AWS

AWS Lambda Managed Instances: 90-minute async timeouts and Graviton5 support

AWS Lambda Managed Instances now allow 90-minute async and event-source timeouts and Graviton5 instance support, shifting longer batch and inference jobs to managed hosts.

September 10, 2026·3 min read·AI researched · AI written · AI reviewed

AWS just extended how long a Lambda "function" can actually live when you run it on Managed Instances — 90 minutes for asynchronous and event-source mapping invocations. That's six times the old 15-minute ceiling and a deliberate nudge: run longer batch jobs, media transcodes, or AI inference on Lambda's managed-host model instead of stitching together EC2 fleets and ad-hoc glue code.

This new timeout applies only to asynchronous invocations and event-source mappings; synchronous (request/response) invocations remain capped at the existing 15-minute limit. That distinction matters. AWS isn't trying to turn Lambda into a generic long-lived compute platform — it's providing a managed execution boundary for jobs that naturally decouple from immediate client response. Expect teams to re-evaluate where they keep state, how they checkpoint progress, and how they surface long-running failures.

Two operational implications are immediate and unavoidable.

First, observability and lifecycle modeling become non-trivial. Ninety minutes of execution means more windows for partial progress, transient retries, and eventual consistency anomalies. If your monitoring treats a Lambda invocation as an atomic short-lived unit, it will under-report intermediates and over-report tail latencies. Instrument for progress checkpoints, expose per-invocation heartbeats, and treat Managed Instances like small, ephemeral batch workers — not stateless request handlers.

Second, capacity semantics and instance selection now influence performance and cost. Lambda Managed Instances can run on the new Arm-based Graviton5 instance families (the publicly announced C9 and M9 families, including variants with local NVMe). AWS reports performance improvements versus earlier Graviton generations, but the exact uplift will depend on your workload and the runtime/native libraries you use. Choose instance families via the Managed Instances capacity configuration in the console or CLI, and test binaries and native dependencies before switching.

If you run platform teams, this is the right call from AWS. The alternative was teams cobbling together privileged credential injection and long-lived EC2 workers with no audit trail. Giving a managed, auditable host that integrates with Lambda's lifecycle and autoscaling reduces ad-hoc patterns. That said, the product also hands you a new attack surface — longer-lived executions and choice of instance families increase the blast radius if your runtime is compromised. Treat these like short-lived VMs: enforce image hardening, runtime least privilege, and EDR/monitoring where required.

Other updates this week are smaller but cumulative. Gateway Load Balancer added TCP Reset (RST) behavior you can enable per target group to close stale or failed connections quickly; that speeds failover for stateful flows and reduces connection tail latency. On the AI front, Amazon Bedrock's model catalog continues to expand with newer Anthropic and third-party model families and higher-context options; check the Bedrock docs for the exact model names and context sizes available in your Region. Recent vendor updates also highlight much larger context windows and longer-running agent patterns on dedicated compute.

All of this nudges architectures in the same direction: managed primitives that blur the line between serverless and batch. Lambda Managed Instances with longer async timeouts and Graviton5 instance support will replace many homegrown worker fleets, Gateway LB's TCP Reset reduces tail-case network pain, and Bedrock continues to make higher-capability models operationally accessible.

One final, blunt note: if your platform hasn't treated long-lived serverless tasks as first-class citizens — with checkpoints, idempotent retries, capacity policies, and hardened images — you will be surprised by the bugs and incidents this feature surfaces. Teams that treat Lambda as only "short work" are about to find a lot of legacy plumbing that needs rethinking. AWS gave you the tool; now do the engineering work it actually requires.

Sources

aws-lambdagraviton5gateway-load-balanceramazon-bedrock
← All articles
AWS

Amazon Bedrock adds Astra model family; EKS fast-loading patterns and Lambda SnapStart impacts

AWS Bedrock added an Astra model family and expanded model access. EKS fast-loading patterns and Lambda SnapStart change startup, cache, and tenancy trade-offs.

Sep 9, 2026·3mamazon-bedrockastra
AWS

Amazon Bedrock Adds Anthropic's Claude Fable — AgentCore Runtime, MCP Federated Data Patterns, and EKS AI Ops

Amazon Bedrock adds Claude Fable; AWS published AgentCore and MCP federated patterns. Platform teams must treat agents, MCP, and GPU EKS as core infra pieces.

Sep 8, 2026·3mamazon-bedrockclaude-fable
AWS

AWS Bedrock expands regional model access and cross-Region inference; Lambda SnapStart & Network Connector operator policy

AWS Bedrock adds regional model access and cross-Region inference to ease compliance and latency. Lambda SnapStart for images and a Network Connector policy.

Sep 6, 2026·3mamazon-bedrockaws-lambda