Sunday, August 02, 2026

Mark Wong is progressing on the PostgreSQL Performance Farm initiative by leveraging AWS Open Source Credits through Kiro Powers. This funding enables large-scale OLTP testing, specifically to update legacy DBT-5 test kits developed during the OSDL era. The effort involves collaboration with AWS maintainers to validate PostgreSQL performance on EC2.

• AWS credits via Kiro Powers are funding the PostgreSQL Performance Farm project.
• Legacy DBT-5 OLTP test kits need updating for modern cloud environments.
• Testing focuses on large-scale OLTP workloads on EC2 instances.
• Collaboration with AWS and maintainers is accelerating the initiative.

A performance benchmark comparing memory allocators in Percona Server for MySQL revealed unexpected memory usage patterns specifically tied to stored procedures. The investigation highlights how routine database setup decisions can lead to surprising resource consumption during execution. These findings suggest that memory allocation strategies significantly impact the overhead of stored procedural logic.

• Unexpected memory spikes occur during stored procedure execution in Percona Server.
• Memory allocator choice directly influences the baseline overhead of procedural logic.
• Benchmarking reveals hidden costs often missed during standard operational monitoring.
• Review memory allocation settings if stored procedures are heavily utilized.

PostgreSQL 19 introduces native support for SQL Property Graph Queries (SQL/PGQ) per the ISO/IEC 9075-16:2023 standard. The update includes the GRAPH_TABLE table function for pattern matching and new DDL commands to manage property graphs. These changes are backed by new system catalogs and information schema views to facilitate graph operations directly within SQL.

• Native graph pattern matching via GRAPH_TABLE eliminates external tool dependency
• ISO/IEC 9075-16:2023 compliance ensures standard SQL graph query syntax
• New DDL commands simplify property graph lifecycle management
• System catalogs provide visibility into graph metadata and structure

PostgreSQL 18 introduces a new Grand Unified Configuration parameter allowing extension control and SQL files to reside outside the server's core directories. This change enables operators to mount extensions as independent OCI container images at runtime, eliminating the need to rebuild the base PostgreSQL image for updates. While this supports leaner fleet management, the decoupling is only effective for extensions that do not rely on deep integration with the server binary.

• PostgreSQL 18 adds a GUC to store extension files outside server directories.
• Extensions can now be packaged as separate OCI images and mounted at runtime.
• Reduces base image size by keeping PostgreSQL unmodified and lean.
• Independent versioning allows rapid updates like pgvector without server rebuilds.
• Decoupling benefits depend on extension depth; tightly coupled extensions gain little.

vip-manager v5 introduces breaking changes that alter cluster behavior during failovers, making pre-upgrade review essential. The tool manages a Virtual IP by monitoring the Distributed Configuration Store for leader elections, attaching the VIP to the current primary. This release requires careful attention to ensure high availability setups do not behave unexpectedly during maintenance windows.

• Review breaking changes before running apt upgrade to avoid unexpected 03:00 failures.
• vip-manager v5 changes how the cluster behaves when leadership transitions occur.
• The tool continues to manage VIPs by watching the DCS for leader info changes.
• Plan your rollout strategy to mitigate risks associated with the new behavior.
• Do not treat this as a simple set-and-forget tooling update.

Amazon Bedrock has aligned its pricing with OpenAI's latest reductions for the GPT-5.6 Luna and GPT-5.6 Terra models. Effective July 30, 2026, on-demand inference for GPT-5.6 Luna drops by 80%, while GPT-5.6 Terra sees a 20% decrease. These adjustments reflect the first-party pricing changes announced by OpenAI.

• GPT-5.6 Luna pricing slashed 80%, targeting high-volume, fast inference tasks.
• GPT-5.6 Terra pricing reduced 20%, balancing speed and reasoning for production.
• Bedrock rates now match OpenAI's direct pricing for these specific model variants.
• Lower costs enable broader application of these models in customer service automation.
• Optimized for content processing, classification, and routine implementation workflows.

Hugging Face has released LFM2.5 encoders designed to accelerate long-context inference tasks on CPU hardware. This release targets practitioners seeking efficient processing for large input windows without relying exclusively on GPU resources. The update provides a practical solution for scaling context handling in cost-sensitive or resource-constrained environments.

• LFM2.5 encoders optimize long-context performance on CPU architecture
• Reduces dependency on GPU resources for large input window processing
• Available via Hugging Face for immediate integration into inference pipelines
• Addresses latency and throughput challenges in CPU-bound LLM deployments

xAI's Grok 4.3 model is now available on Amazon Bedrock within AWS GovCloud (US-West), expanding the selection of model providers for that region. The model is designed as reasoning-first with configurable effort levels ranging from none to high. It emphasizes strong tool use and instruction following to support reliable agentic workflows.

• Grok 4.3 is now live in AWS GovCloud (US-West) via Bedrock.
• Configurable reasoning effort helps balance cost and quality.
• Optimized for enterprise tasks like legal research and support.
• Token efficiency supports high-volume inference cost-effectively.

This open-source project provides a team-level memory hub for AI agents, converting conversations, documents, and code into four reusable assets: Chat Memory, Skill, LLM-Wiki, and Code-Graph. It enables governed, shared, and equipped memory across different agents and frameworks. The tool aims to let agents handle memory tasks while humans focus on innovation, with a beta version currently available for exploration.

• Unifies agent memory into four assets: Chat, Skill, Wiki, and Code-Graph
• Enables cross-agent and cross-framework memory sharing and governance
• Converts unstructured data (docs, code, chats) into reusable agent assets
• Beta release includes quick install script for core, hub, and proxy services

HuggingFace released a modular voice agent pipeline that chains VAD, STT, LLM, and TTS components with low latency. It exposes an OpenAI Realtime-compatible WebSocket API, allowing developers to swap any component while maintaining protocol compatibility. The stack supports local inference via vLLM or llama.cpp, enabling fully open, on-premises voice agents.

• Exposes OpenAI Realtime WebSocket API for easy client integration.
• Fully modular design allows swapping VAD, STT, LLM, or TTS components.
• Supports local LLM inference via vLLM or llama.cpp for privacy.
• Production-ready, currently powering backend for Reachy Mini robots.
• Simplifies building custom voice agents without proprietary vendor lock-in.

GitHub has introduced a CLI extension that automates the workflow for stacked branches and pull requests. The tool handles branch creation, rebasing, and setting correct base branches to break large changes into reviewable chains. It also includes an AI agent skill to help coding assistants understand the stacked PR context.

• Automates tedious branch management and rebasing for stacked PR workflows
• Requires GitHub CLI v2.0+ and installed via gh extension install
• Includes an AI skill to enable coding agents to navigate stacked branches
• Simplifies chaining small, reviewable PRs that build on each other

AWS DataSync Enhanced mode has expanded support to include Amazon EFS and Amazon FSx for Lustre as source or destination locations. This update allows users to leverage parallel processing, removal of file count limits, and detailed transfer metrics for these file systems. Previously restricted to Basic mode, these integrations are now available across all AWS regions.

• Enhanced mode is now available for EFS and FSx for Lustre locations.
• Enables parallel processing and removes file count limitations for these workloads.
• Simplifies large-scale migrations and AI/ML training data movement.
• Provides detailed transfer metrics for better observability.
• Available in all AWS regions where DataSync is supported.

Amazon CloudWatch now offers fully managed Prometheus collectors to ingest metrics from EKS, EC2, ECS, MSK, and OpenSearch without requiring self-managed agents. This replaces the previous need to deploy and scale OpenTelemetry Collectors, as CloudWatch now handles provisioning and scraping automatically via user-provided configurations. Metrics arrive in OpenTelemetry format and are queryable with PromQL alongside native AWS metrics.

• Eliminates the operational overhead of deploying and scaling self-managed OpenTelemetry Collectors.
• Supports zero-agent metric collection from key AWS services including EKS, EC2, and MSK.
• Enables unified PromQL querying across vended AWS metrics and collected Prometheus data.

AWS Direct Connect now exposes BGP route details for private, transit, and public virtual interfaces. Users can inspect accepted and advertised routes, including AS paths and community values, via the console or the ListVirtualInterfaceRoutes API. This capability aids in troubleshooting routing issues and verifying hybrid network propagation.

• Inspect BGP communities and AS paths for all VIF types directly in the console.
• Use ListVirtualInterfaceRoutes API for programmatic route auditing and automation.
• Verify route propagation between on-premises routers and AWS edge nodes.
• Troubleshoot hybrid connectivity issues by comparing accepted vs. advertised routes.

AWS has integrated the IAM Policy Simulator directly into the IAM console, replacing the standalone tool to centralize identity and policy management. The update enables testing of Service Control Policies (SCPs), allowing teams to validate how organizational constraints interact with identity and resource policies. This consolidation supports more realistic scenario modeling for security and platform engineers.

• Simulator is now native to the IAM console, removing the standalone URL.
• New capability to test SCPs alongside identity and resource policies.
• Centralized location simplifies validation workflows for platform teams.
• API support retained for automation and condition key testing.

AWS Directory Service now allows in-place upgrades from Standard to Enterprise Edition for AWS Managed Microsoft AD. This console, CLI, and API-enabled feature eliminates the need to migrate directories or re-join workloads when outgrowing the 5,000 user limit of Standard Edition. Enterprise Edition supports up to 500,000 objects while preserving existing trust relationships, group policies, and DNS configurations.

• Upgrade in-place directly via Console, CLI, or API without migrating directories.
• Preserves existing trust relationships, app integrations, and group policies.
• No DNS changes or workload re-joins required during the transition.
• Scale from 5,000 to 500,000 objects by moving to Enterprise Edition.
• Available in all regions where AWS Directory Service is supported.

Amazon Aurora DSQL now supports multi-Region clusters in Europe (Stockholm), Europe (Spain), Asia Pacific (Mumbai), and Asia Pacific (Singapore). This serverless distributed SQL database offers active-active high availability and multi-Region strong consistency. Each multi-Region cluster provides writable endpoints in both peered Regions, presenting a single logical database that remains available even if one Region becomes unavailable.

• Aurora DSQL multi-Region clusters now cover four additional AWS Regions for broader global coverage.
• Active-active architecture ensures strong consistency and availability across peered Regions.
• Single logical database with writable endpoints in both Regions simplifies multi-Region deployments.
• Serverless design eliminates operational overhead while maintaining high performance.

Amazon Redshift Data API now supports long polling, allowing clients to wait for SQL statement completion before receiving a response, which reduces the need for repeated polling calls. It also introduces ListSessions to enumerate and filter active sessions by status or compute target, improving visibility for application-level session reuse. Additionally, batch execute statements can now run on separate transactions, offering greater flexibility in transactional control.

• Use WaitTimeSeconds in API calls to eliminate busy-wait loops for SQL completion.
• Enumerate and filter active sessions via ListSessions for better operational visibility.
• Execute batch statements on separate transactions for finer-grained control.
• Reduces API call volume and latency for asynchronous SQL execution workflows.

AWS has moved AWS Interconnect for Oracle Cloud Infrastructure to general availability, offering a purpose-built solution for multicloud connectivity. This service replaces complex, do-it-yourself global multi-layered networking setups with a managed approach to interconnect workloads across both clouds. The launch supports customer strategies focused on interoperability, technology selection freedom, and faster application deployment across environments.

• Managed connectivity eliminates the complexity of building DIY global multi-layered networks.
• Enables seamless workload interconnection between AWS and Oracle Cloud Infrastructure.
• Supports multicloud strategies requiring interoperability and technology flexibility.
• Reduces operational overhead for cross-cloud application deployment and speed.

Amazon OpenSearch Service now supports version 3.7, bringing enhanced vector search performance and better search relevance capabilities. Key technical improvements include 1-bit scalar quantization on Faiss and Lucene engines to reduce storage and memory usage for vectors without sacrificing accuracy. Operators can also retrieve vectors faster using doc values without requiring reindexing, while new Query Insights and Search Relevance Workbench tools aid in performance monitoring and optimization.

• OpenSearch 3.7 is now available on Amazon OpenSearch Service.
• 1-bit scalar quantization reduces vector storage and memory usage.
• Vector retrieval uses doc values, eliminating the need for reindexing.
• Search Relevance Workbench adds CSV uploads and hybrid search metrics.
• New Query Insights provide automated recommendations and query caching.

AWS has extended Reserved Instance coverage to R8i and M8i instance families for Amazon RDS for Oracle, offering 1-year and 3-year terms. These new instances, built on custom Intel Xeon 6 processors, provide up to 53% cost savings compared to On-Demand pricing and deliver superior memory bandwidth. The Reserved Instance discount applies to both Single-AZ and Multi-AZ configurations, allowing flexibility for varying workload demands.

• R8i and M8i instances now eligible for Reserved Instances, saving up to 53% vs On-Demand.
• Custom Intel Xeon 6 chips deliver 2.5x memory bandwidth and 15% better price-performance.
• Reservations apply to both Single-AZ and Multi-AZ, enabling config flexibility.
• Supports 1-year and 3-year commitment terms for Oracle database workloads.

Fleet impact: Evaluate R8i/M8i migration for Oracle workloads requiring high memory bandwidth to leverage the 2.5x improvement and significant cost reduction. Ensure Reserved Instance purchases cover both Single-AZ and Multi-AZ needs to maintain architectural flexibility without losing discount benefits.

Amazon EKS has increased the Horizontal Pod Autoscaler (HPA) sync concurrency on Provisioned Control Plane clusters to up to 40 times the default Kubernetes limit. This enhancement allows the control plane to evaluate multiple HPA objects in parallel, significantly reducing the latency between detecting load changes and scaling pod counts. The update specifically targets clusters managing hundreds or thousands of autoscaling objects to improve responsiveness to demand spikes.

• HPA sync concurrency is now up to 40x the default Kubernetes value on Provisioned Control Planes.
• Parallel evaluation of HPA objects reduces the time required to scale pods in response to load.
• Clusters with hundreds or thousands of HPA objects will see the most significant latency improvements.
• No configuration changes are mentioned; this appears to be a backend control plane optimization.