Lambda raises its timeout to 90 minutes, on managed instances only
Victoria Shinder2 min readAWS has raised the Lambda function timeout to 90 minutes — but only for asynchronous and event source mapping invocations, and only on Lambda Managed Instances. The announcement is at https://aws.amazon.com/blogs/compute/announcing-90-minute-function-timeout-on-aws-lambda-managed-instances/.
The ceiling has moved twice in eleven years: 5 minutes at launch in 2014, 15 minutes in 2018, and now six times that for one execution mode. AWS lists the workloads that kept running into the wall — speech-to-text and video transcoding, Monte Carlo simulations and portfolio risk, ETL over multi-gigabyte datasets at peak, model inference that fits Lambda's memory profile but not its clock, and crawls or file transfers held up by a slow counterparty.
Two details decide whether this applies to you. Managed Instances are the on-demand capacity mode where several concurrent requests share an instance and pricing follows EC2 rather than per-invocation billing; the 90 minutes belong to that mode, not to Lambda in general. And within durable functions — which checkpoint progress and replay past completed work after a failure — a multi-step asynchronous execution can now run for up to a year.
What it means
The number that did not change is the interesting one. Synchronous, request-driven functions still stop at 15 minutes. That is not an oversight: a synchronous invocation has a caller holding a connection, and the limit is as much about the caller as the runtime. Anyone reading "Lambda now runs for 90 minutes" and planning an HTTP handler around it will find out the hard way.
This is serverless converging on the thing it was defined against. Multiple concurrent requests per instance, EC2-style pricing, 90-minute executions — each is a reasonable answer to a real complaint, and together they describe a server that someone else patches. The useful question is no longer "serverless or not" but which of the three form factors you are on, because the limits, the billing and the failure modes differ on each.
And the checkpointing is the part worth copying even if you never use Lambda. A job that records progress and skips completed work on replay is the only kind that survives a 90-minute window honestly. Without it, a long timeout just means a longer way to fail and start over.