memgraph / mgcxx

mgcxx (experimental)

A collection of C++ wrappers around non-C++ libraries. The list includes:

• full-text search enabled by tantivy

Requirements:

• cmake 3.15+
• rustup toolchain 1.75.0+

How to build and test?

mkdir build && cd build
cmake ..
make && ctest

text_search

TODOs

• Polish & test all error messages
• Write unit / integration test to compare STRING vs JSON fiels search query syntax.
• Figure out what's the right search syntax for a property graph
• Add some notion of pagination
• Add some notion of backwards compatiblity -> some help to the user
• How to:
  • search all properties
  • fuzzy search // let term = Term::from_field_text(data_field, &input.search_query); // let query = FuzzyTermQuery::new(term, 2, true);
• Add Github Actions
• Add benchmarks:
  • Test what's the tradeoff between searching STRING vs JSON TEXT, how does the query look like?
  • Search direct field vs JSON, FAST vs SLOW, String vs CxxString
  • MATCH (n) RETURN count(n), n.deleted;
  • search of a specific property value
  • benchmark (add|retrieve simple|complex, filtering, aggregations).
  • search of all properties
  • Benchmark (search by GID to get document_id + fetch document by document_id) vs (fetch document by document_id) on 100M nodes + 100M edges
    • Note DocAddress is composed of 2 u32 but the SegmentOrdinal is tied to the Searcher -> is it possible/wise to cache the address (SegmentId is UUID)
      • A searcher per transaction -> cache DocAddress inside Memgraph's ElementAccessors?
• Implement the stress test by adding & searching to the same index concurrently + large dataset generator.
• Consider implementing panic! handler preventing outside process to crash (optionally).

NOTEs

• if a field doesn't get specified in the schema, it's ignored
• TEXT means the field will be tokenized and indexed (required to be able to search)
• Tantivy add_json_object accepts serde_json::map::Map<String, serde_json::value::Value>
• C++ text-search API is snake case because it's implemented in Rust
• Writing each document and then committing (writing to disk) will be expensive. In a standard OLTP workload that's a common case -> introduce some form of batching.

Resources

• https://fulmicoton.com/posts/behold-tantivy-part2
• https://stackoverflow.com/questions/37924383/combining-several-static-libraries-into-one-using-cmake --> decided to have 2 separate libraries user code has to link