My current research focuses observability and tracing in cloud and distributed systems, performance and resource management, and building multi-tenant distributed systems. A few project highlights include the following:

Blueprint is a framework for developing flexible and modular microservice applications. Blueprint provides a programming abstraction for writing microservice applications that makes it easy to later change aspects related to the system's scaffolding and topology -- e.g. concerns such as RPC frameworks or backends to use, placement and load balancing, replication, and so on. Using Blueprint, applications don't early-bind to those choices. Later on, it's easy to change any of these aspects and re-generate a fully functional variant of the application. Blueprint makes comparative evaluation trivially easy -- it helps developers empirically pick the best-performing libraries, frameworks, and backends; and it helps researchers perform rigorous comparative evaluation of new prototypes. Blueprint was published at SOSP 2023 and its code can be found on Github.

Hindsight is a new distributed tracing framework that we have built from the ground-up to support edge-case tracing. Hindsight enables detailed end-to-end tracing for rare and outlier requests without data loss traditionally incurred by sampling-based systems. Hindsight overcomes this by combining a short per-node history of telemetry, programmatic symptom detection, and rapid distributed retrieval. Hindsight was published at NSDI 2023 and its code can be found on GitLab.

Clockwork is a distributed DNN serving system designed for predictable end-to-end performance. Clockwork promotes end-to-end performance predictability as a first-class design concern. To achieve this, Clockwork's design eliminates major sources of performance variability and centralizes choices that lead to variability such as scheduling and admission control. The end result is a system design with extremely tight tail latency. Clockwork received the Distinguished Artifact Award at OSDI 2020 and its code can be found on GitLab.

Pivot Tracing is a cross-component monitoring framework for distributed systems. Often the information needed to troubleshoot cross-component problems is relatively minimal, but inaccessible due to a lack of cross-component visibility. Pivot Tracing combines causal metadata propagation with dynamic instrumentation to overcome this limitation. Using Pivot Tracing, a system operator can use a simple SQL-like interface to define and measure arbitrary system metrics, while grouping, filtering, and aggregating those metrics according to arbitrary identifiers from other system components. Pivot Tracing received the Best Paper Award at SOSP 2015 and its code can be found on GitHub.