Choose serverless, workers or containers from workload constraints

Compute choice follows workload shape. Short stateless bursts can fit a function. Long asynchronous jobs need a queue and worker lifecycle. Persistent connections, custom runtimes or steady state may fit containers. Standardizing blindly on one option moves complexity rather than removing it.

Build the decision artifact

Measure execution duration, arrival pattern, concurrency, memory and accelerator needs, state, network behavior, cold-start tolerance and cancellation. Add security boundaries, deployment frequency, observability and the team’s operating capability. Model retry amplification and quota behavior before estimating cost.

The enterprise delivery lab makes this reasoning executable with synthetic data:

from enterprise_cases import workload_platform_case

result = workload_platform_case()
assert result["choices"]["thumbnail"] == "serverless"
assert result["choices"]["document_batch"] == "worker"
assert result["choices"]["stream_gateway"] == "container"
assert result["one_platform_for_all"] is False

The fixture maps a two-second thumbnail to serverless, a fifteen-minute document batch to a worker and a long-lived streaming gateway to a container. These are reasoned defaults for the authored constraints, not provider-independent laws. A different runtime limit or traffic profile can change the answer.

Test the failure path

A platform decision based only on nominal price omits idle capacity, egress, queueing, failure recovery and engineering effort. A serverless endpoint that synchronously waits on a model can hit timeouts; a worker without bounded concurrency can overwhelm the same model API.

Save evidence that another reviewer can inspect:

• workload envelope and burst model
• platform constraint comparison
• failure and cancellation test
• cost model per successful task

This exercise aligns with the enterprise delivery, secure integration, operations and leadership work in the FDE for Professionals course. The course link describes the pathway; this local draft and its synthetic verifier do not claim that a client deployment, security approval or production result occurred.

Practice task

Classify three steps from one AI workflow. Defend a different compute choice for each, then state the measurement that could make you change it.

Continue learning

This article is part of the Enterprise AI delivery and architecture sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Design compensating actions for partially completed operations.
• Continue with Create an environment and identity separation plan.

Reference: Kubernetes workload documentation.