Adds extract/openweathermap package with daily weather extraction for 8
coffee-growing regions (Brazil, Vietnam, Colombia, Ethiopia, Honduras,
Guatemala, Indonesia). Feeds crop stress signal for commodity sentiment score.
Extractor:
- OWM One Call API 3.0 / Day Summary — one JSON.gz per (location, date)
- extract_weather: daily, fetches yesterday + today (16 calls max)
- extract_weather_backfill: fills 2020-01-01 to yesterday, capped at 500
calls/run with resume cursor '{location_id}:{date}' for crash safety
- Full idempotency via file existence check; state tracking via extract_core
SQLMesh:
- seeds.weather_locations (8 regions with lat/lon/variety)
- foundation.fct_weather_daily: INCREMENTAL_BY_TIME_RANGE, grain
(location_id, observation_date), dedup via hash key, crop stress flags:
is_frost (<2°C), is_heat_stress (>35°C), is_drought (<1mm), in_growing_season
Landing path: LANDING_DIR/weather/{location_id}/{year}/{date}.json.gz
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add extract/extract_core/ workspace package with three modules:
- state.py: SQLite run tracking (open_state_db, start_run, end_run, get_last_cursor)
- http.py: niquests session factory + etag normalization helpers
- files.py: landing_path, content_hash, write_bytes_atomic (atomic gzip writes)
- State lives at {LANDING_DIR}/.state.sqlite — no extra env var needed
- SQLite chosen over DuckDB: state tracking is OLTP (row inserts/updates), not analytical
- Refactor all 4 extractors (psdonline, cftc_cot, coffee_prices, ice_stocks):
- Replace inline boilerplate with extract_core helpers
- Add start_run/end_run tracking to every extraction entry point
- extract_cot_year returns int (bytes_written) instead of bool
- Update tests: assert result == 0 (not `is False`) for the return type change
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The ICE API at /marketdata/api/reports/293/results stores all historical
daily XLS reports date-descending. Previously the extractor only fetched
the latest. New extract_ice_backfill entry point pages through the API
and downloads all matching 'Daily Warehouse Stocks' reports.
- ice_api.py: add find_all_reports() alongside find_latest_report()
- execute.py: add extract_ice_stocks_backfill(max_pages=3) — default
covers ~6 months; max_pages=20 fetches ~3 years of history
- pyproject.toml: register extract_ice_backfill entry point
Ran backfill: 131 files, 2025-08-15 → 2026-02-20
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ICE changed the daily stocks XLS header from 'As of: 1/30/2026' to
'As of: Feb 20, 2026 1:35:39PM'. Expand _build_canonical_csv_from_xls
to try multiple strptime formats (%m/%d/%Y, %b %d, %Y, etc.) on both
single-token and three-token date candidates.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Replace brittle ICE_STOCKS_URL env var with API-based URL discovery via
the private ICE Report Center JSON API (no auth required)
- Add rolling CSV → XLS fallback in extract_ice_stocks() using
find_latest_report() from ice_api.py
- Add ice_api.py: fetch_report_listings(), find_latest_report() with
pagination up to MAX_API_PAGES
- Add xls_parse.py: detect_file_format() (magic bytes), xls_to_rows()
using xlrd for OLE2/BIFF XLS files
- Add extract_ice_aging(): monthly certified stock aging report by
age bucket × port → ice_aging/ landing dir
- Add extract_ice_historical(): 30-year EOM by-port stocks from static
ICE URL → ice_stocks_by_port/ landing dir
- Add xlrd>=2.0.1 (parse XLS), xlwt>=1.3.0 (dev, test fixtures)
- Add SQLMesh raw + foundation models for both new datasets
- Add ice_aging_glob(), ice_stocks_by_port_glob() macros
- Add extract_ice_aging + extract_ice_historical pipeline entries
- Add 12 unit tests (format detection, XLS roundtrip, API mock, CSV output)
Seed files (data/landing/ice_aging/seed/ and ice_stocks_by_port/seed/)
must be created locally — data/ is gitignored.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
extract: wrap response.content in BytesIO before passing to
normalize_zipped_csv, and call .read() on the returned BytesIO before
write_bytes (two bugs: wrong type in, wrong type out)
sqlmesh: {{ var() }} inside SQL string literals is not substituted by
SQLMesh's Jinja (SQL parser treats them as opaque strings). Replace with
a @psd_glob() macro that evaluates LANDING_DIR at render time and returns
a quoted glob path string.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Remove distributed R2/Iceberg/SSH pipeline architecture in favor of
local subprocess execution with NVMe storage. Landing data backed up
to R2 via rclone timer.
- Strip Iceberg catalog, httpfs, boto3, paramiko, prefect, pyarrow
- Pipelines run via subprocess.run() with bounded timeouts
- Extract writes to {LANDING_DIR}/psd/{year}/{month}/{etag}.csv.gzip
- SQLMesh reads LANDING_DIR variable, writes to DUCKDB_PATH
- Delete unused provider stubs (ovh, scaleway, oracle)
- Add rclone systemd timer for R2 backup every 6h
- Update supervisor to run pipelines with env vars
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
## Changes
1. **Support ESC environment variable names**
- Fallback to R2_ADMIN_ACCESS_KEY_ID if R2_ACCESS_KEY not set
- Fallback to R2_ADMIN_SECRET_ACCESS_KEY if R2_SECRET_KEY not set
- Allows script to work with Pulumi ESC (beanflows/prod) variables
2. **Use landing bucket path**
- Changed R2 path from `psd/{etag}.zip` to `landing/psd/{etag}.zip`
- All extracted data goes to landing bucket for consistent organization
3. **Updated Pulumi ESC environment**
- Added R2_BUCKET=beanflows-data-prod
- Fixed R2_ENDPOINT to remove bucket path (now just account URL)
## Testing
- ✅ R2 upload works: Uploaded to landing/psd/316039e2612edc1_0.zip
- ✅ R2 deduplication works: Skips upload if file exists
- ✅ Local mode still works without credentials
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
## Key Changes
1. **Simplified extraction logic**
- Changed from downloading 220+ historical archives to checking only latest available month
- Tries current month and falls back up to 3 months (handles USDA publication lag)
- Architecture advisor insight: ETags naturally deduplicate, historical year/month structure was unnecessary
2. **Flat storage structure**
- Old: `data/{year}/{month}/{etag}.zip`
- New: `data/{etag}.zip` (local) or `psd/{etag}.zip` (R2)
- Migrated 226 existing files to flat structure
3. **Dual storage modes**
- **Local mode**: Downloads to local directory (development)
- **R2 mode**: Uploads to Cloudflare R2 (production)
- Mode determined by presence of R2 environment variables
- Added boto3 dependency for S3-compatible R2 API
4. **Updated raw SQLMesh model**
- Changed pattern from `**/*.zip` to `*.zip` to match flat structure
## Benefits
- Simpler: Single file check instead of 220+ URL attempts
- Efficient: ETag-based deduplication works naturally
- Flexible: Supports both local dev and production R2 storage
- Maintainable: Removed unnecessary complexity
## Testing
- ✅ Local extraction works and respects ETags
- ✅ Falls back correctly when current month unavailable
- ✅ Linting passes
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
