OffNet Newsroom The Brief

Tuesday, July 07, 2026

6 stories worth your scroll — picked from 48 curated today.

1 · Copy  →  2 · Share  →  3 · Paste & post

View the post text

Today's updates redefine LLM inference efficiency, database stability, and cloud cost automation. ▸ SageMaker HyperPod's disaggregated prefill/decode reduces latency for long-context LLM inference via EFA. ▸ Postgres 14-16 updates fix WAL replay deadlocks that blocked replication and point-in-time recovery. ▸ ASK+ resolves context poverty in RL agents, enabling effective guidance from small language models. ▸ Dynata cut DB ops overhead 70% by consolidating to AWS Savings Plans and moving to Graviton. ▸ HubSpot scaled semantic search to 20B vectors, powering agent retrieval and RAG for 38 teams. ▸ AWS ACM supports ACMEv2, enabling automated 45-day TLS certs to meet new CA/Browser standards. Summaries, takeaways and sources → https://offnetchat.com/news/2026-07-07/brief.html #AWS #PostgreSQL #LLM #SiteReliability #CloudCost

LLMs NEW AWS What's New

SageMaker HyperPod adds disaggregated prefill and decode with EFA GPU-Direct RDMA

Amazon SageMaker HyperPod now supports Disaggregated Prefill and Decode (DPD), separating LLM inference phases onto dedicated GPU pools. The system transfers key-value caches between these pools via Elastic Fabric Adapter using GPU-Direct RDMA. This architecture prevents long-context prefill requests from stalling token generation for concurrent decode tasks, improving latency consistency and throughput under mixed workloads.

DPD isolates compute-bound prefill from memory-bound decode onto separate GPU resources.

Database Technology NEW Planet PostgreSQL

Postgres 14-16 Replication Deadlock Bug Affects WAL Replay

Minor releases 14.23, 15.18, and 16.14 introduced a regression causing a MultiXactOffsetSLRU deadlock during WAL replay. This issue can halt streaming replication standbies or block point-in-time recovery operations. The bug was first reported on May 20th and has since generated multiple reports across mailing lists and support channels.

Affected versions: Postgres 14.23, 15.18, and 16.14.

Agentic AI NEW arXiv cs.AI

ASK+ Fixes SLM Guidance for RL Agents by Solving Context Poverty

Reinforcement learning agents often fail to utilize small language models (SLMs) for assistance because vanilla uncertainty-gated approaches result in near-zero overwrite rates. This failure stems from insufficient egocentric context rather than a lack of model capacity. The proposed ASK+ method resolves this by providing SLMs with trajectory-aware context, such as partially revealed maps and visited positions, enabling meaningful reasoning and action guidance.

Vanilla SLM guidance in RL fails due to context poverty, not model capacity.

AWS NEW AWS Database Blog

Dynata cuts DB ops overhead 70% with AWS Savings Plans

Dynata shifted from managing fragmented Reserved Instances to AWS Database Savings Plans to streamline cost optimization. This consolidation reduced operational overhead by 70% and extended coverage to Amazon Aurora Serverless. The move accelerated their migration to Graviton processors while lowering overall TCO.

Consolidate fragmented Reserved Instances into Database Savings Plans for flexibility

AI / ML NEW InfoQ

HubSpot Scales Semantic Search to 20B Vectors for Agents and RAG

HubSpot evolved its semantic search platform from a proof of concept into a core service handling over 20 billion vectors for more than 38 teams. The infrastructure now underpins key capabilities including agent retrieval, RAG workflows, and contact deduplication. Rising agent adoption has shifted the primary engineering focus toward optimizing retrieval quality and reducing latency at massive scale.

HubSpot manages 20B+ vectors internally, supporting 38+ teams.

Automation / DevOps / IaC NEW AWS What's New

AWS ACM adds ACMEv2 support for automated public TLS certificate management

AWS Certificate Manager now exposes a fully managed ACME server endpoint, allowing standard clients like Certbot and cert-manager to issue 45-day public TLS certificates from Amazon Trust Services. This feature addresses the upcoming CA/Browser Forum mandate for shorter certificate lifetimes by enabling standardized, automated issuance and renewal workflows. Administrators can enforce governance through domain scoping and wildcard policies while delegating requests to application teams.

Provision public TLS certs using any ACMEv2 client like Certbot or cert-manager directly from ACM.