Skip to content

Yelp/nrtsearch

Repository files navigation

Coverage Docs

nrtSearch

A high performance gRPC server, with optional REST APIs on top of Apache Lucene version 8.x source, exposing Lucene's core functionality over a simple gRPC based API.

Documentation is available at readthedocs.

Features

  • Relies on Lucene's near-real-time segment replication for data replication. This means, a dedicated primary/writer node takes care of indexing operations and expensive operations like segment merges. This allows the replicas' system resources to be dedicated entirely for search queries. This behavior is in contrast to the document replication approach taken by some other popular search engines based on lucene like elasticsearch where every node is a writer and a reader.
  • Supports concurrent query execution. This is another feature missing from popular search engines based on lucene like elasticsearch.
  • Can be deployed as a "stateless microservice". Indexes are backed up in s3. Clients can choose to commit data outside of this system once their backup is complete. Upon restarts e.g. if you bring up a new container clients can choose to bootstrap indexes from their previous backed up state. Ability to deploy in a stateless manner allows for easy scalability using container tools like kubernetes, mesos etc.
  • Provides gRPC streaming APIs for indexing and searching. Also supports REST APIs.

Design

The design goals are mostly similar to the ones mentioned in the Lucene Server project. This project uses ideas and code from luceneserver and builds on them.

A single node can index a stream of documents, run near-real-time searches via a parsed query string, including "scrolled" searches, sorting, index-time sorting, etc.

Fields must first be registered with the registerFields command, where you express whether you will search, sort etc., and then documents can be indexed with those fields.

There is no transaction log, so you must call commit yourself periodically to make recent changes durable on disk. This means that if a node crashes, all indexed documents since the last commit are lost.

Indexing a stream of documents

NrtSearch supports client side gRPC streaming for its addDocuments endpoint. This means that the server API accepts a stream of documents . The client can choose to stream the documents however it wishes. The example nrtSearch client implemented here reads a CSV file and streams documents from it over to the server. The server can index chunks of documents the size of which is configurable as the client continues to send more documents over its stream. gRPC enables this with minimal application code and yields higher performance compared to JSON. TODO[citation needed]: Add performance numbers of stream based indexing for some datasets.

Near-real-time-replication

This requirement is one of the primary reasons to create this project. near-real-time-replication seems a good alternative to document based replication when it comes to costs associated with maintaining large clusters. Scaling document based clusters up/down in a timely manner could be slower due to data migration between nodes apart from paying the cost for reindexing on all nodes.

Below is a depiction of how the system works in regards to Near-real-time(NRT) replication and durability. alt text

  • Primary node comes up with either no index or reads segments from disk or can restore an index from remote storage if the restore option is specified by the client on the startIndex command. This node will accept indexing requests from clients. It will also periodically publishNrtUpdate to replicas giving them a chance to catch up with the latest primary indexing changes.
  • Replica nodes are also started using the startIndex command. They will sync with the current primary and update their indexes using lucene's NRT APIs. They can also restore the index from remote storage and then receive the updates since the last backup. These nodes will serve client's search queries.
  • Each time client invokes commit on primary, it will save its current index state and related metadata e.g. schemas, settings to the disk. Clients should use the ack from this endpoint to commit the data in their channel e.g. kafka.
  • Client can invoke backupIndex on the primary to backup the index to remote storage.
  • If a replica crashes, a new one can be brought up and will re-sync with the current primary (optionally restoring the index from remote storage first). It will register itself with the primary once it's brought up.
  • If a primary crashes, a new one can be brought up with the restore option on startIndex command to regain previous stored state in the cloud, but since primaries don't serve search requests they can also use network attached storage e.g. Amazon EBS to persist data across restarts. The replicas will then re-sync their indexes with the primary.

Build Server and Client

In the home directory.

./gradlew clean installDist test

Note: This code has been tested on Java21

Run gRPC Server

./build/install/nrtsearch/bin/nrtsearch_server

Build gRPC Gateway

./gradlew buildGrpcGateway

Run REST Server (use the appropriate binary for your platform e.g. for mac os)

./build/install/nrtsearch/bin/http_wrapper-darwin-amd64 <gRPC_PORT> <REST_PORT>

Example to run some basic client commands

Create Index

./build/install/nrtsearch/bin/nrtsearch_client createIndex --indexName  testIdx
curl -XPOST localhost:<REST_PORT>/v1/create_index -d '{"indexName": "testIdx"}'

Update Settings

./build/install/nrtsearch/bin/nrtsearch_client settings -f settings.json
cat settings.json
{             "indexName": "testIdx",
              "directory": "MMapDirectory",
              "nrtCachingDirectoryMaxSizeMB": 0.0,
              "indexMergeSchedulerAutoThrottle": false,
              "concurrentMergeSchedulerMaxMergeCount": 16,
              "concurrentMergeSchedulerMaxThreadCount": 8
}

Start Index

./build/install/nrtsearch/bin/nrtsearch_client startIndex -f startIndex.json
cat startIndex.json
{
  "indexName" : "testIdx"
}

RegisterFields

./build/install/nrtsearch/bin/nrtsearch_client registerFields -f registerFields.json
cat registerFields.json
{             "indexName": "testIdx",
              "field":
              [
                      { "name": "doc_id", "type": "ATOM", "storeDocValues": true},
                      { "name": "vendor_name", "type": "TEXT" , "search": true, "store": true},
                      { "name": "license_no",  "type": "INT", "multiValued": true, "storeDocValues": true}
              ]
}

Add Documents

./build/install/nrtsearch/bin/nrtsearch_client addDocuments -i testIdx -f docs.csv -t csv
cat docs.csv
doc_id,vendor_name,license_no
0,first vendor,100;200
1,second vendor,111;222

Search

./build/install/nrtsearch/bin/nrtsearch_client search -f search.json
cat search.json
{
        "indexName": "testIdx",
        "startHit": 0,
        "topHits": 100,
        "retrieveFields": ["doc_id", "license_no", "vendor_name"],
         "queryText": "vendor_name:first vendor"
}

API documentation

The build uses protoc-gen-doc program to generate the documentation needed in html (or markdown) files from proto files. It is run inside a docker container. The gradle task to generate this documentation is as follows.

./gradlew buildDocs

This should create a src/main/docs/index.html file that can be seen in your local browser. A sample snapshot

Yelp Indexing tool

Reviews

This tool indexes yelp reviews available at Yelp dataset challenge. It runs a default version with only 1k reviews of the reviews.json or you could download the yelp dataset and place the review.json in the user.home dir and the tool will use that instead. The complete review.json should have close to 7Million reviews. The tool runs multi-threaded indexing and a search thread in parallel reporting the totalHits. Command to run this specific test:

./gradlew clean installDist :test -PincludePerfTests=* --tests "com.yelp.nrtsearch.yelp_reviews.YelpReviewsTest.runYelpReviews" --info

About

A high performance gRPC server on top of Apache Lucene

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages