Parallel where it's safe, serial where it matters.

The simplest problem and the one architecture solves: components are containerised with no significant local state, so adding capacity means adding instances behind a load balancer. The relational database remains the point to size carefully, as in any application.

What's useful to know during evaluation is that declared permissions introduce no post-filtering cost: ownership and visibility conditions translate into filters applied at source, not into discards performed in memory after reading more than necessary. It's a difference that shows at volume.

The problem most often underestimated at design time and which appears as soon as the system goes live: a bulk process, a call to a slow external service, generating a complex document.

Two mechanisms address it. The first is asynchronous continuation: a pipeline can answer the requester immediately and continue in the background, reporting completion. It's needed exactly when an external service takes seconds or minutes and there's no sense keeping a person waiting.

The second is a circuit breaker for every external service: when a service starts failing or slowing beyond a threshold, calls to it are cut off instead of piling up. That's what prevents a slow supplier dragging the whole system down, which is the most common failure mode in integrated architectures.

The specific problem of systems executing dynamically generated plans, and one that must be handled by construction because preventing it through care alone isn't realistic.

Every compiled plan runs under explicit limits: maximum step count, nesting depth, loop iterations, query size. Limits are configurable by classification, meaning an internal maintenance automation can have different margins from one exposed to users.

Recursion is denied by default: a pipeline can re-enter itself only if it explicitly opts in, and when it does the re-entry is tracked end to end. It's the protection against an automation that, through a design error or unexpected data, triggers a loop consuming everything.