Virtual Event | OCTOBER 23 + 24, 2024
Explore Rust, C++, Go, event streaming architectures, distributed databases, Linux kernel, observability, K8s & more
Follow us on Twitter @p99conf for announcements!
Chip Huyen, author of AI Engineering and Designing Machine Learning Systems (O’Reilly) shares her tips on how to make the most out of working with multiple agents.
I'm Chip Huyen, a writer and computer scientist. I'm building infrastructure or real-time ML. I also teach Machine Learning Systems Design at Stanford. Previously, I was with Snorkel AI, NVIDIA, Netflix, Primer, Baomoi.com (acquired by VNG). I helped launch Coc Coc - Vietnam’s second most popular web browser with 20+ million monthly active users. In my free time, I travel and write. After high school, I went to Brunei for a 3-day vacation which turned into a 3-year trip through Asia, Africa, and South America. During my trip, I worked as a Bollywood extra, a casino hostess, and a street performer. I’m the author of four bestselling Vietnamese books. I’m working on an English book on machine learning interviews.
The full story of rewriting a feature flag service from Django to Rust at 500k req/min – the decision, the architecture, the results, and what happened after.
Dylan leads the feature flag team lead at PostHog. He loves typed languages, skiing, jazz piano, producing music, and powerlifting.
How do you search 100 billion vectors – over 200 TiB of dense vector data on object storage – with a p99 latency of just 200ms at >1k QPS? We’ll break down how turbopuffer’s vector search index combines hierarchical clustering with RaBitQ binary quantization to enable consistently fast vector search at web scale.
Nathan VanBenschoten is the Chief Architect at turbopuffer, where he works on fast, cheap, scalable vector search. Before joining turbopuffer, Nathan was a principal engineer at CockroachDB, where he led transactions and replication.
Avi Kivity, CTO of ScyllaDB, is known mostly for starting the Kernel-based Virtual Machine (KVM) project, the hypervisor underlying many production clouds. He has worked for Qumranet and Red Hat as KVM maintainer until December 2012. Avi is now CTO of ScyllaDB, a company that seeks to bring the same kind of innovation to the public cloud space.
Gunnar Morling is an open-source software engineer in the Java and data streaming space, currently working as a Technologist at Confluent. Previously, he helped to build a realtime stream processing platform based on Apache Flink and led the Debezium project, a distributed platform for change data capture. He is a Java Champion and has founded multiple open source projects such as Hardwood, kcctl, JfrUnit, and MapStruct. Gunnar is an avid blogger (morling.dev) and has spoken at various conferences like QCon, Java One, and Devoxx. He lives in Hamburg, Germany.
Rust took over infrastructure. Now Topcoat brings its speed and reliability to web apps, turning infrastructure engineers into full-stack developers.
Carl Lerche is a Principal Engineer at AWS. He is best known for his open source Rust libraries, primarily Tokio, the asynchronous I/O runtime for Rust.
This talk covers a production A/B test of HTTP/2 vs HTTP/3 across an 18-market food delivery platform serving millions of users daily. Results were mixed: up to 17% faster p90 response times and 48% fewer network-level failures, but 4%+ throughput degradation in several markets, traced through packet-loss correlation and per-market segmentation. We’ll also cover underreported QUIC operational issues—NAT timeout connection churn and 800ms p90 Cronet init latency that dilutes 0-RTT gains on Android—and a framework for evaluating HTTP/3 readiness.
Elvis Chidera is a staff software engineer at Delivery Hero, based in Berlin. He has previously worked for Careem (acquired by Uber) and for a few startups in the USA. His expertise is in mobile app development, but he has worked across the typical software stacks.
Every Lovable prompt spins up a sandboxed dev environment, and at 50M+ apps, provisioning time and stability become the P99 that defines the product. We’ll cover a year of squeezing latency, cost, and failure out of that path: from third-party sandboxes to a Cloudflare Durable Objects scheduler to our current setup—ten GKE clusters, pre-warmed pod allocation via single JSON PATCH, Envoy/xDS routing—including the outages, ReplicaSet internals that stall refills, and validation at 120k pods and 150 pods/s with sub-second, zero-failure acquisition.
Jonathan Grahl is an engineer at Lovable, where he works on their production infrastructure. He joined through Lovable's acquisition of Molnett, the European cloud provider he founded. Previously, he built game infrastructure at Embark Studios (ARC Raiders) and worked on the team that built Candy Crush's private cloud.
A deep dive how Valkey thinks about balancing vertical and horizontal scaling for in-memory database and caching systems.
I work primarily on the open source Valkey project and evangelize the importance of open source software development.I work primarily on the open source Valkey project and evangelize the importance of open source software development.
Histograms show aggregate latency but not causality; traces show causality but are expensive at scale. Tracegrams fill the gap: a lightweight primitive where requests carry latency state through transition histograms at each checkpoint, showing where p99 is born and amplified with minimal overhead. This talk covers a Rust prototype, synthetic stress tests against known ground truth, and nanosecond-per-checkpoint overhead numbers.
Ivan Goncharov is a performance engineer with 18 years of experience in systems debugging, benchmarking, and observability. He leads PerfLab at Azul Systems, building reproducible JVM benchmarking, regression detection, and large-scale performance analysis infrastructure. Previously, he contributed to PostgreSQL scaling to 1M QPS on a single machine at IBM and led performance testing and infrastructure projects for enterprise payment processing systems.
Coding agents are collapsing the build-vs-buy tradeoff. Instead of buying a library, framework, or platform, you can build one on demand.
The recipe: place the agent inside a closed loop of specification, verification, and optimization. Specification defines the target system, verification guarantees correctness, and optimization guarantees performance.
In this talk, we show how that loop guided an agent from prompt to a production-grade implementation of Resonate on ScyllaDB.
Dominik Tornow is CEO and founder of Resonate HQ, Inc. Dominik focuses on correctness, reliability, and scalability of distributed systems. He is the author of Thinking in Distributed Systems.
NERSC’s Perlmutter uses LDMS for sub-second, near-zero-overhead performance monitoring across its shared network, Lustre, and Spectrum Scale filesystems. We’ll present a userspace daemon that feeds eBPF program output into LDMS for kernel-level metric collection at scale, plus a companion eBPF filesystem latency monitor for real-time visibility into Perlmutter’s shared filesystem performance.
Ershaad Basheer is a HPC Systems Engineer at the National Energy Research Scientific Computing Center (NERSC) in the Lawrence Berkeley National Laboratory. He is part of the group that supports the Lab's largest supercomputing system, Perlmutter. While managing a system such as Perlmutter requires generalist knowledge and skills, Ershaad is particularly interested in debugging and observations at the kernel level. Learning tools such as the drgn debugger and exploring Rust and eBPF are the kind of things that he finds captivating. Ershaad graduated with a Bachelors in Electronics and Communication engineering from Kannur University in India and has a thesis-track Masters in Science degree from the Jawaharlal Nehru Centre for Advanced Scientific Research in Bengaluru, India. He previously was an HPC engineer at Temple University, Philadelphia where he also taught a graduate course in programming.
Everybody knows that AI can code. It can also convert your code to Rust. But can AI write a database? In my session, I’ll share my journey of “vibe coding” the ScyllaDB cache to implement an alternative LRU algorithm and improve our p99 latency. Vibe coding a mature database is very different from vibe coding business logic: intrinsic details like multi-version concurrency control and partitions spanning thousands of rows confuse humans and bots alike, but for those with patience, the payoff is real.
Dor Laor is the CEO of ScyllaDB. Previously, Dor was part of the founding team of the KVM hypervisor under Qumranet that was acquired by Red Hat. At Red Hat Dor was managing the KVM and Xen development for several years. Dor holds an MSc from the Technion and a Phd in snowboarding.
A browser does everything on one main thread, so while your code runs a long task, every tap and keypress has to wait. This talk shows why that makes pages feel slow, and how to keep the main thread free so they stay responsive.
Den Odell is a Staff Frontend Engineer, previously at Canva, Volvo Cars and AKQA, with more than 25 years building for the web. He is the author of Fast by Default (Manning, 2026), which lays out his performance-first approach to building software across web, backend, mobile, and desktop. He has written two earlier books on frontend development and writes about performance and engineering practice at denodell.com.
KERNEL builds crazy fast, open source infra for AI agents to access the internet. To hit our sub‑30ms browser boot times, we evolved our stack from containers to unikernels to microVMs. This talk walks through what we learned with each iteration, and what it actually takes to make cloud browsers not just fast, but crazy fast.
Catherine Jue is the Co-Founder of KERNEL.
- Compiler Engineer at Microsoft, more specifically working on the Triton to MAIA compiler - Enthusiast of fast software - Interested in Compilers, Algorithms, C++, Reflection, Efficient Data Structures, Low-latency, Optimization, LLM Inference, Profiling, Performance and Game Theory - Contributed to a few open-source projects like simdjson, mirror_bridge, karger-klein-tarjan
io_uring is hot but hard to make work well in production. I will show you how I revamped a production database system based on Arrow and DataFusion to get 14GB/sec read speeds end to end, in Rust, using compio + Tokio. We will go over I/O APIs, HUGETLB and kernel page faults, why memory and scheduling are the true bottlenecks.
Evan has been building high performance data systems for 20+ years. He is working on custom databases and streaming systems serving trillions of events per day, using Rust, Apache DataFusion, Arrow, Tantivy, etc. He led a team developing FiloDB, an open source distributed real-time time series database, used as a production telemetry DB at Apple. He has architected, developed, and productionized large scale data and telemetry systems at companies including Apple, and loves solving the most challenging technical problems at both large and small scales, from advanced custom data structures to distributed coordination. He has led the design and implementation of multiple big data platforms based on Rust, DataFusion, Arrow, Spark, Kafka, Cassandra, and Scala/Akka. He has contributed to Spark and DataFusion/Arrow, and was a two-time Datastax Cassandra MVP.
Large Language Models (LLMs) have become core components of modern AI/ML applications – from GenAI to agentic workflows. At the core of this is LLM inference: the process of using a trained large language model to generate text, code, images, embeddings, etc. in response to user prompts. Real-time, multi-user applications of it demand low latency, high throughput and optimized inference. Scalability and performance efficiency are thus of utmost importance in model serving and LLM inference in large-scale deployments. Though this session, we explore scalability of LLM inference along with importance of monitoring the characteristics which affect its performance as bottlenecks.
Simran is a Senior Software Engineer at Ampere Computing developing benchmarks and frameworks to evaluate, scale performance of LLM and ML workloads on current and future aarch64 based compute platforms. She works on identifying bottlenecks, workload characterization, and driving performance improvements in software stacks. She is solving problems at the intersection of software, infrastructure and hardware - pushing the limits of cloud compute!
A customer reported “the network is slow.” Using bpftrace on their multi-layer CDN, we traced random throughput collapses from multi-Gbps to a few Kbps to a memory cgroup reclaim path that flips sockets into TCP memory pressure mode on reclaim failure—invisible because the existing SNMP counter never fires on that path. This talk covers the debugging journey and the kernel patch we upstreamed, exposing the state via a new sock_throttled counter in memory.events since kernel 6.19.
I am a PhD student at the Czech Technical University in Prague, where my main research topic is improving content delivery. Additionally, I work as a Performance Engineer at CDN77, where my job is to address all performance-related issues, ranging from debugging kernel drivers to improving the caching capabilities of the CDN.
This talk explores the “Isolates and Effects” pattern for achieving multi-core concurrency in non-GC languages (C/Zig/Odin) without async/await coloring, thread/mutex contention, or Actor-model GC overhead. Programmers write synchronous-looking handler functions that perform state transitions, mapped onto a shared-nothing, thread-per-core architecture for Actor-like fault tolerance with native performance.
Peter Mbanugo is a Software Engineering Consultant at IT Services Peter Mbanugo.
Binary heaps cache-miss on nearly every pop. This talk covers SimdQuickHeap, which revisits QuickHeap and partitions data QuickSort/QuickSelect-style instead of storing a tree. It gets good I/O-complexity and a data layout that vectorizes with SIMD, making it up to 10x faster than binary and d-ary heaps and 2x faster than radix heaps, with the gap widening as n grows.
Ragnar is a postdoc in high throughput bioinformatics at the Karlsruhe Institute of Technology. His work is on writing high performance libraries for fundamental tasks in bioinformatics, such as minimal perfect hashing, and designing algorithms and datastructures that are optimized for modern hardware. Ragnar is author of the blog curiouscoding.nl.
This talk explores how AI agents can help optimize code through an automated loop of code changes, benchmarking, profiling, and validation with a pros and cons analysis.
Distributed computing and storage engineer for HPC, microservices and ML/Data platforms for many years. Evangelist of type-driven development and category theory for developers.
Kernel-bypass networking assumes NIC features the cloud won’t give tenants. Machnet is a userspace stack that hits near bare-metal tail latency inside cloud VMs using only the minimal vNIC feature set every provider supports – via a shared-memory sidecar and “RSS–,” a trick that probes port combinations to steer flows through the cloud’s opaque, unconfigurable RSS hashing. I’ll cover the design and the tradeoffs.
I am a Ph.D. Candidate in Computer Science at Columbia University, advised by Kostis Kaffes. I focus on hardware-software co-design, kernel-bypass networking, and low-latency systems optimization - squeezing everything out of whatever hardware I'm given, whether that's a CPU, NIC, GPU, or TPU. I previously worked at Microsoft Research (Networking Research Group) and Google (TPU Performance). My research has appeared at ASPLOS, NeurIPS MLSys, and VLDB.
In high-performance distributed systems, every delay at the message streaming layer compounds across the entire platform. When services communicate through a message stream, tail latencies don’t stay isolated. They stack, they amplify, and they eventually surface as degraded end-to-end response times that are nearly impossible to trace back to a single root cause. And when you can’t fix the latency, you scale out instead, throwing more hardware at a problem that’s fundamentally architectural. A few extra milliseconds here, a GC pause there; it adds up fast. The hardware we have today is remarkable, but squeezing out what it’s truly capable of requires designing around it from scratch, not patching over decades-old abstractions. That’s exactly what we did with Apache Iggy, built in Rust from the ground up. Today we’ll dig into the architectural decisions that got us to ultra-low and predictable tail latencies at high throughput.
Founder & CTO of LaserData, Apache Iggy PMC, and a full-time open-source contributor. In a previous chapter, I spent over a decade in the C#/.NET ecosystem building a wide range of applications, including high-performance systems.
Owning the whole stack means owning every bottleneck. A walk up the network path of an Oxide rack, peeling back the layers of wait to optimal performance.
Steve Karam is an engineer at Oxide Computer Company, where he works on product assurance and spends a lot of his time causing problems in complex systems. He brings 25+ years in enterprise infrastructure, including time at Delphix and Perforce across solution engineering, enablement, product management, et al. Outside of work he builds guitars, roasts coffee, and writes steampunk heist novels, among many other hobbies.
Profiling usually means separate tools for CPU time, GPU shaders, and application telemetry. This talk covers Stax, which combines kdebug/kperf, Metal 4 heap counters, and application self-reporting into a single CLI or web interface for engineers and AI agents alike.
I write (and shoot) deep dives about Rust, performance, operating system stuff, videography, and whatever else I feel my wonderful patrons & sponsors will like learning about. I'm extremely lucky to do this full time for now.
*that we know of
TigerBeetle needed object storage. The official SDKs don’t compose with our completion-based event loop or Deterministic Simulation Testing — so we built everything from scratch in Zig: custom DNS, TLS, and HTTP, with static allocation, zero-copy io_uring, and a single thread. The result is 4–10× faster than the official GCP and AWS SDKs. This talk covers the architecture, DST infrastructure, and what it looks like to apply TigerStyle to cloud I/O from first principles.
Georg Kreuzmayr is a Software Engineer at TigerBeetle, where he works on database internals and low-level systems infrastructure. His work spans high-performance networking, deterministic simulation testing, and low-level protocol implementation. He is based in Munich / Germany.
Jakub Czyszczoń graduated with a Bachelor's degree in Computer Science from the University of Warsaw and is pursuing a Master's degree. He is particularly interested in operating systems.
This talk is about a loop that does the full arc of performance engineering end to end: discoveries from runtime evidence, cases as a triaged queue, verdicts that travel with the PR, and knowledge that compounds. It’s a field report.
Tomás Senart wrote Vegeta, the open-source Go HTTP load tester, and presented its design at dotGo. He is the founder of Perfloop. Most recently, as Principal Engineer at Axiom, he scaled EventDB to petabyte workloads across ingestion, compaction, and interactive analytics. He built Haydex, a filtering system that scaled to 178 billion rows per second, enabling trillion-row searches over PB-scale datasets, and led performance work across ingestion and query paths. Before Axiom, four years at Sourcegraph as Staff Engineer, leading the backend and Cloud teams and contributing to scaling sourcegraph.com search across 5 million open-source repositories with sub-second P90 latency. Fifteen-plus years on distributed systems performance, from load generators to large-scale database internals. - GitHub: https://github.com/tsenart - LinkedIn: https://www.linkedin.com/in/tsenart/ - X / Twitter: https://x.com/tsenart
A practical C++23 talk on compile-time perfect hashing for ultra-fast string lookups. Learn how a single-header approach reaches 1B+ queries/second, beats common hash-table baselines, and reduces tail latency in real hot-path parsers and routers.
Daniel Lemire is a full professor of computer science at the Université du Québec (TELUQ) in Montreal. He specializes in software performance, SIMD vectorization, data indexing, parsing, and high-speed data engineering. He ranks among the top 2% of scientists worldwide by citations (Stanford/Elsevier 2025 ranking) and is one of GitHub’s top 1,000 most-followed developers. His open-source contributions have been widely adopted by major systems, including:
Recent Computer Science graduate from the University of Warsaw. Interested in high-performance and distributed systems. Currently working at ScyllaDB.
Szymon Wasik is the Director of Software Engineering at ScyllaDB.
Salesforce’s Tracer platform ingests 300 billion trace spans daily across a federated Elasticsearch deployment, but raw spans are a poor unit of analysis for incident triage. This talk shares how we shifted to trace-level analysis—reconstructing span trees to pinpoint the service responsible for latency, backed by an S3/Parquet caching layer that keeps 84% of queries off Elasticsearch and cuts query times 60-70%.
Seasoned technologist with over 18 years of profound expertise in crafting scalable and resilient distributed systems that handle petabytes of unstructured data daily. Spearheaded the architecture and design of various cloud-native SaaS solutions. His diverse industry exposure encompasses E-commerce, Embedded systems, and Telecom sectors. Sudeep has taken the stage at several international conferences in countries including Japan, the USA, the UK, and India. He is a primary author of various papers featured in IEEE and other international conferences. Sudeep also holds multiple patents in the realms of payments, e-commerce, and distributed systems. Sudeep's technical forte lies in addressing Big data challenges, Monitoring/Telemetry, developing cloud-native platforms and frameworks, client-side programming, and creating scalable server-side backends.
Turso is our Rust rewrite of SQLite. It replaces the single-writer transaction model with row-level MVCC—snapshot isolation, optimistic BEGIN CONCURRENT transactions, commit-time certification—plus an io_uring async I/O stack for non-blocking durability. We’ll cover the architecture, the hard parts (checkpointing, conflict handling, SQLite compatibility), and the resulting p99 write latency under concurrent load.
Pere Diaz is a Senior Software Engineer at Turso.
A decades-old, hidden software bug in Java caused severe instability on large, modern servers running cloud applications. The issue stems from a flaw in how Java manages its internal processes, leading to slow performance and crashes.
This talk reveals how we used specialized tools to find this complex problem and provides essential tips for running reliable, high-speed Java applications on cloud platforms.
Francesco Nigro is a Senior Principal Software Performance Engineer at IBM.
Not every workload fits the same storage shape. While LSM trees optimize for write throughput and range scans, frequently overwritten data and KV workloads require significant compaction I/O to minimize read amplification in a timely fashion. This talk explores how ScyllaDB built a new storage engine to eliminate LSM traversal entirely, enabling a single deployment to handle both wide-column and key-value workloads without compromise. We’ll cover the motivations behind the design, the tradeoffs it explicitly inherits, and early performance results.
Michael Litvak is a Senior Software Engineer at ScyllaDB.
Two services on the same host still pay the full TCP/IP stack just to talk. tcp_splice is a Linux kernel module that transparently bypasses the packet path for co-located connections, halving local-hop latency and multiplying throughput with no application changes. Load this module, attach a cgroup policy, done.
Cong Wang is a professional Linux kernel developer mainly focuses on AI agent infrastructure, he founded Multikernel Technologies and Agentry. He has contributed over 1000 patches to the Linux kernel project and many other open source projects.
Quantized models are smaller and faster than full-precision models. But what actually is quantization? What’s being done to the model to make it smaller? And how can we quantify the damage this process does to a model? In this talk I’ll cover the mathematical process applied to models to make them smaller, and introduce you to tools you can use to gauge how much a quantized model differs to its full-precision original.
Sam Rose is a Developer Educator at ngrok creating deep, visual, and interactive content aimed at helping developers truly understand the topics they're reading about.
A new cohort of Rust profiling tools—Hotpath, Tokio DTrace, dial9—has emerged in the last 12 months, each taking a different approach to instrumenting async and sync code in production. This talk compares their tradeoffs (privilege requirements, overhead, deployment constraints) and shows how AI-assisted analysis, via Hotpath’s MCP server and dial9’s built-in skills, can find bottlenecks faster than a flamegraph.
Hayden is an engineering manager at HERE Technologies, focused on serving real-time traffic information to end users in great detail at low latency (increasingly using Rust). He’s a passionate Rust developer and is a core member of the tokio-rs maintainer team and the defacto maintainer of the Tracing and Tokio Console crates. GitHub: https://github.com/hds Mastodon: https://hachyderm.io/@hds BlueSky: https://bsky.app/profile/hds.cafe
We explore how the new asymmetric io_uring Seastar backend can offload I/O operations from application cores to dedicated networking cores.
Marcin Szopa is a software developer and recent Computer Science graduate from the University of Warsaw, where he is continuing his studies to pursue a Master's degree. He recently completed his bachelor’s thesis in close collaboration with ScyllaDB. Professionally, Marcin works at Foss Polska as an intern, has a rapidly growing interest in Rust, and thrives on collaborative, hands-on problem-solving.
In this talk, we explore what happens when ScyllaDB clusters face connection storms and request overload. We’ll look at real-world failure scenarios, the risks hidden underneath the surface, and the techniques used to keep latency predictable under pressure – from memory management and caching to throttling and prioritization.
Marcin Maliszkiewicz is a Technical Lead, Senior Software Engineer at ScyllaDB.
Open any textbook on databases and you’ll quickly learn that databases are slow because they write to physical disks. But with the latest NVMe drives, correct amortisation patterns, Direct I/O and strategic use of fsync, a well designed write pipeline is actually CPU bound. We’ll walk through the optimisations in Celeriant – an open-source event sourcing database built with io_uring, thread-per-core, Rust, and kernel-offloaded TLS – that push durable, replicated writes to more than 500k/sec; 15x Kafka on identical EC2 hardware.
Tyson Brown is a Lead Consultant at Thoughtworks with 20 years in enterprise systems, from MRI machines to financial platforms. He built Celeriant - an open source Rust event sourcing database - after watching his clients struggle with event sourcing patterns.
Kubernetes P99 latency can burn while CPU, packets, and dashboards look fine. This talk shows how hidden TCP queues inside pods can trap small gRPC or HTTP/2 requests behind bulk traffic, defeating user-space prioritization after sendmsg(). Using eBPF, we expose unsent kernel bytes, connect them to tail latency, and explore how pod-level sysctls like net.ipv4.tcp_notsent_lowat can restore backpressure while revealing surprising risks in “safe” platform defaults.
Linux, Networking, and Go. It’s a somewhat unconventional way to introduce myself, but one that closely reflects who I am. These are the areas that capture my interest in understanding how systems work and what happens beneath their abstractions. That curiosity also extends to the community, where I’m involved as a CNCF Ambassador and an organizer of CNCG in Peru, in addition to having been part of KCD Lima. Professionally, I work as a Platform Engineer.
Computers aren’t getting faster, they’re getting wider. It’s time to rethink basic assumptions about computer performance.
Jared started optimizing Linux in the 20th century. He's hacked Linux phones before Android, SSDs before NAND, PCIe storage before NVME, pmem before Optane, filesystems before they were cool, SQL before Postgres won, and AI before ChatGPT. Now he's excited to see the future of computing he worked for become the present.
Netflix built an autonomous agent that hunts performance inefficiencies in live production services, traces them to source, proposes fixes, and validates results through canary deployment—grounded in measured outcomes, not model confidence. This talk covers what it took to make the agent trustworthy enough to act unsupervised, including a real fix it caught and confirmed with no ticket, oncall, or engineer in the loop.
Rajat is a Staff Software Engineer at Netflix, leading the technical architecture for the global ML Model Serving Infrastructure. With nearly a decade at Netflix and Amazon, he has specialized in building highly available distributed systems and stable, usable ML platforms that power personalization, search, and commerce at scale. A recipient of Amazon’s prestigious "Just Do It" Award from Jeff Bezos for his bias for action, Rajat excels at abstracting the complexities of distributed computing to drive developer velocity and platform reliability. He holds a Master's in Computer Science from North Carolina State University, and has spent his career applying that foundation to massive-scale infrastructure challenges.
Karol Nowacki is a Senior Software Engineer at ScyllaDB
Every Lovable prompt spins up a sandboxed dev environment, and at 50M+ apps, provisioning time and stability become the P99 that defines the product. We’ll cover a year of squeezing latency, cost, and failure out of that path: from third-party sandboxes to a Cloudflare Durable Objects scheduler to our current setup—ten GKE clusters, pre-warmed pod allocation via single JSON PATCH, Envoy/xDS routing—including the outages, ReplicaSet internals that stall refills, and validation at 120k pods and 150 pods/s with sub-second, zero-failure acquisition.
Adrien works on sandbox infra at Lovable in Stockholm. Before that, infra at Hugging Face. Before that, Packer at HashiCorp, back when the templates were JSON. He made containerd pull images much faster. Somewhere somebody is watching a spinner, waiting for a sandbox to come up. He believes that spinner should never be seen.
Rama Teja Repaka is a Principal Software Engineer at Palo Alto Networks with over a decade of experience building large-scale distributed systems. His work spans query processing, search infrastructure, streaming data platforms, and cloud-native systems. He has contributed to the evolution of the Resource Query Language (RQL), developing query parsers, search and indexing capabilities, execution engines, and performance optimizations for complex queries. More recently, he has been building petabyte-scale data ingestion and stream processing systems using Apache Flink, Kafka, Kubernetes, and Java. His interests include distributed systems, information retrieval, stream processing, and AI-powered developer tools.
At Palo Alto Networks, broad security queries against hundreds of millions of cloud resource records produced high response times which lead to lower throughput. Traditional fixes — sharding, materialized views — all require schema changes. This talk presents HSSPS (Heuristic Search Space Partitioning System): a query-time optimization layer that logically partitions search space without touching the database schema. Deployed in production and covered by two granted U.S. patents, HSSPS reduced P95 latency by over 95%, increased throughput 8–10×, and cut infrastructure cost by ~50%. A reusable pattern for any multi-tenant system facing the same problem.
Prashant Pathak is a Senior Principal Software Performance Engineer at Palo Alto Networks, where he leads query optimization and performance architecture for large-scale cloud security infrastructure. He holds two granted U.S. patents in distributed query optimization with production deployments achieving over 95% query latency reduction and 8-10x throughput improvement. Prior to PAN, Prashant was a Performance and Scalability Engineer at Apple, where he was responsible for multiple Apple Pay transit launches across major cities in Asia and North America, as well as performance infrastructure for Apple Card. Prashant's work spans query optimization, distributed systems performance, cloud infrastructure observability, and machine learning for database cost prediction.
go tool trace’s web UI gets overwhelming for long-running or highly concurrent applications. This session covers parsing execution traces programmatically with golang.org/x/exp/trace and the newer tracing APIs: consuming .out files, extracting custom metrics like goroutine scheduling latency and GC impact that don’t show up in the default UI, and building custom tooling to find bottlenecks at scale.
Cristian Velazquez is a Staff Site Reliability Engineer on the Maps Production Engineering team at Uber. He works on multiple efficiency initiatives across multiple organizations. He leads the GC tuning efforts across the company.
Aurora DSQL is a PostgreSQL-compatible distributed database that pays cross-region coordination cost exactly once — at commit — not per statement. This talk walks through the three architectural bets that make that work: per-transaction Firecracker isolation, optimistic concurrency control, and Amazon Time Sync. Then it traces a real multi-statement transaction end-to-end, showing what each millisecond of commit latency buys you and what drives P99 under contention.
Raluca Constantin is a Database Engineer at AWS, working on Amazon Aurora DSQL's distributed SQL solutions for cloud-native environments. With nearly two decades of database experience, including as an Oracle Certified Master, she specializes in solving complex scalability and multi-tenancy challenges. At AWS, she combines deep technical expertise with customer focus to implement cloud-native database solutions, while actively contributing to the technical community.
Chronon is an open-source feature platform serving billions of predictions daily at Stripe, Airbnb, Netflix, and OpenAI, balancing sub-second streaming freshness with under-50ms serving latency over long lookback windows. This talk walks through its Lambda architecture and the optimizations that make it work at scale—tiling, mega/giga-tiles, sawtooth windows, and pushing batch results into Flink state.
Piyush Narang presently the Chief Architect at Zipline.ai and is a PMC member of the Chronon open source project. Previously he worked as a Staff Engineer / Tech Lead of Stripe's Shepherd ML feature computation platform (https://stripe.com/blog/shepherd-how-stripe-adapted-chronon-to-scale-ml-feature-development). At Stripe and prior roles (Criteo, Twitter, Amazon), he has helped build data and ML services and libraries at extremely low latencies, high data scale and with the operational rigor to support a critical part of the company's business. His interests lie in solving hard problems at the intersection of infrastructure, big data and machine learning.
Seastar’s RPC has always been bound to TCP, inheriting head-of-line blocking and a separate TLS handshake. QUIC’s independent streams and built-in TLS 1.3 remove these problems, but most QUIC libraries introduce the locking and context-switching that Seastar was designed to avoid. We built a QUIC transport for Seastar on top of ngtcp2’s sans-I/O state machine, then adapted RPC to it twice: 1) as a one-to-one socket replacement, and 2) a QUIC-aware approach that opens a fresh stream per call. Join us to learn what it takes to drive a callback-driven protocol engine from a coroutine-based reactor, where the costs come from, and the packet-loss threshold at which multiplexed streams start being worth it.
Kamil Dalidowicz is a master's student in Computer Science at the University of Warsaw and a software engineering intern on the GCE autoscaler team at Google. He's passionate about competitive problem solving.
Application Performance Monitoring and Kubernetes monitoring in their current state are pretty expensive. The average VictoriaMetrics installation is processing 2-4 million samples/s on the ingestion path, and 20-40 million samples/s on the read path. The biggest installations account for 100 million samples/s on the ingestion path. This requires being very clever with data pipelines to keep them efficient and scalable by adding more resources. In this session, we’ll explore essential optimizations to maintain database speed such as string interning, caching results, goroutine management and understanding sync.pool. Join us to learn how to optimize your database effectively.
Roman is a software engineer with experience in distributed systems, databases, monitoring, and high-performance microservices. Roman's passion is open source and he's proud to have contributions to Prometheus, Grafana, and ClickHouse. Currently, Roman is working on the open source time series database and monitoring solution VictoriaMetrics.
Discover how DoorDash decoupled the control plane to build a stateless, “clusterless” ML feature store. Learn how we leveraged Apache Kvrocks and a custom Redis Cluster Manager to achieve 50ms P999 latency at 1.6B features/second.
Luigi Tagliamonte focuses on designing and building the highly scalable, distributed backend systems that power DoorDash's global infrastructure. With a deep specialization in distributed datastores and low-latency architectures, his recent work involves completely rethinking how machine learning features are served at massive scale. When he's not optimizing ML predictions or shaving milliseconds off P999 latencies, this authentic Italian is busy passionately defending the sanctity of pizza from its greatest threat: pineapple.
Asynchronous systems improve latency and throughput, but they also introduce uncertainty. Learn practical patterns for managing in-flight state, correlating responses, and maintaining correctness in high-performance distributed systems.
Benjamin Cane is a Distinguished Engineer at American Express, where he plays a pivotal role driving the technical architecture and engineering excellence within the Global Merchant & Network Services Technology team. With a focus on payments, Ben specializes in building high performance, low latency distributed systems. His expertise has been instrumental in driving the evolution of American Express' cloud-native payments platform. Beyond his contributions to American Express, Ben is an active open-source community member and has contributed to various projects.
Not every workload fits the same storage shape. While LSM trees optimize for write throughput and range scans, frequently overwritten data and KV workloads require significant compaction I/O to minimize read amplification in a timely fashion. This talk explores how ScyllaDB built a new storage engine to eliminate LSM traversal entirely, enabling a single deployment to handle both wide-column and key-value workloads without compromise. We’ll cover the motivations behind the design, the tradeoffs it explicitly inherits, and early performance results.
Felipe Mendes is an IT specialist with over a decade of experience working with distributed systems and databases. He has co-authored several technical publications, ranging from Linux, containers, distributed systems and distributed databases. A frequent speaker at events, he works as a Technical Director for ScyllaDB.
Seastar’s RPC has always been bound to TCP, inheriting head-of-line blocking and a separate TLS handshake. QUIC’s independent streams and built-in TLS 1.3 remove these problems, but most QUIC libraries introduce the locking and context-switching that Seastar was designed to avoid. We built a QUIC transport for Seastar on top of ngtcp2’s sans-I/O state machine, then adapted RPC to it twice: 1) as a one-to-one socket replacement, and 2) a QUIC-aware approach that opens a fresh stream per call. Join us to learn what it takes to drive a callback-driven protocol engine from a coroutine-based reactor, where the costs come from, and the packet-loss threshold at which multiplexed streams start being worth it.
Piotr Korcz is a Developer at ScyllaDB.
Salesforce’s Tracer platform ingests 300 billion trace spans daily across a federated Elasticsearch deployment, but raw spans are a poor unit of analysis for incident triage. This talk shares how we shifted to trace-level analysis—reconstructing span trees to pinpoint the service responsible for latency, backed by an S3/Parquet caching layer that keeps 84% of queries off Elasticsearch and cuts query times 60-70%.
Thomas J. Varley is a senior engineer specializing in large-scale distributed systems, observability, and high-performance data platforms. With deep expertise in building resilient, cloud-native architectures, he has consistently driven the design and evolution of systems that operate at massive scale while maintaining strong guarantees around reliability, latency, and data integrity. His work spans advanced telemetry pipelines, real-time analytics, and next-generation tracing platforms, with a focus on transforming raw signals into actionable insights. Known for combining rigorous systems thinking with practical execution, Thomas brings a track record of leading complex initiatives, influencing platform strategy, and advancing the state of observability in modern distributed environments.
In this talk we’ll look at real-world cases where eBPF instrumentation introduced latency, reduced throughput, consumed excessive memory, destabilized applications or even caused crashes. We’ll examine why these problems happen, how to recognize them in production, and practical techniques for avoiding them.
Tanel Poder is a long-time computer performance geek, working on various complex systems like (Oracle) database clusters, modern Big Data & cloud technologies and anything running on Linux/Unix. He has built and fixed enterprise data systems all around the world, this has also resulted in building a few small-but-very-fun tech companies around better tools and methods. He has two patents in the data virtualization space and has realized that this is enough. In addition to his R&D, he occasionally delivers consulting, advisory and training to companies and talks about performance & troubleshooting both at public conferences and his video channels available at his website.
How do you optimize a distributed file system that manages over 500 billion files in a single volume? This talk covers critical AI scenarios: large-file sequential reads (breaking FUSE limits to reach 70+ GB/s), massive small files (scaling metadata IOPS linearly), and high-concurrency random reads (boosting from 7,000 to 20,000+ IOPS via a warm cache). Learn practical tuning strategies that perfect the last mile of AI data access. Because at P99 CONF, we know that there’s always room for optimization and GPUs should never starve.
Joe is an enthusiastic advocate for cloud-native technologies. Having served as the CTO of a dynamic start-up company, Joe excels at directing and inspiring team members to deliver exceptional results. Currently, Joe proudly serves as the Lead Developer Advocate in North America at JuiceFS, a high-performance, cloud-native, distributed file system. He dedicates his role to promoting this innovative solution and enabling AI/ML engineers, software developers, and platform teams to utilize JuiceFS capabilities for optimal performance and efficiency.
How we hunted a 5x P90 regression through batching, zonal affinity, concurrency, networking, and lying metrics — and found 8 milliseconds hiding in Envoy’s Redis buffer flush threshold.
Kishor Kommanaboina is a Staff Software Engineer at Snap Inc., where he leads caching and storage infrastructure serving hundreds of millions of users — Valkey/Redis clusters at 250M+ QPS across AWS and GCP. He led Snap's fleet-wide KeyDB-to-Valkey migration and drove a multi-million dollar infrastructure efficiency roadmap. He spoke at Unlocked Conference 2026 ("Tweaking Valkey for High Write Rate Workloads") alongside Valkey maintainers from AWS and Google. Before Snap, he built data platforms at Twitter and Amazon. He holds a B.Tech from IIIT Hyderabad.
Database systems are tuned to meet the unique requirements of their workloads, so how is it that many of the faster growing database systems all use the same underlying data structure? This talk covers what an LSM tree is, how it works and why such a simple primitive is so powerful — particularly when building the next generation of database systems on object storage. Finally we’ll introduce SlateDB, an OSS implementation of an LSM tree that you can use to build data systems of your own.
Hello, I'm Almog: co-founder of Responsive! I've spent a decade working on industry's hardest problems in distributed systems, starting with search infrastructure at LinkedIn before moving on to realtime data processing. I enjoy sharing insights I've stumbled on throughout my career about the internals of data systems and the future of data infrastructure.
Vector search has become the default first move in RAG pipelines, but for a meaningful class of workloads, the vector database is the slowest part of the system. This talk shows where vector search hurts P99 latency, when alternatives like BM25, learned sparse retrieval, in-context retrieval, and structured indexing outperform it, and how to make the call deliberately based on real measurements.
Computer scientist with 14 years of experience, Jubin is a Senior Software Engineer at Yahoo Finance, where he leads Data Infrastructure and Platforms team. Author of the upcoming Manning book AI-Native Infrastructure, and speaker at conferences including P99 Conf and API Days. Jubin has deep expertise in AI, distributed systems and cloud-native architectures, he focuses on building resilient platforms for real-time and big data processing. Jubin has authored and peer reviewed technical papers such as IEEE, served on jury panels for technology awards, and shares his insights through industry publications. He is passionate about advancing the state of big data and event-driven systems.
PostgreSQL’s analytical extensions—pg_duckdb, Citus Columnar, TigerData Columnar—push it beyond OLTP, but how far does that go in practice? We benchmark these approaches against specialized OLAP engines on TPC-DS, ClickBench, and real-world queries, showing where PostgreSQL competes and where dedicated systems still win, plus a framework for deciding which your workload needs.
Daniel started his career as PhD student in the area of cloud computing with a focus on distributed databases in the cloud. Further interests cover cloud orchestration, model-driven engineering, and performance evaluations of distributed systems. After completing his PhD, Daniel has co-founded the Benchmarking-as-a-Service platform benchANT where he is responsible for the product development.
Seastar’s CPU scheduler divides each shard time between a flat set of scheduling groups proportionally to their shares. This elegant model falls apart as that set grows, since every new group makes the worst-case latency worse. It falls apart for the second time by putting users, who want to configure service-level shares, in direct conflict with developers, who do the same for system-level groups. This talk shows how hierarchical scheduling puts the model back together and keeps it from falling apart further.
Pavel "Xemul" Emelyanov is an ex-Linux kernel hacker whose past experience includes containerizing Linux and the foundation of the project called CRIU. Pavel joined the ScyllaDB core team at the end of 2019 (which probably explains some extra throughput brought to the NoSQL world since then).
In application observability, most of the focus tends to fall on internal service performance and reliability. In reality, however, the network is often the true source of slowdowns and outages. This talk introduces the essential networking-observability concepts needed to determine when the network – not the application – is to blame. You’ll learn how modern tools such as eBPF enable deeper visibility into network behavior, helping you maintain application availability and performance with confidence.
Peter Zaitsev is an entrepreneur and co-founder of Percona, Coroot and other tech companies. As one of the leading experts in Open Source strategy and database optimization, Peter has applied his technical knowledge and entrepreneurial drive to contribute as a board member and advisor to several open source startups. Additionally, Peter is the co-author of the book "High Performance MySQL: Optimization, Backup and Replication," one of the most popular books on MySQL performance.
In this session, I demonstrate what happens when you give an AI agent real power: access to performance profiles via MCP servers, a full multi-core cluster for controlled experiments, and a dedicated budget for iterative code generation and optimization.
Filipe is a Performance Engineer working at the core of Redis, focused on performance analysis and core system optimization. He has designed and automated performance pipelines that integrate profiling, benchmarking, and telemetry data at scale.
Most teams facing the AI agent scale problem give up either isolation or speed — settling for containers or language isolates because VMs are too slow to start. This talk challenges that tradeoff. We’ll dissect isolation primitives (VMs, microVMs, containers, V8/Wasm isolates, gVisor) through the lens of Trusted Computing Base size to show why only VM-based approaches hold up in multi-tenant environments, then get into the techniques — lightweight VMMs, snapshot-and-restore, memory layout optimizations — that push cold-start latency into sub-millisecond territory at million-VM density. Expect real benchmarks and P99 distributions.
Felipe is CEO and Co-Founder of Unikraft, a start-up building the next generation cloud platform. Prior he worked as chief researcher at NEC Laboratories Europe, has published in several top tier conferences such as SOSP, ASPLOS, OSDI, Eurosys, SIGCOMM, NSDI and CoNEXT, and has given talks at Open Source Summit, P99, AI Council and QCon, among others. Finally, Felipe is one of the founders and maintainers of the Linux Foundation Unikraft open source project.
Wednesday, October 23
8:00am – 1:00pm Pacific Time
16:00 – 20:00 UTC
Thursday, October 24
8:00am – 1:00pm Pacific Time
16:00 – 20:00 UTC
Follow us on Twitter @p99conf for the latest updates.
No surprise, but once again #P99CONF from the #ScyllaDB crew is showing everyone how to do it: this is THE model for what a virtual conference should be! (Importantly: FREE!) Day 1 was terrific; looking forward to kicking off Day 2! https://t.co/PUAM7fkb2E
— Bryan Cantrill (@bcantrill) October 18, 2023
The second day of #p99conf and #scylladb. One of the best tech conference nowadays. By engineers for engineers.
— Marcin Rusek (@marcin_rusek) October 20, 2022
This is why I loved the @P99CONF format. Talks were pre-recorded, discord chat with the speaker during their talk and live Q/A after. Low stress for everyone - easy to get at missed talks.
— SMT Solvers (@SMT_Solvers) October 22, 2021
Great discussion on which languages are best for coding for speed (@rustlang or @ziglang) by @glcst, @jarredsumner, and @carllerche at #P99CONF #ScyllaDBhttps://t.co/4j0BZSvHyq
— Samir Alibabic (@samiralibabic) October 18, 2023
That was fun! Great questions from the #p99conf crowd https://t.co/bSRedaDN6x
— Liz Rice 🐝 💙💛 (@lizrice) October 19, 2022
The terrific work by the @P99CONF crew has me wondering if virtual conferences might be the future: the platform fostered collaboration, the price was right (free!), and (best of all?) the videos are already out! Viz. my talk from this morning: https://t.co/qWKsqt43nq
— Bryan Cantrill (@bcantrill) October 7, 2021
True before, true this year too.
— Paul Philleo (@philpauleo) October 18, 2023
Good thing that #P99conf talks are available on demand later!#ScyllaDB #rustlang #NoSql #database #AI #SQL #opensource #memes pic.twitter.com/3qLPBc7ckV
I've been told that cognitive biases are a trend in the #P99CONF community.
— Gwen (Chen) Shapira (@gwenshap) October 19, 2023
If you are not there live, you are missing the world's best live chat. I don't think I ever got so many good insights from so many top experts at the same time. https://t.co/P659WTVyjs
There’s no other event like this — a conference for engineers by engineers, where we’ll share novel approaches for solving complex problems efficiently and at speed. Vendor and tool agnostic, this conference will be for a highly technical audience only. Your boss’s boss is not invited.
Registering for P99 CONF includes 30-day access to the full O’Reilly learning platform!