What the video carries
This is the header of an upload just after the transcoder has finished with it, in the media pipeline of a site where people post photos, videos and documents:
upload: u-61907type: video/mp4slip: transcode, thumbnail, moderate, publishdone: transcode
The file itself sits in object storage; the message carries its id and that list. The transcoder has just added itself to done and sent the message to the thumbnail step's channel, because thumbnail is the first step on the slip that is not done yet. When the thumbnail step finishes it will do the same, and the message will go on to moderation and then to publishing. No component between those steps decides where the message goes. The message already says.
Three file types, three routes
Uploads come in three kinds, and each needs a different path. An image is resized into the display sizes and checked by moderation. A video is transcoded into streaming formats, gets a thumbnail cut from it, and is checked by moderation. A PDF is scanned for viruses and has its text extracted for search. Then each is published.
The first version wired all seven steps into one fixed line, and each began by checking the file type and passing on anything it did not handle. Every upload crossed seven queues to be worked on at three or four, and a new file type meant changing all seven. A router after each step would work too, but it would hold every route for every type and change with every new step.
A Routing Slip decides the route once. When an upload enters the pipeline, the intake step looks at the file type, attaches the ordered list of steps for that type to the message, and sends the message to the first one. From then on there is no central router: each step reads the slip, does its own work, marks itself done, and forwards the message to the next step on the list; the last step ends the route. A step knows its own job and how to read a slip; it knows nothing about which file types take which path.
Intake attaches a different slip to each file type. After that, every step forwards the message by reading the slip it carries; no node in the middle decides.
The steps are still the filters of Pipes and Filters. What changes is the pipe: instead of channels wired once for every message, each message carries the path it will take. And where a Content-Based Router chooses one destination at one point in the flow, the slip settles the whole route at the door.
Adding audio
The payoff shows when the product adds audio uploads, which need moderation and a waveform image for the player. Moderation and publishing exist already; only the waveform step is new. The change is one new step, deployed on its own, plus one new entry in intake's table of routes. No running step is redeployed. Routes change in one place, and the steps are reused in whatever order a route needs.
Two rules hold this together. Every step forwards by the slip and by nothing else: a step that hardcodes "after me comes moderation" quietly turns the pipeline back into a fixed line. And every step keeps the slip: a step that builds a fresh outgoing message, as a transcoder writing a new rendition record might, has to copy the slip and the upload id onto it, or the next step receives a message with no route.
Stuck halfway down the list
One evening the moderation model's server is down for forty minutes, and every image and video that reaches moderation fails. Where do they go? No orchestrator is watching, so the answer has to be built into each step. Moderation retries a few times, then moves the message to a Dead Letter Channel, with its slip intact. That last part makes recovery cheap: the slip still says done: transcode, thumbnail, so when the message is replayed to moderation's channel the route picks up where it stopped, without transcoding the video again, though moderation has to cope with receiving a message it may already have seen.
What nobody holds is the overall picture. A user writes in: the video they uploaded an hour ago is not on their profile. The route was never stored anywhere except on the message, so no single place can say where u-61907 is. Someone searches the moderation queue, the dead letters and each step's logs. Teams that run slips for long often add that place back, with each step writing "u-61907 finished thumbnail" to a status table. That helps, and it is also a sign: a record of every route's progress is a large part of what a central coordinator keeps.
A route can also stop after work has been done. Moderation rejects the video, but the transcoder has already written its renditions to storage and the thumbnail is in the image cache. Moving the message aside does not remove them. Some implementations have each step note on the slip what it did, so the finished steps can be undone in reverse order; how to write an undo that is safe to run is in Compensating Action.
When the route depends on the answers
In the fixed-list form used here, the slip is a plan written before any step has run, so it cannot express a route that depends on what a step finds. Send the video to a human reviewer when the moderation score is uncertain, and wait up to a day for the decision. Add a 4K rendition only if the transcoder finds the source is 4K. Run thumbnail and moderation side by side and continue when both are done. A flat list cannot say any of these: the first two choose the next step from a result, and the third runs two steps at once and joins them.
Some implementations let a step rewrite the rest of the slip, and that is how the trouble usually starts. Branch conditions creep into the step list, several steps start editing it, and the pipeline becomes a workflow engine that nobody designed and that is hard to follow: looking at the slip at intake no longer tells you where the message will go. Keep the slip a flat sequence. When the next step depends on an intermediate result or on a deadline running out, use a Process Manager, one stateful coordinator per upload that receives each step's result and decides what comes next. For long waits that must survive restarts, a durable workflow engine is one way to get that persistence: see Durable workflows: orchestration that survives restarts.
Practising the choice
No Challenge exercises a routing slip yet; the closest, the Process Manager Challenge, practises the other side of this choice, a coordinator that holds each process's state and picks the next command, and its palette has no Routing Slip node.
As an extra, draw the upload pipeline on a blank Grid: a Routing Slip node as intake, a video route and a PDF route of handlers, and edges labelled with what the slip says. Then add the human reviewer who may take a day. If the drawing needs an arrow from a step back to something that remembers the upload and decides what comes next, the route has outgrown the slip, and that something is the Process Manager from the Challenge.