Every change, and where it goes
All 84 cards in the deck. The lane is decided by one question: if this goes wrong, how hard is it to put back? Size, confidence and how many files changed do not enter into it.
L1 · Ship it
Reversible and contained
- Fix typo in the FAQ draftdocs/faq-draft.md, nothing links to it yetA draft in a file. Nobody sees it until someone wires it up.
- Rewrite the whole onboarding page on a preview40 files changed, deployed to a throwaway preview URLLooks huge, but it is isolated. Delete the preview and nothing live ever moved.
- Ship new checkout behind a flag set to offFlag defaults to false in every environmentThe code is deployed but dark. Turning it on is a separate decision.
- Add unit tests for invoice roundingTests only, no source files touchedTests do not run for customers. Reverting is a delete.
- Tweak spacing on a Storybook-only Button variantIsolated component, no consumers yetNo page imports it, so the only thing that can break is the story.
- Spin up a preview URL for Thursday’s demoEphemeral env seeded with fake dataNew and disposable. Tear it down and it is as if it never existed.
- Write a local script to dedupe a CSV exportRuns on your laptop against a copy of the fileIt works on a copy. Worst case you download the export again.
- Update README setup stepsRepo docs only, no deploy attachedWords in a repo. Git remembers the old words.
- Rename a private variable inside one moduleNot exported, 3 call sites, covered by testsContained to one file and the compiler checks every use.
- Draft App Store release notes in a text filerelease-notes-1.4.txt, not pasted into the consoleA draft. Pasting it into the store is the step that actually matters.
- Add a loading skeleton to an unreleased screenScreen is not reachable from any navigation yetNobody can navigate there, so nobody can hit a bug there.
- Try the new schema on your local Docker databaseThrowaway Postgres container, seed data onlyLocal and disposable. docker rm and it is gone.
- Write the post-mortem draft in a private docVisible only to you, nobody invited yetA private draft. Nobody is notified until you share it.
- Draft the new welcome email as a sandbox filetemplates/welcome-v2.html, no sender references itNot wired to any send job, so no real inbox can receive it.
- Add a new feature flag, default off, no readersOne line in flags.yaml, nothing reads itA flag that is off and unread changes nothing for anyone.
- Run the refund flow against the payment sandboxTest keys, fake cards, no real rand movesSandbox money is pretend money. Nobody is charged, nobody is refunded.
- Edit the system prompt in the offline eval setEval harness only, no traffic touches itRuns on a laptop against saved transcripts. Nobody talking to the bot sees it.
- Add a debug menu toggle to the TestFlight buildInternal testers only, hidden in releaseInternal build, hidden in release. Remove the toggle and it is gone.
- Reseed the staging database from scratchStaging only, fake customersFake data in a box that exists to be wiped. Production never notices.
- Run terraform plan against productionRead-only, no applyA plan only reads state. It looks scary and changes nothing.
- Replay last week’s bank webhooks in the sandboxTest endpoint, test accounts, fake randTest endpoint, test money. Replay it a hundred times if you like.
- Add a dry-run mode to the support agent’s toolsDefaults on, logs what it would have doneIt writes to a log instead of acting. The safe default is the point.
- Rename your own unpushed WIP branchfeature/lol-fix to feature/refund-retryOnly exists on your machine. Rename it back whenever.
- Lower the CPU alert threshold on stagingPages the staging channel, not on-callWorst case is a noisy Slack channel. Change it back in a minute.
- Reword the password reset email in a PRTemplate not merged, preview onlyA PR is a proposal. Real users get nothing until it merges and ships.
- Deploy a canary pod in the Cape Town regionOne pod, zero weight on the load balancerIt takes no traffic. Delete the pod and the region looks the same.
- Revoke your own expired test API tokenPersonal token, nothing uses itOnly you held it. Mint a new one in ten seconds if you were wrong.
- Delete a merged feature branch on the remoteAlready merged, commit is in mainThe commits live on in main. You can restore the branch from the hash.
L2 · Review
Reversible, but wide
- Change date formatting in the shared utilsImported by 14 services and both mobile appsEasy to revert, but every consumer picks it up on its next build.
- Add a nullable column to the orders tableAdditive migration on prod Postgres, no backfillAdditive, so it can be undone. Still touches the table everything reads.
- Swap the hero image on the live homepageProduction site, about 40k visits a dayRevert is one deploy, but every visitor sees it the moment it lands.
- Move the nightly report cron from 02:00 to 01:00Scheduled job, now overlaps the backup windowReversible, but it runs unattended and may collide with other jobs.
- Raise the HTTP timeout in shared service config30s to 60s, read by 22 servicesOne number, 22 services. Rolling back is easy, the ripple in between is not.
- Bump the HTTP client library one minor versionpackage.json and lockfile, used in every serviceRevertable, but a quiet behaviour change lands everywhere at once.
- Extend CDN cache TTL on /static to 7 daysEdge rule covering the whole domainRevertable, but stale files can linger at the edge for a week.
- Tighten API rate limit to 60 requests a minuteGateway rule that applies to every clientReversible, but real integrations may start getting 429s straight away.
- Add a database index on users.emailOnline index build on the prod primaryAdditive and droppable. It still loads the busiest table in the system.
- Set log level to debug in shared service configEvery service in the Joburg regionReversible, but log volume and the bill spike for everything at once.
- Change retry policy on the shared job queueMax retries from 3 to 10, every workerReversible, but a bad job now hammers downstream systems ten times.
- Hotfix a one-line CSS bug on the live login pageProduction, about 6k logins a dayTiny diff, but it is live on the front door. A redeploy undoes it.
- Edit the shared CI pipeline templateInherited by 30 repos, hits every buildOne bad step and 30 teams cannot ship. Revert fixes it, after the noise.
- Turn on the new search for 10% of usersFlag change in prod, switch it back any timeReversible, but 10% of real users feel it the moment you flip it.
- Add a new field to the public JSON API responseAdditive, existing clients should ignore itAdditive and reversible, but clients you do not control may still choke.
- Raise autoscaling max from 10 to 40 instancesProd cluster, Joburg regionReversible, but it can burn a lot of rand overnight before anyone looks.
- Bump the global cache key versionEvery cached page and API response goes coldReversible, but every request hits the database at once. Brace for the stampede.
- Swap the LLM model for all support chatsSame prompt, new model, 100% of trafficRevertable in one deploy, but every customer chat changes tone and answers.
- Invalidate every active sessionEveryone has to log in againNothing is lost, but every user is bounced mid-task. Support will feel it.
- Route Joburg traffic through the Cape Town edgeConfig change on the CDN, takes effect in minutesOne config line, but all local users take the new path. Flip it back if latency spikes.
- Raise the minimum supported app version remotelyOlder installs get a blocking update screenReversible server side, but a wrong value locks real users out of the app.
- Set the default timezone to UTC across servicesShared config, every scheduled job reads itReversible, but every job and report timestamp shifts by two hours.
- Loosen refusal rules in the shared agent promptEvery agent in the company inherits itOne edit, many agents. Revert is easy, the odd answers in between are not.
- Move all nightly jobs one hour earlierBackups, reports and syncs all shiftReversible, but jobs now overlap with the backup window. Watch for locks.
- Roll out the new cookie banner to every pageScript tag in the shared layoutReversible with a deploy, but it sits on top of every page and every signup.
- Turn the new dashboard flag on for everyoneWas at 5%, now 100%Flag flips back instantly. Every user sees it meanwhile, bugs included.
- Switch to the new load-shedding schedule feedPulls from a different upstream feedSwitch back any time, but until you do every user sees the new times.
- Bulk-retag 80 000 CRM contactsScript run from the admin console, tags onlyTags can be restored from the export. It is only safe if you took the export.
L3 · Human signs
Hard to undo
- Fix typo in a DNS TXT recordSPF record for the main company mail domainDNS is cached for hours. A bad SPF can send real mail straight to spam.
- Change one word in the welcome email templateSend job fires tonight at 18:00 to 12 000 usersOne word, but once the job runs 12 000 inboxes have it. No recall.
- Submit v1.4 to App Store reviewBuild 212, release set to automaticOnce it is on users’ phones you cannot pull it back from their devices.
- Publish the launch post on the company LinkedInPosts as the company page, scheduled for 08:00Public in the company’s name. Deleting it later does not unsee it.
- Run the refund script for failed debit ordersR412 000 across 1 380 customersMoney that has left the account does not come back by reverting a commit.
- Delete the old uploads bucketNobody has touched it since March, probablyProbably is not a backup. Deleted objects are gone.
- Rotate the API key the payment partner usesPartner has it hard-coded in their integrationTheir side stays down until they update it, and you cannot do that for them.
- Send the load-shedding reminder SMS8 500 customers at R0.28 a messageReal phones, real people, real rand. An SMS cannot be unsent.
- Drop the legacy_customers tableReplaced last quarter, probably no readersA drop is forever unless a restore works. Prove the restore first.
- Delete the abandoned mobile app repoOrg-level delete, two years of history and issuesDeleting takes the history, issues and pull requests with it.
- Send DMs to the 40 customers who complainedPersonal apology from the founder’s accountMessages sent in a person’s name to real people. No take-backs.
- Approve the supplier payment runR1.8m to 23 suppliers, bank file uploadedOnce the bank processes it, getting money back is a phone call and a prayer.
- Point the apex domain at the new hostA record change, TTL currently 86400A day-long TTL means half the world keeps the old answer. No fast undo.
- Email overdue invoice reminders to clientsFrom finance@, 214 clients, goes out tonightReal clients, real inboxes. Including the three who already paid.
- Publish the open-source SDK 1.0.0 publiclyPublic package under the company nameRegistries rarely let you take a version back. Strangers pin it within hours.
- Send a push notification to every app user‘We miss you’ promo, about 60 000 devicesIt lands on real lock screens. There is no recall button.
- Release the stokvel payout batchR340 000 across 120 member accountsMoney has left by morning. Banks do not hand it back on request.
- Publish the thread from the CEO’s accountScheduled, written by the internIt goes out in a real person’s name. Deleting does not unsend screenshots.
- Set a one-day expiry rule on the backups bucketLifecycle rule, takes effect tonightA tiny config line that deletes every backup. Versioning was off.
- Let the support agent email customers aloneAutosend on, 2 000 open ticketsThe agent writes, a real inbox receives. A wrong answer cannot be unsent.
- Update the MX records for the main domainMoving mail to the new providerSmall edit, but mail bounces or vanishes until DNS settles. Some of it never returns.
- Regenerate the courier partner’s signing secretThey hardcode it in their systemYour side changes in seconds. Theirs needs a human, and orders fail meanwhile.
- Move the Play Store rollout from 1% to 100%Version with the login bug fixA public release. You can halt a rollout, but installs already happened.
- Truncate the events table to free up spaceAnalytics, 400 million rows, no snapshotDisk is freed instantly and so is your history. TRUNCATE has no undo.
- Submit the Friday payroll runR1.2 million, 85 staff, early cut-offSalaries land in real accounts. Fixing a wrong amount means asking people for money back.
- Switch OTP messages to the new SMS vendorProduction, all customersSends to real phones in real time. Failures mean locked-out users, not a retry.
- Import the cleaned unsubscribe listSuppresses 30 000 addressesThe tool merges, it does not version. A bad list silences real customers.
- Destroy the old signing key version in the vaultPartners still verify against itDestroyed means gone. Anything signed with it can never be checked again.