Lucene-Benchmarks

A few benchmarks to measure performance of different operations to be able to reason about different Lucene operations.

So far, there is only one test that compares performance of document updates (these are implemented as delete+insert).

Performance results

Updating document vs updating doc value only

This test first ingests 10k documents from CRAN.REL dataset and then runs multiple iterations of 10k updates

Benchmark	(useReplace)	Mode	Cnt	Score	Error	Units
Main.runBenchmark	true	thrpt	6	39.214	± 1.186	ops/s
Main.runBenchmark	false	thrpt	6	1775.49	± 123.876	ops/s

Conclusion:

• Updating doc values is ~45x times faster than updating entire documents.

Fetching stored numeric field vs numeric doc value

This test first ingests 10k documents from CRAN.REL dataset and then runs multiple searches (with having updates and refreshes performed in background)

Benchmark	(useDocValues)	Mode	Cnt	Score	Error	Units
Main.runBenchmark	true	thrpt	6	3930.220	± 380.659	ops/s
Main.runBenchmark	false	thrpt	6	4899.003	± 502.397	ops/s

Conclusion:

• A search with fetching a single number from a doc value is ~20% slower than a search with fetching that number from a stored field
• Fetching doc values vs fetching stored fields have a limited impact on search performance, as the majority of time is spent in query phase compared to fetch phase.

How to run

# Make sure gradle is installed.
git clone https://github.com/nikitiuk0/lucene-benchmarks.git
cd lucene-benchmarks
./gradlew run

Class Structure

Class	Functionality
FileUtils	Create local folder structures, automatically download and decompress Cranfield files. Provide access to all necessary files (Cranfield files, index files).
FileParser	A parser used to parse and load Cranfield files into a Java supported data structre.
DocumentModel	A simple object-oriented way to represent Cranfield documents.
Indexer	Responsible for creating an index and writing/updating the data.
Searcher	Read indices created by Indexer, parse the query string and exectute search. Also runs a background thread the performs a periodic refresh.
Evaluator	Executes a search, can inspect the results for testing.
Main	Entry of program, runs performance tests.

Acknowledgements

This project is partially based on https://github.com/PointerFLY/Lucene-Example.git