Release Notes

v0.6.0 (Unreleased)

  • Changed master_clock, master_clock_dim and master_clock_coords to main_clock, main_clock_dim and main_clock_coords and all occurences of master to main in the rest of the codebase. all master... API hooks are still working, but raise a Futurewarning

  • Added access to main clock in initialize step as main_clock_values and as a xr.DataArray: main_clock_array. for refering to the main clock as a dimension label, the placeholder xs.MAIN_CLOCK can be used. This will be set to the main clock when storing the dataset.

  • Changed default fill_value in the zarr stores to maximum dtype value for integer dtypes and np.nan for floating-point variables.

v0.5.0 (26 January 2021)

Breaking changes

  • Fill values are now masked (NA) when when loading the simulation output store as a xarray Dataset (#148). Note that Zarr sets the fill value to 0 by default for numeric data types, so it is highly recommended to explicitly define another fill value in model variable encodings if 0 is expected to be a valid (non-missing) data value, or alternatively use mask_and_scale=False in the decoding options passed to


  • The process() decorator can now be applied on a class already decorated with attr.s() (#150), using apply_attrs=False.

  • Added %create_setup IPython (Jupyter) magic command to auto-generate code cells with a new simulation setup from a given model (#152). The command is available after executing %load_ext xsimlab.ipython.

  • Added an optional cache for on-demand variables (#156). The @compute decorator now has a cache option (deactivated by default).

  • Added group_dict() variable (#159).

  • Added global_ref() variable for model-wise implicit linking of variables in separate processes, based on global names (#160).

  • Added RuntimeSignal for controlling simulation workflow from process runtime methods and/or runtime hook functions (#161).

Bug fixes

  • Fix saving output variables with dtype=object (#145).

  • Fix issues when saving output datasets to disk that were caused by the _FillValue attribute (#148).

  • Ensure that classes given in a Model are all process-decorated, even those which inherit from process-decorated classes (#149).

  • Fix update_clocks() when only the master clock is updated implicitly (#151).

  • Ensure that the finalize simulation stage is always run, even when an exception is raised during the previous stages (#154).

v0.4.1 (17 April 2020)


  • Added xsimlab.Model.cache public property (#125).

  • Parameter input_vars of create_setup() and xarray.Dataset.xsimlab.update_vars() now accepts array-like values with no explicit dimension label(s), in this case those labels are inferred from model variables’ metadata (#126).

  • Single-model parallelism now supports Dask’s multi-processes or distributed schedulers, although this is still limited and rarely optimal (#127).

  • Improved auto-generated docstrings of variables declared in process classes (#130).

Bug fixes

  • Fix running batches of simulations using dask.distributed (#124).

  • Fix rendering of auto-generated docstrings of process classes (#128).

  • Fix tests with attr v20.1.0 (#129).

v0.4.0 (7 April 2020)

This is a big release which includes many exciting new features built on top of great Python libraries. It is now possible to set default, validate or convert model input values just like regular attributes in attrs, save model input/outputs with zarr, run model(s) in parallel using dask, monitor model runs with a tqdm progress bar, and much more!

Breaking changes



  • It is now possible to assign multiple groups to a single variable (#71).

  • The xarray interface may now handle default values that are defined in model variables (#72). A new method xarray.Dataset.xsimlab.reset_vars() allows to (re)populate an input Dataset with variables and their default values. create_setup() has also a new fill_default parameter.

  • Added static variables, i.e., variables that don’t accept time-varying input values (#73).

  • Added support for the validation of variable values (given as inputs and/or set through foreign variables), reusing attr.validate() (#74, #79). Validation is optional and is controlled by the parameter validate added to

  • Check or automatically transpose the dimensions of the variables given in input xarray Datasets to match those defined in model variables (#76). This is optional and controlled by the parameter check_dims added to

  • More consistent dictionary format for output variables in the xarray extension (#85).

  • %-formatting and str.format() code has been converted into formatted string literals (f-strings) (#90).

  • foreign() has been updated so that it sets its description and its metadata from the variable it refers to (#91, #107).

  • The autodoc parameter of the xsimlab.process() decorator now allows to automatically add an attributes section to the docstring of the class to which the decorator is applied, using the metadata of each variable declared in the class (#67).

  • Added in_bounds() and is_subdtype() validators (#87).

  • xsimlab.variable() has now a converter parameter that can be used to convert any input value before (maybe) validating it and setting the variable (#92).

  • Added xsimlab.index() for setting index variables (e.g., coordinate labels). Using the xarray extension, those variables are automatically added in the output Dataset as coordinates (#94).

  • Added simulation runtime hooks (#95). Hooks can be created by using either the runtime_hook() decorator or the RuntimeHook class.

  • Added some useful properties and methods to the xarray.Dataset.xsimlab extension (#103).

  • Save model inputs/outputs using zarr (#102, #111, #113).

  • Added ProgressBar to track simulation progress (#104, #110).

  • Added the ability to easily run batches of simulations using the batch_dim parameter of (#115).

  • Added ‘object’ variables any_object() for sharing arbitrary Python objects between processes (#118).

  • Run one or multiple simulations in parallel using Dask (#119).

Bug fixes

  • Remove attrs 19.2.0 depreciation warning (#68).

  • Fix compatibility with xarray 0.14.1 (#69).

  • Avoid update in-place attributes in original/input xarray Datasets (#101).


  • Switched to GitHub Actions for continuous integration and Codecov for coverage (#86).

v0.3.0 (30 September 2019)

Breaking changes

  • It is now possible to use class inheritance to customize a process without re-writing the class from scratch and without breaking the links between (foreign) variables when replacing the process in a model (#45). Although it should work just fine in most cases, there are potential caveats. This should be considered as an experimental, possibly breaking change.

  • Model.initialize, Model.run_step, Model.finalize_step and Model.finalize have been removed in favor of Model.execute (#59).


  • run_step methods defined in process classes won’t accept anymore current step duration as a positional argument by default. Use the runtime decorator if you need current step duration (and/or other runtime information) inside the method (#59).


  • Ensure that there is no intent conflict between the variables declared in a model. This check is explicit at Model creation and a more meaningful error message is shown when it fails (#57).

  • Added runtime decorator to pass simulation runtime information to the (runtime) methods defined in process classes (#59).

  • Better documentation with a minimal, yet illustrative example based on Game of Life (#61).

  • A class decorated with process can now be instantiated independently of any Model object. This is very useful for testing and debugging (#63).

Bug fixes

  • Fixed compatibility with xarray 0.13.0 (#54).

  • Fixed compatibility with pytest >= 4 (#56).

v0.2.1 (7 November 2018)

Bug fixes

  • Fix an issue after a change in attrs 0.18.2 (#47).

v0.2.0 (9 May 2018)


This release includes a major refactoring of both the internals and the API on how processes and variables are defined and depends on each other in a model. xarray-simlab now uses and extends attrs (#33).

Also, Python 3.4 support has been dropped. It may still work with that version but it is not actively tested anymore and it is not packaged with conda.

Breaking changes

As xarray-simlab is still at an early development stage and hasn’t been adopted “in production” yet (to our knowledge), we haven’t gone through any depreciation cycle, which by the way would have been almost impossible for such a major refactoring. The following breaking changes are effective now!

  • Variable, ForeignVariable and VariableGroup classes have been replaced by variable, foreign and group factory functions (wrappers around attr.ib), respectively.

  • VariableList has been removed and has not been replaced by anything equivalent.

  • DiagnosticVariable has been replaced by on_demand and the diagnostic decorator has been replaced by the variable’s compute decorator.

  • The provided (bool) argument (variable constructors) has been replaced by intent ({'in', 'out', 'inout'}).

  • The allowed_dims argument has been renamed to dims and is now optional (a scalar value is expected by default).

  • The validators argument has been renamed to validator to be consistent with attr.ib.

  • The optional argument has been removed. Variables that don’t require an input value may be defined using a special validator function (see attrs documentation).

  • Variable values are not anymore accessed using three different properties state, rate and change (e.g., Instead, all variables accept a unique value, which one can get/set by simply using the variable name (e.g., Now multiple variables have to be declared for holding different values.

  • Process classes are now defined using the process decorator instead of inheriting from a Process base class.

  • It is not needed anymore to explicitly define whether or not a process is time dependent (it is now deducted from the methods implemented in the process class).

  • Using class Meta inside a process class to define some metadata is not used anymore.

  • Model.input_vars now returns a list of (process_name, variable_name) tuples instead of a dict of dicts. Model.input_vars_dict has been added for convenience (i.e., to get input variables grouped by process as a dictionary).

  • Model.is_input has been removed. Use Model.input_vars instead to check if a variable is a model input.

  • __repr__ has slightly changed for variables, processes and models. Process classes don’t have an .info() method anymore, which has been replaced by the process_info() top-level function. Another helper function variable_info() has been added.

  • In Model.visualize() and, show_variables=True now shows all model variables including inputs. Items of group variables are not shown anymore as nodes.

  • Model.visualize() and now only accept tuples for show_only_variable.

  • For simplicity, Dataset.xsimlab.snapshot_vars has been renamed to output_vars. The corresponding arguments in create_setup and Dataset.xsimlab.update_vars have been renamed accordingly.

  • Values for all model inputs must be provided when creating or updating a setup using create_setup or Dataset.xsimlab.update_vars. this is a regression that will be fixed in the next releases.

  • Argument values for generating clock data in create_setup and Dataset.xsimlab.update_clocks have changed and are now more consistent with how coordinates are set in xarray. Additionally, auto_adjust has been removed (an error is raised instead when clock coordinate labels are not synchronized).

  • Scalar values from a input xarray.Dataset are now converted into scalars (instead of a 0-d numpy array) when setting input model variables during a simulation.


  • The major refactoring in this release should reduce the overhead caused by the indirect access to variable values in process objects.

  • Another benefit of the refactoring is that a process-decorated class may now inherit from other classes (possibly also process-decorated), which allows more flexibility in model customization.

  • By creating read-only properties in specific cases (i.e., when intent='in'), the process decorator applied on a class adds some safeguards to prevent setting variable values where it is not intended.

  • Some more sanity checks have been added when creating process classes.

  • Simulation active and output data r/w access has been refactored internally so that it should be easy to later support alternative data storage backends (e.g., on-disk, distributed).

  • Added Model.dependent_processes property (so far this was not in public API).

  • Added Model.all_vars and Model.all_vars_dict properties that are similar to Model.input_vars and Model.input_vars_dict but return all variable names in the model.

  • input_vars and output_vars arguments of create_setup and Dataset.xsimlab.update_vars now accepts different formats.

  • It is now possible to update only some clocks with Dataset.xsimlab.update_clocks (previously all existing clock coordinates were dropped first).

Regressions (will be fixed in future releases)

  • Although it is possible to set validators, converters and/or default values for variables (this is directly supported by attrs), these are not handled by xarray-simlab yet.

  • Variables don’t accept anymore a dimension that corresponds to their own name. This may be useful, e.g., for sensitivity analysis, but as the latter is not implemented yet this feature has been removed and will be added back in a next release.

  • High-level API for generating clock coordinate data (i.e., start, end, step and auto_adjust arguments) is not supported anymore. This could be added back in a future release in a cleaner form.

v0.1.1 (20 November 2017)

Bug fixes

  • Fix misinterpreted tuples passed as allowed_dims argument of Variable init (#17).

  • Better error message when a Model instance is expected but no object is found or a different object is provided (#13).

v0.1.0 (8 October 2017)

Initial release.