Package it.unimi.di.big.mg4j.tool

Class PartitionDocumentally

  • public class PartitionDocumentally
extends Object
    Partitions an index documentally.

    A global index is partitioned documentally by providing a DocumentalPartitioningStrategy that specifies a destination local index for each document, and a local document pointer. The global index is scanned, and the postings are partitioned among the local indices using the provided strategy. For instance, a ContiguousDocumentalStrategy divides an index into blocks of contiguous documents.

    Since each local index contains a (proper) subset of the original set of documents, it contains in general a (proper) subset of the terms in the global index. Thus, the local term numbers and the global term numbers will not in general coincide. As a result, when a set of local indices is accessed transparently as a single index using a DocumentalCluster, a call to Index.documents(long) will throw an UnsupportedOperationException, because there is no way to map the global term numbers to local term numbers.

    On the other hand, a call to Index.documents(CharSequence) will be passed each local index to build a global iterator. To speed up this phase for not-so-frequent terms, when partitioning an index you can require the construction of Bloom filters that will be used to try to avoid inquiring indices that do not contain a term. The precision of the filters is settable.

    The property file will use a DocumentalMergedCluster unless you provide a ContiguousDocumentalStrategy, in which case a DocumentalConcatenatedCluster will be used instead. Note that there might be other cases in which the latter is adapt, in which case you can edit manually the property file.

    Important: this class just partitions the index. No auxiliary files (most notably, term maps or prefix maps) will be generated. Please refer to a StringMap implementation (e.g., ShiftAddXorSignedStringMap or ImmutableExternalPrefixMap).

    Warning: variable quanta are not supported by this class, as it is impossible to predict accurately the number of bits used for positions when partitioning documentally. If you want to use variable quanta, use a simple interleaved index without skips as an intermediate step, and pass it through Combine.

    Sizes

    Partitioning the file containing document sizes is a tricky issue. For the time being this class implements a very simple policy: if DocumentalPartitioningStrategy.numberOfDocuments(int) returns the number of documents of the global index, the size file for a local index is generated by replacing all sizes of documents not belonging to the index with a zero. Otherwise, the file is generated by appending in order the sizes of the documents belonging to the index. This simple strategy works well with contiguous splitting and with splittings that do not change the document numbers (e.g., the inverse operation of a Merge). However, more complex splittings might give rise to inconsistent size files.

    Write-once output and distributed index partitioning

    Please see PartitionLexically—the same comments apply.
    Since:
    1.0.1
    Author:
    Alessandro Arrabito, Sebastiano Vigna