Work Experience

• Infrastructure Software Engineer, DouYin Group

  Jul, 2022 - Present

  Responsible for cloud-native development work, covering areas such as container image optimization (lazy loading), container storage (CSI), and container runtimes (Containerd/Kubelet). In addition to regular tasks including feature development, testing, and release, also participated in on-call support and issue troubleshooting.

  • Participated in the community development of the nydus-snapshotter and nydusd projects, contributed approximately 6.5k lines of code, and completed component release and testing. Successfully deployed to production.

  • Synchronized upstream updates of Containerd and Kubelet from the community, followed by testing and release. During user issue investigations, identified and resolved a bug caused by the instability of sort.Sort() in Golang 1.19 that led to container startup failures inside existing containers, and submitted a fix to the community.

  • Delivered feature iterations for CSI components (block storage, file storage, and object storage). Designed, developed, tested, and released a new CSI driver for a proprietary file system from scratch.

  • Investigated issues caused by container images, container runtimes, container storage, and node-level anomalies. Authored troubleshooting documentation and SOPs.

• Infrastructure Software Engineer (Intern), ByteDance

  Jul, 2021 - Sep, 20212 months

  Focused on virtualization storage development, involving feature implementation, upstream code merging, and performance benchmarking.

  • Participated in the development of a storage component based on SPDK and vhost-user, and designed and implemented a hot upgrade mechanism that reduced I/O tail latency from seconds to microseconds during the upgrade process.

Projects Experience

• Container Image Lazy Loading, DouYin Group

  Jul, 2022 - Jun, 202310 months

  n cloud-native scenarios, the time required to pull a container image is typically proportional to the image size. This becomes especially problematic in AI workloads, where large images make container startup times unpredictable. The lazy loading feature addresses this by allowing the container runtime to download only a small amount of metadata during the image pull phase, enabling the filesystem to be ready much earlier and significantly improving container startup performance.

  • Contributed to feature iterations of the Containerd plugin nydus-snapshotter and the FUSE-based client nydusd, with 6.5k lines of code submitted to the open-source community. Completed testing and release of the components. In scenarios leveraging the image lazy loading feature, container startup times were consistently reduced to the level of seconds.

• CSI Driver For Custom Filesystem , DouYin Group

  Jul, 2024 - Jun, 20251 year

  The proprietary filesystem client consists of a user-space process and a kernel module, and supports mounting various storage backends. To integrate with cloud-native environments, a new CSI driver was developed to provide users with a standard way to access storage through the custom filesystem client in Kubernetes scenarios.

  • Responsible for end-to-end project delivery from scratch, covering requirement clarification and validation, overall architecture design, feature development, unit and integration testing, production release, and stability assurance.

• Global Memory With RDMA, HuNan University

  Jun, 2020 - Dec, 20206 months

  The Infiniswap (NSDI’17) approaches the problem from the perspective of memory-constrained hosts, offloading data to the memory of other nodes in the cluster via RDMA when local memory is insufficient. However, this can lead to the issue of busy hosts becoming even more overloaded. Building on Infiniswap, this project leverages one-sided RDMA operations to implement a custom-designed proactive migration strategy, enabling idle hosts to take over most of the workload originally handled by busy ones.

  • Implemented a proactive migration strategy that improved random write performance by 19.12% compared to Infiniswap. For 128KB random read and write operations, latency was reduced by 65.6% and 77.4%, respectively.

Skills

Education