PandasOnRay Execution#
Queries that perform data transformation, data ingress or data egress using the pandas on Ray execution pass through the Modin components detailed below.
To enable pandas on Ray execution, please refer to the usage section in pandas on Ray.
Data Transformation#
When a user calls any DataFrame API, a query starts forming at the API layer
to be executed at the Execution layer. The API layer is responsible for processing the query appropriately,
for example, determining whether the final result should be a DataFrame or Series object. This layer is also responsible for sanitizing the input to the
PandasQueryCompiler, e.g. validating a parameter from the query
and defining specific intermediate values to provide more context to the query compiler.
The PandasQueryCompiler is responsible for
processing the query, received from the DataFrame API layer,
to determine how to apply it to a subset of the data - either cell-wise or along an axis-wise partition backed by the pandas
storage format. The PandasQueryCompiler maps the query to one of the Core Algebra Operators of
the PandasOnRayDataframe which inherits
generic functionality from the GenericRayDataframe and the PandasDataframe.
PandasOnRay Dataframe implementation#
Modin implements Dataframe, PartitionManager, VirtualPartition (a specific kind of AxisPartition with the capability
to combine smaller partitions into the one “virtual”) and Partition classes specifically for the PandasOnRay execution:
Data Ingress#
Data Egress#
When a user calls any IO function from the modin.pandas.io module, the API layer queries the
FactoryDispatcher which defines a factory specific for
the execution, namely, the PandasOnRayFactory. The factory, in turn,
exposes the PandasOnRayIO class
whose responsibility is to perform a parallel read/write from/to a file.
When reading data from a CSV file, for example, the PandasOnRayIO class forwards
the user query to the _read() method of CSVDispatcher, where the query’s parameters are preprocessed
to check if they are supported by the execution (defaulting to pandas if they are not) and computes some metadata
common for all partitions to be read. Then, the file is split into row chunks, and this data is used to launch remote tasks on the Ray workers
via the deploy() method of RayWrapper.
On each Ray worker, the PandasCSVParser parses data.
After the remote tasks are finished, additional result postprocessing is performed,
and a new query compiler with the data read is returned.
When writing data to a CSV file, for example, the PandasOnRayIO processes
the user query to execute it on Ray workers. Then, the PandasOnRayIO asks the
PandasOnRayDataframe to decompose the data into row-wise partitions
that will be written into the file in parallel in Ray workers.