The geospatial engine suite (benchmarks/geospatial/engine_suite.py) drives its remote runs through Coiled Functions. Coiled is convenient (environment sync, one reused VM per machine size) and works well today, but in review Alex raised a fair durability question: with Matthew Rocklin stepping back from Coiled (https://matthewrocklin.com/stepping-back/), it is unclear how durable Coiled Functions are long term, so the suite should not be hard-tied to a single provider.
- Vanilla Cloud Run was suggested as a good fit, since we usually profile against ARCO-ERA5 in-region (us-central1).
- The remote side of the suite is already mostly provider-agnostic: cells are plain functions taking a source tarball, and the native module is provisioned per VM by the harness itself (4b76f31), so porting means swapping the
coiled.function decorator and the VM lifecycle handling.
Context: #227 (comment)
The geospatial engine suite (
benchmarks/geospatial/engine_suite.py) drives its remote runs through Coiled Functions. Coiled is convenient (environment sync, one reused VM per machine size) and works well today, but in review Alex raised a fair durability question: with Matthew Rocklin stepping back from Coiled (https://matthewrocklin.com/stepping-back/), it is unclear how durable Coiled Functions are long term, so the suite should not be hard-tied to a single provider.coiled.functiondecorator and the VM lifecycle handling.Context: #227 (comment)