Sequential (pipeline)
A's output is B's input, all the way down.
Shape：Linear, one-way, no回頭路
When to use it
When stages have clear boundaries, each stage's output can be checked on its own, and you do not need retries. E.g. fetch → clean → structure → report.
When not to
When the task needs trial and error, or when a later stage could recover from an earlier mistake. A pipeline has no way back: if stage one is wrong, everything downstream inherits it.
How it fails
- Error compounds along the chain. 95% correct per stage is roughly 60% after ten stages.
- Schema drift mid-chain. Stage three quietly changes its output shape, stage five still parses the old one, and the failure surfaces far from its cause.
- No place to stop. You reach the end before discovering stage one was wrong, and the cost is already spent.
Design notes
Put a checkable contract (a schema or an assertion) between every pair of stages, and persist inspectable artefacts in between. A pipeline's reliability comes from the interfaces between stages, not from how clever any single stage is.
Origin
管線是既有的軟體工程形態，沒有單一首次公開實作可指向；不硬給日期。