Rates at Hertz changed 10,000 times per second. We didn't cache them.
That sounds wrong. It was the decision that made everything work.
The expensive part of a rate lookup wasn't the rate : it was figuring out which of tens of thousands of rules, discounts, and eligibility criteria applied to that specific request. Hundreds of DB calls per request, every time.
We moved all of that filtering upfront. Pre-grouped rules by location and corporate account into Redis hashed keys. When a request came in: one Redis fetch for the pre-filtered rules, one targeted DB call for the actual rate. Hundreds of round trips became two.
The write path had its own evolution : from REST calls hammering 20+ custom queue workers (LAX and NYC backlogged badly on burst) to Kinesis event streaming. Just dropping the full HTTP stack was a massive speed improvement before we even got to the architecture wins.
Geo-routed updates with two-layer replication : Kinesis for speed, Cloudant multi-master for durability. Kafka would have given us this natively. With Kinesis we built it ourselves. AWS pricing made it worth it.
p95 went from ~3 minutes to 30ms. 6,000% improvement. We fixed latency and throughput. A quarter of the infrastructure : 10 Node instances instead of 40.
I didn't do this alone. Jerry (Gerardo Leon) invented the hash grouping strategy that made the whole cache layer work. Taffy was there from the start, working through the complex system and added SLAYER as a test discount code ; it stuck. Full team callout in the post.
I wrote the whole architecture breakdown here:
https://t.co/2c0pS5gp1P
#Architecture #Engineering #SoftwareDevelopment #HighPerformance #TechLeadership
Booked a car, picked it up from the presidents circle and got something a bit bigger for a road trip. Packed it up, got little one in and...
In line with my usual luck, the rental car was leaking oil. Took the car back, had to swap it. Luckily Hertz swapped it. However, they wanted to swap it down and had to see if they could do an equal exchange because I booked knowing I had President's circle. (Maybe that's on me and I should have forked out for the full thing)
@Hertz what happened to the customer support though? No points, no upgrade, wasted hours and had to plead for the same car 🤣. I asked for the cost of an upgrade and I'm guessing y'all were just hella sold out? Because that didn't really get answered.
Well, this is embarrassing, the automation messed up
I'm a PITA
A pain in the ass,
I'm sure I may be,
but hopefully worth it, to thee.
I'm needy, I'm coarse, I'm callous, I'm sore.
I know sometimes, I'm probably even a bore.
But somewhat smitten, I quite seem to be,
I'm here once again, writing poetry.
I want you, I need you, it is easy to see.
So though quite a pain, please, caress me.
From a pain like myself, it may be asking too much,
But I really desire, enjoy, and need your touch.
Over the moon, head over heels, apples and flames
The idioms abound, they have so much range.
Sayings all over, but they all do the same...
Proclaiming feelings one hopes will not change.
I think that's the right version -.-
A PITA, I'm sure I may be. 🌵
A pain in the ass, I'm sure I may be,
but hopefully worth it, to thee.
I'm needy, I'm coarse, I'm callous, I'm sore.
But somewhat smitten, I quite seem to be,
I'm here once again, writing poetry.
https://t.co/1BDeqloB3Z
#poetry
Google’s 10k+ interview study found zero correlation between scores and performance.
Whiteboarding doesn’t measure skill; it measures anxiety. 📈
Top engineers stand out through what they build, not interview theater. 🛠️
Full breakdown below. 👇
I spent years at EY and Zilker Technology doing 40hr/week consulting engagements where the actual high-value architecture decisions took 10-15 hours a week. The rest was overhead : backlog grooming, PR reviews, data entry, playing three roles because the SOW scoped it that way.
That’s the difference between consulting and fractional. Fractional strips away the noise.
Now I work with companies at every stage. Enterprise teams that need architecture direction for a specific initiative. SaaS startups that need the critical early decisions made right. Non-technical founders building real businesses who need the tech and the growth strategy figured out together.
The pattern is always the same: most companies don’t need 2,080 hours of CTO per year. They need 500. Maybe 200. Maybe 40.
New post breaks down the real cost math (with sources), what each tier of engagement actually looks like week by week, and when you should skip fractional and hire full-time instead.
https://t.co/F28yW2P5Q3
I woke up to 124 open PRs. 90 were duplicates. While I was closing them, the director spun up 7 more.
That's what happens when your agent loop doesn't track what it already did.
I wrote more about this at https://t.co/0KK44mvx1n
The world through a screen:
It can be bright or dark,
filled with color or blank.
You see many people on it,
few you know.
Many sights and many sounds,
landscape,
animals.
The world is in it.
I know one thing,
my TV is always on.
Full piece at https://t.co/0IX60KBwqq ✍️
#poetry #originalpoetry #writersofinstagram
This makes me think of my predictions over time. When I first learned javascript, I NEVER thought I would be writing it for more than moving HTML around. Then jQuery was released later that year and 2 years later I had picked that up. Node and v8 came later to really change things. The pace of change has just continued to increase year over year and now the pace is mindblowing...
It'll be interesting to see what's going on in 6 months considering a year ago AI was reasonable as a code complete but now I'm running agentic swarms that PR, review, help with PRDs, and prep releases for review.
The failures, because that's the real story: 124 open PRs overnight : 90 duplicates. The director never checked whether a PR already existed. 747 sessions ran on Opus before I caught that --model wasn't being passed. $65/day instead of $15. 46 sessions cascade-failed in under 20 seconds. Director marked them all "max attempts exceeded." It was a rate limit. Every one of these was a systems design failure, not an AI failure.
In 2015 I was drawing CI/CD pipeline diagrams at Kohl's. Build → Test → QA → Deploy → Release. Hundreds of people per stage.
Now it's running on two machines in my house. One Node.js script and a vector database.
https://t.co/0KK44mvx1n
The setup: Dragon (Arch Linux, AMD ROCm, 14 containers, 15 session slots) + Hive (Mac Studio, Xcode, iOS simulators, 20 session slots). They talk over SSH and a shared Qdrant instance. The Qdrant instance doubles as a task queue because macOS LaunchAgents get EHOSTUNREACH connecting back to the desktop. Flipping the direction solved it. Peak day: 1,047 agent runs.
To put this before AI: If you stack overflowed it, it is owned by all the people that did the same algorithm and posted something before you...
We all used A-Star when doing pathfinding and jaro winkler for doing fuzzy matching at this point. You're just evolving to the next level
Thank you Brodin, my gainz were good.
I can now lift more than I could.
The endorphin high reminds me to remove my rage.
Forgiveness for all, that is my new page.
At least until tomorrow,
When far off I will go.
Traveling, yet not moving,
the pain will begin shooting.
I'll feel the sore, glad to feel.
I'll embrace the pain and know something real.
However, my attitude may shift,
as I wander around through people adrift.
I'll exit the airport, rage fully renewed...
But then, once again, I'll just have to pray to you. 🏋️♂️
#poetry #originalpoetry #gympoetry #brodin #wheymen
Most DevOps problems aren't DevOps problems. They're people problems.
Stop reaching for Kubernetes too early. Use managed services. Automate the boring stuff with shell scripts.
Stay fast, stay simple.
Full guide: https://t.co/URsK7rXIHD