flinkml: large scale machine learning with apache flink

FlinkML: Large-scale Machine Learning with Apache FlinkTheodore Vasiloudis, SICS

SICS Data Science DayOctober 21st, 2015

Apache Flink

What is Apache Flink?

● Large-scale data processing engine● Easy and powerful APIs for batch and real-time streaming analysis● Backed by a very robust execution backend

○ true streaming dataflow engine○ custom memory manager○ native iterations○ cost-based optimizer

What is Apache Flink?

What does Flink give us?

● Expressive APIs● Pipelined stream processor● Closed loop iterations

Expressive APIs

● Main distributed data abstraction: DataSet● Program using functional-style transformations, creating a Dataflow.

case class Word(word: String, frequency: Int)

val lines: DataSet[String] = env.readTextFile(...)

lines.flatMap(line => line.split(“ “).map(word => Word(word, 1)).groupBy(“word”).sum(“frequency”).print()

Pipelined Stream Processor

Iterate in the Dataflow

Iterate by looping

● Loop in client submits one job per iteration step● Reuse data by caching in memory or disk

Iterate in the Dataflow

Delta iterations

Delta iterations

Learn more in Vasia’s Gelly talk!

Large-scale Machine Learning

What do we mean?

What do we mean?

● Small-scale learning ● Large-scale learning

Source: Léon Bottou

What do we mean?

● Small-scale learning○ We have a small-scale learning problem

when the active budget constraint is the number of examples.

● Large-scale learning

Source: Léon Bottou

What do we mean?

● Small-scale learning○ We have a small-scale learning problem

when the active budget constraint is the number of examples.

● Large-scale learning○ We have a large-scale learning problem

when the active budget constraint is the computing time.

Source: Léon Bottou

What do we mean?

● What about the complexity of the problem?

What do we mean?

● What about the complexity of the problem?

Source: Wired Magazine

Deep learning

What do we mean?

● What about the complexity of the problem?

“When you get to a trillion [parameters], you’re getting to something that’s got a chance of really understanding some stuff.” - Hinton, 2013

Source: Wired Magazine

What do we mean?

● We have a large-scale learning problem when the active budget constraint is the computing time and/or the model complexity.

FlinkML

FlinkML

● New effort to bring large-scale machine learning to Flink

FlinkML

● New effort to bring large-scale machine learning to Flink● Goals:

○ Truly scalable implementations○ Keep glue code to a minimum○ Ease of use

FlinkML: Overview

● Supervised Learning○ Optimization framework○ SVM○ Multiple linear regression

FlinkML: Overview

● Supervised Learning○ Optimization framework○ SVM○ Multiple linear regression

● Recommendation○ Alternating Least Squares (ALS)

FlinkML: Overview

● Supervised Learning○ Optimization framework○ SVM○ Multiple linear regression

● Recommendation○ Alternating Least Squares (ALS)

● Pre-processing○ Polynomial features○ Feature scaling

FlinkML: Overview

● Supervised Learning○ Optimization framework○ SVM○ Multiple linear regression

● Recommendation○ Alternating Least Squares (ALS)

● Pre-processing○ Polynomial features○ Feature scaling

● sklearn-like ML pipelines

FlinkML API

// LabeledVector is a feature vector with a label (class or real value)val trainingData: DataSet[LabeledVector] = ...val testingData: DataSet[Vector] = ...

FlinkML API

// LabeledVector is a feature vector with a label (class or real value)val trainingData: DataSet[LabeledVector] = ...val testingData: DataSet[Vector] = ...

val mlr = MultipleLinearRegression() .setStepsize(0.01) .setIterations(100) .setConvergenceThreshold(0.001)

FlinkML API

// LabeledVector is a feature vector with a label (class or real value)val trainingData: DataSet[LabeledVector] = ...val testingData: DataSet[Vector] = ...

val mlr = MultipleLinearRegression() .setStepsize(0.01) .setIterations(100) .setConvergenceThreshold(0.001)

mlr.fit(trainingData)

FlinkML API

// LabeledVector is a feature vector with a label (class or real value)val trainingData: DataSet[LabeledVector] = ...val testingData: DataSet[Vector] = ...

val mlr = MultipleLinearRegression() .setStepsize(0.01) .setIterations(100) .setConvergenceThreshold(0.001)

mlr.fit(trainingData)

// The fitted model can now be used to make predictionsval predictions: DataSet[LabeledVector] = mlr.predict(testingData)

FlinkML Pipelines

val scaler = StandardScaler()val polyFeatures = PolynomialFeatures().setDegree(3)val mlr = MultipleLinearRegression()

FlinkML Pipelines

val scaler = StandardScaler()val polyFeatures = PolynomialFeatures().setDegree(3)val mlr = MultipleLinearRegression()

// Construct pipeline of standard scaler, polynomial features and multiple linear // regressionval pipeline = scaler.chainTransformer(polyFeatures).chainPredictor(mlr)

FlinkML Pipelines

val scaler = StandardScaler()val polyFeatures = PolynomialFeatures().setDegree(3)val mlr = MultipleLinearRegression()

// Construct pipeline of standard scaler, polynomial features and multiple linear // regressionval pipeline = scaler.chainTransformer(polyFeatures).chainPredictor(mlr)

// Train pipelinepipeline.fit(trainingData)

// Calculate predictionsval predictions = pipeline.predict(testingData)

State of the art in large-scale ML

Alternating Least Squares

R ≅ X Y✕Users

Items

Naive Alternating Least Squares

Blocked Alternating Least Squares

Blocked ALS performance

FlinkML blocked ALS performance

Going beyond SGD in large-scale optimization

● Beyond SGD → Use Primal-Dual framework

● Slow updates → Immediately apply local updates

● Average over batch size → Average over K (nodes) << batch size

CoCoA: Communication Efficient Coordinate Ascent

Primal-dual framework

Source: Smith (2014)

Primal-dual framework

Source: Smith (2014)

Immediately Apply Updates

Source: Smith (2014)

Immediately Apply Updates

Source: Smith (2014)Source: Smith (2014)

Average over nodes (K) instead of batches

Source: Smith (2014)

CoCoA: Communication Efficient Coordinate Ascent

CoCoA performance

Source:Jaggi (2014)

CoCoA performance

Available on FlinkML

SVM

Achieving model parallelism:The parameter server

● The parameter server is essentially a distributed key-value store with two

basic commands: push and pull○ push updates the model

○ pull retrieves a (lazily) updated model

● Allows us to store a model into multiple nodes, read and update it as

needed.

Architecture of a parameter server communicating with groups of workers.

Source: Li (2014)

Comparison with other large-scale learning systems.

Source: Li (2014)

Dealing with stragglers: SSP Iterations

● BSP: Bulk Synchronous parallel○ Every worker needs to wait for the others to finish before starting the next iteration

Dealing with stragglers: SSP Iterations

● BSP: Bulk Synchronous parallel○ Every worker needs to wait for the others to finish before starting the next iteration

● ASP: Asynchronous parallel○ Every worker can work individually, update model as needed.

Dealing with stragglers: SSP Iterations

● BSP: Bulk Synchronous parallel○ Every worker needs to wait for the others to finish before starting the next iteration

● ASP: Asynchronous parallel○ Every worker can work individually, update model as needed.○ Can be fast, but can often diverge.

Dealing with stragglers: SSP Iterations

● BSP: Bulk Synchronous parallel○ Every worker needs to wait for the others to finish before starting the next iteration

● ASP: Asynchronous parallel○ Every worker can work individually, update model as needed.○ Can be fast, but can often diverge.

● SSP: State Synchronous parallel○ Relax constraints, so slowest workers can be up to K iterations behind fastest ones.

Dealing with stragglers: SSP Iterations

● BSP: Bulk Synchronous parallel○ Every worker needs to wait for the others to finish before starting the next iteration

● ASP: Asynchronous parallel○ Every worker can work individually, update model as needed.○ Can be fast, but can often diverge.

● SSP: State Synchronous parallel○ Relax constraints, so slowest workers can be up to K iterations behind fastest ones.○ Allows for progress, while keeping convergence guarantees.

Dealing with stragglers: SSP Iterations

Dealing with stragglers: SSP Iterations

Source: Ho et al. (2013)

SSP Iterations in Flink: Lasso Regression

Source: Peel et al. (2015)

SSP Iterations in Flink: Lasso Regression

Source: Peel et al. (2015)

To be merged soon

into FlinkML

Current and future work on FlinkML

Coming soon

● Tooling○ Evaluation & cross-validation framework○ Predictive Model Markup Language

● Algorithms○ Quad-tree kNN search○ Efficient streaming decision trees○ k-means and extensions○ Colum-wise statistics, histograms

FlinkML Roadmap

● Hyper-parameter optimization● More communication-efficient optimization algorithms● Generalized Linear Models● Latent Dirichlet Allocation

Future of Machine Learning on Flink

● Streaming ML○ Flink already has SAMOA bindings.○ We plan to kickstart the streaming ML library of Flink, and develop new algorithms.

Future of FlinkML

● Streaming ML○ Flink already has SAMOA bindings.○ We plan to kickstart the streaming ML library of Flink, and develop new algorithms.

● “Computation efficient” learning○ Utilize hardware and develop novel systems and algorithms to achieve large-scale learning

with modest computing resources.

Recent large-scale learning systems

Source: Xing (2015)

Recent large-scale learning systems

Source: Xing (2015)

How to get here?

Demo?

Thank you.

@thvasilotvas@sics.se

References

● Flink Project: flink.apache.org● FlinkML Docs: https://ci.apache.org/projects/flink/flink-docs-master/libs/ml/● Leon Botou: Learning with Large Datasets● Wired: Computer Brain Escapes Google's X Lab to Supercharge Search● Smith: CoCoA AMPCAMP Presentation● CMU Petuum: Petuum Project● Jaggi (2014): “Communication-efficient distributed dual coordinate ascent." NIPS 2014.● Li (2014): "Scaling distributed machine learning with the parameter server." OSDI 2014.● Ho (2013): "More effective distributed ML via a stale synchronous parallel parameter server." NIPS

2013.● Peel (2015): “Distributed Frank-Wolfe under Pipelined Stale Synchronous Parallelism”, IEEE BigData

2015● Xing (2015): “Petuum: A New Platform for Distributed Machine Learning on Big Data”, KDD 2015

I would like to thank professor Eric Xing for his permission to use parts of the structure from his great tutorial on large-scale machine learning: A New Look at the System, Algorithm and Theory Foundations of Distributed Machine Learning

“Demo”

“Demo”

“Demo”

“Demo”

“Demo”

“Demo”

“Demo”

“Demo”

Thank you.

@thvasilotvas@sics.se