哈工大 · 数据科学
Data Engineering · Data Warehouse · Stream Processing
Building reproducible data systems around correctness, observability and safe consumption.
Focus: RetailDW → GameStream → SQLGuard → ChangeLake
| Repo | What it is | Run |
|---|---|---|
| RetailDW | Offline retail warehouse on PySpark (ODS→DWD→DWS→ADS): quality gates, refund semantics, dt-partition idempotent reruns, cross-layer GMV reconcile |
bash scripts/run_local.sh |
| GameStream | Realtime game analytics: Kafka → Flink → Doris; event-time, dedup, checkpoint recovery; Golden Path | docker compose up -d && bash scripts/demo_golden_path.sh |
| SQLGuard | Deterministic SQL write/execution gate (AST policy): make seal prints ALLOW/BLOCK + rule_id + evidence |
make seal |
| ChangeLake | Reproducible CDC lakehouse: MySQL → Flink CDC → Paimon on MinIO; schema evolution, recovery, backfill, time travel, reconcile, compaction | make demo |
Other work: DataPilot · DocPilot · AgentOpsLite · RucBaseLab
- RetailDW — 7-day repurchase window + net GMV; quality gate fail-fast; same-
DTrerun reconcile - GameStream — metrics keyed by
metric_id/ ADS tables; Golden Path six beats (incl. late/dup/kill-TM) - SQLGuard — unsupported SQL fail-closed; seal cases are deterministic and evidence-bearing
- ChangeLake — G9 source↔lake Decimal reconcile; G10 compaction fingerprint identity; Golden Path fail-fast (
DEMO_EXIT=0)
