thoth - realtime solr monitor and search analysis engine
DESCRIPTION
Managing a large and diversified Solr search infrastructure can be challenging and there is still a lack of tools that can monitor the entire system and help the scaling process. Trulia wrote and currently uses a real-time Solr monitor and search analysis engine called Thoth. This session will cover how Thoth was designed, why it was chosen to analyze Solr, challenges that were encountered while building and scaling the system and useful features like integration with Apache ActiveMQ and Nagios for real-time paging, generation of reports on query volume, latency, time period comparisons, and the Thoth dashboard. Damiano and Praneet will also summarize their application of machine learning algorithms and its results to the process of query analysis and pattern recognition.TRANSCRIPT
Thoth
Real-time Solr Monitor
Search Analysis Engine
Damiano Braga
Sr. Software Engineer
Praneet Mhatre
Data Mining Engineer
Overview
- What is Thoth ?
- Data Collection and Thoth Core Indexing
- Thoth API & Thoth Dashboard
- Thoth Monitor
- Thoth ML : Prediction and Topic Modeling
- Special Thanks & Q/A
Demo
What is Thoth?
- Innovation project at Trulia
- Understand our search infrastructure without touching logs
- Troubleshoot search performance issues
- Designed as a modular system
- Set of tools that can help gather info, monitor, understand a search infrastructure
- Open source project :
thoth
thoth-ml
thoth-api
thoth-dashboard
thoth-monitor
thoth-demo
Problem: Know Your Search Infrastructure
- Solr logs are a good source. Sometimes partial information
- Decentralized data (at least 1 log per search server)
- Log rotation
- Not searchable
If we could index all the information .. Let’s use Solr !
- We can search on it
- We have some handy features for free: facets, stats etc
- It’s scalable
Thoth Document
1 Solr Request = 1 Thoth (Solr) Document
Server Info
hostname, port number, core name, pool name
Query Info
timestamp, actual query, qtime, hits, exception?
Data Collection (1/2)
- Should be smooth. No traffic slowing down.
- We care about near real-time data
- We care about historical data
- Dataset is growing fast
- Interceptor on each search server
- We use a SolrComponent attached to a Request Handler
- Queue System (E.g: ActiveMQ) to facilitate and temporary store messages
- Each search server has a manifest in the solrconfig.xml
Data Collection (2/2)<requestHandler name="select" class="com.solr2activemq.SolrToActiveMQHandler”>
<arr name="last-components”>
<str>solr2activemq</str>
</arr>
</requestHandler>
<searchComponent name="solr2activemq” class="com.solr2activemq.SolrToActiveMQComponent" >
<str name="activemq-broker-uri">localhost</str>
<int name="activemq-broker-port">61616</int>
<str name="activemq-broker-destination-type">queue</str>
<str name="activemq-broker-destination-name">test-queue</str>
<str name="solr-hostname">localhost</str>
<int name="solr-port">8983</int>
<str name="solr-poolname">default</str>
<str name="solr-corename">collection</str>
<int name="solr2activemq-buffer-size">1000</int>
<int name="solr2activemq-dequeuing-buffer-polling">500</int>
<int name="solr2activemq-check-activemq-polling">5000</int>
</searchComponent>
Sizing of Data
- Need for granular information for near real-time data
- Less granularity for historical data
Too much data = slow search, space problem
- Shrinking feature:
- Create Shrank Document
- Real-time Core cleanup
- Shrinking time is configurable
Thoth Index
- Solr 4.7
- Soft commit for near real-time search
- Soft commit maxTime set to 1s
- Auto commit set to 15s
- Update chain set to enforce UUID as PkID
- Use of Solrj to index data and query
Thoth API
- Abstraction for Thoth index and Thoth data
- Read only REST-like API
- JSON response
- Written in Node.js to accommodate socket.io
Example:
{"numFound":95,"values":[{"timestamp":"2014-09-
16T18:00:02Z","value":45337},{"timestamp":"2014-09-
16T18:15:02Z","value":77325},{"timestamp":"2014-09-
16T18:30:02Z","value":109523},{"timestamp":"2014-09-
16T18:45:02Z","value":112279},{"timestamp":"2014-09-
16T19:00:02Z","value":115334}
thoth:3001/api/server/foo/core/bar/port/portbar/start/NOW-1DAY/end/NOW/count/nqueries
Thoth Dashboard (1/5)
- Visual insight on Thoth data
- Useful graphs divided by server or pool
- Handy list of slow queries and exceptions
- Real-time view for server
- Selecting data based on time
- Sharable URLs (to OPS team, QA team, Release Eng. )
Thoth Dashboard (2/5)
Thoth Dashboard (3/5)
Thoth Dashboard (4/5)
Thoth Dashboard (5/5)
Thoth Monitor
- Continuously monitoring for metrics
- Stateless
- Alerting through email or Nagios
- Examples: QTime, Number of Zero hits,
Predictor Model Health
- Possibility to implement custom monitors
- Reuse StatsComponent
[http://wiki.apache.org/solr/StatsComponent]
if possible
Thoth ML
What can we do with all this data?
• Rich source of information
• Can we turn it into knowledge?
• How about machine learning?
1. Query time prediction
2. Query pattern recognition
3. Server sizing and resource allocation
1. Query Time Prediction (1/4)
• Goal : appropriately route queries to slow/ fast pool
• Look at query attributes
• Query text
• Start parameter
• Facets, range queries, geo spatial searches etc
• Train a supervised learning model
• Use learned model to predict if a query will be slow v/s fast
• H2O Machine Learning Library
1. Query Time Prediction (2/4)
Challenges
• Imbalanced dataset
• Frequency of model training
• Type of model
• Minimal delay requirement
1. Query Time Prediction (3/4)
Challenges Addressed
• Imbalanced dataset
• Stratified sampling
• Frequency of model training
• Auto identify relearning frequency
• Type of model
• Boolean, categorical features -> Tree based
• High accuracy
• Gradient Boosted Machine
• Minimal delay requirement
• User pool queries: 45-50 ms
• Prediction: 1-3 ms
1. Query Time Prediction (4/4)
• 1000 Gradient Boosted Trees
• Slow queries = (>100ms. Configurable)
• Experimental Results
• Training on ~3.1 million
• Test on ~1.4 million
• AUC: 0.94542
• Accuracy: 0.9202223
Query Time Prediction in Action (1/2)
Performance on real time traffic at Trulia
Query Time Prediction in Action (2/2)
Performance on real time traffic at Trulia
2. Query Pattern Recognition
• Exceptions, zero hit queries
• Analyze and find out why
• Probabilistic Topic Modeling
• Using MALLET open source toolkit
Topic Modeling Flow
Topics With Keywords
Future Direction
- Thoth ML improvements:
• Predicting query time buckets
• Regression v/s classification
• Exceptions and zero hit query analysis
• Sizing and resource allocation
- Solr Cloud integration
- Dashboard integration with Solr cloud
- More standard metrics on Thoth Monitor
- More data collection (load, GC)
Contributors and Special Thanks
Damiano : [email protected]
Praneet: [email protected]
Fork us on Github!
github.com/trulia/thoth
JD Cantrell ( API, Dashboard)
Giulio Grillanda (API, Dashboard)
Rajendra Shioramwar (Core)
Ying Wang (Design)
Girish Gudla (Monitor)
Alexander Kanarsky
Alex Burmester