Performance Comparison: Index Raster vs GTI Driver

[elidwa]

Performance Comparison: Index Raster vs GTI Driver

Method	Points Sampled	Total Results Returned	Average Runtime (s)	Relative Speed	Sampling Behavior
Index Raster	20 000	26 322	4.17	1× (baseline)	Samples all overlapping rasters; returns every elevation value for user-side filtering.
GTI Driver	20 000	20 000	25.42	≈ 6× slower	Reads until the first non-NaN value; performance depends on how the .gti catalog was built.

---

Summary

This investigation was run from the working branch 1meter3dep-gti.
The test used a locally placed catalog.gti and its corresponding catalogv2.parquet, while all source GeoTIFFs remained in their AWS S3 buckets.

Timing results show that the index raster method remains substantially faster, about six times more efficient than the GTI driver.
The index sampler achieves high throughput by directly querying all overlapping rasters in parallel and returning every elevation for post-processing or filtering on the client side.

In contrast, the GTI driver introduces additional overhead due to its internal tile-lookup logic.
Because it stops at the first non-NaN value, any optimization must be handled when generating the .gti catalog itself.
Unless the catalog is pre-composed to minimize overlaps and define raster priorities, the driver’s runtime cost remains high.

Overall, using the existing raster index sampling provides the best balance of speed and flexibility.
The most significant performance improvement going forward would come from eliminating the need to query CMR rather than relying on GTI-based access.