Lucene @ Yelp Sudarshan Gaikaiwari
Bio 1. Over a decade of experience in information retrieval 2. Used IR techniques at Symantec's DLP group 3. Search Engineer at Yelp
Outline 1. Overview of search services at Yelp 2. Federation Motivation 3. Lucy Indexing 4. Lucy Searching 5. Efficiently Retrieving top k hits
The services we provide
Lucy: business search
Lucy also powers phone search
Cathy: she 'talks' a lot
Listsearch: it searches lists....
Reviewsearch: it searches reviews....
DYM: did you really mean that?
Suggest: auto completion
Search is too slow
Hard Disk Seek Latency Disk seek 10,000,000 ns
Source Software Engineering Advice from Building Large-Scale Distributed Systems Jeffery Dean
RAM read latency Main memory reference 100 ns
Pinning Index in RAM ● vmtouch ● mlock ● http://hoytech.com/vmtouch/
Problem Index is too large fit in memory on a single machine
Geographical Sharding drawbacks 1. Cumbersome manual process to determine shard boundary 2. No guarantee that a boundary can be found.
Federation 1. Split index across multiple machines 2. Shard on business id 3. TF-IDF scores from different machines should be comparable
Mapping businesses to shards 1. Assigning businesses to shards shard = shardlist[hash(business_id) % len(shardlist)] Problems 1. Involves re-indexing all the businesses if we want to add a new shard
Advantages 1. Flexibility (move vbuckets from one shard to another) 2. Split hot spot shards
Lucy Master Slave Architecture Separate indexing (masters) A master for each shard of a service Searching (slaves) A slave for every replica of a service
Federator: Combining results across shards 1. Once we distribute an index across shards we need a component which will search all these shards and combine their results. 2. Written in Python (runs inside a python web process). 3. Uses Tornado IO loop to send requests to all shards. 4. The transfer protocol for the requests in JSON RPC
Tokens to Business Attributes
Executing queries 1. Gather the top results for a query 2. Collect attribute statitics for attributes like places, categories
Lucene 1. Efficiently executes queries over the index 2. Provides how relevant the business is to the words in the query (word score) 3. Upgrading lucene to 2.9/3.1 is WIP
Successive geobounds relaxation
Successive geobounds relaxation
Efficiently Retrieving top k hits 1. When user moves through multiple pages the number of hits to be returned increases num hits = start + count 2. So if we need to retrieve 500 hits the naive way would be to retrieve 500 hits from each shard and then sort them
Distribution of hits in shards
Probability a hit is in a shard
Binomial Distribution Probability (r of top k hits) are in a particular shard
Formula Std Deviation
Hits selected from each shard k = 100 p = 0.2
Results Missed (%)
Results 1. ~ 50% savings over 100 hits (44 hits requested from each shard) 2. 77% savings over 1000 hits (228 hits requested from each shard)
Future work 1. In memory index 2. Move towards real time search
Come Join Us!
Published on Jun 28, 2011
This talk describes how the Yelp uses Lucene to provide search services. It includes Statistics of Yelp search usage Overview of Yelp search...