Seventy-nine announcements landed in the week of 28 September, five of them frontier models, and the one with the longest tail was a lifecycle notice: Amazon DevOps Guru, Amazon Managed Blockchain, and the Backint Agent for SAP ASE entered sunset, Infrastructure Composer's standalone console goes in December, and Mechanical Turk reached end of support. Two days later AWS Health shipped a version catalog so that kind of news stops arriving as a surprise. On the migration front, AWS Transform added agentic Kafka-to-MSK assessments. EKS Auto Mode opened up kubelet, sysctl, and hugepage tuning, Kubernetes 1.37 arrived, and Aurora PostgreSQL started querying Iceberg and Parquet in S3 through an embedded DuckDB.

The sunset list, and the catalog that should replace it

The AWS Service Availability Updates of 29 September move Amazon Chime SDK SIP Media Application and Amazon WorkSpaces Secure Browser to maintenance (closed to new customers from 29 October 2026, existing customers continue). Entering sunset: Amazon Managed Blockchain (end of support 29 September 2027), Amazon DevOps Guru (30 September 2027), AWS Backint Agent for SAP ASE (29 September 2027), and AWS Infrastructure Composer's standalone console (7 December 2026). Amazon Mechanical Turk reached end of support on 29 September 2026.

DevOps Guru is the one most infra teams will feel. Many turned it on in 2021 and forgot about it, so it is still in CloudFormation templates and still wired to SNS topics that page someone. AWS names no successor; given CloudWatch Omni and the DevOps Agent shipped a week earlier, the agentic investigation path is the obvious reading. Audit for AWS::DevOpsGuru::ResourceCollection and NotificationChannel resources now and decide what replaces the alerts, because twelve months passes fast when the removal needs a change ticket in forty accounts.

Then on 2 October AWS Health introduced the version catalog: a service-wide view of supported software versions and their lifecycle timelines, in the Health Dashboard for everyone and through the Health API for Business Support Plus, Enterprise Support, and Unified Operations customers. At launch it covers RDS, EKS, and Lambda, with more services to follow. Wire the API into whatever tracks your upgrade backlog; an EKS version with four months left should be a ticket, not an email.

AWS Transform assesses Kafka to MSK

On 29 September AWS Transform added agentic migration assessments for on-premises Apache Kafka to Amazon MSK. You upload a cluster inventory file or describe the clusters in chat, and the agent checks each cluster for MSK compatibility across topology, version, configuration, authentication, and quotas, sizes MSK Express brokers, and produces a TCO business case with pricing options and next steps.

The slow part of a Kafka migration is not the data copy, it is the weeks of spreadsheet work to find out whether it fits and what it costs. Compressing that to minutes changes who gets to ask the question.

Read the word "assessment" literally. It produces a case, not a migration. Consumer group offsets, schema registry, MirrorMaker topology, client-side auth changes, and the partition count decisions Express brokers impose are all still yours. Treat the TCO as an opening number and verify the Express broker sizing against your retention and throughput peaks.

EKS Auto Mode drops its last excuse

On 1 October EKS Auto Mode added advanced compute configuration: a new advancedCompute field in the NodeClass resource lets you set kubelet options, Linux sysctls, and hugepages declaratively: user namespaces for rootless image builds, eviction thresholds, ARP cache limits, 2Mi or 1Gi hugepages for databases and HPC. EKS validates the settings, applies them at boot, and keeps them through scaling and upgrades, with no launch templates, custom AMIs, or privileged DaemonSets.

"We need a custom AMI for sysctl tuning" was the most common honest reason for staying on self-managed node groups. That reason is gone, in the right way: the tuning lives in a versioned Kubernetes object, not an image pipeline nobody wants to own. The failure mode is the same as any node-level knob: a bad sysctl in a NodeClass rolls out to every node that class provisions. Give the tuned NodeClass its own node pool and canary it with a single workload.

On 2 October EKS and EKS Distro added Kubernetes 1.37: the Metrics API is GA as metrics.k8s.io/v1, DRA device taints and tolerations are GA, and HPA scale-to-zero is beta and on by default. The same day ECS added blue/green, linear, and canary deployments for services behind VPC Lattice.

Aurora PostgreSQL reads the data lake in place

On 30 September Aurora PostgreSQL gained direct querying of Apache Iceberg and Parquet data. You create PostgreSQL foreign tables over data in S3, S3 Tables, or the Glue Data Catalog (including external Iceberg REST catalogs federated through Glue), and Aurora runs those queries with DuckDB embedded in the engine. Existing applications and BI tools keep talking plain PostgreSQL; for latency-sensitive paths you can materialize the lake data locally.

This kills a class of nightly "copy the dimension tables into Postgres" jobs, and DuckDB is a sensible engine for scanning columnar files. The trade-off is that the scan now runs on your OLTP instance. A reporting query over a year of Parquet is a lot of CPU and memory on the box that also serves checkouts; put the lake-facing workload on a reader, and watch the connection you give to the BI tool as closely as you would a cron job.

Also this week

The throughline

The week's real theme is lifecycle. A service you adopted in 2021 is leaving. The version catalog exists so the next one does not catch you out. AWS Transform now writes the business case for leaving on-prem Kafka. EKS Auto Mode removed the configuration gap that kept node groups alive past their time, and Aurora removed the copy job that kept a dozen ETL pipelines alive past theirs. Platform work is mostly deciding what to stop running, and this week AWS shipped tools for both halves of that: finding out what is ending, and making the replacement cheap enough to actually do.

Read this next

For the AI and agent reading of the same week, see ercan.ai. If your team needs hands-on help with AWS, cloud architecture, CI/CD pipelines, Linux, GitHub or GitLab, and the platform work around them, the consulting side lives at ercanermis.com.

References