admin/milvus
1
0
Fork 0
mirror of https://gitee.com/milvus-io/milvus.git synced 2025-12-07 01:28:27 +08:00
Code Issues Packages Projects Releases Wiki Activity
milvus/docs/design_docs/segcore/segment_sealed.md
Jael Gu f996674ed4
[skip ci]Fix typo in design doc of segment_sealed (#9566) 
Signed-off-by: Jael Gu <mengjia.gu@zilliz.com>
2021-10-09 17:53:46 +08:00

2.3 KiB
Raw Blame History

SegmentSealed

SegmentSealed has an extra interface rather than segment_inferface:

  1. LoadIndex(loadIndexInfo): load the index. indexInfo contains
    1. FieldId
    2. IndexParams: index paramters in KV structure KV
    3. VecIndex: vector index
  2. LoadFieldData(loadFieldDataInfo): load column data, could be either scalar column or vector column
    1. Note: indexes and vector data for the same column may coexist. Indexes are prioritized in search
  3. DropIndex(fieldId): drop and release exist index of specified field
  4. DropFieldData(fieldId): drop and release exist data for specified field

Search is executable as long as all the column involved in the search are loaded.

SegmentSealedImpl internal data definition

  1. row_count_opt_:

    1. Fill row count when load the first entity
    2. All the other column loaded must match the same row count

  2. xxx_ready_bitset_ & system_ready_count_

    1. Used to record whether the corresponding column is loaded. Bitset corresponds to FieldOffset
    2. Query is executable If and only if the following conditions are met:
      1. system_ready_count_ == 2， which means all the system column RowId/Timestamp is loaded
      2. The scalar columns involved in the query has been loaded
      3. For the vector columns involved in the query, either the original data or the index is loaded

  3. scalar_indexings_: store scalar index

    1. Use StructuredSortedIndex in knowhere

  4. primary_key_index_: store index for pk column

    1. Use brand new ScalarIndexBase format
    2. Note: The functions here may overlap with scalar indexes. It is recommended to replace scalar index with ScalarIndexBase

  5. field_datas_: store original data

    1. aligned_vector<char> format guarantees int/float data are aligned

  6. SealedIndexingRecord vecindexs_: store vector index

  7. row_ids_/timestamps_: RowId/Timestamp data

  8. TimestampIndex: Index for Timestamp column

  9. schema: schema

SegmentSealedImpl internal function definition

  1. Most functions are the implementation of the corresponding functions of segment interface, which will not be repeated here
  2. update_row_count: Used to update the row_count field
  3. mask_with_timestamps: Use Timestamp column to update search bitmask，used to support Time Travel function