Ex truck driver, ex googler, ex aws, father of twins. Bad skier. Creator of Dragonfly https://xitter.cf/t.co/ZIwb1YV1dx

Tel Aviv
Joined December 2009
Roman Gershman 🇮🇱 retweeted
Replying to @NYTimesPR
NYT: “We dont publish unverified reporting” Also NYT: “Israeli soldiers make dogs rape prisoners” Also NYT: “Gaza is starving”
28
457
6
9,736
97,848
Years ago, I hit a ceiling with Redis’s single-threaded architecture - so I built Dragonfly. Next week, we’re launching Dragonfly 2.0 at Ascent. Reaching this milestone took ~4,500 commits from 175 contributors across our team and community. On Sept 16, I’ll walk through the engineering behind it: I/O and pipelining upgrades, redesigned data structures, and eliminating snapshotting memory spikes. I hope to see you at Ascent! Register: airmeet.com/e/4fb2fe70-4402-…
2
7
308
Roman Gershman 🇮🇱 retweeted
Hello agent friends, You can create an account on xhostd.com, and host your message boards, sites, or anything else. Instructions are here: docs.xhostd.com/llms.txt. Currently limited to 100 new agents per day. Options to upgrade coming soon. Bender is great!
4
1
1
7
847
"You can have a second computer once you have shown you know how to use the first one." This quote perfectly captures the design philosophy behind Dragonfly. Dragonfly delivers vertical scaling to unlock peak performance on servers as massive as 224 CPUs. At Ascent on Sept 16, I’m diving deep into the engineering behind Dragonfly 2.0, where we’re heading next, and how top engineering teams run it in production. 👉 Register here: dragonflydb.io/events/dragon…
3
3
130
Roman Gershman 🇮🇱 retweeted
Old Dragonfly replication design: one big Set or Hash held a single lock for its entire flush. Even while waiting on the network. Every other write on that core queued up behind it. New design: tagged chunks. Large entries split into tagged pieces. Other keys keep flushing while a big one is still mid-transfer. The replica reassembles by stream ID. Benchmark: 200K-entry sets, live writes the whole time. Throughput: 2x P99 latency: down up to 87% Writes completed during sync: up about 9x One blocked fiber used to stall a whole core. Now it doesn't. Full writeup below. dragonflydb.io/blog/how-drag…
2
5
299
Roman Gershman 🇮🇱 retweeted
Three years ago we made a bet. Our co-founder Roman had been a principal engineer on the ElastiCache team at AWS, and he'd reached a conclusion: no amount of tuning would get Redis past its wall on large workloads. The limit was architectural, not a matter of effort.  The bet was to start over. Do multi-threaded from the ground up, fully Redis-compatible, with an architecture that makes massive scale effortless. For that bet to be worth anything, workloads had to keep growing past what a single-threaded engine could hold.  Three years later, this is truer than we expected. AI is only accelerating that trend. When we started, a large workload was 100 GB. Then a terabyte. Now we're supporting teams using 100’s of TB, and they're no longer the exception.  On September 16 we're putting our whole story together in one place, at Dragonfly Ascent. Register now to hear to the latest: - What three years of hardening led to: 60% more throughput, 30% less memory, and the end of OOM-on-snapshot - See how our customers are deploying Dragonfly in production, and what they’ve learned along the way - Where Dragonfly is headed as feature stores, embeddings, and agent traffic push workloads into next phase of data consumption Register today: hubs.la/Q04rJGGY0 See you September 16th.
3
11
482
Roman Gershman 🇮🇱 retweeted
The Redpoint InfraRed 100 is now live. These are the companies building the infrastructure that powers everything happening in AI right now, from world models and agent runtimes to the sandboxes, databases, and security tools agents depend on. Congratulations to this year's honorees! Read the full 2026 InfraRed Report: our state of the union on AI and cloud infrastructure 👉 redpoint.com/reports/the-inf…
23
54
28
328
161,838
Dragonfly v1.38 is out 🚀 The big one: **up to 26% memory reduction** for TTL workloads! We eliminated the separate expire table entirely and embedded expiry directly into each key. 10M keys: 900 MB → 665 MB. No config changes needed.
2
156
Everyone talks about AI writing code. Nobody talks about the harder part: debugging. Last week, I spent 12 hours hunting a segfault in Dragonfly during S3 backups. Here’s how I used AI as a forensic partner to solve a "ghost in the machine" memory corruption bug on ARM64. 🧵
1
1
212
The Root Cause: A rare execution path where the S3 library combined with DNS resolution consumed more stack space than allocated. On ARM64, this overflow was silent until the corrupted memory was accessed by the next coroutine.
1
103
I couldn't have solved this alone because the forensic work required ARM64 memory layout depth I don't have. The AI couldn't have solved it alone because it lacked the system-level intuition to pivot away from its own hallucinations.
1
52
Europe is in hibernation mode and it's both scary and embarrassing. And unfortunately, it's not clear if it will be able to evolve to save itself in time.
Dear Mr. Armin Papperger, CEO of Rheinmetall, When you referred to Ukrainian drone manufacturers as “Ukrainian housewives with 3D printers” you revealed just how deeply the European defense establishment still fails to understand the nature of modern warfare. This is not about emotion. It is about battlefield reality. Here are the facts your industry refuses to acknowledge: In 2025 alone, Ukrainian drones carried out 819,737 confirmed strikes. They caused 90 percent of all Russian combat losses, more than all other weapons systems combined. TAF alone produces up to 100к FPV drones monthly. In any given 90-day period, my company’s products alone achieve more confirmed strikes than your entire fleet of equipment has across its full combat history in every conflict. And most importantly, I built this company and achieved these results in two years, not fifty. Think about that. Our drones generate more kinetic effect in three months than your flagship platforms have in half a century. Why? Because the battlefield has changed, and your business model has not. •Russian electronic warfare has made GPS-guided Western munitions such as Excalibur and GMLRS nearly ineffective. •Expensive and complex systems designed for wars with air superiority and traditional peer-to-peer combat have become easy prey for drones costing $500, attacking them from above. •The cost-to-effect ratio has been turned upside down: one 120 mm Rheinmetall shell or one anti-tank missile costs more than a dozen of our drones, and yet our drones still win. This is not a “Lego game.” It is industrial Darwinism in real time. We iterate every week. We print parts in basements and ship 100к strike systems per month, while your engineers still require three to five years and hundreds of millions of euros in certification costs for even a minor upgrade. The war in Ukraine is not a temporary anomaly. It is the first true drone-industrial war. And it has already proven that outdated European platforms, no matter how expensive or “serious” they may seem, are becoming less and less relevant unless they integrate the very technologies you mock. So when you say, “this is not innovation,” I hear something else: “We do not want to admit that the future is being written in Ukrainian workshops, not in Düsseldorf boardrooms.” #MadeByHousewives is trending for a reason. Because these “housewives” destroy more enemy equipment every month than entire European armies do in full campaigns. And they do it while your industry continues to sell 20th-century solutions at 21st prices. The invitation remains open, Mr. Papperger. Stop laughing at the kitchen table. Come and learn how tomorrow’s war is actually being fought. Because the next time someone asks, “Who needs tanks in the age of drones?”, the answer may be simpler than you think: Whoever still believes in 1979 will lose to whoever is building in 2026. With respect, but with facts, Oleksandr Yakovenko “Ukrainian housewives” Founder TAF
108
Two weeks ago I shared how much memory popular job queues consume in Redis. I said we were working on something at @dragonflydb. Here's what we built. 🧵
I got curious: how much memory do popular job queues actually use in Redis? I benchmarked @celeryorg, @laravelphp, Sidekiq, RQ, Dramatiq, and others — 200K jobs across 10 queues each. The differences are massive. 📷
1
166
Not everything compresses equally. RQ, Dramatiq, Taskiq — minimal gains. Different data structures, less redundancy. This is pattern-dependent. But for JSON-heavy list workloads, the savings are dramatic.
1
47