Dependency resolution for datasets

depgraph is a tiny (<500 LOC) Python library for expressing networks of datasets and their relationships. In this way, it is superficially similar to Airflow and Luigi, although those tools contain significantly more functionality.

Networks are declared in terms of the relationships (graph edges) between source and target datasets (graph nodes). Target datasets can then report sets of precursor datasets in the correct order. This makes it simple to throw together a build script and construct dependencies, sequentially or with parallelization.

Traditionally, each Dataset is designed to correspond to a file. A DatasetGroup class handles cases where multiple files can be considered a single file (e.g. a binary data file and its XML metadata).

Different kinds of resources, such as database tables, can be used as long as they can be queried to determine whether they exist (how how old they are, in order to tak advantage of age-based incremental building).

When a Dataset requires a different dataset to be built to satisfy its dependencies, it provides a reason, such as:

• the Dataset is missing, and so must be built
• the Dataset is out of date

depgraph is intended to be a reusable component for constructing scientific dataset build tools. Important considerations for such a build tool are that it must:

• permit reproducible analysis
• be documenting so that a workflow can be easily reported
• perform fast rebuilds to enable experimentation

Beyond the standard library, depgraph has no dependencies of its own, so it is easy to include in projects running on a laptop, on a large cluster, or in the cloud. depgraph supports modern Python implementations (Python 3, PyPy3), and works on Linux, OS X, and Windows.

Important parts

Dataset defines an individual data product, represented by a filename, name. Additional keyword arguments may be provided in order to facilitate the build process.

The ancestors of a dataset can be retrieved with Dataset.parents(n), where n is the number of generations to include. n=0 means include only the direct parents, while n=1 includes grandparents. n=-1 includes every ancestor. Dataset.roots() returns the top-level ancestors, i.e. those with no additional parents.

Similarly, Dataset.children(n) yields the descendants of a dataset, if any.

Relationships are defined with Dataset.dependson(obj), where obj is another Dataset instance. Relationships can be defined programmatically to construct large dependency graphs.

A user defined build(dataset, reason) function (name unimportant) takes a dataset and constructs it based on its ancestors and any other attributes of the Dataset. The reason is a Reason object that specifies the motivation for a build step.

The depgraph.buildall() function or Dataset.buildnext() method can be used to obtain ancestor datasets and reason pairs to feed to the build() function. Alternatively, the build() function can be decorated with the buildmanager decorator, which creates a function that automatically constructs a dataset by assembling its dependencies in order (see the examples below).

Complex dependency graphs can be visualized by using the graphviz() function, which returns a DOT language string encoding the visual graph.

Example

Declare a set of dependencies resembling the graph below:

     R0      R1      R2      R3         [raw data]
       \     /       |       |
         DA0         DA1    /
             \      /  \   /
                DB0     DB1
                 \     / |  \
                  \   /  |   \
                   DC0  DC1  DC2        [products]

from depgraph import Dataset, buildmanager

# Define Datasets
# Use an optional keyword `tool` to provide a key instructing our build tool
# how to assemble this product. Here we've used strings, but another pattern
# would be to provide a callback function
R0 = Dataset("data/raw0", tool="read_csv")
R1 = Dataset("data/raw1", tool="read_csv")
R2 = Dataset("data/raw2", tool="database_query")
R3 = Dataset("data/raw3", tool="read_hdf")

DA0 = Dataset("step1/da0", tool="merge_fish_counts")
DA1 = Dataset("step1/da1", tool="process_filter")

DB0 = Dataset("step2/db0", tool="join_counts")
DB1 = Dataset("step2/db1", tool="join_by_date")

DC0 = Dataset("results/dc0", tool="merge_model_obs")
DC1 = Dataset("results/dc1", tool="compute_uncertainty")
DC2 = Dataset("results/dc2", tool="make_plots")

# Declare dependency relationships so that depgraph and determine the order of
# the build
DA0.dependson(R0, R1)
DA1.dependson(R2)
DB0.dependson(DA0, DA1)
DB1.dependson(DA1, R3)
DC0.dependson(DB0, DB1)
DC1.dependson(DB1)
DC2.dependson(DB1)

# Option 1:
# Define a function that builds individual dependencies. The *buildmanager*
# decorator transforms it into a loop that builds all dependencies above a
# target
@buildmanager
def batchbuilder(dependency, reason):
    # [....]
    return exitcode

batchbuilder(DC1)

# Option 2:
# Implement the build loop manually
from depgraph import buildall

def build(dependency, reason):
    # This may have the same logic as `batchbuilder` above, but we
    # will call it directly rather than wrapping it in @buildmanager
    # [....]
    return exitcode

for stage in buildall(DC1):

    # A build stage is a list of dependencies whose own dependencies are met and
    # that are independent, i.e. they can be built in parallel

    for dep, reason in stage:

        # Each target is a dataset with a 'name' attribute and whatever
        # additional keyword arguments where defined with it.
        # The 'reason' is a depgraph.Reason object that codifies why a
        # particular target is necessary (e.g. it's out of date, it's missing
        # and required by a subsequent target, etc.)
        print("Building {0} with {1} because {2}".format(dep.name, dep.tool,
                                                         reason))

        # Call a function or start a subprocess that will result in the
        # target being built and saved to a file
        return_val = build(dep, reason)

        # Perform logging, clean-up, or error handling operations
        # [....]

Changes

0.4

• Performance improvements
• buildall generator function, which is more efficient than repeatedly calling Dataset.buildnext()

0.3

• Cyclic graph detection
• Graphviz export

0.2

• Rewrite, dropping DependencyGraph and making Dataset the primary class

0.1

• First version, copied from depchain module of asputil package