A fix you cannot diff is a fix you should not approve. A fix you cannot reverse is a fix you should not let a directory feed apply on its own. Kiweely Roots, coming next, puts those two requirements in one line: every change shows what it does to the data, and goes back in one step. The line is short because the feature is specific. It is not a source-control history of the integration code, though you should have that too. It is a diff of the data the other system will see, on the rows that made you change anything, and a return to the previous behavior without a project.
Roots is not available, so this article also says how to get a weaker version of the same control from Workday today, and where that version stops.
What the diff is of
There are three diffs people mix up. Only one of them supports an approval on a payroll or benefits change.
| Diff | What it compares | Who it convinces | What it misses |
|---|---|---|---|
| DiffCode or assembly diff | What it comparesThe map, the Studio step, the recipe, before and after the edit. | Who it convincesThe person who wrote it. | What it missesWhether the carrier will accept the row. A one-line map change can add a tier or blank a required segment. |
| DiffFull-file diff | What it comparesLast night’s entire output against the new output. | Who it convincesSomeone with an hour and a file viewer. | What it missesThe rejected population, which is buried. It also tempts you to store two full files of worker data. |
| DiffOutcome diff on the failed rows | What it comparesThose rows as the other system judged them, against the same rows after the fix, on Sandbox. | Who it convincesThe approver, and the owner who gets the note. | What it missesNothing that the approval actually needs, if the rows are the ones that failed. |
The Roots page points at the third. Payroll, benefits, and tax fixes are proven on Sandbox with the rows that failed, then they wait. The product illustration is 14 benefits rows, rejected because a coverage tier is now required, mapped from benefit elections, rerun on Sandbox. The before is “14 rows rejected.” The after is “14 rows accepted.” That pair is the diff. It is an illustration, not a customer metric. The shape is what to demand: a count, the reason, the population, the Sandbox result. Not a transcript of the 834 in the approval card, and not a model’s summary of the file with no rerun behind it.
A code diff still belongs in the record for the person who will debug the next failure. It is not a substitute. Workday’s benefits testing chapter is why. File-format failures — missing segments, blank required fields, wrong codes, wrong coverage status — are visible in the output and in the vendor’s rejection, not in the fact that a map was edited. If you approve the map and do not rerun the rows, you approved an intention.
What “shows what it does to the data” must include
A useful before-and-after for an integration fix has six fields. Fewer than this and the approver is trusting a sentence.
| Field | Example that matches the Roots illustration | Why it is there |
|---|---|---|
| FieldIntegration | Example that matches the Roots illustrationBenefits enrollment to the carrier, nightly | Why it is thereSo the rule (wait versus resume) is obvious |
| FieldPopulation | Example that matches the Roots illustrationThe 14 rows the carrier rejected, not the full file | Why it is thereSo the proof matches the failure |
| FieldBefore | Example that matches the Roots illustrationRejected at 01:04, missing coverage tier | Why it is thereThe external outcome, not only Workday’s event |
| FieldChange | Example that matches the Roots illustrationTier taken from benefit elections | Why it is thereThe rule, in one line |
| FieldAfter | Example that matches the Roots illustrationThe same 14 rows accepted on Sandbox | Why it is thereA rerun, not a prediction |
| FieldWhere it has not gone | Example that matches the Roots illustrationProduction send is waiting | Why it is thereThe diff is not a claim that Production already changed |
Directory feeds need the same six when they resume without a prior approval. The note to the owner is the diff, sent after the fact. “We fixed the Okta map” is not a note. “Twelve joiners were missing a department attribute; Sandbox now sends the supervisory organization; Production was updated; here is how to reverse” is a note.
What the diff should not include is the payload as a habit. The Roots trust line is that the platform keeps times, counts, and outcomes, not the records. A before-and-after of “rejected versus accepted,” with identifiers your operators are already allowed to see, fits that line. A before-and-after that attaches both 834 files does not. If an auditor needs the file, Workday’s benefits template has document retention, default 30 days, inside Workday. That is the store for the file. The approval record can point at the event and the counts.
What one step means
“Goes back in one step” means one action returns the integration to the previous behavior, and that action is available to the person who was asked to approve or who received the note. It does not mean:
- a redeploy from a laptop by the person who wrote the fix,
- a revert of a git commit that still needs a promotion through three tenants,
- a comment that says the old map is in a ticket,
- a second fix that undoes the first and itself needs a change review.
The test is operational. The owner of a directory feed reads the note, disagrees with the after, and can put the previous behavior back without assembling a project. The approver of a payroll fix, if they approved and then learned the Sandbox population was wrong, has the same button. If the only person who can reverse the change is the agent that made it, you do not have a rollback. You have a request.
Workday today gives you pieces of this and not the button.
Studio and connector integrations can be cloned before you change them. The benefits course says to clone, then change the template, so the previous configuration still exists as its own integration system. That is a real rollback path. It is not one step. You have two integration systems, schedules to swap, and a risk of both firing. It is the right habit until you have a product that treats the previous version as a switch.
Orchestrate’s lifecycle replaces the Implementation or Sandbox release when you promote again. Read that as a warning: promoting a new version replaces the release you had in that lane. Your previous version has to be something you can promote again on purpose. If you cannot name the previous app version and put it back on Sandbox, you cannot roll back, regardless of how small the canvas edit was.
Source control of an assembly is necessary and, by itself, it is a code diff. The data diff still has to be rerun after you revert, on the same rows, or you do not know the revert worked. One step that restores code and does not retest is a hopeful step. The Roots line pairs the two: the change shows the data effect, and the return is one step. A return that skips Sandbox would violate the other line on the same page, that every fix runs on Sandbox first. The coherent reading is that rollback restores the previous tested behavior, not that it edits Production in place with no test. Early access should be checked against that reading.
How this sits on the two repair rules
The diff is what makes the wait rule an approval instead of a delay. A person looking at “14 rejected, 14 accepted on Sandbox, tier from elections, Production not sent” can say yes or no. A person looking at a ticket that says “please approve the benefits fix” cannot. If the card does not contain the after, do not approve it. Ask for the Sandbox rerun. Payroll waits and directory feeds resume only works if both paths have this card. The directory path sends it as a note. The payroll path sends it as a question.
The rollback is what makes the resume rule safe enough to be a default. A directory feed that goes back on at 02:10 and can be put back by 02:20 is a different risk from a directory feed that goes back on and stays wrong until the next business day because the author is asleep. If the one-step action does not exist yet, do not operate the resume rule automatically. Operate it by hand, with a cloned previous version you have already tested. That is slower, and it is the actual control.
Move an integration to the wait rule when the diff is hard to read. A full-population directory replace, a compensation report, a tax file: if the before-and-after cannot be stated as a small population with an external outcome, a person should see it first. The page allows that move at any time. Difficulty of the diff is a good reason.
A manual diff you can do this week
Until Roots exists, this is the procedure that matches the page as closely as Workday’s tools allow.
- Save the vendor’s rejection identifiers and the reason. Do not forward the whole file to a general channel.
- Clone the integration system before you edit it, when the tool offers clone. You now have a previous version, even if switching back is clumsy.
- Reproduce the rejected workers. On a benefits connector, the debug service can include only those workers. Run that on Sandbox, not as the first change in Production.
- Write the six fields down in the ticket. Before, change, after, population, and the statement that Production has not been sent.
- For payroll, benefits, and tax, stop there until a named person accepts the six fields. Workday’s outbound business process can require an approval before delivery if you want the file itself held. The approval of the map is still the six fields, not only the business process step.
- For a directory feed, a named on-call role may continue after the six fields exist, and must send them to the owner.
- If the after is wrong, switch the schedule back to the clone. Rerun the same workers. That is your one step until a product makes it one button. Record how long it took. If it took a day, you do not have a one-step rollback, and the resume rule is not justified for that integration.
What to ask in early access
| Question | Answer that matches the page |
|---|---|
| QuestionIs the diff the failed rows, or the whole file? | Answer that matches the pageThe rows that failed, for payroll, benefits, and tax. |
| QuestionIs the after a Sandbox rerun? | Answer that matches the pageYes. Real data. |
| QuestionCan Production be sent before the diff exists? | Answer that matches the pageNo. The wait rule is the diff plus a person. |
| QuestionWho can roll back? | Answer that matches the pageThe person who approves, or the owner who was notified. One step. |
| QuestionDoes rollback itself get a Sandbox check? | Answer that matches the pageThe page says every fix runs on Sandbox first. Treat a rollback as a change and ask. |
| QuestionAre both files stored on the platform? | Answer that matches the pageThey should not be. Outcomes, not records. |
An early access build that shows a code diff and calls it the data diff has not met the line. An early access build that can reverse only by asking the agent in chat has not met “one step” for an owner who is not in that chat. The assistant can request a change. The rollback of a change that already applied is a control in the integration, not a prompt.
Frequently asked questions
Is a git revert one step?
Only if the owner can do it, it restores the previous behavior, and you can show the data diff again on the same rows. A revert that requires a developer, a pipeline, and a promotion is a runbook. It may be fast. It is not what the page is specifying for the person who got the note.
Should the diff show pay amounts?
No, unless the approver’s job is to check those amounts and the channel is restricted to them. The outcome diff for a file-format failure is accepted versus rejected, plus the element that changed, not a dump of rates. Compensation fixes that really are about the amount need an approver who is allowed to see the amount, in a system that is allowed to hold it. That is a narrower channel than the integration platform’s run record.
What if the vendor does not send a row-level rejection?
Then you do not have a population, and you cannot claim the Roots-shaped diff. You have a file-level failure. The before is “file rejected,” the after is “file accepted on the vendor’s test environment,” and the population is the file you actually sent to that test. Do not invent row counts. Workday’s connectivity test — deliver a file, or send a test file and require a success message — is the before you can document. The format test is still separate.
Register interest in Kiweely Roots if you want the six fields and the one-step return on the platform rather than in a ticket. Bring one real rejection, with the population you can name, and ask to see the before and the after on Sandbox before you talk about Production. That is the whole feature.