MongoDB World 2019: Fast Machine Learning Development with MongoDB
1 like1,033 views
The document discusses 'Spoke,' a machine learning-based workplace management software that uses MongoDB for online learning to adapt and deliver real-time responses to user requests. It highlights the challenges of customized ML models for different clients, the benefits of MongoDB's scalability, and details various data models and indexing strategies to optimize performance. Future directions include enhancing query understanding and automating the labeling of user requests, alongside MongoDB's advancements in full-text search and serverless data lake solutions.
![Example: Text Classification
Data Model 2: tabular
matrix: one row per
article, one column
per word in article
Data Model 3:
JSON documents
extract keywords and
topics and enrich
word1 word2 word3
article1 1 0 2
article2 0 1 0
article3 0 1 1
LDA Topic
extraction
{
"_id" : “0b917217ae”,
"title" : "Document Model Design Patterns",
“text”: blob,
"topics" : [ "Models", "MVC" ],
“top_words”: [“join”, “embed”, “one-to-many”]
“model”:
{
“location”:
“last_updated”: Timestamp(“05-29-19
00:00:00”)
“confidence”: Decimal128("0.9123")
...
}
...
}](https://image.slidesharecdn.com/spokeatmongodb2019-190717214644/85/MongoDB-World-2019-Fast-Machine-Learning-Development-with-MongoDB-17-320.jpg)
![Example: Text Classification & Graph Traversal
Data Model 3:
JSON documents
extract keywords and
topics and enrich
Data Model 4: graph
tree/hierarchy of
topics modeled as a
graph
Hierarchical
Clustering
{
"_id" : “0b917217ae”,
"title" : "Document Model Design Patterns",
“text”: blob,
“parent”: “Databases”,
"topics" : [ "Models", "MVC" ],
“top_words”: [“join”, “embed”, “one-to-many”]
“model”:
{
“location”:
“last_updated”: Timestamp(“05-29-19
00:00:00”)
“confidence”: Decimal128("0.9123")
...
}
...
db.topics.insert( { _id: "Models", parent: "Databases" } )
db.topics.insert( { _id: "Storage", parent: "Databases" } )
db.topic.insert( { _id: "MVCC", parent: "Databases" } )
db.topic.insert( { _id: "Databases", parent: "Programming" } )
db.topic.insert( { _id: "Languages", parent: "Programming" } )
db.topic.insert( { _id: "Programming", parent: null } )
Programming
Languages Databases
ModelsStorage
MVCC
$graphlookup](https://image.slidesharecdn.com/spokeatmongodb2019-190717214644/85/MongoDB-World-2019-Fast-Machine-Learning-Development-with-MongoDB-18-320.jpg)
MongoDB World 2019: Fast Machine Learning Development with MongoDB
- 1. Fast Machine Learning Development with MongoDB Rajhans Samdani - Spoke Jane Fine - MongoDB
- 3. Demo Flow
- 4. Spoke: A simpler, smarter way to manage requests ● Problems: natural language processing problems ● Challenge: customized machine learning models for every client ○ Need to learn quickly (near real time) from user interactions ○ 1000s of ML models ● MongoDB: very useful in scaling up our ML ● Spoke is workplace management software that uses machine learning to answer questions and assign requests to right teams ○ Started in August, 2016. Funded by Greylock and Accel.
- 7. Machine Learning Problem: Team Triaging Problem: pick right team based on the text and context of the request Challenge: Each client has different teams so pretraining not possible; must learn from demonstration Implication => Separate ML model for each client
- 8. Traditional ML vs Adaptive Approach Claim: most ml-driven startups are in the second bucket Low data and low query volume domain Startups must build quickly and adapt to users to show utility Traditional ML Pipeline Adaptive ML Pipeline
- 9. Adapting with Online Machine Learning ● Online learning: Update the model at each time step as the data sequentially arrives ● For the first year Spoke built quickly by using online learning to deliver a slick product experience ○ Users see the utility because the system learns in real time! ● Easy to serve and scale using MongoDB
- 10. Serving flow Online Learning with MongoDB Training flow
- 11. Online Learning: Tips ● Use feature hashing to get bounded model size ○ E.g. a linear model with a hash of size 10k for a 5 way classification problem => model size = 200KB. ● Load test your setup to ensure it works for your QPS ● We modified Scikit Learn’s online learning algorithm to change the number of class labels on the fly ● Gotchas: concurrency is a problem. Possible that for two training events arriving at the same time, one will be ignored ○ Possible to avoid using queues
- 12. Augmenting TensorFlow with MongoDB ● Years later, we developed a batch training environment using Tensorflow ○ Maintainable ○ Retrainable ● Still a tension between online learning and batch training ○ Good UX from online learning ○ Engineering scalability from batch training ● Achieved a good compromise: use Mongo-based online learning model for first few hundred responses and then silently switch to Tensorflow
- 14. Multiple Data Models and Access Patterns in MongoDB Rich Queries Point | Range | Geospatial | Faceted Search | Aggregations | JOINs | Graph Traversals JSON Documents Tabular Key-Value Text GraphGeospatial
- 15. Example: Text Classification Data Model 1: key-value raw text input whole corpus 0b917217ae 7fef14c0b3 cb9eadad9a
- 16. Example: Text Classification Data Model 1: key-value raw text input whole corpus 0b917217ae 7fef14c0b3 cb9eadad9a Data Model 2: tabular matrix: one row per article, one column per word in article word1 word2 word3 article1 1 0 2 article2 0 1 0 article3 0 1 1 TF-IDF Vectorization
- 17. Example: Text Classification Data Model 2: tabular matrix: one row per article, one column per word in article Data Model 3: JSON documents extract keywords and topics and enrich word1 word2 word3 article1 1 0 2 article2 0 1 0 article3 0 1 1 LDA Topic extraction { "_id" : “0b917217ae”, "title" : "Document Model Design Patterns", “text”: blob, "topics" : [ "Models", "MVC" ], “top_words”: [“join”, “embed”, “one-to-many”] “model”: { “location”: “last_updated”: Timestamp(“05-29-19 00:00:00”) “confidence”: Decimal128("0.9123") ... } ... }
- 18. Example: Text Classification & Graph Traversal Data Model 3: JSON documents extract keywords and topics and enrich Data Model 4: graph tree/hierarchy of topics modeled as a graph Hierarchical Clustering { "_id" : “0b917217ae”, "title" : "Document Model Design Patterns", “text”: blob, “parent”: “Databases”, "topics" : [ "Models", "MVC" ], “top_words”: [“join”, “embed”, “one-to-many”] “model”: { “location”: “last_updated”: Timestamp(“05-29-19 00:00:00”) “confidence”: Decimal128("0.9123") ... } ... db.topics.insert( { _id: "Models", parent: "Databases" } ) db.topics.insert( { _id: "Storage", parent: "Databases" } ) db.topic.insert( { _id: "MVCC", parent: "Databases" } ) db.topic.insert( { _id: "Databases", parent: "Programming" } ) db.topic.insert( { _id: "Languages", parent: "Programming" } ) db.topic.insert( { _id: "Programming", parent: null } ) Programming Languages Databases ModelsStorage MVCC $graphlookup
- 19. Indexing in MongoDB • Primary Index – Every Collection has a primary key index • Compound Index – Index against multiple keys in the document • MultiKey Index – Index into arrays • Text Indexes – Support for text searches • GeoSpatial Indexes – 2d & 2dSphere indexes for spatial geometries • Hashed Indexes – Hashed based values for sharding Index Types • TTL Indexes – Single Field indexes, when expired delete the document • Unique Indexes – Ensures value is not duplicated • Partial Indexes – Expression based indexes, allowing indexes on subsets of data • Case Insensitive Indexes – Supports text search using case insensitive search • Sparse Indexes – Only index documents which have the given field Index Features
- 20. Scalability & Distributed Processing Process large volumes of data in parallel queries and aggregations run in parallel data is returned in parallel • Automatically scale beyond the constraints of a single node • Optimized for query patterns and data locality • Transparent to applications and tools ≤ ∑ ⟕ " sharded cluster
- 21. Intelligent Data Distribution: Workload Isolation Enable different workloads on the same data ANALYTICAL ML & AI A single replica set • Combine operational and analytical workloads on a single platform • No data movement or duplication • Extract insights in real-time to enrich applications • MongoDB Atlas - Analytics Nodes TRANSACTIONAL Operational Analytics S S S Application
- 22. Text Search
- 23. MongoDB Text Search db.restaurants.find( { $text: { $search: "java coffee shop" } } ) db.restaurants.find( { $text: { $search: "java coffee shop" } }, { score: { $meta: "textScore" } } ).sort( { score: { $meta: "textScore" } } ) Match Text Score and Sort Results Create Text Index1 2 3 db.restaurants.createIndex( { description: "text" } )
- 24. Index any field whose value is a string or an array of string elements A collection can only have one text search index, but that index can cover multiple fields Optionally Specify Language: Text Indexing
- 25. Text Matching $text will tokenize the search string using whitespace and most punctuation as delimiters, and perform a logical OR of all such tokens in the search string. Search for a Single Word Match Any of the Words Search for a Phrase Negations
- 26. Scoring and Sorting: Control Search Results with Weights Weight is the significance of the field (default = 1) For each indexed field, MongoDB multiplies the number of matches by the weight and sums the results → score of the document Use “textScore" metadata for projections, sorts, and conditions subsequent the $match stage that includes the $text operation.
- 27. Spoke: Knowledge Base Search with ML ● User asks a question to Spoke and expects real time response ○ Search best knowledge answer from 1000s of answers ● We use a combination of ML algorithms in determining the right answer ○ Scoring each answer independently is not an option due to latency ● Candidate generation to rescue!
- 28. How Spoke uses Text Search ● Use MongoDB text search to select top k highest scoring articles ● Only run extensive ML-based search on k articles ○ Works as long as the right answer is in top k ○ Allows us to build latest ML algos without worrying too much about latency ● Tip: set your MongoDB text index weights carefully by fine tuning { title: 10, body: 2, keywords: 6...}
- 30. What Spoke is working on ● Understand user queries and take actions ○ “I need access to Salesforce” => “issue_license(user, software=salesforce)” ● Assign custom labels to user questions ○ Allow customers to add labels to their requests ■ {“hardware”, “software”, “licensing”, “urgent”}, ■ {“benefits”, “payroll”, “immigration”} ○ Specific custom labels for each client stored in MongoDB ○ Automatically predict the right labels for requests
- 31. What MongoDB is working on: Full Text Search (Beta) ● Based on Apache Lucene 8 ● Integrated into MongoDB Atlas ● Separate process co-located with mongod ● Shard-aware ● Indexing = collection scan -> steady state How Do I use it? Create a cluster on MongoDB Atlas using 4.2 RC (M30+) Create an Full Text Index via the MongoDB Atlas UI or API Query Index via $searchBeta operator using MongoDB Compass or shell, add to your existing aggregation pipelines
- 32. What MongoDB is working on: Atlas Data Lake (beta) ● Serverless: no infrastructure to set up and manage ● Usage-based pricing: only pay for the queries your run ● On-demand: no need to load data; bring your own S3 bucket ● Auto-scalable: parallel execution delivers performance for large and complex queries across multiple user sessions ● Multi-format: JSON, BSON, CSV, TSV, Avro, Parquet ● Integrated with Atlas: users are managed by Atlas, enabled via Atlas console ● The best tools to work with your data: MongoDB Query language enable flexible and efficient data access; integrates with Compass, MongoDB Shell and MongoDB drivers
- 33. What MongoDB is working on: Atlas Data Lake (beta) Operational Analytics Aggregations Machine Learning and AIData Lake in-app analytics Transactional in-app analytics Primary Secondary Secondary AnalyticsAnalytics
- 34. Q&A