Let's celebrate the power of community, performance engineering, and Valkey Ecosystem. Learn why Khawaja is beyond excited for today, what to expect, and how to make the most out of this event.

Valkey established itself as a fast moving and innovative open-source project through an engineering focused mindset. By focusing on low-level optimization, we were able to improve Valkey to meet the needs of our most demanding users and contributors. In this session, we dive into the core engineering principles and the decision making driving the project today and the future.

Snap manages a fleet of over 350 caches powering mission-critical features. Moving this workload from KeyDB to Valkey meant replacing deep-seated features including IO multithreading, forkless background saves, and cross-slot MGets, without impacting the user experience.

We’ll share the engineering reality of this migration: how we bridged these feature gaps and the operational safeguards we implemented to maintain cluster health throughout the transition. This session provides a blueprint for executing a large-scale, high-stakes migration with zero drama.

Time for some networking!

Uber's cache platform is the foundational state engine for mission-critical operations, sustaining peaks of one billion requests per second. This session covers our current architecture and lessons learned from connection storms, response desynchronization, and high availability under extreme load. We'll also outline the roadmap for Uber's next-generation cache architecture, designed to harden platform resilience at scale. 

Valkey Search makes building AI applications easier. Explore a reference architecture for semantic caching with Valkey, get practical guidance, and see its impact on cost, performance, and response quality.

Sponsored by EloqData - Thank you for supporting Unlocked Conf!

Resharding is one of the most dangerous operations in a distributed system. It often involves complex coordination, network churn, redirects, and data duplication, along with an acceptance of latency or downtime risk. For always-on systems, this trade-off is unacceptable. 

Valkey's Atomic Slot Migration changes the model. It delivers predictable behavior, faster migrations, less churn, and resharding that no longer feels like a boss fight. 

This talk breaks down how atomic slot migration works, the issues with traditional slot migration, and what this design means for performance and day-two operations in mission critical clusters. The goal is simple: level up sharding by making resharding boring. 

Valkey accepts request and response payloads up to a global bulk-size limit. In practice, however, even moderate increases in payload size can introduce significant overhead. Larger payloads take longer to transmit over the network, require more memory allocation, increase replication pressure, and block the event loop for longer periods. 

When these effects stack, overall system performance can degrade well before payloads approach the configured maximum size. 

This talk proposes a set of high-level guardrails to make large-payload workloads easier to observe and manage in production. 

It covers: Adding lightweight server-side signals to track request and response payload size trends and the frequency of large operations (e.g., Issue #2911, PR #2918); Practical guidance on choosing bulk-size limits that better match different workload patterns; Exploratory runtime safeguards that use signals such as latency or replication lag to temporarily reject or delay expensive operations when the server is under stress.

Nu is one of the largest digital financial services platforms in the world, serving 127 million customers across Brazil, Mexico, and Colombia. Datomic is the primary transactional database at Nu. Datomic stores a record of immutable facts, which gives applications strong consistency combined with horizontal read scalability, auditability and built-in caching.

This talk describes the leverage an immutable architecture brings to bear on transactional systems, the performance of Datomic's caching tower and the economics of immutable caching at scale.

Achieving 1 billion requests per second (RPS) was the target, but verifying it at a 2,000-node scale was a challenge. This session dives deep into the Valkey 9.0 cluster improvements including new failure ranking, connection throttling, and optimized cluster bus gossip that ensure bounded recovery time during massive primary failures.

We discuss the architectural changes and the benchmarking that made this milestone possible. We'll also share the operational lessons learned from pushing the system to its absolute limits, revealing exactly why we won't attempt this specific verification again.

Infinity called, they want their Big-O back. So, what’s left? Well, we will have to dig down a bit from these beautiful abstractions…

ElastiCache Serverless offers a high-performance multi-tenant Rust proxy to abstract the client-side complexities of a fast-scaling cluster-mode enabled Valkey backend. Explore how we designed the proxy for low latency and high throughput - from its network and I/O architecture that enables local AZ routing with an optimized number of backend multiplexed sockets to its threading model that eliminates unnecessary synchronization while still ensuring appropriate Quality of Service (QoS).

Hot keys and uneven data distribution can silently degrade performance and trigger outages, while diagnosing them often risks making things worse. Valkey Admin provides configurable diagnostic approaches to inspect keys, identify hot slots, and understand cluster layout using operationally sound techniques, helping operators debug and fix issues with appropriate impact considerations. We'll cover how it works internally and show a brief demo.

Mercado Libre operates one of the largest and most demanding commerce platforms in the world. As the platform team, we manage a shared, multi-tenant caching platform built on top of Valkey. This talk details the migration from Memcached to Redis to Valkey, the technical realities of multi-tenancy, noisy-neighbor mitigation, capacity planning, and weighs the pros and cons of client-side caching in a high-traffic environment.

We’ll include some of the architectural challenges we had along the way, failure modes we encountered, and techniques used to deliver a robust caching infrastructure that is predictable at scale.

As investments shift toward AI models and applications, infrastructure is often dismissed as a commodity. This perspective overlooks a historical reality: standard utilities, including the CPU, only became "commodities" after years of deep research and architectural optimization.

In practice, interactions between services and agents are multiplying, and latency budgets are shrinking. This keynote makes the case that systems engineering is more critical than ever. We examine specific system internals - cores, scheduling, NUMA boundaries, the network stack - to show how small design choices create outsized effects on performance, utilization, and user experience.