The inverted index

Key points

• A forward index maps a document to its text; the inverted index maps a term to its documents.
• Terms are held sorted, so locating elasticsearch is a dictionary seek, not a scan.
• Each term points at a postings list: matching doc ids, term frequency, and token positions.
• Those positions are what makes match_phrase possible - order and adjacency are recorded.
• Lucene writes this into immutable segments, so a delete only flips a bit in a liveDocs file.
• Sorting and aggregating read a separate column store called doc values, not the inverted index.

Example

Documents
  1: "Spring Boot tuning"
  2: "Boot time and JVM tuning"
  3: "Elasticsearch tuning guide"

Inverted index (after the standard analyzer)
  and            -> [2 (pos 2)]      <- kept: the standard analyzer removes NO stopwords
  boot           -> [1 (pos 1), 2 (pos 0)]
  elasticsearch  -> [3 (pos 0)]
  guide          -> [3 (pos 2)]
  jvm            -> [2 (pos 3)]
  spring         -> [1 (pos 0)]
  time           -> [2 (pos 1)]
  tuning         -> [1 (pos 2), 2 (pos 4), 3 (pos 1)]

Matching reads the inverted index; sorting and aggregating read doc values - different structures.