Tableau Prep alternative: a flow is a file, a pipeline is code you can release

A Prep flow is a diagram you edit. A pipeline here is a JSON document of SQL nodes with explicit dependencies, which happens to render as a diagram. The distinction sounds academic until the second week, when the questions change from "can I build it?" to "can I review it, diff it, hand it to someone else, and say which version of it produced the number on the board?". The flow answers those in person and from memory. The pipeline answers them structurally: the diff is a pull request, the review is a draft a person can run before releasing, and the version is immutable once released.

Both join across sources without a warehouse. Prep joins in memory on the machine running the flow; a pipeline stages each source's rows into a scratch database created for the run and joins them there, with ordinary SQL any reviewer can read. When a source is slow, the difference gets loud: the pipeline's node carries a query timeout, and the SQL probe returns the plan, so the agent fixes the shape from evidence rather than from a spinner.

There is a quieter difference underneath, and it decides who can help you. The flow's knowledge lives in the person who built it and in the file only that tool opens; hand the work to a colleague, a contractor or an agent, and the first week is archaeology. The pipeline's knowledge lives in the document: the nodes are named for what they do, the SQL is in the file, the dependencies are declared — so the second person (or the second agent) starts from reading, not from reverse-engineering. Teams do not adopt pipelines because the first build is faster; often it is not. They adopt them because the second week, the second person and the second year are all cheaper.

Visual profiling — the distributions, the nulls, the value counts a flow shows while you drag — is a real strength, and this server does not ship an equivalent: the agent reads catalogue statistics and the probe's rows instead, which is evidence, not a histogram. The click-to-pivot is faster than writing the SQL by hand if you do not have an agent to write it. And Tableau-native output — the flow that lands directly where a Tableau author expects — is friction this does not attempt to match; the honest path here is the write-back node or the endpoint.

If the preparation is a one-off reshape by a person who lives in Prep, none of the six above pays for the move. The pipeline earns its keep when the preparation is repeated, reviewed, handed between people and machines, or served — which is the moment a file one tool opens stops being enough.

Both panels are the demo workspace's run of record — demo/top_company_by_borough, a recorded acceptance result over the published sample data (the reference zones and the lake's ride ledger on S3), reproduced from the bucket. The same sentence, the same steps and the same numbers the home page shows.