Data, metadata, and configuration

Many, varied kinds of data are used to prepare and modify MESSAGEix-GLOBIOM scenarios. Other data are produced by code as incidental or final output.

These can be categorized in several ways. One is by the purpose they serve:

  • data—actual numerical values—used or produced by code,

  • metadata, information describing where data is, how to manipulate it, how it is structured, etc.;

  • configuration that otherwise affects how code works.

Another is by whether the data are input, output, or both.

This page describes how to store and handle such files in message_ix_models and message_data. [1]

Choose locations for data

These are listed in order of preference.

(1) Not in message_ix_models

Data that are available from public, stable sources should not be added to the message_ix_models repository. Instead:

  1. Fetch the code from their original location. If possible, this should be done by extending or using message_ix_models.util.pooch.

  2. If message_ix_models relies on certain adjustments to the data, do not commit the adjusted data. Instead:

    1. Commit code that performs the adjustments. This makes methods for data transformation (and any assumptions involved) transparent.

    2. If necessary, cache the result—see below.

(2) message_ix_models/data/

  • Files in this directory are public.

  • In standard Python terms, these are “package data”.

  • This is the preferred location for:

    • General-purpose metadata for the MESSAGEix-GLOBIOM base global model or variants.

    • Configuration.

    • Data for publicized model variants and completed/published projects.

  • These files are packaged, published, and installable from PyPI with message_ix_modelsunless specifically excluded via (see Large/binary input data, below).

  • These data can be reached with package_data_path(), load_package_data(), or other, more specialized code.

  • Documentation files like doc/pkg-data/*.rst describe the contents of these files, and appear in the automatically-built documentation. For example: Node code lists.

(3) data/ directory in the message_data repository

  • Files in this directory are private and not installable from PyPI (because message_data is not packaged for or installable from PyPI).

  • This is the preferred location for:

    • Data for model variants and projects under current development.

    • Specific data files that cannot (currently, or ever) be made public, for instance because of restrictive licenses.

  • These data can be reached with private_data_path(), load_private_data() or other, more specialized code.

(4) Other, system-specific (“local”) directories

These are the preferred location for:

  • Outputs, such as data or plot files generated by reporting.

  • Data files not distributable with message_ix_models, for instance those with access conditions (registration, payment, etc.).

  • Caches: temporary data files used to speed up other code by avoiding repeat of slow operations.

These kinds of data must not be committed to message_ix_models. Caches and output should not be committed to message_data.

(4A) Local data

Each user may configure a location for these data, appropriate to their system, and then use Context.get_local_path() and/or local_data_path() to construct paths under this directory.

This setting can be made in multiple ways. From lowest to highest precedence:

  1. The default location is the current working directory: the directory in which the mix-models Command-line interface is invoked, or in which Python code is run that imports and uses message_ix_models.

  2. The ixmp configuration file setting message local data.

  3. The MESSAGE_LOCAL_DATA environment variable.

  4. The --local-data CLI option and related options such as the --output option to the report command.

  5. Code that directly modifies the local_data setting on Context.

This location should be outside the Git-controlled directories for message_ix_models or message_data. In other words, users should at least use (2) or (3) to specify such directories. If not, they may use .gitignore files to hide these from Git.

(4B) Cache data

Code should use platformdirs.user_cache_path() to identify a system-specific path to a cache directory. For example:

from platformdirs import user_cache_path

# Always use "message-ix-models" as the `appname` parameter
ucp = user_cache_path("message-ix-models")

# Construct the sub-directory for the current module
dir_ = ucp.joinpath("my-project", "subdir")
dir_.mkdir(parents=True, exist_ok=True)

# Construct a file path within this directory
p = dir_.joinpath("data-file-name.csv")

General guidelines

Always consider: “Will this code work on another researcher’s computer?”

Prefer text formats

…such as CSV, over binary formats like Excel. CSV files up to several thousand lines are compressed by Git automatically, and Git can handle diffs to these files easily.

Do not hard-code paths

Data stored with (2–4) above can be retrieved with the utility functions mentioned, instead of hard-coded paths.

For system-specific paths (4) only, get a Context object and use it to get an appropriate Path object pointing to a file:

# Store a base path
project_path = context.get_local_path("myproject", "output")

# Use the Path object to generate a subpath
run_id = "foo"
output_file = project_path.joinpath("reporting", run_id, "all.xlsx")
Keep input and output data separate

Where possible, use (1–3) above for input data, and (4A) for output data.

Use a consistent scheme for data locations

For a submodule for a specific model variant or project named, for instance, message_ix_models.model.[name] or message_ix_models.project.[name], keep input data in a well-organized directory under:

  • [base]/[name]/ —preferred, flatter,

  • [base]/model/[name]/,

  • [base]/project/[name]/,

  • or similar,

where [base] is (2) or (3), above.

Keep project-specific configuration files in the same locations, or (less preferable) alongside Python code files:

# Located in `message_ix_models/data/`:
config = load_package_data("myproject", "config.yaml")

# Located in `data/` in the message_data repo:
config = load_private_data("myproject", "config.yaml")

# Located in the same directory as the code
config = yaml.safe_load(open(Path(__file__).with_name("config.yaml")))

Use a similar scheme for output data, except under (4A).

Re-use configuration

Configuration to run a set of scenarios or to prepare reported submissions should re-use or extend existing, general-purpose code. Do not duplicate code or configuration. Instead, adjust or selectively overwrite its behaviour via project-specific configuration read from a file.

Large/binary input data

These data, such as Microsoft Excel spreadsheets, must not be committed as ordinary Git objects. This is because the entire file is re-added to the Git history for even small modifications, making it very large (see issue #37).

Instead, use one or more of the following patterns, in order of preference. Whichever pattern is used, code for handling large input data must be in message_ix_models, even if the data itself is private, for instance in message_data or another location.

Fetch directly from a remote source

This corresponds to section (1) above. Preferably, do this via message_ix_models.util.pooch:

  • Extend pooch.SOURCE to store the Internet location, file name(s), and hash(es) of the file(s).

  • Call pooch.fetch() to retrieve the file and cache it locally.

  • Write code in message_ix_models that processes the data into a common format, for instance by subclassing ExoDataSource.

This pattern is preferred because it can be replicated by anyone, and the reference data is public.

This pattern may be applied to:

  • Data published and maintained by others, or

  • Data created by the IIASA ECE program to be used in message_ix_models, such as Zenodo records.

Use Git Large File Storage (LFS)

Git LFS is a Git extension that allows for storing large, binary files without bloating the commit history. Essentially, Git stores a 3-line text file with a hash of the full file, and the full file is stored separately. The IIASA GitHub organization has up to 300 GB of space for such LFS objects.

To use this pattern, simply git add ... and git commit files in an appropriate location (above). New or unusual binary file extensions may require a git lfs command or modification to .gitattributes to ensure they are tracked by LFS and not by Git itself. See the Git LFS documentation at the link above for more detail.

For large files stored in message_ix_models/data/ (2, above) using Git LFS, these:

  • must be added to This avoids including the files in Python distributions published on PyPI.

  • should be added to util.pooch. This allows users who install message_ix_models from PyPI to easily retrieve the data. This usage must be included in the documentation that describes the data files.

Retrieve data from existing databases

These include the same IIASA ENE ixmp databases that are used to store scenarios. Documentation must be provided that ensures this data is reproducible: that is, any original sources and code to create the database used by message_data.

Other patterns

Some other patterns exist, but should not be repeated in new code, and should be migrated to one of the above patterns.

  • SQL queries against a Oracle/JDBC database. See message_data:data-iea (in message_data) and issue #53 for a description of how to replace/simplify this code.


Context objects are used to carry configuration, environment information, and other data between parts of the code. Scripts and user code can also store values in a Context object.

# Get an existing instance of Context. There is always at
# least 1 instance available
c = Context.get_instance()

# Store a value using attribute syntax = 42

# Store a value with spaces in the name using item syntax
c["PROJECT data source"] = "Source A"

# my_function() responds to 'foo' or 'PROJECT data source'

# Store a sub-dictionary of values
c["PROJECT2"] = {"setting A": 123, "setting B": 456}

# Create a subcontext with all the settings of `c`
c2 = deepcopy(c)

# Modify one setting = 43

# Run code with this alternate setting

For the CLI, every command decorated with @click.pass_obj gets a first positional argument context, which is an instance of this class. The settings are populated based on the command-line parameters given to mix-models or (sub)commands.

Top-level settings

These are defined by message_ix_models.Config.

Specific modules for model variants, projects, etc. should:

  • Define a single dataclass to express the configuration options they understand. See for example:

    • model.Config for describing existing models or constructing new models,

    • report.Config for reporting,

    • message_data.model.buildings.Config (for the MESSAGEix-Buildings model variant / linkage).

  • Store this on the Context at a simple key. For example model.Config is stored at context.model or context["model"].

  • Retrieve and respect configuration from existing objects, i.e. only duplicate settings with the same meaning when strictly necessary.

  • Communicate to other modules by setting the appropriate configuration values.

class message_ix_models.Config(local_data: ~pathlib.Path = <factory>, platform_info: ~typing.MutableMapping[str, str] = <factory>, scenario_info: ~typing.MutableMapping[str, str] = <factory>, scenarios: ~typing.List[~message_ix_models.util.scenarioinfo.ScenarioInfo] = <factory>, dest_platform: ~typing.MutableMapping[str, str] = <factory>, dest_scenario: ~typing.MutableMapping[str, str] = <factory>, url: str | None = None, dest: str | None = None, cache_path: str | None = None, debug_paths: ~typing.Sequence[str] = <factory>, dry_run: bool = False, verbose: bool = False)[source]

Top-level configuration for message_ix_models and message_data.

cache_path: str | None = None

Base path for cached data, e.g. as given by the --cache-path CLI option. Default: the directory message-ix-models within the directory given by platformdirs.user_cache_path().

debug_paths: Sequence[str]

Paths of files containing debug outputs. See Context.write_debug_archive().

dest: str | None = None

Like url, used by e.g. clone_to_dest().

dest_platform: MutableMapping[str, str]

Like platform_info, used by e.g. clone_to_dest().

dest_scenario: MutableMapping[str, str]

Like scenario_info, used by e.g. clone_to_dest().

dry_run: bool = False

Whether an operation should be carried out, or only previewed. Different modules will respect dry_run in distinct ways, if at all, and should document behaviour.

local_data: Path

Base path for system-specific data, i.e. as given by the --local-data CLI option or message local data key in the ixmp configuration file.

platform_info: MutableMapping[str, str]

Keyword arguments—especially name—for the ixmp.Platform constructor, from the --platform or --url CLI option.

scenario_info: MutableMapping[str, str]

Keyword arguments—model, scenario, and optionally version—for the ixmp.Scenario constructor, as given by the --model/ --scenario or --url CLI options.

scenarios: List[ScenarioInfo]

Like scenario_info, but a list for operations affecting multiple scenarios.

url: str | None = None

A scenario URL, e.g. as given by the --url CLI option.

verbose: bool = False

Flag for causing verbose output to logs or stdout. Different modules will respect verbose in distinct ways.