Community Weekly Report
Jul 13 – Jul 19, 2026
00Overview
Doris merged 122 PRs this week from 55 contributors, with real-time DWH and kernel engineering as the twin focal axes — clear signal that the project is doubling down on a lakehouse-native, real-time analytics foundation. Vector search, Iceberg/open lake formats, external caching, and the cloud-native control plane all moved forward in lockstep, while the optimizer keeps tightening operators and rewrites, pointing momentum toward a unified real-time lakehouse built on a cloud-native kernel.
01Highlights
02By Scenario
Multi-modal Lakehouse
01Multi-modal lakehouse work advanced across correctness, performance, and observability: Iceberg row-level deletes across schema evolution, ORC v2 pushdown/fallback, complex Parquet struct counting, and Iceberg system-table projection were repaired; ORC v2 stripe-merged reads and optimized filtered Parquet scans were added; Python UDF Arrow memory is now tracked through Doris allocators.
- [improvement](be) Track Python UDF Arrow memory allocations
- [improvement](be) Optimize filtered Parquet scans and condition cache
- [fix](arrow-flight) Keep coordinator alive across GetFlightInfo/DoGet for external table scan
- [fix](hive) invalidate stale file cache on partition refresh when partition cache misses
- [improvement](be) Add ORC v2 stripe merged reads
Multi-modal lakehouse work is closing gaps in Iceberg compatibility and external table scanning. Recent fixes preserve nested struct field name case in Iceberg and route position_deletes system-table scans to the native scanner instead of JNI defaults. Parquet INT96 decoding is decoupled from the session timezone and export defaults now mirror
