Graduate Data Engineer (TikTok Recommendation Ecosystem, Singapore)

About the role

ByteDance's Arch-Data Ecosystem team builds the offline and real-time data storage solutions behind the TikTok Recommendation System, which serves over 1 billion users. The team aims for high system reliability, uninterrupted service and smooth data processing, and is looking for a graduate to join its recommendation ecosystem architecture team in Singapore with a 2027 start.

Key responsibilities

• Build a unified infrastructure integrating the "training database" and "training/inference state system" for multimodal foundation models in search, recommendation and advertising, optimising data lakes, caching, distributed computing and GPU IO to lower training and inference costs.
• Design and implement real-time and offline data architecture for large-scale recommendation systems, and build scalable, high-performance streaming Lakehouse systems for feature pipelines, model training and real-time inference.
• Collaborate with ML platform teams to support PyTorch-based model training workflows and design efficient data formats and access patterns for large-scale samples and features.
• Own core components of the distributed storage and processing stack, from file format to stream compaction to metadata management.

Required profile

• Completing or recently completed a PhD in AI, software development, computer science, computer engineering or a related discipline.
• Experience building large-scale distributed systems, preferably in storage, stream processing or ML infrastructure.
• Commitment to an onboarding date by the end of the year; availability and graduation date should be stated in the resume.

Required skills

• Understanding of Apache Flink internals, with hands-on experience in state management, connectors or UDFs.
• Familiarity with Lakehouse technologies such as Apache Paimon, Iceberg, Delta Lake or Hudi, especially incremental ingestion, schema evolution and snapshot isolation.
• Preferred: experience with Flink + Paimon architectures, feature storage and training data pipelines integrated with PyTorch, columnar formats (Parquet, ORC, Lance), Java/Scala/C++ and knowledge of HBase/Kudu.