Notes on Building Software
Notes on software engineering: product-minded code, TypeScript, React, PHP, Laravel, and how to keep systems simple.
53 posts matching
| Post | Topics | Published | Link |
|---|---|---|---|
| The Last Response Isn't the Latest One Fire two requests and the network gets to pick which one finishes first. Search boxes, filters, and webhooks all break the same way. Cancel what you can, ignore what's stale, and let newer data win on purpose. |
| Read | |
| Every Error Should Look the Same If your API returns five different error shapes, every client ends up guessing. Pick one shape, render every failure into it, and give the frontend a code it can switch on instead of a sentence it has to parse. |
| Read | |
| You Can't Redeploy Someone's Phone Once a mobile app ships, every version of it keeps calling your API until someone taps update. Treat the response shape as a contract, change it by adding, and give yourself a way to say "this version is too old." |
| Read | |
| If It Ships to the Browser, It's Public A prefix in your .env file doesn't make a key safe. It makes it public on purpose. Anything in your bundle, your page props, or your inline scripts belongs to whoever opens DevTools. Here's how to keep secrets on the server, in TypeScript, Laravel, and WordPress. |
| Read | |
| Promise.all Is Not a Plan Firing every request at once works great with twenty items in dev. With five thousand in production, you are load testing someone else's API. Put a number on your concurrency, in TypeScript and in PHP. |
| Read | |
| Your Worker Remembers the Last Job Queue workers, Octane, and Node servers boot once and keep running. Anything you leave in a static, a singleton, or module scope is still there for the next job or request. Sometimes for the next user. |
| Read | |
| Let It Finish Before You Kill It Every deploy, scale-down, and restart starts with a signal asking your process to stop. If your app ignores it, the platform kills it mid-request or mid-job. Handle SIGTERM on purpose. |
| Read | |
| Don't Let the Request Pick the Columns Passing the whole request body into a create or update call means the client decides which columns change. Name the fields you accept, and make the framework yell when something sneaks in. |
| Read | |
| Don't Redirect Wherever They Tell You A ?next= parameter feels harmless until your login page starts sending people to someone else's site. Only redirect to paths you own, and check them the way a browser reads them. |
| Read | |
| You Wrapped the Column DATE(created_at) and helpers like whereDate() look tidy. In MySQL they usually keep a plain created_at index from doing a range seek. Compare a range against the raw column instead. |
| Read | |
| Page Two Already Moved OFFSET pagination skips and duplicates rows when the table changes under you. Use keyset/cursor pagination for live feeds, and keep OFFSET for stable admin lists. |
| Read | |
| Don't Trust X-Forwarded-For Rate limits, audit logs, and bans all lean on the client IP. Behind a load balancer that header is easy to misconfigure, and easy to spoof if you trust the wrong hop. |
| Read | |
| A Unique Index Beats Your If Statement Checking that a row does not exist, then inserting it, is two steps racing the same table. Put the real guard in the database and treat the duplicate as a normal outcome. |
| Read | |
| Changing the Password Should Log Them Out A password reset that leaves every old session alive only fixed the problem for the attacker's next login. Here's how to make sessions and tokens actually end in Laravel, Node, and WordPress. |
| Read | |
| A String Isn't as Long as You Think Bytes, code points, UTF-16 units, and what a human calls a character are four different numbers. Your frontend, backend, and database each pick one, and they usually don't pick the same one. |
| Read | |
| Don't Fetch Whatever URL They Give You Link previews, webhook settings, and "import from URL" all mean your server makes requests on a user's behalf. Make sure it can't be pointed at your own network. |
| Read | |
| Escape It on the Way Out Validation keeps bad data out. Escaping keeps the data you let in from turning into code on someone else's screen. They are different jobs, and the second one depends on where the data is going. |
| Read | |
| Stop Logging the Password Logs and error trackers end up holding passwords, tokens, and card data because nobody decided they shouldn't. Redact at the source, log on purpose, and check what's actually in there. |
| Read | |
| A Flaky Test Is Telling You Something Flaky tests aren't bad luck. They're usually time, order, shared state, or the network leaking into your suite. Here's how to find the leak instead of hitting rerun. |
| Read | |
| Staging Should Never Email a Real Customer Copy production data into staging, point it at a real mail server, and wait. Sooner or later a customer gets an invoice that doesn't exist. Make it impossible, not just unlikely. |
| Read | |
| You Don't Have Backups Until You've Restored One A backup job that says 'success' every night proves a file got written. It doesn't prove you can get your data back. Test the restore, on purpose, before you need it. |
| Read | |
| CORS Is Not Your Firewall CORS decides which web pages can read your API's responses in a browser. It doesn't stop anyone from calling your API. Configure it like that's true. |
| Read | |
| Commit the Lockfile. Then Actually Use It. composer.json and package.json describe what you'd accept. The lockfile is what you actually tested. Treat it that way in CI and on deploy. |
| Read | |
| Every Endpoint Needs a Speed Limit If an endpoint can be called once, it can be called ten thousand times. Rate limit by the right key, count in a shared store, return a real 429, and put the tightest limits on the things that cost you money. |
| Read | |
| Never Trust the Filename File uploads are user input with extra steps. Validate the contents, rename everything, keep it out of the webroot, and serve it on your terms. |
| Read | |
| Soft Deleted Isn't Deleted Soft deletes keep the row around, which is the point. It also means the row still holds its email, still hits your unique index, and still shows up in any query that skips the scope. Decide what deleted means before the database decides for you. |
| Read | |
| Anyone Can POST to Your Webhook A webhook endpoint is a public URL that changes data in your app. Verify the signature on the raw body, compare it safely, reject old timestamps, and don't trust the payload more than you have to. |
| Read | |
| Your Cron Job Runs on Every Server Scheduled tasks are easy on one box. Add a second server or a third container and your nightly job runs three times. Pick one runner, lock the work, and make it safe to repeat anyway. |
| Read | |
| The Job Ran Before the Row Existed A database transaction isn't done until it commits. Jobs, emails, and API calls fired from inside one can run too early, or run for something that never happened. Wait for the commit, and keep the slow stuff out. |
| Read | |
| Ship It Turned Off Deploying code and releasing a feature don't have to happen at the same time. Merge small, ship behind a flag, roll out on purpose, and delete the flag when you're done. |
| Read | |
| Logged In Isn't the Same as Allowed Authentication tells you who someone is. It says nothing about whether this record is theirs. Check ownership on the server, on every route, and scope your queries so the wrong row never comes back. |
| Read | |
| 200 OK Is Not a Health Check A health route that always returns 200 only proves the process is on. Split "is it alive" from "can it serve traffic," check the things that actually break, and don't let one slow dependency take every server out of rotation. |
| Read | |
| Retrying Harder Won't Fix a 400 Retries help when the network hiccuped. They hurt when the request was wrong, the dependency is down for real, or fifty workers all try again at once. Retry the transient ones, back off, and know when to quit. |
| Read | |
| Last Write Wins. That's the Bug. Read it, change it, save it works fine with one user. Put two requests on the same row and one of them quietly disappears. Let the database do the math, lock when you have to, and check a version when a human is in the middle. |
| Read | |
| Don't Load the Whole Table Code that grabs every row works great in dev and falls over in production. Process big data in pieces, and pick pieces that don't move under you. |
| Read | |
| The Cache Is Lying to You Caching makes slow things fast and correct things wrong. Decide how stale is okay, who clears it, and what happens when it expires, before you add it. |
| Read | |
| Validate It at the Door Bad data is cheapest to reject the moment it shows up. After that, every layer has to guess. |
| Read | |
| The AI Wrote It. You Still Own It. AI tools write plausible code fast. Plausible is not the same as correct, and your name is still on the commit. |
| Read | |
| A Log Line Nobody Can Search Is a Diary If your logs are sentences, you are writing a diary. Log events with context so you can actually find the problem at 2am. |
| Read | |
| Your Old Code Is Still Running Deploys aren't instant. Old servers, old queued jobs, and old browser tabs keep running against your new schema and API. Change things in steps they can survive. |
| Read | |
| It Was Fast With Ten Rows Most slow pages aren't slow because of PHP or React. They're slow because of how many queries and requests they make once real data shows up. Count them. |
| Read | |
| Your Server Doesn't Live in Your Timezone Store moments in UTC, keep the user's timezone as a real IANA name, and convert at the edges. A lot of date bugs come from skipping one of those. |
| Read | |
| Money Is Not a Float Floats are fine for a lot of things. Prices, totals, and refunds are not on the list. Store integer cents, round once, and do the math in one place. |
| Read | |
| Your .env File Is Not a Global Variable Reading environment variables wherever you need them works until a cached config, a typo, or the string "false" bites you. Read config once, validate it, and pass it around. |
| Read | |
| Hang Forever Is Not a Strategy Every outbound call has a timeout, even when you never set one. Pick numbers on purpose so hung remotes fail fast instead of tying up workers and users. |
| Read | |
| Make It Safe to Run Twice Retries, webhooks, and double clicks will run your write path again. Design the work so a second pass does not double-create, double-charge, or double-email. |
| Read | |
| Catching It Doesn't Mean You Handled It An empty catch block (or one that only logs "failed") hides the failure from everyone who needs it. Catch with a plan, or let it surface. |
| Read | |
| A Boolean Is a Terrible Status true and false feel simple until you have to explain what actually happened. Name the outcome and your code gets easier to read, retry, and debug. |
| Read | |
| Treat Every Integration Like It Will Break Stripe, email, shipping, CRMs, webhooks. The other system will flake. Design the handoff so your app stays honest when it does. |
| Read | |
| If It Can Wait, Put It In a Queue Web requests are for answering people. Emails, exports, webhooks, and image work can wait in a job. Here’s the boring habit that keeps apps feeling fast. |
| Read | |
| Stop Turning Every WordPress Problem Into a Plugin Problem Another plugin is often the wrong fix. Sometimes the cleaner path is custom code, a setting you already have, or deleting something. |
| Read | |
| PHP Has Been Dying for 20 Years. It’s Doing Fine. The obituaries keep coming. The web keeps running on PHP anyway. Here’s why the “dead language” story keeps missing the point. |
| Read | |
| I Still Like Boring Software New frameworks are fun. Most products still need a stack that ships features, takes payments, and stays up without a PhD to deploy. |
| Read |