Sunday, September 20, 2026

Henrietta Dombrovskaya has released significant updates to the pg_acm extension, addressing bugs she identified after leaving DRW. The update includes several requested new functions and conceptual improvements to the tool's core logic. While the code is now published, the author notes that comprehensive documentation is still pending.

• pg_acm receives critical bug fixes and new function additions.
• Conceptual changes improve the extension's underlying logic.
• Documentation is currently incomplete despite code release.
• Updates stem from post-DRW development backlog.

This AWS blog post details a three-tier architecture combining Amazon ElastiCache for Valkey, an exact-match cache, and Amazon Aurora PostgreSQL to perform membership lookups. By layering a Bloom filter with strict verification against the relational source of truth, the pattern eliminates false-positive risks inherent in probabilistic structures. The design achieves sub-millisecond decision times at peak throughput while ensuring data correctness.

• Use Valkey for high-speed probabilistic filtering before hitting Aurora PostgreSQL
• Always verify Bloom filter hits against an exact-match layer to prevent false positives
• Architect for sub-millisecond latency without compromising data integrity guarantees
• Leverage this pattern for high-throughput membership queries in read-heavy workloads

Fleet impact: For Aurora PostgreSQL and ElastiCache fleets, this pattern offers a robust way to offload high-volume membership checks while maintaining strict consistency. Ensure your Valkey instances are sized for the filter load and that the exact-match fallback path remains performant under peak concurrency.

A PostgreSQL agent using the MCP protocol can inadvertently read another tenant’s data when connecting through PgBouncer in transaction mode. This occurs because the pooled connection retains session state from a previous client, bypassing the expected isolation. The issue highlights a critical flaw in relying on session variables for multi-tenant isolation when using connection pooling.

• PgBouncer transaction mode reuses connections without resetting session state
• MCP protocol lacks inherent session isolation for agent tool calls
• RLS policies reading session vars are vulnerable to connection reuse
• Agents may access unauthorized tenant data via pooled connections
• Reset session state explicitly before each transaction or tool call

PostgreSQL 19 will miss its expected fall release window due to a heavy review cycle and significant feature reversions during beta. The development team is prioritizing quality over the original timeline, reducing the release scope to stabilize the codebase. Beta 4 is scheduled for September 24, 2026, with the final release pushed back by weeks or potentially months.

• Postgres 19 is delayed beyond the traditional fall release schedule.
• Major features reverted during beta to maintain code quality.
• Release scope is being reduced to meet stability requirements.
• Beta 4 is set for September 24, 2026, with final release pending.
• Expect weeks to months of delay before the official launch.

The PostgreSQL 19 stable branch saw 71 commits in its final week, but this was offset by 74 reverts across six separate pull requests. The most significant revert involves the online data checksum transition feature, which was pulled by Daniel Gustafsson on September 16. This reflects a conservative stabilization approach as the release window closed.

• PostgreSQL 19 stabilized with a net positive of 71 commits after reverting 74 changes.
• Online data checksum transition was removed from v19, likely due to stability concerns.
• Six features remain tracked for inclusion, with two still actively shipping.
• PostgreSQL 20 master branch received 46 commits in the same reporting window.

A new study reveals that large language model weights can be extracted from cloud inference endpoints using timing side channels. The attack exploits subtle latency variations in the model's response generation to reconstruct the underlying parameter data. This method bypasses traditional API access controls by leveraging the physical execution characteristics of the hardware.

• Timing side channels can leak proprietary model weights even with secure APIs.
• Physical execution latency correlates with specific model parameter access patterns.
• Current API rate limiting and access controls do not mitigate this extraction vector.
• Model providers should evaluate hardware-level noise and jitter as defense mechanisms.
• High-value models may need obfuscation or noise injection to prevent reconstruction.

An engineer shares results from building non-autoregressive decision models using reinforcement learning a year ago. The approach deviates from standard token-by-token generation by producing decisions in parallel. The post highlights experimental outcomes and technical hurdles encountered during development.

• Non-autoregressive architectures can be viable for decision-making tasks.
• Reinforcement learning provides a training signal for parallel decision models.
• Parallel generation may offer latency benefits over sequential token prediction.
• Experimental results suggest potential but also highlight significant challenges.

Moonshot AI's Kimi K3 is now generally available on Amazon Bedrock, offering a 2.8 trillion parameter open-weight model with native vision and a 1-million-token context window. The model delivers approximately 2.5x better scaling efficiency than its predecessor and supports long-running coding sessions and multi-document analysis. It operates within the same security boundaries as other Bedrock models, enabling safe integration into enterprise agent workflows.

• Kimi K3 features a 1M token context window, ideal for large codebases and document analysis.
• Native vision capabilities allow processing of scanned pages and screenshots alongside text.
• Runs within standard Bedrock security boundaries for consistent governance and compliance.
• Reports 2.5x scaling efficiency improvement over Kimi K2 for better cost-performance ratio.
• Supports extended agent workflows requiring long-horizon reasoning and memory.

Percona evaluated how well large language models handle specialized database administration workloads by building a custom evaluation harness. The study focused on the ability of these models to execute actual DBA tasks on remote systems rather than just generating code. This provides a practical benchmark for the current state of LLM capabilities in systems administration.

• Percona built a custom harness to test LLMs on real remote DBA tasks
• Focus is on practical execution ability rather than theoretical code generation
• Results offer a baseline for LLM readiness in specialized database roles

Higgsfield is an open-source framework combining GPU workload management with machine learning capabilities for training massive models. It provides fault-tolerant orchestration and supports ZeRO-3 DeepSpeed and PyTorch fully sharded data parallel APIs. The system handles exclusive and non-exclusive resource allocation to enable efficient scaling for billions to trillions of parameters.

• Built for fault-tolerant, highly scalable GPU orchestration across clusters.
• Supports ZeRO-3 DeepSpeed and PyTorch FSDP for efficient model sharding.
• Manages both exclusive and non-exclusive compute resource allocation.
• Designed specifically for training LLMs with billions to trillions of parameters.

Amazon SageMaker AI now allows users to submit prioritized lists of instance types for training and processing jobs instead of specifying a single instance. The service automatically selects the first available capacity from the provided list, significantly reducing wait times for high-demand GPUs. This update removes the need for complex custom retry logic or concurrent job submissions to secure compute resources during peak periods.

• Submit prioritized instance lists to let SageMaker auto-select available capacity.
• Eliminate custom retry logic previously required for GPU availability.
• Reduce job scheduling latency during peak compute demand periods.
• Simplify job definitions by removing the need for multiple concurrent submissions.

Cua provides open-source infrastructure for training and evaluating AI agents that interact with desktop environments. It supports cross-OS fleets, isolated cloud desktops, and local macOS VMs to facilitate data generation and testing. The platform includes CUA-S1 for specialized decision models and Cua Bench for simulating and verifying agent tasks.

• Provision Linux desktops in the cloud to run commands and capture screenshots for agent training data.
• Use CUA-S1 to deploy small, specialized models optimized for computer-use decision making.
• Leverage Cua Bench to create and verify simulated tasks for evaluating agent performance.
• Operate specific applications like Calculators to test precision and verification capabilities.
• Connect to Tahoe VMs via SSH to integrate custom agents into the automation workflow.

AWS has generally available the next-generation AgentCore Runtime, a serverless microVM compute environment within Amazon Bedrock AgentCore. This update introduces elastic memory management that dynamically allocates and reclaims memory to optimize costs and ensures consistent cold start times independent of container size or concurrency levels. The runtime maintains serverless benefits like scale-to-zero and hardware-enforced isolation while reducing overall compute expenses.

• Elastic memory allocation reclaims unused RAM, shifting costs from peak to actual usage.
• Cold start latency is now consistent regardless of container image size or concurrency.
• Retains serverless features: no pre-provisioning, scale-to-zero, and hardware isolation.
• Reduces operational costs through efficient, on-demand memory profiling per session.

Enterprise teams are increasingly choosing on-prem PostgreSQL to support AI workloads, driven by cost savings from running open models and regulatory requirements for data control. This shift challenges the vendor assumption that open-source software is inherently easy to self-host at scale. The article highlights that production-grade deployment demands significantly more rigor than hobbyist setups.

• AI inference costs and data sovereignty are primary drivers for returning to on-prem infrastructure.
• Self-hosting open-source databases requires robust operational readiness beyond basic installation.
• Keep data internal to meet regulated industry compliance and control AI training inputs.

A recent analysis highlights that standard benchmarks for Btrfs, ZFS, and bcachefs frequently fail to represent actual production scenarios. The author argues that typical test suites skip critical performance characteristics relevant to database and cloud engineering workloads. This gap means synthetic scores may not accurately predict behavior under complex, concurrent access patterns.

• Standard benchmarks for Btrfs/ZFS/bcachefs may not reflect real production performance.
• Skip synthetic tests that ignore complex, concurrent access patterns.
• Validate filesystem choices against workload-specific metrics, not just raw IOPS.
• Be cautious when using public benchmark scores for critical database infrastructure.

AWS Step Functions now automatically provisions AWS SDK integrations for newly released services and capabilities within weeks of launch, eliminating manual wait times. The initial release includes support for AWS Lambda MicroVMs and AWS Lambda Core, enabling immediate orchestration of these resources. This automation allows developers to build distributed applications using the latest AWS features without custom coordination code.

• New AWS services get Step Functions integrations automatically within weeks of release.
• Initial support covers AWS Lambda MicroVMs and AWS Lambda Core.
• Orchestrate agentic workflows without writing custom coordination logic.
• Lambda MicroVMs offer isolated execution with built-in retry handling for failed starts.

AWS Batch now allows operators to cancel or terminate up to 50 jobs in a single API call using CancelJobs, TerminateJobs, and TerminateServiceJobs. This update reduces operational overhead for large-scale workloads by enabling group actions and returning per-job status in one response. Enhanced ListJobs and ListServiceJobs endpoints now expose isCancelled and isTerminated fields for better lifecycle tracking.

• Use CancelJobs for jobs in SUBMITTED, PENDING, or RUNNABLE states.
• TerminateJobs and TerminateServiceJobs handle any state, including RUNNING.
• All three bulk APIs accept up to 50 job IDs per request.
• ListJobs now returns isCancelled and isTerminated for status checks.
• Works with both individual jobs and array job components.

AWS has integrated Amazon Q Console with CloudTrail, enabling natural language queries for API activity logs. Users can investigate security events, troubleshoot operational issues, and audit trail configurations without writing code or manually parsing JSON. The feature supports checking logging coverage, identifying gaps, and tracing specific IAM role access.

• Eliminates need for manual log parsing or query writing in CloudTrail.
• Enables natural language troubleshooting for operational and security incidents.
• Automates validation of trail configuration and data event source coverage.
• Simplifies access investigations by querying specific IAM role usage.

Amazon ECS Express Mode now allows specifying ARM64 as the CPU architecture, enabling direct deployment of ARM-based container images on Graviton instances. This update simplifies the launch of containerized applications by handling networking, load balancing, and auto-scaling while offering up to 40% better price-performance than x86 alternatives. Users can now select the compute architecture that best fits their workload requirements without manual infrastructure configuration.

• Deploy ARM64 container images directly via ECS Express Mode on Graviton compute.
• Achieve up to 40% improved price-performance compared to x86-based instances.
• Automated orchestration includes networking, load balancing, and auto-scaling.
• Select compute architecture to match specific workload needs easily.
• No manual infrastructure setup required for rapid application launch.

AWS has released Amazon EC2 T8i instances, a new generation of low-cost burstable compute powered by custom sixth-generation Intel Xeon 6 processors. Built on the latest AWS Nitro cards, these instances offer up to 30% better price performance and up to 70% higher compute performance compared to previous T3 generations. They are available in nano, micro, small, and medium sizes, targeting workloads with low to moderate CPU utilization such as small databases, CI/CD pipelines, and microservices.

• T8i uses custom 6th Gen Intel Xeon 6 processors on AWS-exclusive Nitro hardware for improved efficiency.
• Up to 30% better price performance and 70% higher compute than T3 instances.
• Available in nano, micro, small, and medium sizes for cost-sensitive, low-to-moderate CPU workloads.
• Offers 1.25x network bandwidth and 2.4x EBS bandwidth compared to predecessors.
• Ideal for small databases, batch processing, event-driven functions, and CI/CD pipelines.

AWS has enabled S3 Express One Zone in seven additional regions, including Singapore, Sao Paulo, N. California, Canada Central, Paris, Sydney, and Seoul. This single-Availability Zone storage class provides consistent single-digit millisecond latency for latency-sensitive workloads. It offers up to 10x faster data access and 80% lower request costs compared to S3 Standard. The service is optimized for machine learning training, interactive analytics, and AI search engine caching.

• S3 Express One Zone is now live in 7 more global regions for low-latency needs.
• Delivers single-digit millisecond access with up to 10x speed over S3 Standard.
• Request costs drop by up to 80% compared to the standard storage class.
• Ideal for ML training, interactive analytics, and key-value caching in AI apps.

AWS Resilience Hub has introduced three new features to help platform and SRE teams assess workload resilience. The update allows users to scope service inputs using existing EKS namespace labels, providing dependency insights for better visibility, and enables sharing resilience policies across organizations via AWS Organizations. These additions integrate with existing capabilities like automated dependency discovery and generative AI failure analysis.

• Scope service inputs in Resilience Hub using EKS namespace labels instead of manual ARN lists.
• Gain deeper visibility into workload resilience through new dependency insights.
• Standardize and share resilience policies across multiple AWS accounts using AWS Organizations.

Amazon Keyspaces, the managed serverless service compatible with Apache Cassandra, has reached general availability in 11 additional AWS Regions. This expansion enables customers to deploy Cassandra-compatible applications closer to end-users, reducing latency while ensuring data remains within specific geographic boundaries to satisfy residency mandates. As a serverless offering, it provides virtually unlimited throughput and storage, charging only for the resources consumed.

• Keyspaces is now GA in 11 new regions, expanding geographic coverage for Cassandra workloads.
• New regions help meet strict data residency requirements by keeping data within local boundaries.
• Serverless architecture scales automatically to handle thousands of requests per second without provisioning.
• Cost model remains pay-per-use, covering both throughput and storage consumption efficiently.