Kubernetes Kai
“Place it wherever the diagram has whitespace.”
Who decides where a pipeline runs?
Elevator answer
Place work together until ownership, scaling, security, latency, or failure isolation gives you a concrete reason not to. Package names are not placement strategy.
3 coffee-machine misconceptions
Kubernetes Kai
“Autoscaling is an architecture decision with metrics.”
Microservice Mike
“Network distance is a sign of independence.”
The real explanation
If PDF rendering needs ten times the memory of order validation, or Fulfillment belongs to another team with its own lifecycle, a runtime boundary may earn its cost. If two steps merely sit in different packages, sending them over HTTP gives the package structure latency and a pager. Runtime placement is a production decision, not diagram feng shui.
TPF treats transport and platform as orthogonal dimensions. gRPC, REST, and LOCAL are transport modes. Function-style and container-oriented deployments are platform or deployment choices. A flow may be local today, gain a remote boundary tomorrow, or use a function pattern for a particular entry point. The business semantics should remain coherent while the generated adapters and runtime mapping reflect the chosen placement.
The right placement emerges from trade-offs. Co-locate work that needs low latency, shared state, or simple operation. Split work that has separate ownership, security, scaling, availability, or failure isolation needs. Place a high-volume connector where it can be operated without forcing unrelated business behavior to scale with it. Do not split merely because an internal step looks important on a diagram.
The framework makes this explicit because accidental placement is expensive. If a local method silently becomes a remote call, its failure, cardinality, tracing, and retry behavior change. If build topology is mistaken for runtime layout, a refactor can produce a deployment claim nobody actually validated. TPF’s model allows the compiler and generator to participate in the decision rather than letting a package boundary invent it.
The cost is that teams must make the decision in plain language. A runtime mapping is not an infrastructure footnote. It is part of how the application behaves under latency, failure, and load.
Trade-offs
TPF gains an explicit place to reason about deployment behavior. It gives up the accidental simplicity of “the module decides.” Teams must maintain the mapping as architecture changes.
When TPF is not a good fit
If a system has one small local runtime and no credible reason to vary placement, keep the layout simple. Explicitness should not become topology theatre.