Home› Blog› AWS Weekly Intelligence #16 - 16-20 November 2026
AWS Weekly Intelligence AWS

AWS Weekly Intelligence #16 - 16-20 November 2026

Verified against current vendor documentation on 21 November 2026. Pricing, limits and API behaviour were checked against the official docs on that date. Cloud services change fast — if you are reading this much later, treat the specifics as a starting point and re-check the linked sources.

The pre-re:Invent wave, and it broke late: Thursday and Friday alone carried ninety-eight announcements. The through-line across the whole week is that AWS spent it raising ceilings and teaching existing services to describe themselves — S3 read rates, CloudFormation stacks, Lambda timeouts, and three separate launches that give networking and GPU workloads telemetry they did not have.

The week in one paragraph

Monday to Wednesday were steady and specific: hybrid search in S3 Vectors, a second KMS key on EBS snapshots, queues for Kinesis Data Streams, resource context shared from Transit Gateway into firewalls, GPU Event Functions on Lambda, and network path monitoring in CloudWatch. Then Thursday and Friday delivered ninety-eight announcements between them, Friday carrying seventy-four on its own.

Two themes hold it together. The first is quota inflation at a scale that changes designs rather than conveniences — CloudFormation moving from 2,000 stacks per account and Region to 300,000 is the removal of a constraint that shaped how teams partitioned infrastructure. The second is telemetry for things that were previously opaque: network paths, VPC Lattice resources, catalog-wide data quality, and degraded-but-running GPUs. Four of this week's seven daily posts are about that second theme, and in three of them the useful detail was a documented behaviour the announcement did not mention.

Covered in depth — seven posts

Every weekday of this window produced something worth a full post, which is unusual.

DayPostWhat the announcement left out
Mon 16 #77 — A second key on the snapshot Associating a key does not re-encrypt; the original key still governs the data, so the sharing path and the decryption path are different grants
Tue 17 #78 — The queue gives up the sequence number Queue mode trades strict per-shard ordering for parallelism, so the position guarantee consumers were built on no longer holds
Wed 18 #79 — The image must outlive the version The runtime publisher takes on the patching half of the responsibility, and a deleted image breaks cold starts rather than deployments
Thu 19 #80 — A fifth status check A failed performance check also increments the aggregate StatusCheckFailed, so recovery and Auto Scaling replacement now fire on a nonterminal condition whose documented remedy is a reboot
Thu 19 #81 — The tag filter is an OR The filter matches when at least one key matches, so a multi-key filter widens rather than narrows, and it cannot separate a resource tag from its account's
Fri 20 #82 — A zero-flow alarm The metrics publish only in minutes carrying traffic, so the suggested zero-flow alarm needs missing data treated as breaching; service network owners get none of it
Fri 20 #83 — Catalog-scale anomaly detection A detected anomaly becomes normal in later runs unless explicitly excluded, and the default observation mode assumes the regular schedule the headline trigger is not

Compute: EC2 and Lambda

The most consequential compute launch of the week landed on Wednesday and is easy to miss: “AWS Lambda now supports GPU-accelerated Event Functions on Lambda Managed Instances using EC2 G-series instances.”

What makes it more than a bigger instance type is the scaling model. “Lambda runs multiple invocations concurrently in each execution environment and scales responsively based on GPU utilization, CPU utilization, and concurrency to keep your GPUs busy and reduce costs” — GPU utilisation as a scaling input is the thing teams build badly by hand. AWS also ships the awkward part: “Lambda provides GPU base images preconfigured with frameworks such as vLLM, SGLang, and PyTorch”, with “NVMe ephemeral storage of up to 32 GB each” for large artifacts and SOCI lazy loading to cut cold starts.

On EC2 proper, T9g instances on Graviton5 reached GA on Friday, and on Tuesday C8gn arrived in Israel (Tel Aviv) — worth noting for the number rather than the Region, since C8gn instances “offer up to 600 Gbps network bandwidth, the highest network bandwidth among network optimized EC2 instances” and “provide up to 30% better compute performance than Graviton3-based Amazon EC2 C7gn instances”.

Lambda also gained Node.js 26 and Python 3.15, a 30-minute timeout for on-demand Event Functions, SQS FIFO batches up to 10,000 messages, cost optimization recommendations, customer managed KMS keys for MicroVMs, and OpenTelemetry tracing for durable functions. Anyone who read Wednesday's post on customer-managed runtimes now has two new runtime versions to plan onto.

Networking: VPC, Transit Gateway and DNS

Tuesday's Transit Gateway launch is the week's most interesting networking change, because it attacks a problem every segmented network has: “AWS Transit Gateway now supports context sharing with security services, such as attached firewalls, allowing you to use Amazon Virtual Private Clouds (VPCs), VPC subnets, VPC security groups, and transit gateway attachments in firewall policies alongside IP addresses.”

AWS is unusually candid about what was wrong before: “Until now, attached firewalls segmented traffic by IP address, which is hard to plan and maintain, or by resource tags, which firewalls periodically synchronize with IP addresses and can miss new workloads in between.” That second clause is the real defect — a tag-synchronised firewall policy has a window in which a new workload is unprotected, and nobody sees it.

The fix is in-band rather than periodic: “Transit Gateway embeds resource context in the traffic it sends to your firewall, which can enforce the right policy from the first packet.” From the first packet is the claim that matters. Setup is “create a context store in AWS Network Manager and associate your transit gateways”, and both AWS Network Firewall and Palo Alto Networks Software Firewalls support it alongside IP rules.

Route 53 had a large week: Availability Zone (AZ) affinity routing, which “routes DNS queries from your VPCs to resources in the same AZ as the client that sends the query” — a cross-AZ data transfer reduction as much as a latency one; “Advanced Hosted Zones, a new hosted zone tier with additional guardrails, visibility, and resilience features”; public DNS fallback for private hosted zones; and health checks for private endpoints. Network Firewall announced “the public preview of automated Internet firewall”, with VPC IPAM policies extended to cover it.

Storage: S3 and EBS

Monday opened with two storage items in the priority list. The S3 one is a genuine capability change: “Amazon S3 Vectors now supports hybrid search that combines vector similarity search with full-text search within a single query.” The implementation detail is the useful part — you add the searchable text as another metadata field on PutVectors, and “S3 Vectors then automatically indexes that field for full-text search with BM25 relevance scoring”. It “maintains subsecond latency for infrequent queries and latency around 100 milliseconds or less for frequent queries”, which is the honest two-tier latency statement that RAG designs need.

By Friday S3 Vectors had also raised filterable metadata “up to 40 KB of filterable metadata per vector, up from 2 KB” — a 20-fold increase in the same week as hybrid search, which together make it a materially different product than it was on the 13th.

The EBS snapshot launch — associating additional KMS keys so encrypted snapshots can be shared and copied across accounts without re-encrypting — was Monday's, and is covered in depth in #77. On Friday, S3 general purpose buckets went to “at least 11,000 GET/HEAD requests per second, up from 5,500”.

Observability: CloudWatch

Wednesday's CloudWatch launch is the one to read if you have ever argued with AWS about whether the network was the problem. Performance monitors “report per-path metrics such as latency, a network impact score that shows network health, and traffic volume across all flows in your VPCs”, and the acquisition model is the appealing part: “Using telemetry AWS already collects, with no additional services to enable, performance monitors determine the paths your traffic takes, within and across Availability Zones, across AWS Regions, and to AWS services, for TCP, UDP, and ICMP traffic.”

The scoring method is worth understanding before trusting it: “They then score each path by comparing it with its own recent history and with all other paths AWS measures at the same time.” Two baselines — your own history and AWS's simultaneous fleet-wide measurement — which is what lets a monitor distinguish your path degraded from everything degraded. That is precisely the question a VPC owner cannot answer alone, and it is the strongest thing in this launch.

Alongside it: the PromQL query window went “a 30-day query window for PromQL queries against OpenTelemetry metrics, up from 7 days”, Managed Prometheus went to “up to 2 billion active metric time series and up to 300,000 total recording and alerting rules per workspace”, and EKS Container Insights and the Observability add-on gained custom OpenTelemetry and Prometheus telemetry with filtering. Friday's VPC Lattice metrics are in #82.

Agents: Bedrock AgentCore

Monday brought a Region expansion that is more interesting than Region expansions usually are: “Amazon Bedrock AgentCore runtime instances are now available in six additional AWS Regions”. The reason to care is what runtime instances are — “Runtime instances run your agents on Amazon EC2 instances while AgentCore handles provisioning, patching, scaling, and lifecycle management” — with GPU-accelerated, memory-optimized and compute-optimized types and agent sessions of up to 14 days.

A 14-day session is not a request-response model, and it puts the data-residency question in a different place: the instance, its storage volumes and its session data all live in the Region you pick, which is why the expansion matters beyond latency. Friday added AgentCore runtime instances on EC2 G7 instances, and the Marketplace gained agentic payments and an agent assessment mode.

The raised ceilings

Grouped because they are the same kind of event, and each invalidates a workaround somebody is still maintaining:

ServiceWasNow
CloudFormation stacks per account and Region2,000300,000 by default
S3 read requests, general purpose buckets5,500 GET/HEAD per secondat least 11,000 per second
Lambda timeout, on-demand Event Functions15 minutes30 minutes
Lambda batch size, SQS FIFO ESMs in provisioned mode10 messages10,000 messages
Managed Prometheus per workspace—2 billion active series, 300,000 rules
CloudWatch PromQL query window7 days30 days
S3 Vectors filterable metadata per vector2 KB40 KB

The CloudFormation one deserves a second look, because the old limit was a design input. Teams split workloads across accounts, collapsed many logical stacks into fewer physical ones, or built stack-reaping automation specifically because 2,000 was reachable. A 150-fold increase does not make those designs wrong retroactively, but the next review should ask whether the partitioning still earns its complexity.

The Lambda SQS FIFO jump from 10 to 10,000 is the most extreme ratio here and the most conditional — FIFO event source mappings in provisioned mode. A consumer written around a batch of ten has per-message error handling that behaves very differently three orders of magnitude up.

What I would act on

  1. Audit recover actions and Auto Scaling health checks against the aggregate status-check metric. The week's one change that alters existing behaviour with nobody editing anything. Covered in #80.
  2. Look at Transit Gateway resource context if you run a segmented network. It closes a real gap — tag-synchronised firewall policies that miss new workloads between syncs — and it is enforcement from the first packet rather than after a sync interval.
  3. Turn on CloudWatch network performance monitors. No software to install, no service to enable, and it answers the question you cannot answer alone during an incident: is this path bad, or is everything bad.
  4. Reopen the CloudFormation partitioning decision, and check S3 prefix-sharding workarounds. Both are maintenance you may now be able to retire.

Complete inventory — Thursday and Friday

Every announcement AWS published on 19 and 20 November, with AWS's own one-line summary, newest day first. Built from the What's New feed rather than by hand, and every link was fetched and validated before publishing. The Monday-to-Wednesday launches in this roundup's focus areas — compute, storage, networking, observability and agents — are covered in the sections above.

Friday 20 November — 74 announcements

Thursday 19 November — 24 announcements

Official AWS references

The announcements the quotes and figures in this post were checked against. Every announcement from Thursday and Friday is linked in the complete inventory above.

Looking for one service rather than a whole week?

Every AWS, Azure and Google Cloud announcement is browsable by service and date, each linked to the vendor’s own page — and the other weekly roundups are collected in one place.

Browse announcements →

Comments

How was your experience?
Your feedback helps improve this site.
PoorExcellent