Sunday, October 18, 2026

The pg_hint_plan extension has released version 1.9.0, introducing compatibility with PostgreSQL 19 and support for path generation strategies (PGS). This major update also delivers bug fixes and improvements to the existing hinting framework. Minor releases have been issued concurrently for PostgreSQL 14 through 18 to maintain stability across older versions.

• PostgreSQL 19 users must upgrade to pg_hint_plan 1.9.0 for compatibility
• New path generation strategies (PGS) allow more granular control over plan creation
• Backward compatibility maintained via v1.8.1 (PG 18) and v1.7.2 (PG 17)
• Extension relies on SQL comments for hints; review existing query annotations after upgrade

Fleet impact: For AWS Aurora PostgreSQL clusters, verify if the pg_hint_plan extension is enabled and tested against the new PGS features before upgrading to PostgreSQL 19. Oracle ExaCC users running PostgreSQL in containers or side-by-side deployments should check compatibility matrices, as hint behavior changes in v1.9.0 may impact execution plans for critical workloads.

Percona introduced mongo-health, a utility designed to analyze MongoDB Full Time Diagnostic Data (FTDC) for historical performance investigations. The tool addresses scenarios where application teams report past latency issues that did not trigger real-time alerts. By parsing the diagnostic.data samples recorded by mongod every second, it enables engineers to reconstruct performance states from specific past windows.

• Enables retrospective analysis of MongoDB performance using existing FTDC logs
• Resolves 'slow yesterday' tickets without requiring real-time alerting coverage
• Leverages the 1-second metric sampling inherent to mongod diagnostic.data
• Provides a structured workflow for diagnosing intermittent latency issues

A Percona case study reveals that upgrading PXC from 5.7 to 8.0 can cause sporadic, high-volume disk writes despite unchanged workload parameters. The root cause is linked to a specific Linux kernel improvement that interacts differently with Galera-based replication mechanisms. This highlights how underlying OS optimizations can inadvertently alter I/O behavior in distributed database clusters.

• Kernel updates can change I/O patterns in Galera clusters even without config changes
• PXC 8.0 upgrades may expose latent write amplification issues from kernel optimizations
• Monitor disk I/O bursts post-upgrade, not just CPU or query latency
• Reproduce upgrade scenarios in staging to catch kernel-related regressions

Fleet impact: While this specifically affects PXC/Galera, DBAs running Aurora or RDS should remain vigilant: kernel-level I/O scheduler changes in newer Linux distributions can similarly alter write patterns for InnoDB-based engines, potentially impacting IOPS provisioning and storage performance on RDS instances.

Christophe Pettus argues that the name random_page_cost is misleading because the PostgreSQL planner actually charges this cost per index page based on the degree of index correlation, not purely on randomness. This distinction highlights how the optimizer weighs physical data layout against index structure when estimating access costs.

• The term random_page_cost is a misnomer; it applies to index pages
• Planner costs depend on index correlation, not just I/O randomness
• Understanding correlation helps explain unexpected plan choices
• Reviewing correlation stats may clarify cost estimates for index scans

Fleet impact: For Aurora PostgreSQL and RDS fleets, review pg_stats correlation for hot tables to ensure planner costs align with actual SSD I/O behavior, potentially adjusting random_page_cost or relying on correlation stats to avoid inefficient sequential scans.

XERJ introduces an autoindexing engine designed to optimize AI agent data retrieval by indexing code, docs, and logs locally. It claims to reduce token consumption by 40x compared to grep while serving as a durable, cost-effective alternative to ElasticSearch for RAG and agent memory tasks.

• Reduces AI token waste by indexing data locally instead of relying on grep-style file scanning
• Supports diverse inputs including code, PDFs, logs, and documentation for unified search
• Positioned as a low-cost alternative to ElasticSearch with claimed 10x infra savings
• Enables agents to reference open-source repositories for code reuse via automated indexing

Liquid AI has introduced LFM2.5-24B-A2B, a large language model designed specifically for running agents locally on devices. The model emphasizes speed and privacy, enabling on-device execution without cloud dependency.

• LFM2.5-24B-A2B targets on-device agent workloads
• Focuses on fast inference and data privacy
• Available via Hugging Face ecosystem
• Reduces need for cloud API calls for agent tasks

NVIDIA’s Nemotron 3.5 Super VL vision-language model is now accessible via Amazon SageMaker JumpStart. This hybrid Mixture-of-Experts model combines a 120B-parameter backbone with 12B active parameters and a vision encoder to process images, video, and text. It supports context lengths up to 1 million tokens and includes a Multi-Token Prediction head to speed up generation.

• Deploy multimodal AI on AWS without external vendor lock-in via JumpStart
• Hybrid MoE architecture balances capacity (120B) with efficiency (12B active)
• Supports 1M-token context, ideal for long-document and video analysis tasks
• Built-in Multi-Token Prediction head accelerates text generation via speculation
• Targets specific use cases: document intelligence, GUI understanding, and agentic workflows

Reflection has released Beam, its first open-weight model, into Amazon SageMaker JumpStart for deployment on AWS infrastructure. The model targets reasoning, coding, and agentic tasks, allowing teams to build automated workflows while maintaining data sovereignty. It is also optimized for AWS Trainium3, providing a path to scale inference alongside existing SageMaker AI controls.

• Beam is Reflection's first open-weight model, focusing on reasoning and coding benchmarks
• Deployable via SageMaker JumpStart for integration with existing ML pipelines
• Optimized for AWS Trainium3 to leverage specific hardware acceleration
• Enables data control and audit trails for regulated or sensitive environments

ai-memory provides a persistent context layer for AI coding agents, allowing seamless handoffs between different tools like Claude Code and Codex. It stores architectural decisions, failed attempts, and open questions in a format accessible across multiple agent harnesses. This solves the fragmentation issue where existing agent memories are siloed within specific platforms or local machines.

• Enables seamless context transfer between different AI coding agents (e.g., Claude to Codex)
• Persists architectural decisions and failed approaches outside of single-agent memory limits
• Supports over 20 agent harnesses including Cursor, Gemini CLI, and OpenCode
• Decouples agent memory from local machine state, facilitating team collaboration
• Reduces redundant re-explaining of project context when switching tools

Uber built an automated system that correlates caller and downstream failures using RPC middleware and production metrics. The tool classifies dependencies as fail-close, fail-open, or unknown to overcome the limitations of sampled tracing for rare incidents.

• Correlates caller and downstream failures via RPC middleware and production metrics
• Classifies dependencies as fail-close, fail-open, or unknown for better handling
• Addresses gaps in sampled distributed tracing for rare failure scenarios
• Automates dependency mapping across Uber’s large-scale service mesh

AWS Transform now supports automated conversion of .NET application code from Azure SDKs to AWS SDKs. The service identifies Azure-specific calls and maps them to equivalent AWS services, such as translating Azure Functions with Cosmos DB into Lambda with DynamoDB. Users review and adjust proposed mappings in an assessment report before executing the final code transformation.

• Automates .NET code migration from Azure SDK to AWS SDK
• Maps Azure services to AWS equivalents like Cosmos DB to DynamoDB
• Includes human-in-the-loop review of conversion options before execution
• Targets cloud-native .NET applications for accelerated re-platforming

AWS has released the general availability of the Lambda Durable Execution SDK for Go, allowing developers to build long-running, resilient workflows directly in Go. The SDK leverages Lambda durable functions to automatically checkpoint progress and pause execution for up to a year while waiting for external events, removing the need for custom state management code. This supports complex use cases like deployment pipelines with approval gates and AI model orchestration that requires human review.

• Eliminates custom progress tracking code for long-running Go workflows on Lambda
• Supports execution pauses up to one year for waiting on external events or human input
• Enables native Go development for multi-step applications like deployment rollouts
• Integrates with Lambda's event-driven model via automatic checkpointing

AWS has expanded Lambda's built-in recursive loop detection to cover invocations triggered by Amazon EventBridge Custom event buses. This feature automatically halts execution when a misconfiguration or code defect causes a Lambda Event Function to recursively trigger itself via EventBridge rules. The update aims to prevent runaway workloads and unexpected billing spikes by stopping infinite loops at the source.

• Prevents unexpected costs from recursive Lambda-EventBridge invocation loops
• Automatically stops runaway workloads triggered by Custom event bus misconfigurations
• Feature is enabled by default on AWS accounts, requiring no manual setup
• Protects automation workflows and scheduled tasks from code defects

Amazon Application Recovery Controller's Region Switch now supports four new capabilities for orchestrating multi-Region failovers. These include an AWS Systems Manager Run Command execution block, custom action Lambda payloads, service quota parity checks, and a wait-for-target-group-health feature for compute scaling. The updates allow for more complex automation and validation steps during the recovery process.

• Execute SSM Run Command actions on managed instances directly within failover plans
• Attach custom payloads to Lambda execution blocks for tailored recovery logic
• Automatically verify service quota parity across Regions before switching
• Pause scaling actions until target group health checks pass successfully

AWS has expanded Lambda durable functions to the GovCloud (US-West and US-East) regions. This feature adds checkpointing and pause/resume primitives to the Lambda event handler, enabling reliable orchestration of multi-step workflows and AI tasks. It allows regulated industries to build complex applications with automatic failure recovery and no compute charges during wait states, all within strict compliance boundaries.

• Adds 'step' and 'wait' primitives to Lambda handlers for checkpointing
• Enables automatic recovery from failures in multi-step workflows
• Pauses executions without incurring compute charges for on-demand functions
• Supports AI-assisted tasks and complex orchestration in regulated environments

Oracle Database@AWS now allows advertising aggregate CIDRs and default routes when peering with Amazon VPCs. This eliminates the need to manually update specific CIDR advertisements for every new network extension. The change simplifies routing management for environments using AWS Transit Gateway or on-premises connections.

• Advertise 0.0.0.0/0 or broad aggregates instead of individual CIDRs for ODB peering
• Reduces operational overhead when adding new VPCs or on-prem networks
• Leverages AWS Transit Gateway for scalable connectivity to ODB databases
• Removes previous limitation requiring separate CIDR advertisements per network

Fleet impact: For DBAs managing Oracle ExaCC on AWS, this simplifies network topology changes by allowing broad route advertisements. Review current ODB peering configurations to potentially consolidate CIDR lists and reduce maintenance tasks when extending VPCs or on-prem networks.