Package org.apache.lucene.codecs.lucene912

Class Lucene912PostingsFormat

  • All Implemented Interfaces:
    NamedSPILoader.NamedSPI
    public final class Lucene912PostingsFormat
extends PostingsFormat
    Lucene 9.12 postings format, which encodes postings in packed integer blocks for fast decode.

    Basic idea:

    • Packed Blocks and VInt Blocks:

      In packed blocks, integers are encoded with the same bit width (packed format): the block size (i.e. number of integers inside block) is fixed (currently 128). Additionally blocks that are all the same value are encoded in an optimized way.

      In VInt blocks, integers are encoded as VInt: the block size is variable.

    • Block structure:

      When the postings are long enough, Lucene912PostingsFormat will try to encode most integer data as a packed block.

      Take a term with 259 documents as an example, the first 256 document ids are encoded as two packed blocks, while the remaining 3 are encoded as one VInt block.

      Different kinds of data are always encoded separately into different packed blocks, but may possibly be interleaved into the same VInt block.

      This strategy is applied to pairs: <document number, frequency>, <position, payload length>, <position, offset start, offset length>, and <position, payload length, offsetstart, offset length>.

    • Skipdata:

      Skipdata is interleaved with blocks on 2 levels. Level 0 skip data is interleaved between every packed block. Level 1 skip data is interleaved between every 32 packed blocks.

    • Positions, Payloads, and Offsets:

      A position is an integer indicating where the term occurs within one document. A payload is a blob of metadata associated with current position. An offset is a pair of integers indicating the tokenized start/end offsets for given term in current position: it is essentially a specialized payload.

      When payloads and offsets are not omitted, numPositions==numPayloads==numOffsets (assuming a null payload contributes one count). As mentioned in block structure, it is possible to encode these three either combined or separately.

      In all cases, payloads and offsets are stored together. When encoded as a packed block, position data is separated out as .pos, while payloads and offsets are encoded in .pay (payload metadata will also be stored directly in .pay). When encoded as VInt blocks, all these three are stored interleaved into the .pos (so is payload metadata).

      With this strategy, the majority of payload and offset data will be outside .pos file. So for queries that require only position data, running on a full index with payloads and offsets, this reduces disk pre-fetches.

    Files and detailed format:

    Term Dictionary

    The .tim file contains the list of terms in each field along with per-term statistics (such as docfreq) and pointers to the frequencies, positions, payload and skip data in the .doc, .pos, and .pay files. See Lucene90BlockTreeTermsWriter for more details on the format.

    NOTE: The term dictionary can plug into different postings implementations: the postings writer/reader are actually responsible for encoding and decoding the PostingsHeader and TermMetadata sections described here:

    • PostingsHeader --> Header, PackedBlockSize
    • TermMetadata --> (DocFPDelta|SingletonDocID), PosFPDelta?, PosVIntBlockFPDelta?, PayFPDelta?
    • Header, --> IndexHeader
    • PackedBlockSize, SingletonDocID --> VInt
    • DocFPDelta, PosFPDelta, PayFPDelta, PosVIntBlockFPDelta --> VLong
    • Footer --> CodecFooter

    Notes:

    • Header is a IndexHeader storing the version information for the postings.
    • PackedBlockSize is the fixed block size for packed blocks. In packed block, bit width is determined by the largest integer. Smaller block size result in smaller variance among width of integers hence smaller indexes. Larger block size result in more efficient bulk i/o hence better acceleration. This value should always be a multiple of 64, currently fixed as 128 as a tradeoff. It is also the skip interval used to accelerate DocIdSetIterator.advance(int).
    • DocFPDelta determines the position of this term's TermFreqs within the .doc file. In particular, it is the difference of file offset between this term's data and previous term's data (or zero, for the first term in the block).On disk it is stored as the difference from previous value in sequence.
    • PosFPDelta determines the position of this term's TermPositions within the .pos file. While PayFPDelta determines the position of this term's <TermPayloads, TermOffsets?> within the .pay file. Similar to DocFPDelta, it is the difference between two file positions (or neglected, for fields that omit payloads and offsets).
    • PosVIntBlockFPDelta determines the position of this term's last TermPosition in last pos packed block within the .pos file. It is synonym for PayVIntBlockFPDelta or OffsetVIntBlockFPDelta. This is actually used to indicate whether it is necessary to load following payloads and offsets from .pos instead of .pay. Every time a new block of positions are to be loaded, the PostingsReader will use this value to check whether current block is packed format or VInt. When packed format, payloads and offsets are fetched from .pay, otherwise from .pos. (this value is neglected when total number of positions i.e. totalTermFreq is less or equal to PackedBlockSize).
    • SingletonDocID is an optimization when a term only appears in one document. In this case, instead of writing a file pointer to the .doc file (DocFPDelta), and then a VIntBlock at that location, the single document ID is written to the term dictionary.
    Term Index

    The .tip file contains an index into the term dictionary, so that it can be accessed randomly. See Lucene90BlockTreeTermsWriter for more details on the format.

    Frequencies and Skip Data

    The .doc file contains the lists of documents which contain each term, along with the frequency of the term in that document (except when frequencies are omitted: IndexOptions.DOCS). Skip data is saved at the end of each term's postings. The skip data is saved once for the entire postings list.

    • docFile(.doc) --> Header, <TermFreqs>TermCount, Footer
    • Header --> IndexHeader
    • TermFreqs --> <PackedBlock32> PackedDocBlockNum/32, VIntBlock?
    • PackedBlock32 --> Level1SkipData, <PackedBlock> 32
    • PackedBlock --> Level0SkipData, PackedDocDeltaBlock, PackedFreqBlock?
    • VIntBlock --> <DocDelta[,Freq?]>DocFreq-PackedBlockSize*PackedDocBlockNum
    • Level1SkipData --> DocDelta, DocFPDelta, Skip1NumBytes?, ImpactLength?, Impacts?, PosFPDelta?, NextPosUpto?, PayFPDelta?, NextPayByteUpto?
    • Level0SkipData --> Skip0NumBytes, DocDelta, DocFPDelta, PackedBlockLength, ImpactLength?, Impacts?, PosFPDelta?, NextPosUpto?, PayFPDelta?, NextPayByteUpto?
    • PackedFreqBlock --> PackedInts, uses patching
    • PackedDocDeltaBlock --> PackedInts, does not use patching
    • Footer --> CodecFooter

    Notes:

    • PackedDocDeltaBlock is theoretically generated from two steps:
      1. Calculate the difference between each document number and previous one, and get a d-gaps list (for the first document, use absolute value);
      2. For those d-gaps from first one to PackedDocBlockNum*PackedBlockSizeth, separately encode as packed blocks.
      If frequencies are not omitted, PackedFreqBlock will be generated without d-gap step.
    • VIntBlock stores remaining d-gaps (along with frequencies when possible) with a format that encodes DocDelta and Freq:

      DocDelta: if frequencies are indexed, this determines both the document number and the frequency. In particular, DocDelta/2 is the difference between this document number and the previous document number (or zero when this is the first document in a TermFreqs). When DocDelta is odd, the frequency is one. When DocDelta is even, the frequency is read as another VInt. If frequencies are omitted, DocDelta contains the gap (not multiplied by 2) between document numbers and no frequency information is stored.

      For example, the TermFreqs for a term which occurs once in document seven and three times in document eleven, with frequencies indexed, would be the following sequence of VInts:

      15, 8, 3

      If frequencies were omitted (IndexOptions.DOCS) it would be this sequence of VInts instead:

      7,4

    • PackedDocBlockNum is the number of packed blocks for current term's docids or frequencies. In particular, PackedDocBlockNum = floor(DocFreq/PackedBlockSize)
    • On skip data, DocDelta is the delta between the last doc of the previous block - or -1 if there is no previous block - and the last doc of this block. This helps know by how much the doc ID should be incremented in case the block gets skipped.
    • Skip0Length is the length of skip data at level 0. Encoding it is useful when skip data is never needed to quickly skip over skip data, e.g. if only using nextDoc(). It is also used when only the first fields of skip data are needed, in order to skip over remaining fields without reading them.
    • ImpactLength and Impacts are only stored if frequencies are indexed.
    • Since positions and payloads are also block encoded, the skip should skip to related block first, then fetch the values according to in-block offset. PosFPSkip and PayFPSkip record the file offsets of related block in .pos and .pay, respectively. While PosBlockOffset indicates which value to fetch inside the related block (PayBlockOffset is unnecessary since it is always equal to PosBlockOffset). Same as DocFPSkip, the file offsets are relative to the start of current term's TermFreqs, and stored as a difference sequence.
    • PayByteUpto indicates the start offset of the current payload. It is equivalent to the sum of the payload lengths in the current block up to PosBlockOffset
    • ImpactLength is the total length of CompetitiveFreqDelta and CompetitiveNormDelta pairs. CompetitiveFreqDelta and CompetitiveNormDelta are used to safely skip score calculation for uncompetitive documents; See CompetitiveImpactAccumulator for more details.
    Positions

    The .pos file contains the lists of positions that each term occurs at within documents. It also sometimes stores part of payloads and offsets for speedup.

    • PosFile(.pos) --> Header, <TermPositions> TermCount, Footer
    • Header --> IndexHeader
    • TermPositions --> <PackedPosDeltaBlock> PackedPosBlockNum, VIntBlock?
    • VIntBlock --> <PositionDelta[, PayloadLength?], PayloadData?, OffsetDelta?, OffsetLength?>PosVIntCount
    • PackedPosDeltaBlock --> PackedInts
    • PositionDelta, OffsetDelta, OffsetLength --> VInt
    • PayloadData --> bytePayLength
    • Footer --> CodecFooter

    Notes:

    • TermPositions are order by term (terms are implicit, from the term dictionary), and position values for each term document pair are incremental, and ordered by document number.
    • PackedPosBlockNum is the number of packed blocks for current term's positions, payloads or offsets. In particular, PackedPosBlockNum = floor(totalTermFreq/PackedBlockSize)
    • PosVIntCount is the number of positions encoded as VInt format. In particular, PosVIntCount = totalTermFreq - PackedPosBlockNum*PackedBlockSize
    • The procedure how PackedPosDeltaBlock is generated is the same as PackedDocDeltaBlock in chapter Frequencies and Skip Data.
    • PositionDelta is, if payloads are disabled for the term's field, the difference between the position of the current occurrence in the document and the previous occurrence (or zero, if this is the first occurrence in this document). If payloads are enabled for the term's field, then PositionDelta/2 is the difference between the current and the previous position. If payloads are enabled and PositionDelta is odd, then PayloadLength is stored, indicating the length of the payload at the current term position.
    • For example, the TermPositions for a term which occurs as the fourth term in one document, and as the fifth and ninth term in a subsequent document, would be the following sequence of VInts (payloads disabled):

      4, 5, 4

    • PayloadData is metadata associated with the current term position. If PayloadLength is stored at the current position, then it indicates the length of this payload. If PayloadLength is not stored, then this payload has the same length as the payload at the previous position.
    • OffsetDelta/2 is the difference between this position's startOffset from the previous occurrence (or zero, if this is the first occurrence in this document). If OffsetDelta is odd, then the length (endOffset-startOffset) differs from the previous occurrence and an OffsetLength follows. Offset data is only written for IndexOptions.DOCS_AND_FREQS_AND_POSITIONS_AND_OFFSETS.
    Payloads and Offsets

    The .pay file will store payloads and offsets associated with certain term-document positions. Some payloads and offsets will be separated out into .pos file, for performance reasons.

    • PayFile(.pay): --> Header, <TermPayloads?, TermOffsets?> TermCount, Footer
    • Header --> IndexHeader
    • TermPayloads --> <PackedPayLengthBlock, SumPayLength, PayData> PackedPayBlockNum
    • TermOffsets --> <PackedOffsetStartDeltaBlock, PackedOffsetLengthBlock> PackedPayBlockNum
    • PackedPayLengthBlock, PackedOffsetStartDeltaBlock, PackedOffsetLengthBlock --> PackedInts
    • SumPayLength --> VInt
    • PayData --> byteSumPayLength
    • Footer --> CodecFooter

    Notes:

    • The order of TermPayloads/TermOffsets will be the same as TermPositions, note that part of payload/offsets are stored in .pos.
    • The procedure how PackedPayLengthBlock and PackedOffsetLengthBlock are generated is the same as PackedFreqBlock in chapter Frequencies and Skip Data. While PackedStartDeltaBlock follows a same procedure as PackedDocDeltaBlock.
    • PackedPayBlockNum is always equal to PackedPosBlockNum, for the same term. It is also synonym for PackedOffsetBlockNum.
    • SumPayLength is the total length of payloads written within one block, should be the sum of PayLengths in one packed block.
    • PayLength in PackedPayLengthBlock is the length of each payload associated with the current position.
    WARNING: This API is experimental and might change in incompatible ways in the next release.