
Founder & the dev behind @refract_app — one app for everything you watch, free & no ads. Shipping the engineering, the numbers, the mistakes. Paris.
Paris, France
Joined December 2011
- Tweets44
- Following282
- Followers89
- Likes223
Pinned Tweet
Hey X. I'm the Discord caveman behind Refract. 👋
Refract started in March 2026 as a side project, because TV Time couldn't load my homepage anymore. Seven months later:
→ 200k signups, 100k daily actives
→ 17k people on Discord, 10k messages a day
→ 1.5 billion episodes tracked
→ free, no ads, and no feature behind a paywall. Ever.
Movies, series and anime now. Games in November. Books, manga and podcasts down the road.
I'm the founder who writes the code. Here you'll get the engineering: how we survived July, what breaks at 100k daily users, the numbers, the mistakes.
Building it with @Thomaslarrieu and @juanraymondoPOV. Link in the first comment 👇
Florent L. retweeted
🆕 UPDATE: We had to ship games early 🎮
GTA VI comes out next week, and after that nobody on the team will be doing anything else, so it was now or never🤷♂️
Movies, TV Shows, anime and now games, now all on Refract💜
What’s the first game you’re adding?
Florent L. retweeted
🎂 24 turns 25 today ⏱️
One day per season, told in real time, and Jack Bauer never once stopped for lunch.
How many seasons did you make it through?
Florent L. retweeted
John Wick is going animated 🐶
Lionsgate just greenlit an animated prequel, and Keanu Reeves is back to voice him. It's the story of the "impossible task" he pulled off to quit the assassin life the first time.
Hyped, or would you rather some stories stay untold?
Florent L. retweeted
We lost a legend today. Sir Patrick Stewart has passed away at 86.
Captain Picard, Professor X, and decades of Shakespeare on stage.
Thank you for everything, Sir Patrick 💜
Before writing a line of Refract, I spent January and February doing data science on the competition. 🔬
→ gathered the App Store and Play Store reviews of hundreds of tracking apps
→ ran sentiment analysis on the whole dataset
→ sorted every feature into "people love this" and "people suffer from this"
That became the roadmap. Then I built the app, put it on the Play Store for one friend, and here we are, 200k signups later.
10 years in data, and the best model I ever shipped was reading what people were already saying.
One oversized query was nuking our entire offline cache. Only for our best users. Post-mortem 👇
Refract persists its query cache on device, so the app opens instantly and works on a bad connection. The cache has a size cap, as it should.
What we got wrong: when one query's result went over the cap, the persistence layer gave up on the whole snapshot. Not that query. Everything. Next cold start: spinners on every screen.
Which queries go over the cap? The ones from people who track the most. So the bug hit exactly our most engaged users, and only them. 🤌 (not the good kind)
The fix:
→ exclude the oversized entry
→ persist the rest
→ log which key got excluded so we can shrink it
The cache is now partial instead of empty, which is what a cache should be.
General version: when a limit is hit, degrade one item, not the system. "All or nothing" is almost never the right failure mode for a cache.
3
48
How Refract survived July, with the actual graphs. 📈
Day 0: 3,000 users, imports done in minutes, 1 to 5 bug reports a day. Then TV Time announced the end, and import files started hitting us at 950 an hour. Our pipeline was clearing 5. Per hour.
What we did, in order:
→ Day 2: bigger server, imports rebuilt on dedicated pools, the queue secured so a file uploaded once is never lost
→ Day 3: full database migration at dawn, zero data loss. Dumping the old DB took 1h30 because it was that slow. Restoring on the new one took 1 minute.
→ Day 5: multithreading inside the same container. Half-worked: network hiccups killed threads. Lesson taken.
→ Day 6: workers isolated on their own instance, three lanes (small, standard, large). Imports went from 5 to 250 an hour. 10 minutes per import became 90 seconds.
→ Day 8: multiprocessing, 5 app instances behind one switch. Logouts divided by 500. The red/green ratio finally flipped (third picture).
Day 9: 550 imports an hour, 120k users, and I slept. Twice.
What it cost, since people ask: about $250/month more on the database, $320/month of extra machines, and a $720 Expo bill for 130k downloads in a week. Free app, real invoices.
The honest part: none of this was planned. Two devs on one container. The plan was whatever we could ship between two outages, with a mod team and a few Avengers who stayed up with us.
Pragmatism > Perfectionism. Still true at 100k daily users.
Push notifications bug that hid for weeks: the app re-registered its push token on every foreground. 🙃
Open the app 20 times a day → 20 writes on the same row → 20 chances of a race with the server → a token that ends up "unconfirmed" and silently stops receiving.
At 100k daily users, that's 2 million pointless writes a day. For a token that never changes.
Now:
→ register once a day, max
→ a token is live only once the server has confirmed it
Notifications: the feature nobody notices when it works, and everybody notices when it doesn't.
4
62
Refract's roadmap is not decided in a meeting. It's voted on Discord, by 17k people. 🗳️
Next two, by their vote:
→ Rewatching: start a show or anime again without erasing the first run
→ Web app: Refract in your browser, same account, same friends
Want a say in what comes after? The server is open, link in the first comment. Or reply here, I'll be around.
Culture should be free. Everyone should have access to it. That's not a tagline, it's the rule we build Refract on. 🔓
→ Free. No ads. We will never put a functional feature behind a paywall.
→ Subscriptions exist only to support the nights and weekends of work. They unlock nothing you need.
→ Your data is yours: backup and restore anytime. You will never be soft-locked.
TV Time died in July and a wave of apps rushed in. Good apps, built by people I respect. We just answer a different question: not "how do we monetize tracking", but "how do we make culture travel further".
Movies, series and anime today. Games in November. Books, manga and podcasts down the road. 100k people a day already came for it.
We were losing scheduled jobs every night at midnight. Nobody noticed for days. Post-mortem 👇
Refract runs its background work on a cron pool: catalogue refreshes, notification digests, cleanup, stats. A lot of them on the hour. Way too many on 00:00.
At midnight they all wake up together. The pool has a fixed size. The first N grab a worker, the rest wait. Fine in theory.
Except some of those jobs are long. By the time a worker frees up, the waiting jobs had drifted past their own window, and the scheduler went "already late, skip". No error. No retry. Just a run that never happened.
Nothing in our dashboards showed it, because "a run that didn't happen" doesn't emit a metric. 🙃
The fix:
→ a delayed start is not a missed run: a job queued on time runs, even if it starts late
→ spread the pile-up: jobs that don't need 00:00 exactly get a per-job offset
→ a CI check reads the cron inventory and fails if midnight gets crowded again
Lesson: the absence of an event is the hardest bug to monitor. Count the runs you expected, not just the ones you got.
31
We cut Refract's database traffic by 90% in one week. Here's the trick, it's embarrassingly simple. 📉
Every screen you opened was re-downloading way more than it needed. Open your profile: full reload. Come back 10 seconds later: full reload again. Same data, same bill.
So we taught the app to ask one question first: "did anything actually change?" If not, use what's already on the phone.
What it gave us:
→ database traffic down ~90%
→ TV Time imports 3× faster, and lighter
→ new-episode detection, which choked at peak hours, now runs clean
→ a server bill that stops growing with every tap
Free app, no ads, 100k people a day. Every pointless request is money we'd rather spend on features.
Lesson: before buying a bigger server, check how many times you're asking the same question.
Early July. I'm at a restaurant. My phone dies. Not the battery: the notifications. 📱💀
I get home, plug it in, and get it: TV Time just announced it was shutting down, users had less than 15 days to grab their data, and 100k of them signed up to Refract overnight.
We were a side project on one database. You can picture the queue lines. (Einstein said time is relative. Our users disagreed.)
What we did in the weeks after:
→ migrated the database
→ multiprocessed the app
→ optimized every hot path we could find
Now we can take several million daily users without a server blinking. A bit earlier than planned. 🙃
Lesson: your scaling plan is whatever you can actually execute while it's on fire.
Florent L. retweeted
Aujourd'hui on vous partage un article de Florent Laval, @TSEinfo Alumni. Un article informatif regroupant les chiffres clefs sur la situation actuelle engendrée par le Covid-19.
Merci Florent pour cette initiative et le temps consacré à la rédaction! 👏
medium.com/@florent.laval.tl…
Florent L. retweeted
#Webinar 📆 Mercredi 25/03 à 🔟h.
Besoin d’y voir plus clair dans les prix de #transport de #fret ? Ne manquez pas la présentation de notre outil Comparer et Analyser, suivi d’une session Questions/Réponses avec les experts Upply.
Inscription ici 👉 hubs.ly/H0nGHjz0
Florent L. retweeted
. @Thomaslarrieu & @LavalFlorent présentent Upply et le travail quotidien réalisé par l’équipe #data, à l’occasion du .Local organisé par @MongoDB.
#ia #supplychain #MachineLearning #DataScience
Florent L. retweeted
Join our live Q&A session around #Upply, #ML, #MongoDB on #mdblocal at @BridgeParis
@Thomaslarrieu @LavalFlorent
@upply_FR
#mongodb #upply #transformationdigitale #logistique #supplychainoptimization #BridgeParis

















