about

About Me

My name is Calvin Luo (R.Z. Luo, 骆融臻). I’m currently a KVCache engineer in MiniMax. Before, I was a database kernel engineer at PingCAP, working on TiFlash and TiCI. My work focuses on distributed storage, transaction processing, cloud architecture, and full-text search. Before, I worked in Tencent with several AI projects.

Basic information

  • GITHUB: CalvinNeo
  • EMAIL: calvinneo@calvinneo.com
  • BLOG: www.calvinneo.com
  • EDUCATION:
    • ME: Software Engineering at Nanjing University, 2017 - 2019.
      • By recommended admission without an entrance examination.
    • BE: Civil Engineering at Hohai University, 2013 - 2017.
      • Awarded with outstanding undergraduate thesis.
  • OPEN-SOURCED PROJECTS:
    • Nuft implements the Raft protocol, including leader election, replication, snapshots, and joint consensus.
    • ATP provides reliable transport sockets over UDP, with retransmission, flow and congestion control, port multiplexing, and connection migration. It has been used as sample code for undergraduates at Nanjing University’s Software Institute.
    • CFortranTranslator is a line-by-line translator from Fortran77/90 to C++14, supporting arrays, I/O, and standard library functions. It has been used by research institutions and has an active contributor community.
    • PolyClip clips millions of circles and lines with a polygon in seconds.
    • Citation4Word is a word citation format for thesis.
    • NuCut is a distributed streaming graph partitioning system, using Bloom filters for edge deduplication and distributed state to reduce communication.

Career (2026 Version)

MiniMax — Shanghai, China. Sep 2026 - Present

PingCAP — Shanghai / Remote, China. Aug 2021 - Jun 2026

Database kernel engineer for TiFlash and TiCI, owning multiple modules while collaborating across regions. TiFlash Committer, TiFlash Proxy Maintainer, and TiKV Reviewer.

TiCI: Full-text search and inverted index engine

As a founding member, I worked on TiCI throughout its lifecycle, from design and development to production rollout.

  • Built the write path, including incremental and baseline data merging, persistence, sharding, compaction, garbage collection, scheduling, and high availability.
  • Implemented read replica caching for S3, keeping data readable during shard splits and merges.
  • Customer POCs achieved 3,000+ QPS with 60 ms p99.9 latency on static data, with room to scale further. Validated stable incremental ingestion across multiple workloads with 10 TB of real data and 100 TB of simulated data.
  • Built the project framework, actor-based asynchronous execution, FFI, and CI/CD, with safeguards against deadlocks and panics.
  • Standardised prompt.md conventions and designed agent skills to keep AI-assisted development focused, backed by multiple levels of testing.

TiFlash Cloud: Storage and compute separation, and using S3 to rebuild the storage layer

  • Designed and implemented cloud Raft log storage, balancing high availability, elasticity, strong consistency, latency, durability, and cost. See TiDB and TiFlash in the cloud.
  • Reused data on S3 for fast scale-out and recovery, reducing replication and row-to-column conversion overhead. Achieved 17.5× scale-out throughput, 20% lower memory usage, and 96% lower CPU usage.
  • Introduced unified storage for data from multiple engines to provide consistent snapshots on S3, with caching adapted to metadata access patterns.
  • Improved observability across languages and components. Built write rate limiting based on table- and object-level memory tracking, alongside heap profiling across deployment environments using jeprof and eBPF.

TiFlash: Replication, transactions, and storage

  • Implemented the Raft Learner protocol and resolved replication issues such as lost merge events. Supported strongly consistent reads through Learner Read and ReadIndex, with sub-second replication lag.
  • Implemented large transaction writes in the Percolator transaction model, verified correctness with TLA+, and reduced memory usage by 35%. Added support for stale reads.
  • Parallelised row-to-column conversion for large snapshots, reducing synchronisation time linearly with parallelism. Asynchronous flushing of incremental data reduced random write latency by 60×.
  • Supported asynchronous replica bootstrapping from heterogeneous data sources, replica management, and dynamic scaling. Adapted TiFlash to multiple RocksDB storage engines, including Partitioned Raft KV.
  • Maintained TiFlash Proxy, aligning TiKV and TiFlash persistence strategies and integrating them through FFI. Decoupled TiFlash Proxy from TiKV to reduce upstream merge conflicts during upgrades.

Tencent IEG — Shenzhen / Shanghai, China. July 2019 - Aug 2021

Backend engineer in the shared R&D and operations organisation and TiMi Studio Group, working on data mining and Spark optimisation.

  • Delivered an online MOBA match win probability prediction system with 88–92% real-time prediction accuracy. Processed tables with tens of billions of rows using Spark, trained traditional machine learning and deep learning models, and built the online inference service.
  • Built mobile game level completion rate prediction, reducing prediction error by 34%.
  • Analysed Event, GC, and Spark logs to recommend optimisations, including identifying repeated RDD computations and redundant caching.
  • Developed machine learning components and fused JAR/SQL execution, inferring intermediate table schemas to avoid disk writes and reduce scheduling overhead.

Tencent IEG — Shenzhen, China. June 2018 - Sept 2018

Software engineering intern.

Honours and community

  • PingCAP company-wide Engineering Excellence Award for reducing TiKV upstream merge conflicts and enabling bug-free TiFlash upgrades.
  • 2023 OpenAtom 活力开源贡献者.
  • Multiple ACM/ICPC regional bronze medals; third place in RoboCup China 3D Simulation; second prize in the National Computer Design Competition. Other competition awards include CCPC and the Zhou Peiyuan Mechanics Competition.
  • Over 200 blog posts across 10 years, with articles published by Tencent Technical Engineering and PingCAP, as well as public technical talks. I use English in my daily work.

For my up-to-date CV, please email calvinneo@calvinneo.com with a brief introduction.

Personal

  • I have an extensive collection of wooden modules. Check 系列拼装模型攻略 to find how I built them.
  • I play a wide range of sports, such as table tennis, archery, badminton, and I write blogs about them. I love hiking and cycling. I have been cycling for over 5000KM+ from 2022 to 2024.
  • I am fond of board games. My favorite board game includes: Gaia Project, ArkNova, Dune, Splendor(especially the pokemon edition), 11 nimmt, Rummikub, Arge of War, Las Vegas, CABO, Camel Cup, AZUL, PowWow, The Castles of Burgundy. I have vibed an Open Board Game project in which I can customize different board game rules, and playing with AI bots designed by myself.
  • Je commencerai à apprendre le français en 2025. J’apprends cette langue parce que pense que c’est facile. Mais maintenant je le trouve difficile, je pense que c’est vraiment intéressant.

About Calvin’s Marbles

本博客的写作方式

版权声明

版权声明