The next article initially appeared on Medium and is being republished right here with the writer’s permission.
There’s a pretend confidence you possibly can carry round while you’re studying a brand new know-how. You watch just a few movies, skim some docs, get a toy instance working, and inform your self, “Yeah, I’ve acquired this.” I’ve performed that. It by no means lasts. A tough lesson typically accompanies the one expertise that issues.
You study via failure—falling flat in your face, trying on the mess, and determining why it broke. Something that feels too simple? It in all probability was, and also you didn’t exit the method with something price studying.
Ask About Failure: Failure === Expertise
Once I’m hiring somebody who claims relational database experience, I ask a “trick” query:
Inform me concerning the worst database schema you ever created. What did it train you to keep away from?
It’s not likely a trick. Anybody who’s been knee‑deep in relational databases is aware of there’s no good schema. There are competing use instances that always pull towards one another. You design for transaction workloads, however inevitably, somebody tries to make use of it for reporting, then everybody wonders why queries crawl. One other developer on the group inadvertently optimizes the schema (normally years later) for the reporting use case solely to make the transactional workload unworkable.
The proper reply normally appears like:
We constructed for transactional throughput—one of many founders of the corporate thought MySQL was a database, which was our first mistake. The enterprise then used it for reporting functions. The system modified fingers a number of occasions over the course of a number of years. Joins grew to become gnarly, indices didn’t match the entry patterns, and nightly jobs began interfering with consumer visitors. We needed to cut up learn replicas, ultimately introduce a warehouse, and after 5–6 years, we ended up simplifying the transactions and shifting them over to Cassandra.
That’s an individual who has lived the trade-offs. They’ve skilled a drawn-out existential failure associated to working a database. Whereas they won’t know the best way to resolve among the foolish logic questions which might be more and more standard in job interviews, that is the kind of expertise that carries way more weight with me.
The Schema That Practically Broke Me
I as soon as shipped a transactional schema that appeared superb on paper: normalized, neat, the whole lot in its correct place.
Then analytics confirmed up with “simply a few fast dashboards.” Subsequent factor , my fairly 3NF mannequin, now related to each elementary classroom in America, was getting used like a million-row Excel spreadsheet to summarize an accounting report. For just a few months, it was superb till it wasn’t, and the database had made a sluggish‑movement faceplant as a result of it was spending 80% of its time updating an index. It wasn’t as if I may repair something, as a result of that will imply a number of days of downtime coupled with a rewrite for a undertaking whose contract was virtually up.
And the way have been we making an attempt to repair it? For those who’ve been on this state of affairs, you’ll perceive that what I’m about to jot down is the signal that you’ve reached a brand new degree of determined failure. As an alternative of contemplating a rational method to reform the schema or separating what had turn out to be a “web-scale” workload in 2007 from a NoSQL database, we have been making an attempt to determine the best way to buy quicker onerous drives with increased IOPS.
I discovered loads of issues:
- I discovered that upgrading {hardware} (shopping for a quicker machine or dropping one million {dollars} on onerous drives) will solely delay your disaster. The actual repair is unavoidable—large horizontal scaling is incompatible with relational databases.
- I discovered the that means of “question plan from hell.” We band‑aided it with materialized views and skim replicas. Then we did what we should always’ve performed from day one: arrange an precise reporting path.
- In case you are having to optimize for a question plan each week? Your database is sending you an essential sign, which it’s best to translate to, “It’s time to begin in search of an alternate.”
Lesson burned in: Design for the use case you even have, not the one you hope to have—and assume the use case will change.
What Does This Should Do with Cursor and Copilot?
I’m seeing lots of people writing on LinkedIn and different websites about how superb vibe coding is. These celebratory posts reveal extra concerning the folks posting them than they notice, as they hardly ever acknowledge the truth of the method—it’s not all enjoyable and video games. Whereas it’s astonishing how a lot progress one could make in a day or every week, these of us who’re really utilizing these instruments to jot down code are the primary to inform you that we’re studying loads of tough classes.
It’s not “simple.” There’s nothing “vibey” concerning the course of, and in case you are doing it proper, you might be beginning to use curse phrases in your prompts. For instance, a few of my prompts in response to a Cursor Agent yesterday have been: “You’ve got to be kidding me, I’ve a rule that acknowledged that I by no means needed you to do this, and also you simply ignored it?”
At any time when I see folks get excited concerning the newest, biggest fad factor that’s altering the world, I’m additionally the primary to note that perhaps they aren’t utilizing all of it. In the event that they have been, they’d perceive that it’s not as “simple” as they’re reporting.
The failure muscle you construct with databases is identical one you want with AI coding instruments. You may’t tiptoe in. You must push till one thing breaks. Then you determine the best way to method a brand new know-how as knowledgeable.
- Ask an agent to refactor one file—nice.
- Ask it to coordinate adjustments throughout 20 recordsdata, rethink error dealing with, and hold checks passing—now we’re studying.
- Watch the place it stumbles, and study to body the work so it might succeed subsequent time.
- Spend a complete weekend on a “wild goose chase” as a result of your agentic coder determined to disregard your Cursor guidelines fully. ← That is costly, however it’s the way you study.
The trick isn’t avoiding failure. It’s failing in a managed, reversible method.
The Meta Lesson
For those who’ve by no means damaged it, you don’t actually understand it. That is true for coding, budgeting, managing, cooking, and snowboarding. For those who haven’t failed, you don’t understand it. And the general public speaking about “vibe coding” haven’t.
The folks I belief most as engineers can inform me why one thing failed and how they adjusted their method because of this. That’s the whole sport with AI coding instruments. The quicker you possibly can run the loop—attempt → break → examine → refine—the higher you get.