init

docs/source/api/dataframe.rst

@@ -0,0 +1,104 @@
+.. _dataframe:
+
+DataFrame
+=========
+
+DataFrame is the main class in smallpond. It represents a lazily computed, partitioned data set.
+
+A typical workflow looks like this:
+
+.. code-block:: python
+
+   import smallpond
+
+   sp = smallpond.init()
+
+   df = sp.read_parquet("path/to/dataset/*.parquet")
+   df = df.repartition(10)
+   df = df.map("x + 1")
+   df.write_parquet("path/to/output")
+
+Initialization
+--------------
+
+.. autosummary::
+   :toctree: ../generated
+
+   smallpond.init
+
+.. currentmodule:: smallpond.dataframe
+
+.. _loading_data:
+
+Loading Data
+------------
+
+.. autosummary::
+   :toctree: ../generated
+
+   Session.from_items
+   Session.from_arrow
+   Session.from_pandas
+   Session.read_csv
+   Session.read_json
+   Session.read_parquet
+
+.. _partitioning_data:
+
+Partitioning Data
+-----------------
+
+.. autosummary::
+   :toctree: ../generated
+
+   DataFrame.repartition
+
+.. _transformations:
+
+Transformations
+---------------
+
+Apply transformations and return a new DataFrame.
+
+.. autosummary::
+   :toctree: ../generated
+
+   Session.partial_sql
+   DataFrame.map
+   DataFrame.map_batches
+   DataFrame.flat_map
+   DataFrame.filter
+   DataFrame.limit
+   DataFrame.partial_sort
+   DataFrame.random_shuffle
+
+.. _consuming_data:
+
+Consuming Data
+--------------
+
+These operations will trigger execution of the lazy transformations performed on this DataFrame.
+
+.. autosummary::
+   :toctree: ../generated
+
+   DataFrame.count
+   DataFrame.take
+   DataFrame.take_all
+   DataFrame.to_arrow
+   DataFrame.to_pandas
+   DataFrame.write_parquet
+   DataFrame.write_parquet_lazy
+
+Execution
+---------
+
+DataFrames are lazily computed. You can use these methods to manually trigger computation.
+
+.. autosummary::
+   :toctree: ../generated
+
+   DataFrame.compute
+   DataFrame.is_computed
+   DataFrame.recompute
+   Session.wait

docs/source/api/dataset.rst

@@ -0,0 +1,28 @@
+.. currentmodule:: smallpond.logical.dataset
+
+Dataset
+=======
+
+Dataset represents a collection of files.
+
+To create a dataset:
+
+.. code-block:: python
+
+   dataset = ParquetDataSet("path/to/dataset/*.parquet")
+
+DataSets
+--------
+
+.. autosummary::
+   :toctree: ../generated
+
+   DataSet
+   FileSet
+   ParquetDataSet
+   CsvDataSet
+   JsonDataSet
+   ArrowTableDataSet
+   PandasDataSet
+   PartitionedDataSet
+   SqlQueryDataSet

docs/source/api/execution.rst

@@ -0,0 +1,85 @@
+.. currentmodule:: smallpond.execution
+
+.. _execution:
+
+Execution
+=========
+
+Submit a Job
+------------
+
+After constructing the LogicalPlan, you can use the JobManager to create a Job in the cluster to execute it. However, in most cases, you only need to use the Driver as the entry point of the entire script and then submit the plan. The Driver is a simple wrapper around the JobManager. It reads the configuration from the command line arguments and passes it to the JobManager.
+
+.. code-block:: python
+
+   from smallpond.execution.driver import Driver
+
+   if __name__ == "__main__":
+      driver = Driver()
+      # add your own arguments
+      driver.add_argument("-i", "--input_paths", nargs="+")
+      driver.add_argument("-n", "--npartitions", type=int, default=10)
+      # build and run logical plan
+      plan = my_pipeline(**driver.get_arguments())
+      driver.run(plan)
+
+
+.. autosummary::
+   :toctree: ../generated
+
+   ~driver.Driver
+   ~manager.JobManager
+
+Scheduler and Executor
+----------------------
+
+Scheduler and Executor are lower-level APIs. They are directly responsible for scheduling and executing tasks, respectively. Generally, users do not need to use them directly.
+
+.. autosummary::
+   :toctree: ../generated
+
+   ~scheduler.Scheduler
+   ~executor.Executor
+
+.. _platform:
+
+Customize Platform
+------------------
+
+Smallpond supports user-defined task execution platforms. A Platform includes methods for submitting jobs and a series of default configurations. By default, smallpond automatically detects the current environment and selects the most suitable platform. If it cannot detect one, it uses the default platform.
+
+You can specify a built-in platform via parameters:
+
+.. code-block:: bash
+
+   # run with your platform
+   python script.py --platform mpi
+
+
+Or implement your own Platform class:
+
+.. code-block:: python
+
+   # path/to/my/platform.py
+   from smallpond.platform import Platform
+
+   class MyPlatform(Platform):
+      def start_job(self, ...) -> List[str]:
+         ...
+
+.. code-block:: bash
+
+   # run with your platform
+   # if using Driver
+   python script.py --platform path.to.my.platform
+
+   # if using smallpond.init
+   SP_PLATFORM=path.to.my.platform python script.py
+
+.. currentmodule:: smallpond
+
+.. autosummary::
+   :toctree: ../generated
+
+   ~platform.Platform
+   ~platform.MPI

docs/source/api/nodes.rst

@@ -0,0 +1,89 @@
+.. currentmodule:: smallpond.logical.node
+
+.. _nodes:
+
+Nodes
+=====
+
+Nodes represent the fundamental building blocks of a data processing pipeline. Each node encapsulates a specific operation or transformation that can be applied to a dataset.
+Nodes can be chained together to form a logical plan, which is a directed acyclic graph (DAG) of nodes that represent the overall data processing workflow.
+
+A typical workflow to create a logical plan is as follows:
+
+.. code-block:: python
+
+   # Create a global context
+   ctx = Context()
+
+   # Create a dataset
+   dataset = ParquetDataSet("path/to/dataset/*.parquet")
+
+   # Create a data source node
+   node = DataSourceNode(ctx, dataset)
+
+   # Partition the data
+   node = DataSetPartitionNode(ctx, (node,), npartitions=2)
+
+   # Create a SQL engine node to transform the data
+   node = SqlEngineNode(ctx, (node,), "SELECT * FROM {0}")
+
+   # Create a logical plan from the root node
+   plan = LogicalPlan(ctx, node)
+
+You can then create tasks from the logical plan, see :ref:`tasks`.
+
+Notable properties of Node:
+
+1. Nodes are partitioned. Each Node generates a series of tasks, with each task processing one partition of data.
+2. The input and output of a Node are a series of partitioned Datasets. A Node may write data to shared storage and return a new Dataset, or it may simply recombine the input Datasets.
+
+Context
+-------
+
+.. autosummary::
+   :toctree: ../generated
+
+   Context
+   NodeId
+
+LogicalPlan
+-----------
+
+.. autosummary::
+   :toctree: ../generated
+
+   LogicalPlan
+   LogicalPlanVisitor
+   .. Planner
+
+Nodes
+-----
+
+.. autosummary::
+   :toctree: ../generated
+
+   Node
+   DataSetPartitionNode
+   ArrowBatchNode
+   ArrowComputeNode
+   ArrowStreamNode
+   ConsolidateNode
+   DataSinkNode
+   DataSourceNode
+   EvenlyDistributedPartitionNode
+   HashPartitionNode
+   LimitNode
+   LoadPartitionedDataSetNode
+   PandasBatchNode
+   PandasComputeNode
+   PartitionNode
+   ProjectionNode
+   PythonScriptNode
+   RangePartitionNode
+   RepeatPartitionNode
+   RootNode
+   ShuffleNode
+   SqlEngineNode
+   UnionNode
+   UserDefinedPartitionNode
+   UserPartitionedDataSourceNode

docs/source/api/tasks.rst

@@ -0,0 +1,76 @@
+.. currentmodule:: smallpond.execution.task
+
+.. _tasks:
+
+Tasks
+=====
+
+.. code-block:: python
+
+   # create a runtime context
+   runtime_ctx = RuntimeContext(JobId.new(), data_root)
+   runtime_ctx.initialize(socket.gethostname(), cleanup_root=True)
+
+   # create a logical plan
+   plan = create_logical_plan()
+
+   # create an execution plan
+   planner = Planner(runtime_ctx)
+   exec_plan = planner.create_exec_plan(plan)
+
+You can then execute the tasks in a scheduler, see :ref:`execution`.
+
+RuntimeContext
+--------------
+
+.. autosummary::
+   :toctree: ../generated
+
+   RuntimeContext
+   JobId
+   TaskId
+   TaskRuntimeId
+   PartitionInfo
+   PerfStats
+
+
+ExecutionPlan
+-------------
+
+.. autosummary::
+   :toctree: ../generated
+
+   ExecutionPlan
+
+
+Tasks
+-----
+
+.. autosummary::
+   :toctree: ../generated
+
+   Task
+   ArrowBatchTask
+   ArrowComputeTask
+   ArrowStreamTask
+   DataSinkTask
+   DataSourceTask
+   EvenlyDistributedPartitionProducerTask
+   HashPartitionArrowTask
+   HashPartitionDuckDbTask
+   HashPartitionTask
+   LoadPartitionedDataSetProducerTask
+   MergeDataSetsTask
+   PandasBatchTask
+   PandasComputeTask
+   PartitionConsumerTask
+   PartitionProducerTask
+   ProjectionTask
+   PythonScriptTask
+   RangePartitionTask
+   RepeatPartitionProducerTask
+   RootTask
+   SplitDataSetTask
+   SqlEngineTask
+   UserDefinedPartitionProducerTask
+