aws-lambda-managed-instances

Assess traffic and costs first, then create a capacity provider with VPC config and an operator IAM role, attach it to the function, and publish a version to trigger provisioning. Review code for concurrency safety and shift traffic gradually with weighted aliases.

LMI becomes cost-competitive at roughly 50-100M+ requests per month with steady, predictable traffic. It is not cheaper for low-volume, bursty, very short-duration, or scale-to-zero workloads because three instances run continuously at minimum.

No. Python on LMI uses process-based isolation, so globals and module-level state are not shared between invocations. You only need to handle shared /tmp file paths, per-process connection pools, and memory sizing for multiplied processes.

The most common cause is private subnets missing a NAT gateway route, since instances need outbound internet access for image pulls and Lambda service communication. Also verify subnets span 3+ AZs with available IPs and the operator role has required permissions.

Maximum concurrency per vCPU is 64 for Node.js, 32 for Java and .NET, and 16 for Python. The actual setting is PerExecutionEnvironmentMaxConcurrency per execution environment, and CPU-bound workloads should use 1-2 per vCPU.

No. LMI requires a minimum of three execution environments for AZ resiliency and cannot scale to zero. If your workload needs to pay nothing when idle, standard Lambda is the better choice.