Troubleshooting and Maintenance

There is no error message in this section.

Every troubleshooting guide you have ever read opens with something going red. A failed job. A broken build. A stack trace, a 500, a console full of warnings. Something that stands up and announces itself.

Classifications do not do that. A broken classification produces a report. A complete, well-formatted, perfectly plausible report, with totals that reconcile and columns that add up and a chart that renders beautifully. It just happens to be wrong.

Which changes what you are actually doing here. You are not debugging. Debugging is what you do when a machine tells you it is unhappy. You are auditing, which is what you do when a machine tells you nothing at all, and you have to go and find out whether it was telling the truth.

Where a classification can break

Four places, and every symptom in this section traces back to one of them.

The index

Come back to this table. That is what it is for.

Start by checking whether it is even a classification problem

Before touching a set, look at the parent dimension.

If the club's Membership Tier report is empty, open the User ID report. If that is also empty, then no User ID was collected, nothing was there to classify, and the classification behaved perfectly. You have an implementation problem wearing a classification costume, and no amount of uploading will fix it.

This one check would resolve a large share of the classification tickets ever raised. It takes eleven seconds. People skip it because the symptom appears in a classification report, and the eye follows the symptom rather than the cause.

The mismatch family, and why they are so hard to see

Now the group that produces the most wasted afternoons. The key is there. The file was uploaded. And the row is still Unspecified.

Which means Adobe looked up your key, and did not find it, because the key in the file is not the key in the data. They look identical on a screen. They are not identical to a machine.

Two of those come from the source system and cannot be fixed in Adobe at all. If your tracking codes arrive URL encoded, no rule, no file, and no amount of cleverness inside the classification interface will save you. That is a conversation with whoever builds the campaign links, and it is better to have it on a Tuesday than during a quarterly review.

Which is where Unspecified deserves a second look

Practitioners spend their careers trying to get rid of Unspecified. It looks like failure. It looks untidy. Stakeholders ask about it.

Stop. Look at what it is actually doing.

So the goal is not zero Unspecified. The goal is a number you can explain. Unspecified holding steady at 3% because that is the share of traffic that genuinely arrives before login is a healthy report. Unspecified at 0% because somebody classified everything with a catch-all rule is a report that has stopped telling you anything.

The values are wrong, which is worse than missing

Missing data raises a question. Wrong data answers one, incorrectly, and nobody thinks to check.

Two causes, and both were built in Classification Rule Builder.

Rule precedence is inverted from your instinct. With Rules overwrite set to apply to all values, the last matching rule wins, so moving a rule to the top of the table takes its priority away. With apply to only unset values, the first wins. Same rules, same order, opposite answer. When a column holds a value you cannot account for, check that dropdown before you check anything else.

Something is fighting your file. If curated values keep reverting, a rule is rewriting them every four hours, and it will keep doing so forever. Nobody is at fault. The rule is doing exactly what it was told.

Four instruments, and almost nobody uses them

Everything above is diagnosable, and the tools have been sitting in the interface the whole time.

The Jobs tab is the one to build a habit around. It is the only place in this entire feature where a failure is written down in plain language, and it is the first place to look when something is missing and you cannot say why. If an import silently failed at 2am, the Jobs tab knows. Nothing else does.

And Download with Data Returned set to Any columns empty is the single most useful diagnostic in the module. It turns "something feels wrong with the tier data" into a list of exactly which keys are missing exactly which columns. Vague suspicion becomes a spreadsheet you can act on.

The fourth instrument is the only one that lives outside the classification interface, and the only one that raises a problem before somebody reports it. Put Unspecified on a trend, as a percentage of whatever metric the business actually watches, and leave it somewhere it will be seen. The level is already covered above: a steady number you can explain is a healthy report. What a trend adds is the shape of the movement, and the shape names the cause.

The drift is the one worth building the habit for, because it is the one the level check cannot catch. A step is obvious to anybody who opens the report that week. A drift stays inside the band you would still call normal for months, which is comfortably long enough for the person who used to run the upload to move on and for nobody to learn that the job existed. It is the same instrument, and the same reasoning, as watching the Direct share in Marketing Channels: two cheap alarms that catch upstream failures in days rather than quarters, and both routinely missing from otherwise careful reporting setups.

Maintenance, which is really just deciding things once

Almost every failure in this section is a failure of ownership rather than a failure of technology. Two systems wrote to one column because nobody said which one should. A file stopped arriving because the person who uploaded it left. A bucket filled up because nobody agreed to empty it.

So the maintenance discipline is small, and it fits on one page.

That document is the bookend to the tracking code convention from Designing the Key. One decides what the key looks like. This one decides who is allowed to write to it. Between them they prevent most of what this section exists to diagnose.

What was actually built here

Step back from the interface for a second.

This module took a value the business could not read, M-10482 or eml_2026_07_winback_in, and gave it meaning it never carried. No developer shipped code. No server call changed. Not one byte of collected data was altered. And yet a report that could only say "one hundred members did things" can now say which of them were over sixty, on which nights, in which halls.

That is what report-time shaping buys you, and it is why classifications sit where they do. Processing rules and VISTA rules shaped the data before it landed, permanently, with no way back. Classifications shaped it afterward, reversibly, retroactively, forever. Same goal, opposite end of the pipeline, and now you know when to reach for each.

And with that, the shaping is done. That closes Classifications, and with it the whole of the Shape the data phase. Foundations built the understanding, collection captured the behavior, and shaping made it legible: processing rules, VISTA rules, marketing channels, and classifications, each adding a layer of meaning the raw hit never carried. Three of the five phases are behind you, and the data is as meaningful as it is ever going to get.

Which raises a question this module cannot answer

Everything from processing rules to classifications has been about one question: what does this value mean?

Not once has anything asked a different question, and it is the question every stakeholder actually has.

Which of these people are worth looking at?

Members over sixty who booked a hall but never a gym slot. Visitors who arrived on a Winback campaign, came back twice, and then vanished. People who did one thing and not another, in that order, within a week.

No amount of shaping answers that. Shaping tells you what a value means. It does not let you carve the audience. For that you need something that can hold a condition, nest it, sequence it, and apply it across every report you own, and that is a different instrument entirely.

It is called a segment, and it lives in the next phase of this curriculum: Analyze the data, where the question stops being what a value means and becomes which of the people behind those values are worth your attention. That phase begins with Calculated Metrics, the measures you define for yourself, and then reaches Segments, the instrument built to carve an audience into exactly the groups this section kept asking about.