Move climate history to ERA5 served from our own object storage, with a
prod-only query service in front:
- data/era5lake.py: lake layout (per-point whole-record 1940+ serving files,
a tile/year/month hive table, an Iceberg-style manifest) plus grid math and
the read client (local dir -> lake service -> bucket).
- gen_era5_lake.py: tile-aligned extractor from the Earthmover Icechunk ERA5
archive (one 86-year pull covers all 144 points of a chunk tile; resumable
from the manifest; --cities / --tiles / --land).
- lake_app.py + THERMOGRAPH_ROLE=lake: the lake service. /history serves a
point's parquet off a disk cache (11ms cold / 3ms warm in rehearsal);
/query runs SELECT-only SQL on DuckDB over the hive table (79ms pruned
aggregate, 88ms full scan of a 4.5M-row tile). httpfs is baked at image
build.
- climate.py: history chain is now era5-lake -> NASA POWER -> Open-Meteo; the
lake slice starts at START_DATE so grading windows are unchanged, and an
unconfigured lake costs nothing (beta/LAN unchanged).
- stack: lake service (1..2 replicas behind the VIP, own cache volume) plus a
second autoscaler instance (autoscale.sh gains TARGET_SERVICE); web/worker
get THERMOGRAPH_LAKE_URL.
- seed_era5.py: constants corrected against the live store (icechunkV2,
single/temporal, valid_time, ECMWF short names, pcodec).
Bucket creds (THERMOGRAPH_LAKE_S3_ACCESS_KEY/_SECRET_KEY) go in the prod
vault; until they land the lake stays healthy and everything falls through to
NASA exactly as before.
