Checkpoints, replay… and that yields Lambda functions that are “durable”.
AWS highlighted this option at its re:Invent 2025 conference. The promise: executions can run for up to a year, with reliable resumption after interruption or pause.
An SDK to integrate into the function code enables implementing the primitives that manage this mechanism. Pauses can be set for a fixed duration, and you can also condition the resume on a specific event.
The pricing is structured across three tiers:
- “Durable” operations (stages, pauses, callbacks): $8 per million
- Data written: $0.25/GB
- Data stored: $0.15/GB/month
Lambda, a bit less serverless
Another option highlighted: managed Lambda instances. This approach involves selecting the EC2 configurations on which to run the functions.
To do this, you create “capacity providers.” These providers run within the AWS account, inside a VPC (and at least one subnet). You can configure several aspects:
- CPU architecture
- Allowed instance types (whitelist, blacklist, or unrestricted)
- Maximum number of instances
- Scaling mode (manual or automatic)
- EBS encryption key (potentially custom)
Another concurrency model…
When you publish a version of a function associated with a capacity provider, Lambda launches managed instances (three by default for resilience). Alternatively, it uses existing ones if resources are sufficient to host the runtime environment.
Similarly, a runtime environment can handle several invocations in parallel (64 maximum). The concurrency model is therefore different from that of standard Lambda (one invocation = one environment).
… on security…
This system assumes that thread safety, state management, and context isolation must be handled differently depending on the context.
Security providers effectively form the trust boundary. With managed Lambda instances, functions run in containers, which do not offer the same level of security as the standard Firecracker micro-VM technology.
… on scaling…
With managed Lambda instances, there is no cold start. Scaling is asynchronous, based on CPU usage signals. In this sense, the option should be reserved for workloads with predictable traffic. AWS also does not guarantee stability if the load more than doubles within a five-minute window.
Four parameters influence scaling:
- Memory and vCPU allocated to a function
- Maximum concurrency per environment
- Target resource utilization
- Allowed instance types
… and pricing
Managed Lambda instances are billed at on-demand EC2 prices, with a 15% surcharge. The option nonetheless allows taking advantage of potential discounts (savings plans, reserved instances…). There is an additional charge of $0.20 per million requests.
