Skip to content

Lab 3.4 - Three Pull Requests in one release: review the merge order

Level: 3 Release Manager

Time: ~15 min (observed: median 16 min, 15 learners, fastest 6 min)

You will: go back over the order two Pull Requests on the same file went in, check that neither lost anything, and send back a third one that does not work.

The situation

Friday afternoon. Three stories were planned for this week's release:

Pull Request Author Checks What it touches Where it is now
US-018 Cap the crew size Mariia Pyvovarchuk green the assign flow, Helios_Delivery_Manager merged, in Level 2
US-019 Quote PDF Romain Panda green Helios_Delivery_Manager merged, in the Level 2 capstone
US-020 Refactor InstallationScheduler Mariia Pyvovarchuk failing InstallationScheduler not opened yet

Two of them edited the same permission set, and they are already in integration: you merged them as a contributor in Level 2, without a release manager deciding anything. Reset this level starts Level 3 from a branch that holds them too. So nothing collides in front of you in this lab. What you do is the release manager's half of it, after the fact: was the order right, did the second merge keep what the first one brought, and what happens to the third, which does not work.

On a real project the first two questions come before the merge. Asking them about two merges you can still read in full is the cheapest way to learn to ask them.

Before you start

  • Lab 3.3 finished
  • Nothing uncommitted

Steps

1. Receive the one Pull Request still to come

Goal the Pull Request of US-020 open into integration, its checks failing, and the two merged Pull Requests of US-018 and US-019 open beside it.

Do

  1. Open Training: Level 3 > Simulate my teammates, then US-020 Refactor InstallationScheduler. It is on the Welcome page, where each level menu is a page of cards, and in the SFDX HARDIS view of the left bar, where each level is a folder you unfold. Either route runs the same thing

    The Level 3 training menu on the Welcome page

  2. Find the merged Pull Requests of US-018 and US-019, in the Pull Requests tab of the integration window of the DevOps Pipeline panel, or in the Closed list of the Pull Requests of your fork on GitHub. Keep both open, the next steps read them

  3. Wait for the checks of US-020. They will fail

Careful

Do not simulate US-018 or US-019. Each teammate scenario is used once, and both are already in integration: the command would say "Nothing to commit" and open nothing.

You should now see

  • Mariia's Pull Request into integration, opened for you: the command prints its address, and ends with an Open the Pull Request button at the bottom of the panel
  • The checks of US-020 failing

2. Triage before you touch anything

Goal the order a release manager would have picked, set against the order the merges really went in.

Sort the three, in this order of questions.

Do

  1. Which ones are green? A failing Pull Request is not a decision, it is a task for its author
  2. Which ones touch the same file? US-018 and US-019 both edit Helios_Delivery_Manager
  3. Which one is smaller? All else equal, merge the smaller one first
  4. Compare with what happened: the Pull Requests tab of the integration window lists them in the order they went in

Why

Do not spend your Friday fixing Mariia's refactor. Whichever of US-018 and US-019 merged second inherited whatever the first one did to that file, and step 5 is about what "inherits" turned out to mean. The smaller one goes first because its author has less to redo if the other one lands badly.

The order a release manager would have picked: US-019 first (small, green, no dependants), then US-018 (green, bigger), and US-020 goes back to Mariia. In Level 2, US-018 went first, because Lab 2.7 needed it there. Nothing broke, and step 5 shows why. Whether the order mattered is a question you can only answer by reading what the second merge did.

You should now see

  • The order you would have picked: US-019, then US-018, and US-020 back to Mariia
  • The order it really happened in, in the Pull Requests tab: US-018 first

3. Send US-020 back, properly

Goal one review comment on US-020 that names the failure, says who owns it and what happens to the release, and the Pull Request left open.

Do

  1. Open US-020 and read the failure, under Deployment errors in the sfdx-hardis comment. Read it properly before you write anything, because it is not what the words "failing tests" suggest
  2. Leave one review comment that does three things:

    The check fails to compile InstallationSchedulerTest: earliestInstallDate returns a Datetime now and the test still assigns it to a Date, in two places. No tests ran. Not blocking anything else, so US-019 and US-018 go out in this week's release and this one can land on Monday.

  3. Leave the Pull Request open and move on

Do not fix a contributor's Pull Request yourself

It is faster once and expensive every time after. The author does not learn the failure, and you become the person every failing check is escalated to.

Why

It is a genuine failure, in her code, and it is hers to fix. earliestInstallDate now returns a Datetime, and InstallationSchedulerTest still assigns it to a Date, in two places. The class does not compile, so no test runs at all. What the sfdx-hardis comment gives you, under Deployment errors, is a compilation error with a class name, twice, not a failed assertion.

The comment names the failure, says who owns it, says what happens to the release. On a real project you would send it as Request changes. GitHub does not offer that on a Pull Request opened from your own account, and in this fork the teammate Pull Requests are opened from yours, so a plain comment does the job here.

You should now see

  • US-020 still open, with your comment naming the compilation error of InstallationSchedulerTest

4. Read how US-019 went in

Goal you know what US-019 changed in Helios_Delivery_Manager, and how little there was to it.

Refresher Lab 3.2

Do

  1. Look at the diff of US-019 on Helios_Delivery_Manager
  2. Read it the way Lab 3.2 taught you to review one, and note how little there was to it

Why

Green, small, nothing in its way. Had it still been open, you would have merged it on that reading.

Nothing here was hard, which is exactly why it is worth knowing what happened next.

You should now see

  • One new <fieldPermissions> block for Panel_Batch__c.Quote_Pdf_Url__c, added near the bottom of the file where the Panel_Batch__c grants live

5. Find the conflict that never happened

Goal you can say why git merged two edits of the same file without a word, and what to do when it does stop.

Refresher Lab 2.7

US-018 granted Installation__c.Crew_Capacity_Cap__c, near the top of the same file. US-019 granted Panel_Batch__c.Quote_Pdf_Url__c, near the bottom. Two people, same file, same week.

Do

  1. Open the file on integration

Why

Git merged them without a word. No conflict, no resolution, no second look.

That is not luck and it is not git being clever. A conflict needs the two edits to land in the same place, within the few lines of context git compares. These two are about seventy lines apart in an alphabetically sorted file, so git took both and moved on. Lab 2.7 gave you the other case: your Crew_Notes__c grant sat one line from Mariia's Crew_Capacity_Cap__c, git could not choose, and it stopped and asked.

The release manager's takeaway is not "permission sets rarely conflict". It is this: a silent merge is the common case, and a conflict is the rare one. Git warns you about the rare one. Nothing warns you about the common one, so if you want to know that both grants survived, you have to go and look. That is step 6, and on a real project it is a habit rather than a step.

When git does stop and ask, permission set conflicts are almost always take both: one field each, and the permission set holds as many as it needs. Resolve it keeping both <fieldPermissions> blocks in alphabetical order, then wait for the checks to run again. A conflict resolved in the web editor is a new commit, and it has never been validated. Merging without re-validating is how a resolution that dropped a closing tag reaches an org.

You should now see

  • Both grants in the file on integration, in order, as if one person had written them

6. Look at the permission set in the org

Goal Helios_Delivery_Manager in helios-integration carrying both new field permissions.

Do

  1. In helios-integration, check Helios_Delivery_Manager has both new field permissions: Crew_Capacity_Cap__c on Installation and Quote_Pdf_Url__c on Panel Batch

You should now see

  • Crew_Capacity_Cap__c on Installation and Quote_Pdf_Url__c on Panel Batch, both granted by Helios_Delivery_Manager

7. Understand what kept this survivable, and what is not there

Goal you can point at the three mechanisms, and say which one is doing anything here.

Refresher Lab 3.3

Three mechanisms are worth being able to point at.

Do

  1. Open Pipeline Settings from the DevOps Pipeline panel, and look at each one, because only the first is actually doing anything here

    Global Pipeline Settings, with the tabs that hold the cleaning, overwrite and delta settings

  2. Cleaning: open the Salesforce Project tab (1) of Global Settings, and find Auto-Cleaning types (autoCleanTypes)

  3. Delta deployment: open the global Deployment tab (2), and read Use Delta Deployment (3)

Why

Cleaning (autoCleanTypes) is on, and it is why these two grants were two small blocks in a permission set rather than two edits in a thousand-line Profile. A Profile conflict is a bad afternoon. This is most of the reason the project bans permissions on Profiles.

The overwrite manager protects components that are deliberately different per org. Anything listed in manifest/package-no-overwrite.xml is removed from the package when the target org already has it, so a deployment cannot flatten a named credential that points at a different endpoint in each environment. The project starts with the list every sfdx-hardis project gets. It covers reports, dashboards, connected apps and credentials, Helios deploys none of those, so nothing it deploys is protected yet: Lab 3.5 adds the first entry that matters, before the first promotion to uat. The list is a file: you open it from the Deployment packages menu of the DevOps Pipeline panel, not from a tab of this one.

Delta deployment is Use Delta Deployment, and Lab 3.3 showed you it is Disabled here. With it on, each merge deploys the components that changed rather than the declared package. It is a speed and blast-radius decision, not a safety net: it does not stop one merge overwriting another, because both deployments send what their own commit contains.

Under the hood: the three mechanisms and where each one lives
Mechanism Configuration What it does
Cleaning autoCleanTypes in config/.sfdx-hardis.yml Rewrites the sources at commit time, in the contributor branch
Overwrite manager manifest/package-no-overwrite.xml Removes components from the package at deploy time, when the org already has them
Delta useDeltaDeployment Reduces the package to what changed since the last deployed commit

The distinction between the first two is worth keeping straight, because they fail differently.

Cleaning is a source decision: what the repository is allowed to contain. When it drops something, the component is genuinely not in the repository any more, and the diff shows it.

Overwrite is a deployment decision: what this particular org is allowed to receive. The component stays in the repository, and it is simply not sent to an org that already has its own version. A new org, which has nothing, receives it.

That is why a named credential belongs in the overwrite list and not in the cleaning rules: a fresh org must get one, and an existing org must keep its own.

You should now see

  • Cleaning on, on the Salesforce Project tab
  • Use Delta Deployment reading Disabled, on the Deployment tab
  • No tab for the overwrite list, which is a file

What you should see

  • US-019 and US-018 both in the integration history
  • US-020 open, with a review that names the compilation error
  • Helios_Delivery_Manager in helios-integration carrying both grants
If it goes wrong

Simulate my teammates says "Nothing to commit". That scenario is already in your integration. Expected for US-018 and US-019, which Level 3 starts with. For US-020, it means you received it already: its Pull Request is open in your fork.

One of the two grants is missing from the permission set. Something took one side of the file wholesale, most likely a manual merge during Level 2. Put it back in a follow-up Pull Request, not by editing the org.

Two Pull Requests are waiting and both are green. Merge them one at a time, waiting for each deployment to finish. Merging two at once into the same branch is how a release manager loses an evening.

Check your work

Welcome page > Training: Level 3 > Check my work, then pick Lab 3.4.

Go deeper

Next: Lab 3.5 - Promote to UAT and write the release notes

Questions and comments