Sixteen minutes or twelve
A podcast episode can go out once three jobs have run on the master recording. Transcribing takes 12 minutes, normalising the loudness takes 3, and rendering a waveform image for the episode page takes 1. One after another, that is 16 minutes before anyone can publish. Started together with room to run, the last one finishes at minute 12.
The saving is the sum against the longest, and that tells you more than the four minutes do. Once the jobs run side by side, only the longest one sets the time. Make the waveform instant and the episode still takes 12 minutes; among these three, only a faster transcriber shortens the wait.
A flowchart draws this with a Split: one input and up to four outputs, out-1 to out-4. Every output you connect starts, which is the difference from a Decision, where the flow takes one exit.
With only one server
"Parallel" in a flowchart means that no order is implied between the branches. It does not mean a machine runs them at the same instant. If the studio's one server handles one job at a time, the three jobs queue up and take 16 minutes, and the diagram above is still correct. Twelve minutes needs room to run the other two during Transcribe, and a second server doing them one after the other is enough. The Split permits that; it does not provide it.
The branch that was reading another branch
Ask what each job reads, not only what it does. Transcribe reads the master. Normalise reads the master. The waveform is different: this studio wants it to show the audio listeners will hear, and normalising changes the levels, so it reads the normalised file. The diagram above lets Render waveform start before that file is finished, and then the image can show the old levels or come from a file still being written.
If one job needs another's result, those two aren't parallel, they're sequential. So Render waveform moves under Normalise audio, into one branch.
Do the sum again. The right branch takes 3 + 1, four minutes, against Transcribe's 12, so the episode still takes 12. The correction cost nothing here because the job that moved was not on the longest branch. Had the transcriber needed the normalised audio too, the total would be 15, and most of the saving would be gone.
Coming back together
Publish episode needs the transcript and the image, and a Process takes one incoming edge, so the branches have to come back into one node before it, placed after the last step of every branch. The diagrams above use a Summing Junction, which states the wait for every branch, as a Parallel Gateway does where the palette offers one (see Junctions: Or & Summing and Gateways: parallel, inclusive & event). A Merge goes in the same place, but it only converges paths and does not state the wait; Merge & path convergence covers the node.
A branch nothing later needs, like copying the master to an archive, can end at its own End; a job only some episodes need belongs behind an Inclusive Gateway, because every output of a Split starts; and Error handling & error boundaries takes up a branch that fails.
One question before you draw the Split
File Upload & Processing Flow asks you to validate an uploaded file, fan out into parallel steps, merge the results and generate an output document. Its palette has Split and Merge, with no gateway or Summing Junction, so the Merge is how your branches come back. Choose the steps, then test each pair: does this step read anything the other one writes, or write the same file? Any yes goes into one branch, in order, so that no branch is left waiting on another mid-flight. The output document comes after the Merge, and the Merge does not state the wait for every branch. So name what the document needs in its title or on the edge into it, and a reader sees the assumption.
ETL Pipeline Flow asks the same question at a harder level, with the load split across parallel targets and a Merge to bring them back.