Data Engineering Podcast Scaling Graph Analytics Without ETL: Inside PuppyGraph’s Architecture
32 snips
Jun 1, 2026 
Weimo Liu, co-founder of PuppyGraph and former engineer at TigerGraph and Google F1, builds zero-copy graph querying for lakehouses. He discusses an edge-sharded, vectorized MPP design that avoids ETL. The conversation covers mapping tables to graphs, Iceberg-aware caching, integrations with MongoDB, and real-world uses like cybersecurity and entity resolution.
AI Snips
Chapters
Transcript
Episode notes
Graph Workloads Need Graph Native Computation
- General-purpose SQL engines (Spark, Trino) aren't optimized for graph workloads; PuppyGraph implements its own graph-native computation to avoid mismatch and expensive shuffles.
- PuppyGraph focuses on graph-aware MPP and shuffle handling because graph computation needs specialized parallel patterns beyond SQL engine optimizations.
Model Graphs From Normalized Or Denormalized Tables
- When source tables are normalized, map them directly to node and edge tables; if denormalized, create logical views or flexible mappings to dedupe attributes into nodes.
- PuppyGraph can dedupe attributes (e.g., zip code in a wide table) into nodes and suggests graph schemas based on query patterns and use cases.
Flatten Documents To Preserve Structure Over Embeddings
- Unstructured documents often contain repeatable internal structure; extracting entities into tables then running graph queries preserves that structure better than naive chunk+embedding approaches.
- Weimo cites PDF bank statements where fixed fields (account number, address) are extracted to tables, enabling richer graph queries than embeddings alone.