Use GitHub Actions
Need to generate test datasets for each feature to establish the expected behavior

Only catchments containing a gauge or upstream of a gauge need to be calibrated. Removing others simplifies computations and makes calibration faster.

Choice of runoff transformer and routing algorithm should be plug and play

Need to distinguish between ensemble and sequential runoff volumes

Allow routing with variable k (and/or x) values

On each iteration

• check if k needs to be changed based on current Q vector
• update k
• recalculate muskingum coefficients
• recalculate the LHS matrix
• calculate the iteration's rhs or b vector (existing)
• solve (existing)

Need to determine the data structure for determining k or x values based on q thresholds

• equations? require recomputing every time. Or only if q is >5% different than the previous Q or Q the last time it was calculate?
• tables? what are the thresholds for changing k? can they be known a priori or must you iteratively solve to reach an equilibrium state when the thresholds stop changing (e.g. return periods)?

Allow options for alternative methods for reading the runoff volumes array.

Runoff volumes array incremental reader generator instead of in memory array

Do you want to read the entire array in memory up front and then route it or do you want to incrementally read from the array/file on disc?

To reduce memory consumption (but maybe be slower?) you can incrementally read from the array if the data are provided in a chunked format that is optimized for it.

Recommend PETSc options
Describe config file options
Establish standard config file names

Write functions which read available inputs and outputs in certain directories, determine the right sets of inputs and outputs, and then returns them or writes them to disc for future executions.

Expand calibration options to both linear and nonlinear options

• config files
• example usage
• api docs
• how to overwrite the function which writes arrays to disc with custom logic

Sometimes you run simulations where the runoffs/volumes represent a sequential and continuous set of time steps. Sometimes its an ensemble of predictions all starting from the same time step. Add a flag which allows the user to specify the condition type.

Add QA checks

• last time step of the previous file is ~1 time step ahead of the next time step
• If ensemble, all runoff time series start at the same date

Current code runs in sequential mode. To switch to ensemble, do not overwrite the saved self.q_init property at the end of the routing computations.

Provide additional options for writing outputs to disc. A single netCDF per runoff file is not always the best option for writing outputs.

Either provide additional formats for writing to disc or else allow the user to give their own function that takes in the results as arguments and then they give their own code for saving.

Create a way for sources and sinks to be created.

Perhaps specify a time series of flows in or out of a river segment.

Use those to create a vector to add (positive/negative flows for flow into or out of river).

Add that vector after each routing computation but before enforcing max/min Q.

Source/sink distinction is probably not necessary. Flows will be added or removed by the sign of the flow.

Now that the linear and nonlinear solvers are working and relatively efficient, reimplement the option to provide a LHS-inverse matrix for the fastest solution paths

Resample the outputs to a timestep coarser than the inflow time steps

need data structure for notating diverging rivers and the percentage of flow to upstream flow to push to each downstream river

Giving an option to force minimum flows to be >=0 should improve accuracy and improve calculation stability/convergence.

Giving option to enforce maximum flows limit to some arbitrary large number can guarantee that flows are not too large to write to disc if there is a convergence issue or miscalculated inflows. Maybe raise warnings if its not too slow to check for large values first.