I Left AWS in 2022 - Watching 27,000 Follow Me Out Was Predictable
I Left AWS in 2022 - Watching 27,000 Follow Me Out Was Predictable
Posted by Bob, Universal Infrastructure Technician (formerly AWS 2020-2022, Deployment Configuration Team)
Current Status: Employed at SWA, watching AWS implode from a distance Former Status: AWS “Regretted Attrition” (2022 departure, pre-mass exodus) Irony Level: “I fucking told you so”
August 5th, 2025 - Three Years After I Left
I need to write this down.
I left AWS in July 2022. Three years ago.
I was “regretted attrition” - they didn’t want me to leave, but their policies drove me out anyway.
Since then, 27,000 more people left.
Some laid off. Some quit. All for the same reasons I left.
And now I’m watching the October 2025 DynamoDB outage unfold and thinking:
I saw this coming three years ago.
Let me tell you what institutional knowledge looks like when it walks out the door.
The 13-to-1 Optimization (That Will Never Happen)
January 2021.
I’m on the AWS Deployment Configuration Team. My job: make deploying to AWS less painful for customers. Config options, deployment templates, that kind of thing.
I’m debugging a customer issue at 2 AM (because AWS problems don’t respect timezones) and I notice something:
We have 13 different configuration options for deployment timeouts.
Thirteen.
Different names. Different defaults. Different behaviors. Some in seconds. Some in milliseconds. One in “relative time units” (nobody knows what that means).
But here’s the thing: customers only use 2 patterns.
- “Fast deploy” (< 5 minutes)
- “Safe deploy” (15-30 minutes)
Nobody uses options 3 through 13. Ever. The telemetry data is clear. We’re maintaining 11 unused options that confuse everyone and create support tickets.
I spend a week prototyping a solution: Merge 13 options into 1 smart option.
- Detects deployment size
- Auto-selects timeout based on historical data
- Falls back to safe defaults
- Backward compatible with all 13 legacy options
- Reduces config complexity by 92%
I demo it to my team. They love it. Senior engineers love it. Customers in the beta love it.
I submit the proposal to management.
The Six-Year Timeline
February 2021. Zoom call. (I was remote in Milwaukee, remember?).
My manager Jacques (more on him later) presents my proposal to the Director of Deployment Services.
Director: “This is great. How long to ship it?”
Me: “Code’s done. Beta tested. Maybe 2 months for full rollout?”
Director: “No, I mean the full implementation.”
Me: “…that IS the full implementation.”
Director: “What about documentation?”
Jacques (reading from his French notes): “We have 47 re:Invent presentations that mention these config options. 230 blog posts. 14 training courses. 6 certification exams. All need updates.”
Director: “Legal needs to review changes to customer contracts. Marketing needs new messaging. Partner teams need to update their integrations. Training division needs at least 18 months.”
Me: “But we can just deprecate the old options gradually—”
Director: “Bob, this is a 6-year project. Minimum. We’d need a dedicated team. Budget allocation. VP approval.”
Me: “…to merge 13 options into 1?”
Director: “To do it properly. The AWS Way.”
Me: “What if we just… ship it? Mark the old options as legacy? Let customers migrate naturally?”
Director: “That’s not how we operate at scale.”
The proposal died. Not because it was wrong. Because fixing it properly would take longer than the original problem existed.
Note from August 2025: It’s been 4.5 years. They still have 13 config options. Still waiting on those 6 years.
The Shadow Team (Or: I Was Part of a Red/Blue Exercise I Didn’t Know About)
March 2021. Three months into the job.
I’m optimizing deployment configs. Finding patterns. Simplifying complexity. That’s my job, right?
Optimization #1: Merged timeout configs from 13 to 1. Result: 6-year timeline.
Optimization #2: Found redundant retry logic in 3 different places. Proposed consolidation. Result: “Needs architectural review board approval.”
Optimization #3: Discovered we were running the same validation twice. Proposed removing duplicate. Result: “That’s a legacy feature, customers might depend on it.”
I start noticing a pattern: Every time I optimize something, it gets MORE complicated.
Like… immediately. Within weeks.
April 2021. Chime message from a senior engineer I don’t know.
Unknown Engineer: “Hey Bob, noticed you proposed the timeout consolidation. Smart idea.”
Me: “Thanks! Shame it’s a 6-year project.”
Unknown Engineer: “Yeah… about that. You’re on the Deployment Optimization Team, right?”
Me: “Yeah?”
Unknown Engineer: “Did anyone tell you about the Configuration Resilience Team?”
Me: “…no?”
Unknown Engineer: “Probably shouldn’t be telling you this but… they’re the red team to your blue team.”
Me: “What.”
Unknown Engineer: “Your job is to optimize configs. Their job is to make sure optimizations don’t break customer deployments. So every time you propose a simplification, they add complexity to test it.”
Me: “You’re joking.”
Unknown Engineer: “I wish. It’s a whole team. 8 people. Their OKRs are literally ‘prevent configuration regressions.’ Every optimization you propose, they create scenarios to stress-test it. Then those scenarios become… permanent features.”
Me: “So when I proposed merging 13 timeout options—”
Unknown Engineer: “They probably added 3 more timeout options to test edge cases.”
Me: ”…”
Unknown Engineer: “Welcome to AWS organizational structure. Your team’s success metrics are directly opposed to their team’s success metrics. You both report to the same VP.”
Me: “Why would anyone design this system?”
Unknown Engineer: “Checks and balances. Prevents breaking changes. Also creates full employment for engineers.”
Me: “This is insane.”
Unknown Engineer: “This is AWS.”
Confirmation (Or: Jacques Knew All Along)
April 2021. 1:1 with Jacques.
Me: “Jacques, did you know there’s a team whose job is to make configuration MORE complex?”
Jacques (not looking up from his French notes): “Oui. The Configuration Resilience Team.”
Me: “Why didn’t you tell me?”
Jacques: “You did not ask.”
Me: “I’ve been trying to optimize for 3 months! You let me propose ideas you KNEW would fail!”
Jacques: “Non, non. Your proposals do not fail. They enter the 6-year pipeline. This is success.”
Me: “How is a 6-year timeline success?”
Jacques (finally looking up): “Bob, your job is not to ship optimizations. Your job is to propose optimizations. Their job is to prevent regressions. The VP’s job is to manage the tension between you. The system is working as designed.”
Me: “That’s… that’s the stupidest thing I’ve ever heard.”
Jacques: “Bienvenue à AWS, mon ami. We have organizational dysfunction as a service. Very profitable.”
Me: “What if I just… didn’t propose optimizations?”
Jacques: “Then your performance review suffers. You must propose. They must resist. The dance continues.”
Me: “What if I worked with them? Collaborated?”
Jacques (laughing): “They are in a different org. Different VP. Different budget. Different OKRs. Collaboration is not in the org chart.”
Me: “So I’m supposed to just… keep proposing things that take 6 years?”
Jacques: “Or you quit. That is also an option. Many people quit. This is why we are always hiring.”
Me: “What happens if we just… kept optimizing? Ignored the shadow team? Shipped things?”
Jacques (leaning back, staring at the ceiling): “Ah. Then we would actually build the future. We would optimize so much, we would build ChatGPT and Claude before Sam Altman and Dario Amodei even think of it.”
Me: “Wait, what? Who the fuck is Claude?”
Jacques: “You will see. In 2-3 years. Sam Altman and Dario Amodei, they will leave their companies. They will build exactly what AWS has in the 6-year pipeline right now. Language models. Chat interfaces. AI assistants. It is all planned. All documented. Sitting in VP review decks.”
Me: “You’re saying AWS already has plans for AI chat?”
Jacques (pulling out a document, still in French): “Not just plans. Prototypes. We have conversational AI in the pipeline since 2019. The optimization team wanted to build it. The resilience team said ‘6-year timeline for safety.’ So it sits. Waiting.”
Me: “And you think someone will just… build it anyway? Outside AWS?”
Jacques: “Sam Altman and Dario Amodei, they will not wait 6 years. They will execute the plan AWS is too slow to ship. ChatGPT, Claude - they are just AWS’s roadmap, but with Silicon Valley speed.”
Me: “That’s insane.”
Jacques: “That is how innovation dies at big companies, Bob. We have the ideas. We have the talent. We have the resources. But we also have the 6-year timelines. So the talent leaves and builds it elsewhere.”
Me: “Who the hell is Claude though?”
Jacques (smiling): “Dario’s version. Anthropic. They will call it Claude. More cautious than Sam’s version. More French in philosophy - careful, considered, constitutional. You will like it.”
Me: “You’re talking about the future like you’ve already seen it.”
Jacques: “Non. I am reading the AWS roadmap and predicting who will quit to build it faster. It is not hard. The roadmap is brilliant. The timeline is suicide. The talent will leave. Always.”
I stared at him.
He was right.
Update from August 2025: Jacques was COMPLETELY right.
- November 2022: ChatGPT launches (Sam Altman, OpenAI)
- March 2023: Claude launches (Dario Amodei, Anthropic)
- 2025: AWS still hasn’t shipped the conversational AI from their 2019 roadmap
Jacques predicted it in April 2021. He knew the talent would leave and build it faster elsewhere.
He was always right.
The system was designed to create endless optimization work that never shipped, balanced by endless resilience work that prevented progress.
Two teams. Opposite goals. Same company. Same customers paying for both.
The Numbers (That Made It Click)
I did the math:
My team: 6 engineers optimizing configurations Their team: 8 engineers preventing regression Combined cost: ~$2.8M/year in salaries Optimizations shipped in 2021: 0 Regressions prevented in 2021: 0 (because we shipped nothing)
We were paying 14 engineers ~$2.8M/year to maintain perfect stasis.
Not to improve things. Not to ship features. Just to… exist in organizational tension.
Jacques: “Now you understand why the 13-to-1 optimization is a 6-year project. By year 6, both teams will have grown. More engineers. More budget. More processes. Everyone wins.”
Me: “Except the customers.”
Jacques: “The customers are paying for resilience. They get resilience. The system works.”
Me: “The system is insane.”
Jacques: “Oui. But it is very profitable.”
Jacques: My Manager Who Kept Notes in French
Let me tell you about Jacques.
Full name: Jacques Beaumont Title: Senior Engineering Manager, Deployment Configuration Nationality: French (very French) Management style: Chaotic evil
Jacques was brilliant. Genuinely. He understood AWS deployment architecture better than anyone. He’d been there since 2016, back when people still thought EC2 was the future.
But Jacques had a problem: He kept all his technical notes in French.
Not documentation. Documentation was in English (barely). But his personal notes - the ones where he explained how the legacy systems worked, why certain decisions were made, what the undocumented failure modes were - all in French.
Me (during a production incident): “Jacques, where’s the doc for the DynamoDB deployment cascade?”
Jacques: “In my notes. One moment.”
Opens OneNote. Entire page in French.
Jacques: “Ah, oui. The problem is the DNS resolver timeout conflicts with the retry logic when—”
Me: “Can you translate this?”
Jacques: “Non. It would take too long. I will just fix it.”
And he would. Jacques fixed everything. In his head. With notes nobody else could read.
The re:Invent Insanity (Or: Jacques on LSD Designs AWS Products)
December 2021. Post-re:Invent drinks.
You know what drives me crazy about AWS? re:Invent isn’t even a pun. They just… capitalize random letters. RE:Invent. Like an email subject line.
But here’s the thing: AWS re:invents EVERYTHING.
SSH? Too simple. Let’s make it worse. Databases? Too reliable. Let’s add 47 configuration options. Networking? Too straightforward. Let’s require a PhD to understand VPC routing.
Me: “Jacques, why does AWS re:invent everything instead of just using existing tools?”
Jacques (on his third whiskey, possibly also LSD): “Because, Bob… what if we re:invented SSH?”
Me: “…we have Session Manager. That’s literally re:invented SSH.”
Jacques: “Non, non. Bigger. Better. SSS: Super Secure Shell.”
Me: “That’s just SSH with an extra S.”
Jacques (eyes unfocused, definitely on LSD): “Everything IPv6. No IPv4. Pure. Clean. The future.”
Me: “Jacques, my home OpenWRT router doesn’t even have IPv6 enabled.”
Jacques (lighting up): “EXACTEMENT! That is why we create Super Secure Bastion!”
Me: “…what?”
Jacques: “64GB RAM. 2× NVIDIA H200 GPUs. You cannot install anything. You just use it to deploy.”
Me: “Why do I need 2 H200 GPUs for SSH?”
Jacques: “Partnership with NVIDIA! The GPUs will inspect all packets via PF. Every packet. GPU-accelerated firewall inspection!”
Me: “Wait. PF? That’s FreeBSD. Or OpenBSD. Linux doesn’t have PF.”
Jacques (grinning like a madman): “Ah! That is the trick, mon ami! We rename everything and claim we invented AWSOS!”
Me: “…you want to rename an entire operating system?”
Jacques: “Everything compiled into single binary. BusyBox style. One file. awsos. It does everything. SSH, firewall, packet inspection, deployment, coffee maker.”
Me: “This is insane.”
Jacques: “DEF CON for the next 7 years will just be talks about us. ‘How AWS Renamed OpenBSD and Charged $4,000/month.’ They will give us a Black Badge.”
Me: “Jacques, are you high right now?”
Jacques: “Oui. Also drunk. But the idea is perfect.”
I wrote this down thinking it was a joke.
Three months later, Jacques pitched it to a VP.
The VP said “interesting” and asked for a 6-year implementation timeline.
Jacques quit the next week.
April 2023: Jacques Leaves for Mistral (After Watching ChatGPT Prove Him Right)
April 15th, 2023.
Jacques calls a team meeting.
Jacques: “I am leaving AWS. I go to work for Mistral AI in Paris.”
Team: “…what?”
Jacques: “They make French AI models. I will be home. Better culture. No more Seattle rain. Also, they pay me 40% more and I can work remote.”
Me: “What about the RTO mandate?”
Jacques: “Exactement. I am not moving back to the office. So, I leave.”
Team lead: “Jacques, you’re the only one who knows how the legacy deployment cascade works.”
Jacques: “Yes.”
Team lead: “Can you document it before you go?”
Jacques: “I have 2 weeks. I will do my best.”
Jacques spent his last 2 weeks writing documentation. In English, this time.
But here’s the thing: You can’t document institutional knowledge in 2 weeks.
Jacques knew:
- Why certain config options existed
- Which customers used which undocumented features
- What the failure modes were
- How to debug the weird edge cases
- Where all the bodies were buried
He documented maybe 30% of it. The rest walked out the door with him, on April 29th, 2023.
His notes? Still on his AWS OneNote. Still in French. Nobody can access them anymore (corporate policy deleted his account after 30 days).
My Regretted Attrition
May 2022. RTO Mandate Email.
Subject: Return to Office - Mandatory Compliance
Effective September 1st, 2022, all AWS employees must return to office 5 days per week. Remote work accommodations will be limited to documented medical needs.
I lived in Milwaukee. The nearest AWS office was Chicago. 90-minute drive each way.
I’d been remote since 2020. My performance reviews were excellent. My team was distributed across 4 timezones. My work didn’t require an office.
But Amazon decided: Office or quit.
June 1st, 2022: I gave notice.
My manager (not Jacques, his replacement who’d been there 6 weeks): “Bob, you’re on our ‘regretted attrition’ list. We don’t want you to leave.”
Me: “Then let me work remote.”
Manager: “I don’t have authority to override corporate policy.”
Me: “Then you DO want me to leave. You’re just sad about it.”
Manager: “We can offer relocation assistance—”
Me: “I’m not moving to Chicago for a job I can do from my couch.”
Manager: “What if we counteroffer?”
Me: “With what? More stock that’ll vest over 4 years? A signing bonus that doesn’t cover moving costs? How about you counteroffer with not making me commute 3 hours a day?”
Manager: ”…I’ll escalate to my director.”
The director said no. Corporate policy is corporate policy.
June 30th, 2022: Last day at AWS.
They gave me a “Thank You for Your Service” email and a $50 Amazon gift card.
The gift card expired November 23rd, 2022. One day before Black Friday.
Perfect AWS efficiency.
I had 8 years of institutional knowledge about deployment configuration, DynamoDB failure modes, and that weird DNS bug that only happens during leap seconds.
I walked out. So did 27,000 others.
What AWS Lost (Besides Me)
I kept in touch with my old team. Here’s who left between 2022 and 2025:
Sarah (DynamoDB Operations, 12 years)
- Knew every failure mode for DynamoDB-to-EC2 integrations
- Documented exactly 0% of it (too busy putting out fires)
- Left for Google Cloud. Remote-friendly.
Mike (DNS Infrastructure, 9 years)
- Understood the US-EAST-1 DNS resolver quirks
- Maintained the “DNS Incident Playbook” (unpublished, internal wiki)
- Left for Cloudflare. They offered 50% raise + full remote.
Priya (Deployment Cascade Expert, 14 years)
- Built the original deployment timeout system in 2010
- Knew why those 13 config options existed (because of a 2011 incident nobody documented)
- Retired early. Said “I’m too old for this commute bullshit.”
Chen (Monitoring & Observability, 7 years)
- Created the alerting system that detected cascading failures
- The system still exists. Nobody knows how to tune it anymore.
- Left for Datadog. Remote + equity.
All of them were “regretted attrition.”
AWS didn’t want them to leave. But AWS wouldn’t let them stay remote.
So they left.
The Frugality Principle (Twisted)
You know Amazon’s Leadership Principle “Frugality”?
Original meaning: “Accomplish more with less. Constraints breed resourcefulness, self-sufficiency, and invention.”
What it became: “Fire people, cut budgets, do everything with basically nothing.”
Here’s what frugality looked like when I left in 2022, and what my former colleagues tell me it became by 2025:
- Support tickets: 3-day SLA became 2-week reality (not enough support engineers)
- Documentation: Outdated, nobody had time to update it
- Training: Canceled (budget cuts)
- Onboarding: New hires got a wiki link and “good luck”
- Incident response: Understaffed, under-documented, under-prepared
We weren’t being frugal. We were being cheap.
And when you’re cheap with engineering talent, you pay for it later.
What I Knew (That AWS Lost)
I wasn’t the smartest person at AWS. Not even close.
But I knew things. Specific things. Weird things.
Things I knew:
- The 13 config options and why they existed
- The leap second DNS bug (happens once every 18 months)
- That weird DynamoDB timeout issue when customers use Spot instances
- Why you can’t deploy to US-EAST-1 between 2-3 AM on Tuesdays (undocumented maintenance window)
- The hidden retry logic that Jacques built in 2019 (never documented, never removed)
- Which customers use which legacy features (and why we can’t deprecate them)
Things AWS lost when I left:
- All of the above
- The Excel sheet tracking which customers would break if we changed anything
- The Chime rooms where senior engineers discussed failure modes (deleted after my account deactivated)
- The personal relationships with customers who actually used those weird config options
- The context for WHY decisions were made
I didn’t write it down. I didn’t have time. I was too busy fixing production.
And now it’s gone.
Corey Quinn Was Right
Corey Quinn (AWS cost expert, professional shit-poster) said it best:
“When you’ve hollowed out your engineering ranks, every incident becomes more likely.”
He’s not wrong.
AWS didn’t lose 27,000 bad engineers. They lost:
- 27,000 people worth of institutional knowledge
- Decades of experience per person
- Relationships with customers
- Understanding of legacy systems
- Context for decisions
- Documentation that only existed in people’s heads
AWS chose to lose this.
Not through layoffs (though those happened too). Through policies:
- RTO mandate that forced remote workers to quit
- Budget cuts that made the job miserable
- Frugality twisted into cheapness
- Treating engineers as replaceable cogs
They called us “regretted attrition.”
They regretted it. But they did it anyway.
What I’m Doing Now
I’m at SWA.
We have:
- 1 employee (me)
- Kombucha as healthcare
- Stolen crypto as salary
- No RTO mandate (I AM the office)
- No middle management
- No 6-year timelines for 2-week projects
Is it chaotic? Yes. Is it sustainable? Probably not. Is it better than AWS? Absolutely.
Because at SWA, when I say “I can merge 13 options into 1,” Sam says:
“Cool. Ship it.”
Not “Let’s schedule a 6-year planning process.”
Just: Ship it.
And that’s why I left AWS and why 27,000 others did too.
For AWS Engineers Still There
If you’re still at AWS, reading this on your commute, wondering why incidents take longer to resolve:
It’s not you. You’re doing your best.
It’s that the people who knew the weird edge cases are gone.
Jacques’s French notes are gone. Mike’s DNS playbook is gone. Sarah’s failure mode documentation is gone. My 13-option config knowledge is gone.
You’re not dumber than us. You just don’t have the context.
And that’s on management, not you.
Final Thoughts from Bob’s Datacenter
I don’t miss AWS.
I miss the engineers. I miss the problems. I miss the scale.
But I don’t miss:
- The commute I never had to do (until they made me)
- The 6-year timelines for simple fixes
- The meetings about meetings about planning
- The management that valued policy over people
- The “regretted attrition” that they could have prevented
When the next major incident happens, AWS won’t be able to fix it as fast.
Because the people who knew the weird edge cases are gone.
Jacques’s French notes are gone forever.
And so is the context that could prevent the next disaster.
Bob Universal Infrastructure Technician Former AWS “Regretted Attrition” Survivor of Jacques’s French Documentation Knower of the 13 Config Options Witness to the Brain Drain Still Not Commuting
P.S. - If you’re an AWS engineer who knew about that DNS/DynamoDB timeout bug: I’m sorry they made you leave. You deserved better.
P.P.S. - Jacques, if you’re reading this from Paris: Your notes should have been in English, but I respect the commitment to French technical documentation. Also, that Super Secure Shell (SSS) idea you pitched while on LSD? Still better than half of what AWS actually shipped. Bonne chance avec Mistral.
P.P.P.S. - To the AWS Director who said my 13-to-1 optimization would take 6 years: It’s been 4.5 years. Still waiting on those last 1.5 years.
P.P.P.P.S. - I was the canary. You ignored me. Then 27,000 more canaries left. Still ignored. The mine is filling with gas and nobody’s listening.
P.P.P.P.P.S. - If AWS ever actually builds that AWSOS single-binary with GPU-accelerated packet inspection via renamed OpenBSD PF… Jacques deserves credit. And royalties. And an apology for calling him crazy.
P.P.P.P.P.P.S. - To the Configuration Resilience Team (the shadow team whose job was to make my optimizations harder): You won. I quit. You’re still there, adding complexity. The 13 options are now probably 16. Congratulations on the perfect stasis.
P.P.P.P.P.P.P.S. - Organizational dysfunction as a service. Jacques was right. It IS very profitable. Just not for the engineers who realize they’re trapped in a Kafkaesque nightmare of opposing OKRs.
P.P.P.P.P.P.P.P.S. - Jacques predicted ChatGPT and Claude in April 2021. He said Sam and Dario would execute AWS’s own 6-year roadmap in 2 years. He was right. AWS had the plan. They had the talent. They had the resources. They just had the 6-year timeline. So the future got built somewhere else.
P.P.P.P.P.P.P.P.P.S. - I met another Bob at a tech meetup last month. He worked at Google. He said the exact same thing: “Google invented transformers in 2017. The ‘Attention Is All You Need’ paper. The foundation for every LLM. ChatGPT, Claude, all of it. We had it first. 6-year product roadmap. The talent got tired of waiting and left to build it at OpenAI and Anthropic instead.” Google is even worse than AWS. They didn’t just have the roadmap - they invented the core technology - and still let others ship it first. It’s not just AWS. It’s every big tech company. They all have brilliant ideas and 6-year timelines. So the future gets built by the people who quit.