Скачать или смотреть [CNG 2025] Why We Don't Use Zarr (Yet) – Sarah Zweip

Sarah Zwiep, geospatial data engineer at Floodbase presents on the complexities of building a near-real-time flood monitoring system from CNG Conference 2025.

Sarah introduces Floodbase's US Daily Flood Model, a machine learning model that produces high-resolution flood extent maps across the continental US and also talks about the data and infrastructure challenges the team faces when dealing with multi-modal geospatial data sources. She then explains why Floodbase has chosen to stick with Cloud Optimized GeoTIFFs (COGs) instead of using Zarr, which includes:

The benefits of COGs for a small engineering team and the ease of debugging and data exploration they provide
The pain points of their current system, including a PostgreSQL bottleneck and the limitations of processing on a daily scale.
The trade-offs and remaining challenges with Zarr, such as its impact on a modular data model and the complexity of integrating it with existing infrastructure.