WHAT IS IN A LUCENE INDEX Adrien Grand Software engineer at Elasticsearch @jpountz 

• Lucene/Solr committer • Software engineer at Elasticsearch • I like changing the index file formats! – stored fields – term vectors – doc values – ... About me 

Why should I learn about Lucene internals? 

• Know the cost of the APIs – to build blazing fast search applications – don’t commit all the time – when to use stored fields vs. doc values – maybe Lucene is not the right tool • Understand index size – oh, term vectors are 1/2 of the index size! – I removed 20% of my documents and index size hasn’t changed • This is a lot of fun! Why should I learn about Lucene internals? 

• Make data fast to search – duplicate data if it helps – decide on how to index based on the queries • Trade update speed for search speed – Grep vs full-text indexing – Prefix queries vs edge n-grams – Phrase queries vs shingles • Indexing is fast – 220 GB/hour for 4K docs! – http://people.apache.org/~mikemccand/lucenebench/indexing.html Indexing 

• Tree structure – sorted for range queries – O(log(n)) search Let’s create an index sql index term data Lucene in action Databases Lucene 

Lucene doesn’t work this way 

Lucene term 2 3 • Store terms and documents in arrays – binary search Another index Lucene in action Databases 0 1 data index 0 1 0,1 0,1 0 0 sql 4 1 

Lucene term 2 3 • Store terms and documents in arrays – binary search Another index Lucene in action Databases 0 1 data index 0 1 0,1 0,1 0 0 Segment doc id document term ordinal terms dict postings list sql 4 1 

• Insertion = write a new segment • Merge segments when there are too many of them – concatenate docs, merge terms dicts and postings lists (merge sort!) Insertions? Lucene term 2 3 Lucene in action Databases 0 1 data index 0 1 0,1 0,1 0 sql 4 0 1 Lucene term 2 3 Lucene in action 0 data index 0 1 0 0 0 0 Databases 0 data index 0 1 0 0 sql 2 0 

• Insertion = write a new segment • Merge segments when there are too many of them – concatenate docs, merge terms dicts and postings lists (merge sort!) Insertions? Lucene term 2 3 Lucene in action Databases 0 1 data index 0 1 0,1 0,1 0 sql 4 0 1 Lucene term 2 3 Lucene in action 0 data index 0 1 0 0 0 0 Databases 1 data index 0 1 1 1 sql 2 1 

Lucene term 2 3 • Deletion = turn a bit off • Ignore deleted documents when searching and merging (reclaims space) • Merge policies favor segments with many deletions Deletions? Lucene in action Databases 0 1 data index 0 1 0,1 0,1 0 0 sql 4 1 1 0 live docs: 1 = live, 0 = deleted 

• Updates require writing a new segment – single-doc updates are costly, bulk updates preferred – writes are sequential • Segments are never modified in place – filesystem-cache-friendly – lock-free! • Terms are deduplicated – saves space for high-freq terms • Docs are uniquely identified by an ord – useful for cross-API communication – Lucene can use several indexes in a single query • Terms are uniquely identified by an ord – important for sorting: compare longs, not strings – important for faceting (more on this later) Pros/cons 

Lucene can use several indexes Many databases can’t 

Index intersection 1, 2, 10, 11, 20, 30, 50, 100 2, 20, 21, 22, 30, 40, 100 red shoe Many databases just pick the most selective index and ignore the other ones 1 2 3 4 5 6 7 8 9 Lucene’s postings lists support skipping that can be use to “leap-frog” 

• We just covered search • Lucene does more – term vectors – norms – numeric doc values – binary doc values – sorted doc values – sorted set doc values What else? 

• Per-document inverted index • Useful for more-like-this • Sometimes used for highlighting Term vectors Lucene term 2 3 Lucene in action 0 data index 0 1 0 0 0 0 Databases 1 data index 0 1 0 0 sql 2 0 Lucene term 2 3 data index 0 1 0,1 0,1 0 0 sql 4 1 

• Per doc and per field single numeric values, stored in a column-stride fashion • Useful for sorting and custom scoring • Norms are numeric doc values Numeric/binary doc values Lucene in action Databases 0 1 Solr in action Java 2 3 42 1 3 10 afc gce ppy ccn field_a field_b 

• Ordinal-enabled per-doc and per-field values – sorted: single-valued, useful for sorting – sorted set: multi-valued, useful for faceting Sorted (set) doc values Lucene in action Databases 0 1 Solr in action Java 2 3 1,2 0 0,1,2 1 distributed Java search 0 1 2 Ordinals Terms dictionary for this dv field 

• Compute value counts for docs that match a query – eg. category counts on an ecommerce website • Naive solution – hash table: value to count – O(#docs) ordinal lookups – O(#doc) value lookups • 2nd solution – hash table: ord to count – resolve values in the end – O(#docs) ordinal lookups – O(#values) value lookups Faceting Since ordinals are dense, this can be a simple array 

• These are the low-level Lucene APIs, everything is built on top of these APIs: searching, faceting, scoring, highlighting, etc. How can I use these APIs? API Useful for Method Inverted index Term -> doc ids, positions, offsets AtomicReader.fields Stored fields Summaries of search results IndexReader.document Live docs Ignoring deleted docs AtomicReader.liveDocs Term vectors More like this IndexReader.termVectors Doc values / Norms Sorting/faceting/scoring AtomicReader.get*Values 

• Data duplicated up to 4 times – not a waste of space! – easy to manage thanks to immutability • Stored fields vs doc values – Optimized for different access patterns – get many field values for a few docs: stored fields – get a few field values for many docs: doc values Wrap up 0,A 1,A 2,A 0,A 0,B 0,C 1,A 1,C 2,A 2,B 2,C 1,B 0,B 1,B 2,B 0,B 1,B 2,B Stored fields Doc values At most 1 seek per doc At most 1 seek per doc per field BUT more disk / file-system cache-friendly 

File formats 

• Save file handles – don’t use one file per field or per doc • Avoid disk seeks whenever possible – disk seek on spinning disk is ~10 ms • BUT don’t ignore the filesystem cache – random access in small files is fine • Light compression helps – less I/O – smaller indexes – filesystem-cache-friendly Important rules 

• File formats are codec-dependent • Default codec tries to get the best speed for little memory – To trade memory for speed, don’t use RAMDirectory: – MemoryPostingsFormat, MemoryDocValuesFormat, etc. • Detailed file formats available in javadocs – http://lucene.apache.org/core/4_5_1/core/org/apache/lucene/codecs/package- summary.html – Codecs 

• Bit packing / vInt encoding – postings lists – numeric doc values • LZ4 – code.google.com/p/lz4 – lightweight compression algorithm – stored fields, term vectors • FSTs – conceptually a Map – keys share prefixes and suffixes – terms index Compression techniques 

What happens when I run a TermQuery? 

• Lookup the term in the terms index – In-memory FST storing terms prefixes – Gives the offset to look at in the terms dictionary – Can fast-fail if no terms have this prefix 1. Terms index l/4 u c r y/3 b/2 r a/1 br = 2 brac = 3 luc = 4 lyr = 7 

• Jump to the given offset in the terms dictionary – compressed based on shared prefixes, similarly to a burst trie – called the “BlockTree terms dict” • read sequentially until the term is found – 2. Terms dictionary [prefix=luc] a, freq=1, offset=101 as, freq=1, offset=149 ene, freq=9, offset=205 ky, frea=7, offset=260 rative, freq=5, offset=323 Jump here Not found Not found Found 

• Jump to the given offset in the postings lists • Encoded using modified FOR (Frame of Reference) delta – 1. delta-encode – 2. split into block of N=128 values – 3. bit packing per block – 4. if remaining docs, encode with vInt 3. Postings lists 1,3,4,6,8,20,22,26,30,31 1,2,1,2,2,12,2,4,4,1 [1,2,1,2] [2,12,2,4] 4, 1 2 bits per value 4 bits per value Example with N=4 vInt-encoded 

• In-memory index for a subset of the doc ids – memory-efficient thanks to monotonic compression – searched using binary search • Stored fields – stored sequentially – compressed (LZ4) in 16+KB blocks 4. Stored fields 0 1 2 3 4 5 6 16KB 16KB 16KB docId=0 offset=42 docId=3 offset=127 docId=4 offset=199 

• 2 disk seeks per field for search • 1 disk seek per doc for stored fields • It is common that the terms dict / postings lists fits into the file-system cache • “Pulse” optimization – For unique terms (freq=1), postings are inlined in the terms dict – Only 1 disk seek – Will always be used for your primary keys Query execution 

Quizz 

What is happening here? #docs in the index qps 1 2 

What is happening here? #docs in the index qps 1 2 Index grows larger than the filesystem cache: stored fields not fully in the cache anymore 

What is happening here? #docs in the index qps 1 2 Index grows larger than the filesystem cache: stored fields not fully in the cache anymore Terms dict/Postings lists not fully in the cache 

Thank you!