How to Really Scale Autocomplete

Source: bonsai.io (Part 2 of series)

Problem

Part 1’s approach (significant_terms on shingles) collapsed at 6M Wikipedia documents: 4+ second queries, CPU at 100%, widespread timeouts. The significant_phrases aggregation with shingles explodes cardinality at millions of docs.

Solution Architecture

Schema

• completions: search_as_you_type field (max_shingle_size: 3) — fast prefix matching for page titles
• suggest_word_all: text field with fielddata (for significant_terms aggregation on words)
• suggest_phrase: text field with shingles analyzer + fielddata (for phrase aggregation)
• Views column as primary relevance signal (log1p boosting)

Query Pattern

{
  "query": {
    "function_score": {
      "query": {"match_phrase_prefix": {"completions": {"query": q, "boost": 0.01}}},
      "field_value_factor": {"field": "views", "modifier": "log1p"}
    }
  },
  "aggs": {
    "significant_words": {"significant_terms": {"field": "suggest_word_all", "include": "q+.*"}},
    "significant_phrases": {"terms": {"field": "suggest_phrase", "include": "q+.*"}}
  }
}

Key change from Part 1: significant_phrases replaced with plain terms aggregation (frequency-ordered). This is fast because the prefix filter narrows to thousands of docs.

Performance Results

• 416 req/s at p99 ≤ 45ms, warm index
• With simultaneous hybrid search on same cluster: ~96 autocomplete req/s + 96 hybrid req/s both within SLA

Implementation Notes

• Use a separate index for autocomplete — different scaling profile than main search
• Debounce 25-50ms before firing (saves requests, only fires on user pause)
• Watch fielddata memory if corpus grows very large

Co-occurrence Sugar

After a user selects a suggestion, a follow-up query using significant_terms with exclude: q reveals semantically related terms (co-occurrence):

• “elasticsearch” → ['kibana', 'lucene', 'opensearch', 'solr']
• “bidirectional encoder” → ['nlp', 'transformer', 'devlin', 'bert']

Useful for guiding discovery/navigation without manual curation.

Related Concepts

• Autocomplete
• Search Architecture
• Faceted Search