Adaptive execution
How can a database choose better physical behavior at microsecond timescales?
I study execution engines that learn from fine-grained runtime feedback and change strategy as data and workloads shift.
Piece of CAKEComputer & Information ScienceUniversity of Pennsylvania
ZijieZhao.
Current focus — Self-evolvable software
I build adaptive software systems that learn from execution and improve how they behave, with a current focus on database engines.
Advised by Ryan Marcus (opens in a new window) at the Penn Database Group (opens in a new window)
Leveraging AI · Codex17.2B
lifetime tokensUpdated 2026-10-05
Research threads
03 — Three questions I’m working on
Workload understanding
What do production traces hide about the work users actually wanted to run?
I examine the feedback loop between platforms and users so workload data can be interpreted without mistaking survivors for demand.
Survivorship BiasSelf-evolvable software
Can software adapt inputs, policies, and implementations instead of waiting for retraining or manual tuning?
My broader goal is to combine machine learning with program structure to build systems that improve while they run.
CodeImprove04 — Selected · reverse chronological
Publications
Updates
06 — Recent activity- Released the Piece of CAKE preprint on microsecond-scale adaptive execution. (opens in a new window)
- Survivorship Bias in Industrial Database Workloads received the CIDR 2026 Best Paper Award. (opens in a new window)
- CodeImprove appeared at the 47th IEEE/ACM International Conference on Software Engineering. (opens in a new window)
- Joined the Penn Database Group to work on adaptive and self-tuning systems. (opens in a new window)
Education
07PhD, Computer and Information Science
University of PennsylvaniaMA, Statistics and Data Science
The Wharton School University of PennsylvaniaBE, Computer Science
Tianjin University