies and maintain fragile mappings between raw events and KPIs.

• Automates the translation of raw, noisy telemetry into structured business insights.
• Uses LLMs augmented with domain knowledge to identify high-level features first.
• Derives fine-grained business nodes in a second stage for detailed abstraction.
• Reduces engineering overhead in maintaining fragile raw-to-KPI mappings.

The Jemalloc team has published version 5.4.0 of their high-performance memory allocator. This release introduces specific updates to the allocation logic and internal structures to enhance efficiency. The update is now available for download and integration into existing software stacks.

• Jemalloc 5.4.0 is now available for production use
• Includes internal logic updates for better allocation efficiency
• Review release notes for specific behavioral changes
• Test upgrade path in staging before fleet deployment

AWS Transfer Family now supports source IP preservation for SFTP servers behind a Network Load Balancer using Proxy Protocol v2. Previously, the NLB masked the client's actual IP, forcing logs and identity providers to see only the load balancer's private address. This update enables administrators to retain visibility of the true client IP for auditing, access controls, and compliance requirements.

• Enables Proxy Protocol v2 to pass original client IP through NLB to Transfer Family SFTP endpoints.
• Restores ability to perform IP-based auditing and logging with accurate source addresses.
• Allows custom identity providers to authorize users based on their true source IP instead of the NLB IP.
• Supports stricter compliance and security policies requiring granular access control by client origin.
• Applies specifically to VPC-hosted SFTP endpoints placed behind an NLB.

AWS has released general availability for T8i instances, featuring custom sixth-generation Intel Xeon 6 processors and AWS Nitro hardware. These new micro, small, medium, and nano sizes offer up to 30% better price performance than T3 instances, alongside significant gains in compute, network, and EBS bandwidth. They are optimized for low-to-moderate CPU workloads such as small databases, CI/CD pipelines, and event-driven functions.

• T8i instances provide up to 70% higher compute performance compared to previous T3 generations.
• Network bandwidth increases by 1.25x and EBS bandwidth by 2.4x over prior burstable options.
• Available in nano, micro, small, and medium sizes for cost-sensitive, variable workloads.
• Ideal for small databases, CI/CD pipelines, and low-traffic microservices requiring burst capacity.

AWS Direct Connect now offers a single fixed monthly price for 10 Gbps and 100 Gbps dedicated connections, eliminating per-gigabyte data transfer out charges within the selected tier. This model addresses the unpredictability of costs for workloads with large, sustained data egress, replacing the previous pay-as-you-go structure. The change targets network architects and FinOps teams seeking stable, predictable networking expenses without long-term contracts.

• Flat-rate pricing removes variable DTO charges for 10G and 100G dedicated connections.
• Monthly costs become predictable for workloads with large, sustained data egress volumes.
• No upfront commitment or long-term contracts required for this new pricing tier.
• Ideal for FinOps teams stabilizing network spend on high-throughput hybrid architectures.

AWS has updated root user authentication to route traffic across US East (N. Virginia), US East (Ohio), and US West (Oregon), eliminating the single-region dependency on N. Virginia. The system automatically handles this distribution transparently, requiring no changes to how users sign in. ConsoleLogin events are now logged in whichever of these three regions processes the request, shifting the logging location from a fixed point to a dynamic one.

• Root sign-in is now distributed across three regions to improve resilience during outages.
• CloudTrail ConsoleLogin events for root users now appear in the processing region.
• Update monitoring and alerting to cover N. Virginia, Ohio, and Oregon for full visibility.
• No action is required to change sign-in behavior as routing is automatic.

AWS Glue zero-ETL now tracks which integration owns target table properties, preventing accidental overlaps across Amazon S3 Tables and SageMaker Lakehouse catalogs. When a user attempts to create or modify an integration that conflicts with an existing one, the service identifies the owner and guides resolution. This ensures predictable control over data landing zones and keeps pipelines isolated.

• Glue now associates table properties with the owning zero-ETL integration for better governance.
• Conflicts are detected when multiple integrations target the same table without coordination.
• Support spans both Amazon S3 Tables and SageMaker Lakehouse catalogs.
• Users receive guidance to choose different targets or update existing integrations.
• Data teams gain predictable control over where source tables land in the catalog.

AWS STS now enforces a unified 4,096-byte limit on session tokens, replacing previous separate caps for tokens and input parameters like inline policies. The service returns token size and utilization metrics in responses, logging them to CloudTrail and publishing them to CloudWatch. An optional API parameter allows generating tokens up to the new limit to test infrastructure readiness.

• Unified 4,096-byte limit replaces separate caps for tokens and inline policies.
• CloudTrail and CloudWatch provide visibility into token size and utilization.
• New API parameter enables testing larger tokens against application limits.
• Simplifies capacity planning for complex session policy combinations.

Octop is an open-source, self-hosted AI assistant designed for multi-user and multi-agent environments. It enables parallel operation and collaborative intelligence for teams or individuals while running entirely on local infrastructure. The system supports single-process startup with access via web dashboard, CLI, and instant messaging platforms like Feishu and DingTalk.

• Self-hosted architecture ensures data privacy by running entirely on user infrastructure.
• Multi-agent design allows parallel processing and collaborative task execution.
• Supports multiple interfaces including Web, CLI, Feishu, and DingTalk integrations.
• Single-process startup simplifies deployment and management for teams.

Alibaba has open-sourced Open Code Review, an internal AI code review tool used by tens of thousands of developers. The CLI combines deterministic pipelines for static analysis with LLM agents for nuanced feedback, supporting OpenAI and Anthropic endpoints. It natively detects issues like NPE, XSS, and SQL injection via built-in multi-language rulesets.

• Hybrid architecture blends static rule checks with LLM reasoning for precise line-level comments.
• Supports OpenAI and Anthropic model endpoints for flexible AI integration.
• Pre-configured rules cover critical risks like SQL injection and thread-safety.
• Operates as a CLI tool reading Git diffs for easy CI/CD pipeline integration.

Bend is a new programming language designed to prevent errors introduced by AI coding assistants by using formal proofs. It supports execution on both CPU and GPU architectures, aiming to provide a safety net for generated code. The project highlights a shift toward verifying AI output rather than just trusting it.

• Bend uses formal proofs to catch AI-generated code errors before runtime.
• Supports both CPU and GPU execution, broadening hardware compatibility.
• Aims to mitigate hallucinations and logic bugs common in AI assistants.
• Signals growing industry focus on verifiable AI code safety.
• Available for testing at bend-lang.com with active community discussion.