P99 CONF 2025 Recap: Latency to LLMs

Another year has flown by — in the blink of an eye, I found myself back in the US hosting P99 CONF again. This makes it my third time on stage and the fifth in this incredible series of talks from engineers around the world, all sharing their stories about chasing that elusive P99 latency. Beyond raw speed and tail latency, we explored modern systems programming, kernel innovation, databases and storage at scale, observability, testing, performance insights… and of course, this year’s big wave: artificial intelligence, with vector search and LLMs taking center stage. In this blog post, I’ll help you chart a course through it all (though honestly, every session is worth your time). Watch 60+ P99 CONF Talks On Demand Starting with low latency “Taming tail latency” has been a running theme at P99 for years, and this time PayPal and TigerBeetle both took the stage to show how they deal with unpredictable outliers that wreck user experience. PayPal showed how tiny inefficiencies multiply under load. And (as we’ve come to expect from the past optimization talks), TigerBeetle engineered determinism, single-threaded scheduling, predictable I/O, and batching everywhere. ScyllaDB joined the party with their low-latency vector search engine, proving that AI queries don’t have to experience high latency. If your database architecture already handles the long tail at the storage layer, your vector workloads inherit that speed for free. And since I can’t resist some speculative trading (Aussies love a bet), I enjoyed the talks from Maven Securities on lock-free queues for trading systems and Bloomberg on building scalable, end-to-end latency metrics from distributed traces. If you’re chasing nanoseconds instead of milliseconds, check out Steve Heller’s Design Considerations for P99-Optimized Hash Tables. Rust was everywhere (again) Turso is rewriting SQLite in Rust. ClickHouse tried converting 1.5 million lines of C++ … or at least part of it. Neon rebuilt its I/O stack with tokio and io_uring, Datadog squeezed out extra juice from Lambda extensions in Rust, and Trigo visualized async abstractions (also Rust). Maybe we’re the unofficial Rust conf? But Go and C++ held their ground too. Miguel Young de la Sota presented a faster protobuf implementation in Go, and Manya Bansal (MIT PhD student) gave a highly detailed talk on how not to program GPUs with some C++ insights there. On the database and storage side Avi Kivity kicked off the conference by sharing how ScyllaDB’s Seastar CPU scheduler prioritizes complex workloads while keeping latency predictable. Nadav Har’El showed how ScyllaDB manages client ingestion to prevent memory blowouts, Dor Laor connected theory to real-world tail behavior, and Felipe explained how tablet replication delivers true elasticity. And Andy Pavlo took us through the very real challenge of both humans and autonomous systems tuning databases. Other databases showed up strong too. Turso is pushing SQLite into new dimensions, DragonflyDB nailed sorted sets with B+ trees, and Qian Li from DBOS shared how they merged app logic and state right into the database. These are all good reminders that performance starts at the data layer. Performance engineering, testing, and observability My old mate Ashutosh Agrawal shared how he built and tested systems for 32 million concurrent cricket fans. Also on testing, I loved the double feature on deterministic simulation testing: Resonate’s approach to catching Heisenbugs and Antithesis running fuzzing workloads at scale in near real time. eBPF continues to fascinate. Cosmic elaborated on reliability versus memory trade-offs, and Tanel Poder showed thread-level observability with some seriously impressive tooling. AWS’s Geoff Blake introduced aperf for profiling the nitty-gritty, Arm unlocked new insights with the PMUv3 plugin, and Raphael Carvalho took us on a wild hunt through a Linux kernel bug triggered by io_uring. Proper detective work, that one. And then there’s AI and ML, the new frontier Chip Huyen delivered a standout keynote on LLM inference optimization, tying together hardware, software, and architecture choices. Eshcar Hillel went deep into KV cache offloading, Microsoft’s Magdalen Manohar showed how to make vector search cost-effective and fast, and ScyllaDB’s Pawel Pery explained how we decoupled a Rust-based vector engine for high-performance ANN queries. We wrapped it up with a cracking conversation between Rachel Stephens from RedMonk and Adrian Cockcroft, exploring AI-assisted analytics and that eternal hunt for predictability at the tail. Hosting this conference is always a privilege. Big thanks to Natalie Estrada, who keeps the whole thing running (and somehow curates killer playlists). Cynthia Dunlop, who finds the best speakers and replies to messages faster than anyone I know, and Felipe and the lounge crew, daring the demo deities and answering audience questions live. Plus the many ScyllaDB engineers who present, coordinate, and quietly make it all happen behind the scenes. And to all 30,000 registrants and 100,000+ participants so far… You’re what makes this community so special. From all of us at ScyllaDB: a huge thank you. See you at the next one. Join is for P99 CONF’s sister conference, Monster SCALE Summit