SciDB PartitioningSchema

/**
 * Partitioning schema shows how an array is distributed among the SciDB instances.
 *
 * Guidelines for introducing a new partitioning schema:
 *   - Add to enum PartitioningSchema (right above psMax).
 *   - Modify the doxygen comments in LogicalSG.cpp.
 *   - Modify redistribute() to handle the new partitioning schema.
 *   - Modify std::ostream& operator<<(std::ostream& stream, const RedistributeContext& dist). (See Operator.cpp)
 *   - If the partitioning schema uses extra data:
 *   -    Modify doesPartitioningSchemaHaveOptionalData.
 *   -    Derive a class from PartitioningSchemaData.
 *   -    When modifying redistribute(), consider the extra data for the new partitioning schema.
 */
enum PartitioningSchema
{
    psUninitialized = -1, // e.g. _ps after ArrayDesc::ArrayDesc()
    psMIN = 0,
    psReplication = psMIN,
    psHashPartitioned,
    psLocalInstance,
    psByRow,
    psByCol,
    psUndefined,        // a range of meanings, including sometimes "wildcard"
                        // TODO: replace with psWildcard and others as required
    psGroupby,
    psScaLAPACK,
    // A newly introduced partitioning schema should be added before this line.
    psEND
};


/**
 * defaultPartitioning
 *
 * Most of the code was assuming that arrays are scanned and stored in
 * psHashPartitioned Partitioning (currently a category of ArrayDistribution).
 * Therefore there were many "psHashPartitioned" scatter about the code that
 * logically mean "whatever the default PartitioningSchema is".  We are moving on
 * a path toward generalizing what distributions can be stored.  The first attempt
 * will be to extend tempArrays to psReplication, psByRow, and psByCol.  When that
 * is working we will enable it for dbArrays.  Later we will move on to distributions
 * which require additional parameterizations to be stored, such as the 2D
 * ScaLAPACK.
 *
 * As scaffolding for that effort, we are making the "default" PartitioningSchema
 * configurable.  This will allow us to (1) experiment with performance of alternate
 * distributions for certain workflows and (2) start fixing the query compiler/optimizer
 * to insert SGs without assuming that psHashPartitioned is the universal goal.
 *
 * Then we can add a create array parameterization for particular distributions
 * and drive the query tree's output toward that. (e.g. right now, for iterated spgemm,
 * psByRow and psByCol are far more efficient than psHashed)
 *
 * Once this is working for TEMP arrays, we can allow it for general arrays
 *
 * Once this all works correctly via create statements, we will no longer be managing
 * "ps" outside of ArrayDesc's and ArrayDistributions, this function should no longer be referenced
 * and this scaffolding removed.
 */
inline PartitioningSchema defaultPartitioning()
{
    // you MAY NOT change this from psHashPartitioned at this time
    // this is scaffolding code that is locating all the places where
    // psHashPartitioned is assumed which need to be generalized.
    // there are further changes in the optimizer that must be planned first
    // and then we can decide which PartitioningSchema variable is the
    // definitive one.  Until then, the code will read "defaultParitioning()"
    // and the reader should think "psHashPartitioned"


    return psHashPartitioned;


}