Workflow & Execution

Overview

The food-opt model uses Snakemake for workflow orchestration. If you have never used Snakemake before, consider having a look at the official tutotial to get familiar with the basic concepts. The workflow follows these main stages:

  1. Downloads (GAEZ, GADM, UN WPP, FAOSTAT)

  2. Preprocessing (regions, resource classes, yields, population, health)

  3. Model Building (PyPSA network construction)

  4. Solving (LP optimization with health costs)

  5. Visualization (plots, maps, CSV exports)

Each stage is defined by Snakemake rules that specify inputs, outputs, and a script or piece of code. Paths shown below use the default roots (processing/, results/, logs/, benchmarks/). These roots are configurable via config.paths.

Validation Hook

Before Snakemake resolves any rules, the workflow/Snakefile uses the onstart hook to run configuration/data validation powered by Pydantic and Pandera. The checks live in workflow/validation/ and currently ensure, for example, that every category in data/curated/food_groups.csv is listed under config.food_groups.included. Add new validators by dropping another module in that package and registering it in workflow/validation/__init__.py—the hook aggregates all errors and aborts the workflow if any check fails.

The complete workflow dependency graph is shown below. Each node represents a Snakemake rule, and edges show dependencies between rules.

Workflow dependency graph

Complete workflow dependency graph showing all Snakemake rules and their relationships

Key Snakemake Rules

Data Preparation Rules

simplify_gadm

  • Input: data/downloads/gadm.gpkg

  • Output: processing/shared/gadm-simplified.gpkg

  • Script: workflow/scripts/simplify_gadm.py

  • Purpose: Simplify administrative boundaries for faster processing

build_regions

  • Input: Simplified GADM

  • Output: processing/{name}/regions.geojson

  • Script: workflow/scripts/build_regions.py

  • Purpose: Cluster administrative units into optimization regions

prepare_population

  • Input: data/downloads/WPP_population.csv.gz

  • Output: processing/{name}/population.csv, processing/{name}/population_age.csv

  • Script: workflow/scripts/prepare_population.py

  • Purpose: Extract population for planning horizon and countries

compute_resource_classes

  • Input: All GAEZ yield rasters, regions

  • Output: processing/{name}/resource_classes.nc

  • Script: workflow/scripts/compute_resource_classes.py

  • Purpose: Define yield quantile classes within each region

aggregate_class_areas

  • Input: Resource classes, suitability rasters, regions

  • Output: processing/{name}/land_area_by_class.csv

  • Script: workflow/scripts/aggregate_class_areas.py

  • Purpose: Compute available land area per (region, class, water, crop)

build_crop_yields

  • Wildcards: {crop} (crop name), {water_supply} (“r” or “i”)

  • Input: Resource classes, GAEZ rasters (yield, suitability, water, growing season)

  • Output: processing/{name}/crop_yields/{crop}_{water_supply}.csv

  • Script: workflow/scripts/build_crop_yields.py

  • Purpose: Aggregate yields by (region, class) for each crop

build_multi_cropping

  • Input: Resource classes, regions, the RES01 multiple-cropping zone rasters, and the required GAEZ RES05 rasters (yield, suitability, growing season start/length, plus water requirement for irrigated variants) for every crop referenced in config.multiple_cropping

  • Output: processing/{name}/multi_cropping/eligible_area.csv (eligible hectares, irrigated water requirement), processing/{name}/multi_cropping/cycle_yields.csv (per-cycle yields)

  • Script: workflow/scripts/build_multi_cropping.py

  • Purpose: Filter pixels by the RES01 class (ignoring relay-only options), confirm crop calendars fit within the year, aggregate eligible hectares to regions/resource classes, and compute per-cycle yields

build_grassland_yields

  • Input: ISIMIP grassland yield NetCDF, resource classes, regions

  • Output: processing/{name}/grassland_yields.csv

  • Script: workflow/scripts/build_grassland_yields.py

  • Purpose: Aggregate grassland yields for grazing production

process_blue_water_availability

  • Input: Water Footprint Network shapefile + Excel workbook

  • Output: processing/{name}/water/blue_water_availability.csv

  • Script: workflow/scripts/process_blue_water_availability.py

  • Purpose: Build monthly basin-level blue water availability

build_region_water_sustainable

  • Input: Blue water availability, regions, crop yields

  • Output: processing/{name}/water/sustainable/monthly_region_water.csv, processing/{name}/water/sustainable/region_growing_season_water.csv

  • Script: workflow/scripts/build_region_water_availability.py

  • Purpose: Allocate basin availability to regions and growing seasons

build_region_water_current_use

  • Input: Huang irrigation withdrawals, regions, crop yields

  • Output: processing/{name}/water/current_use/monthly_region_water.csv, processing/{name}/water/current_use/region_growing_season_water.csv

  • Script: workflow/scripts/process_huang_irrigation_water.py

  • Purpose: Aggregate current irrigation withdrawals to regions and growing seasons

select_water_scenario

  • Input: Scenario-specific water outputs

  • Output: processing/{name}/water/monthly_region_water.csv, processing/{name}/water/region_growing_season_water.csv

  • Purpose: Copy the configured water supply scenario into the unified paths used by the model

build_current_grassland_area

  • Input: Resource classes, land-cover fractions (processing/{name}/luc/lc_masks.nc), regions

  • Output: processing/{name}/luc/current_grassland_area_by_class.csv

  • Script: workflow/scripts/build_current_grassland_area.py

  • Purpose: Derive observed managed grassland area by region/resource class from the same land-cover fractions used for LUC calculations; clamps grazing links during validation runs

build_grazing_only_land

  • Input: Resource classes, land-cover fractions (processing/{name}/luc/lc_masks.nc), rainfed GAEZ suitability rasters

  • Output: processing/{name}/land_grazing_only_by_class.csv

  • Script: workflow/scripts/build_grazing_only_land.py

  • Purpose: Estimate marginal pasture area (grassland that is unsuitable for cropland even after accounting for convertible cropland/forest cover) so grazing can draw from a dedicated land pool without competing with cropland-suitable hectares

prepare_health_costs

  • Input: Regions, DIA health data, population

  • Output: processing/{name}/health/*.csv (risk breakpoints, dose-response, clusters)

  • Script: workflow/scripts/prepare_health_costs.py

  • Purpose: Compute health cluster parameters for DALY calculations

Model Building and Solving

build_model

  • Input: All crop yields, multi-cropping aggregates, grassland yields, land areas, population, water availability, static data files (crops.csv, foods.csv with pathway-based processing, etc.)

  • Output: results/{name}/build/model.nc

  • Script: workflow/scripts/build_model.py

  • Purpose: Construct PyPSA network with all components, links, and constraints. Creates multi-output links for regular crops, configured multi-cropping sequences, and crop→food processing pathways defined in foods.csv.

solve_model

  • Input: Built model, health data, food-to-risk mapping

  • Output: results/{name}/solved/model.nc

  • Script: workflow/scripts/solve_model.py

  • Purpose: Add health costs, solve LP, save results. Health impacts are encoded as per-cluster, per-cause Store assets (carrier yll_<cause>) whose level equals million YLL derived from dietary risk. The store cost encodes the value of a life year so the health burden shows up in standard PyPSA statistics and in the objective breakdown plot without custom objective edits.

Visualization Rules

Plots and maps (see workflow/rules/plotting.smk):

  • plot_regions_map: Optimization region boundaries

  • plot_resource_classes_map: Resource class spatial distribution

  • plot_crop_production_map: Land use intensity gridcell map

  • plot_water_value_map: Water shadow prices (economic value)

  • plot_health_impacts: Health risk and baseline maps

  • plot_results: Production, resource usage, objective breakdown

  • plot_food_consumption: Dietary composition

  • plot_crop_use_breakdown: How crops are used (food vs. feed vs. waste)

Execution Commands

Running the Full Workflow

Build, solve, and visualize everything:

tools/smk -j4 --configfile config/my_scenario.yaml all

  • -j4: Use 4 parallel cores (adjust to your CPU count)

  • --configfile config/my_scenario.yaml: Specify which configuration file to use

  • all: Target rule that depends on all major outputs (strictly speaking optional)

This will:

  1. Download datasets (if not already cached)

  2. Process data for configured scenario

  3. Build and solve the model

  4. Generate all plots and exports

Running Specific Stages

Build model only (no solving):

tools/smk -j4 --configfile config/my_scenario.yaml -- results/my_scenario/build/model_scen-default.nc

Solve model:

tools/smk -j4 --configfile config/my_scenario.yaml -- results/my_scenario/solved/model_scen-default.nc

Build all scenarios (no solving):

tools/smk -j4 --configfile config/my_scenario.yaml -- build_all_scenarios

Solve all scenarios (no plots):

tools/smk -j4 --configfile config/my_scenario.yaml -- solve_all_scenarios

Regenerate specific plot:

tools/smk --configfile config/my_scenario.yaml -- results/my_scenario/plots/scen-default/crop_production.pdf

Prepare data without building model:

tools/smk -j4 --configfile config/my_scenario.yaml -- processing/my_scenario/regions.geojson processing/my_scenario/resource_classes.nc

For any of the above targets, Snakemake will first run any other previous rules in order to generate the necessary inputs for the specified target output/rule.

Checking Workflow Status

Dry-run (show what would be executed without running):

tools/smk --configfile config/my_scenario.yaml -n

Dependency graph: See the workflow dependency graph figure at the top of this page. To generate a detailed job-level DAG for a specific configuration:

tools/smk --dag all | dot -Tpdf > dag.pdf

List all rules:

tools/smk --list

Memory Management

The tools/smk Wrapper

It is possible to run the workflow directly with the snakemake command. Food-opt, however, provides a simple shell script, tools/smk, which:

  1. Runs Snakemake in a systemd cgroup with hard memory limit (default 10 GB), killing the process group if memory limit is exceeded

  2. Disables swap to prevent system instability

  3. Sets the -j1 argument (running only one job at a time) by default unless the user sets the -j<n> option explicitly.

  4. When SMK_MEM_MAX is set, forwards it as --resources mem_mb=<...> so Snakemake scheduling respects the same memory ceiling.

Default memory limit: 10 GB (configurable via SMK_MEM_MAX environment variable)

Override memory limit:

SMK_MEM_MAX=12G tools/smk -j4 all

Parallelization

Snakemake automatically runs rules concurrently (e.g., downloading multiple GAEZ files, processing yields for different crops), depending on the configured number of parallel rules allowed. This is set with the -j<n> option, where n is the number of parallel rules. Note that individual rules (such as the model solving rule) may use more than one processor core.

Snakemake automatically detects dependencies and runs tasks in correct order.

Incremental Development

Workflow philosophy: Snakemake tracks file modification times and only reruns rules whose inputs changed. This includes rule input files, the script associated with the rule as well as rule parameters (relevant configuration sections).

Example workflow:

  1. Run full workflow: tools/smk -j4 all

  2. Modify crop list in config → only crop yield rules rerun

  3. Modify solver options → only solve_model reruns (build model reused)

  4. Modify visualization script → only plotting rules rerun

Rerun specific rule:

tools/smk -j4 --configfile config/my_scenario.yaml --forcerun solve_model -- results/my_scenario/solved/model_scen-default.nc

Mark all existing outputs as up to date (preventing rules from being run due to modification times, etc.):

tools/smk --configfile config/my_scenario.yaml --touch

Remote Solving (Experimental)

Warning

Remote solving is experimental. It works for the common case but has rough edges around SSH reliability, error reporting, and edge-case recovery. Expect to troubleshoot SSH/rsync issues yourself when first setting things up.

When models are too large to solve on a laptop, the remote_solve feature delegates only the solve_model step to a remote machine (typically an HPC login node with SLURM) while keeping model building, analysis, and plotting local. The workflow transparently syncs inputs, submits the solve, polls for completion, and pulls results back.

Remote Setup

On the remote host:

  1. Clone this repository (or create the workdir directory).

  2. Install pixi and run pixi install (or pixi install -e gurobi if you want Gurobi).

  3. Ensure the tools/smk wrapper is executable.

  4. Verify you can run tools/smk --help from the remote workdir.

On the local machine:

  1. Ensure passwordless SSH to the remote host works (key-based auth, or an active ControlMaster session). The workflow makes many SSH/rsync calls and cannot handle interactive password prompts.

  2. An SSH ControlMaster (configured in ~/.ssh/config) is strongly recommended to avoid repeated connection setup. Example:

    Host mycluster
    HostName login.cluster.example.com
    User myuser
    ControlMaster auto
    ControlPath ~/.ssh/sockets/%r@%h-%p
    ControlPersist 4h
    ServerAliveInterval 60

    Create the sockets directory: mkdir -p ~/.ssh/sockets.

  3. If the remote host requires two-factor authentication, open an SSH connection manually before starting the workflow so the ControlMaster session is active.

Configuration

Enable remote solving in your config file:

remote_solve:
  enabled: true
  host: "mycluster"           # SSH host or alias from ~/.ssh/config
  workdir: "~/food-opt"       # Remote project root
  pixi_env: "gurobi"          # Remote pixi environment (passed to tools/smk -e)
  use_slurm: true             # Submit via sbatch (false = direct SSH execution)
  slurm_account: "myaccount"  # SLURM --account
  slurm_partition: "normal"   # SLURM --partition
  local_scenarios: ["baseline"]  # Scenarios to always solve locally

See Configuration for the full list of remote_solve options (sync_workflow, sync_pixi_files, ssh_options, rsync_options, preflight_check).

How It Works

The workflow uses three rules, executed in order:

  1. sync_remote_workspace (once per config): creates the remote workdir, syncs workflow/, config/, and tools/smk via rsync, and writes a config snapshot with remote_solve.enabled: false (to prevent recursive remote dispatch on the remote end).

  2. submit_remote_solve (per scenario): syncs the built model and solve inputs, then either submits an sbatch job (if use_slurm: true) or runs Snakemake directly over SSH (blocking).

  3. collect_remote_solve (per scenario): polls the SLURM job for completion via a shared background daemon, re-submits with scaled resources on SLURM failure (OOM, timeout), and pulls the solved network and logs back via rsync.

When use_slurm: true, a single background poll daemon batch-queries all active SLURM jobs in one squeue/sacct call per cycle, rather than each scenario opening its own SSH session. The daemon starts automatically and exits when all jobs finish.

Concurrency is controlled by the remote_solves Snakemake resource (default: 1). Override on the command line:

tools/smk -j4 --resources remote_solves=4 --configfile config/my_scenario.yaml

Interruption

When you press Ctrl-C:

  • Snakemake sends SIGTERM to all collect_remote_solve scripts.

  • The first collect script to handle the signal sends SIGTERM to the poll daemon.

  • The daemon cancels all tracked SLURM jobs in a single scancel call and exits.

  • Each collect script removes its .jobid file so the next workflow run submits fresh.

Troubleshooting

SSH connection issues: The workflow detects stale SSH ControlMaster sockets and attempts automatic recovery. If recovery fails, you will see an error asking you to run ssh <host> manually to re-authenticate.

Checking daemon logs: The poll daemon writes to .snakemake/remote/<config>/jobids/.poll_daemon.log. Check this file if jobs appear stuck.

Orphaned SLURM jobs: If the daemon could not be signalled (e.g. it had already exited), remote jobs may still be running. Check with ssh <host> squeue -u $USER and cancel manually if needed.

“No .jobid files” in daemon log: Normal — the daemon self-terminates after two consecutive empty cycles. It will be restarted on the next workflow run.

Direct SSH mode (use_slurm: false): Runs the solve in a single blocking SSH call. Simpler, but ties up your terminal and does not support SLURM resource scaling on failure. Useful for small models or non-SLURM remote machines.