Geospatial SQL benchmark suite: prove core geo/climate ops are relational - #177

Merged

alxmrs merged 26 commits into

mainfrom

geospatial-sql-benchmarks

Jun 29, 2026

Commits on Jun 29, 2026

Add geospatial SQL benchmark suite (benchmarks/geospatial/)

alxmrs
and
claude
committed
NDVI: open Sentinel-2 idiomatically (pystac + open_datatree), drop manual scaling/CASE/sort

alxmrs
and
claude
committed
Make benchmark cases idiomatic xarray: fluent opens, Dataset results, coordinate-aligned comparison

alxmrs
and
claude
committed
Lazy registration, parameterized queries, model-concat: address round-3 review

alxmrs
and
claude
committed
Cases 07/08: use Earth Engine via Xee (independent reproj reference; real SRTM regrid)

alxmrs
and
claude
committed
README: add "Does it work?" section (addresses #46); drop "honest"

alxmrs
and
claude
committed
EE cases: initialize via ADC (no blocked OAuth); fix 07 coord names — both pass live

alxmrs
and
claude
committed
Fix run_all.sh: run from any directory, run all cases, use the local xarray-sql

alxmrs
and
claude
committed
Point each case's uv metadata at the local xarray-sql checkout

alxmrs
and
claude
committed
Print each case's result (xarray repr) after the match is verified

alxmrs
and
claude
committed
Docs: write for the library's reader, not the review thread

alxmrs
and
claude
committed
docs: cite the coiled/benchmarks #1545 survey the operations come from

alxmrs
and
claude
committed
Add opt-in performance profiling to the harness (CSV perf table)

alxmrs
and
claude
committed
Profiling via a `for _ in measured(...)` loop instead of re-running main()

alxmrs
and
claude
committed
Address review: partial-result guard, lazy NDVI, no-reshape regrid

alxmrs
and
claude
committed
Add cold-vs-cold perf driver (fair SQL-vs-xarray timing)

alxmrs
and
claude
committed
docs: publish perf results, analysis, and the SQL-vs-arrays tradeoff

alxmrs
and
claude
committed
simpler conclusion title.
alxmrs
committed
Address review: fair cold-vs-cold (.compute), to_pandas headline, doc fixes

alxmrs
and
claude
committed
Address review: drop "whole archive" overclaim, inline perf summary

alxmrs
and
claude
committed
Re-collect cold-vs-cold timings with the .compute() fix

alxmrs
and
claude
committed
Address review: lazy zonal-stats reference, split perf summary back out

alxmrs
and
claude
committed
Address review: make case 08 a symmetric laziness test

alxmrs
and
claude
committed
Case 05: one partition each (both chunks=100); document the chunk finding

alxmrs
and
claude
committed
Add case 09: warp (reproject + resample) as UDF → weight-table JOIN

alxmrs
and
claude
committed
docs: refresh perf table — full 9-case run on one in-region VM with EE

alxmrs
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Geospatial SQL benchmark suite: prove core geo/climate ops are relational - #177

Geospatial SQL benchmark suite: prove core geo/climate ops are relational#177
alxmrs merged 26 commits into
mainfrom
geospatial-sql-benchmarks

Commits on Jun 29, 2026

Add geospatial SQL benchmark suite (benchmarks/geospatial/)

NDVI: open Sentinel-2 idiomatically (pystac + open_datatree), drop manual scaling/CASE/sort

Make benchmark cases idiomatic xarray: fluent opens, Dataset results, coordinate-aligned comparison

Lazy registration, parameterized queries, model-concat: address round-3 review

Cases 07/08: use Earth Engine via Xee (independent reproj reference; real SRTM regrid)

README: add "Does it work?" section (addresses #46); drop "honest"

EE cases: initialize via ADC (no blocked OAuth); fix 07 coord names — both pass live

Fix run_all.sh: run from any directory, run all cases, use the local xarray-sql

Point each case's uv metadata at the local xarray-sql checkout

Print each case's result (xarray repr) after the match is verified

Docs: write for the library's reader, not the review thread

docs: cite the coiled/benchmarks #1545 survey the operations come from

Add opt-in performance profiling to the harness (CSV perf table)

Profiling via a `for _ in measured(...)` loop instead of re-running main()

Address review: partial-result guard, lazy NDVI, no-reshape regrid

Add cold-vs-cold perf driver (fair SQL-vs-xarray timing)

docs: publish perf results, analysis, and the SQL-vs-arrays tradeoff

simpler conclusion title.

Address review: fair cold-vs-cold (.compute), to_pandas headline, doc fixes

Address review: drop "whole archive" overclaim, inline perf summary

Re-collect cold-vs-cold timings with the .compute() fix

Address review: lazy zonal-stats reference, split perf summary back out

Address review: make case 08 a symmetric laziness test

Case 05: one partition each (both chunks=100); document the chunk finding

Add case 09: warp (reproject + resample) as UDF → weight-table JOIN

docs: refresh perf table — full 9-case run on one in-region VM with EE