Aug 2024 – Present · 1 yr 7 mos

• Co-designed and implemented a Copy-on-Write (CoW) B-Tree for Distributed Block Storage to manage file system metadata. Leveraged CoW semantics to enable O(1) instant snapshots and built an end-to-end Garbage Collection workflow that performs orphaned block reclamations using delta-based CoW B-Tree scans.
• Co-led the design of the patent-pending 1N1D feature for a 3-node cluster that supports 1 node and 1 disk failure simultaneously, a major advancement over the industry-standard 4-node solution. Led a team of 3 engineers in the implementation of this feature and collaborated with sales, PM, UX/UI, and QA to deliver it. This effort helped secure Nestlé as the first customer to adopt the feature, which is now available to all customers.
• Developed a Pause-Resume functionality in MapReduce scans, addressing the inefficiency of discarding progress upon failure. This allows scans to resume from failure (e.g., node crash, worker crash, orchestrator failure), preventing the loss of up to 15 hours of work in long-running scans on dense node clusters, which run back-to-back with 3-hour intervals.

Feb 2022 – Aug 2024 · 2 yrs 6 mos

• Co-led the design and development of a framework that decoupled fault tolerance (FT) configurations across different domains, in collaboration with cross-functional teams like data path, PM, and QA. This solution enabled RF3 on 3-node clusters instead of the traditional 5-node setups, optimizing cluster resource utilization by reducing the number of spare nodes required for RF3 configuration (from 2 spare nodes to 1) resulting in a 50% reduction in reserved capacity and operational overhead.
• Redesigned the resilient capacity algorithm to eliminate performance bottlenecks, reducing execution time from 1 second to 1 microsecond (benchmarked on a 1.5PB cluster), which previously caused cluster upgrades to stall due to delays during high-frequency computations. Also improved calculation precision, reducing error margins from 2GB to 2 bytes.
• Mentored and guided new grads and interns on the development of long-term emitter storage that enables data to be carried over from one MapReduce scan to the next. This technique has been fully implemented and is now a critical part of the Nutanix’s distributed file system infrastructure, facilitating pause-resume scan based deduplication.

Aug 2020 – Feb 2022 · 1 yr 6 mos

• Optimized VM snapshot management, achieving a 40% reduction in metadata size for telescopic snapshot schedules, while balancing garbage collection aggressiveness with metadata load dynamically, adapting to system requirements.
• Enhanced the I/O integrity testing framework to support deterministic and uniformly distributed deduplication-oriented writes. The framework identified 3 critical bugs within a week of deployment.
• Developed a framework to derive the cluster’s data resiliency status based on data fault tolerance indicators across various components, such as disk, node, and rack, as well as the availability of I/O and metadata services.