Using Spark Mllib Models in a Production Training and Serving Platform: Experien

下载 2

Spark开源社区

发布于

8235

人观看

#信息技术

Overview Uber’s Michelangelo is a machine learning platform that supports training and serving thousands of models in production. Most Michelangelo customer models are based on Spark Mllib. In this talk, we will describe Michelangelo’s experiences with and evolving use of Spark Mllib, particularly in the areas of model persistence and online serving. Extended Description Michelangelo [https://eng.uber.com/michelangelo/] was originally developed to support scalable machine learning for production models. Its end-to-end support for scheduled Spark-based data ingestion and model training, along with model evaluation and deployment for batch and online model serving, has gained wide acceptance across Uber. More recently, Michelangelo is evolving to handle more use cases, including evaluating and serving models trained outside of core Michelangelo, e.g., on a distributed tensorflow platform providing Horovod [https://eng.uber.com/horovod/] or using PySpark in a Jupyter notebook on Data Science Workbench [https://eng.uber.com/dsw/] To support evaluation and serving of models trained outside of Michelangelo, Michelangelo’s use of Spark Mllib needed updating, to generalize its mechanisms for model persistence and online serving. In this talk, we will describe these mechanisms and explore possible avenues for open-sourcing them.

展开查看详情

1 .WIFI SSID:SparkAISummit | Password: UnifiedAnalytics

2 .Using Spark MLlib Models in a Production Training and Serving Platform: Experiences and Extensions Anne Holler, Michael Mui Uber #UnifiedAnalytics #SparkAISummit

3 .Introduction ● Michelangelo is Uber’s Machine Learning Platform ○ Supports training, evaluation, and serving of ML models in production ○ Uses Spark MLlib for training and serving at scale ● Michelangelo's use of Spark MLlib has evolved over time ○ Initially used as part of a monolithic training and serving platform, with hardcoded model pipeline stages saved/loaded from protobuf ○ Initially customized to support online serving, with online serving APIs added to Transformers ad hoc #UnifiedAnalytics #SparkAISummit 3

4 .What are Spark Pipelines: Estimators and Transformers Estimator: Spark abstraction of a learning algorithm or any algorithm that ﬁts or trains on data Transformer: Spark abstraction of an ML model stage that includes feature transforms and predictors #UnifiedAnalytics #SparkAISummit 4

5 .Pipeline Models Encode Operational Steps #UnifiedAnalytics #SparkAISummit 5

6 .Pipeline Models Enforce Consistency ● Both Training and Serving involve pre- and post- transform stages in addition to raw ﬁtting and inferencing from ML model that need to be consistent: ○ Data Transformations ○ Feature Extraction and Pre-Processing ○ ML Model Raw Predictions ○ Post-Prediction Transformations #UnifiedAnalytics #SparkAISummit 6

7 . Pipeline Models Encapsulate Complexity ML Workﬂow In Practice #UnifiedAnalytics #SparkAISummit 7

8 .Complexity arises from Diﬀerent Workﬂow Needs #UnifiedAnalytics #SparkAISummit 8

9 . Complexity arises from Diﬀerent User Needs Data Analysts / Data Engineers / Software Engineers Research Scientists / Data Scientists / Research/ML Engineers ML Engineers / Production Engineers #UnifiedAnalytics #SparkAISummit 9

10 .Evolution Goal: Retain Performance and Consistency ● Requirement 1: Performant distributed batch serving that comes with the DataFrame-based execution model on top of Spark’s SQL Engine ● Requirement 2: Low-latency (P99 latency <10ms), high throughput solution for real-time serving ● Requirement 3: Support consistency in batch and real-time prediction accuracy by running through common code paths whenever practical #UnifiedAnalytics #SparkAISummit 10

11 .Evolution Goal: Increase Flexibility and Velocity ● Requirement 1: Flexibility in model deﬁnitions: libraries, frameworks ○ Allow users to deﬁne model pipelines (custom Estimator/Transformer) ○ Train and serve those models eﬃciently ● Requirement 2: Flexibility in Michelangelo use ○ Decouple its monolithic structure into components ○ Allow interoperability with non-Michelangelo components / pipelines ● Requirement 3: Faster / Easier Spark upgrade path ○ Replace custom protobuf model representation ○ Formalize online serving APIs #UnifiedAnalytics #SparkAISummit 11

12 .Evolve: Replacing Protobuf Model Representation ● Considered MLeap, PMML, PFA, Spark PipelineModel: all supported in Spark MLlib ○ MLeap: non-standard, impacting interoperability w/ Spark compliant ser/de ○ MLeap, PMML, PFA: Lag in supporting new Spark Transformers ○ MLeap, PMML, PFA: Risk of inconsistent model training/serving behavior ● Wanted to choose Spark PipelineModel representation for Michelangelo models ○ Avoids above shortcomings ○ Provides simple interface for adding estimators/transformers ○ But has challenges in Online Serving (see Pentreath’s Spark Summit 2018 talk) ■ Spark MLlib PipelineModel load latency too large ■ Spark MLlib serving APIs too slow for online serving #UnifiedAnalytics #SparkAISummit 12

13 .Spark PipelineModel Representation ● Spark PipelineModel format example ﬁle structure ├── 0_strIdx_9ec54829bd7c │ ├── data part-00000-a9f31485-4200-4845-8977-8aec7fa03157.snappy.parquet │ ├── metadata part-00000 ├── 1_strIdx_5547304a5d3d │ ├── data part-00000-163942b9-a194-4023-b477-a5bfba236eb0.snappy.parquet │ ├── metadata part-00000 ├── 2_vecAssembler_29b5569f2d98 │ ├── metadata part-00000 ├── 3_glm_0b885f8f0843 │ ├── data part-00000-0ead8860-f596-475f-96f3-5b10515f075e.snappy.parquet │ └── metadata part-00000 └── 4_idxToStr_968f207b70f2 ├── metadata part-00000 ● Format Read/Written by Spark MLReadable/MLWritable trait MLReadable[T] { def read : org.apache.spark.ml.util.MLReader[T] def load(path : scala.Predef.String) : T } trait MLWritable { def write: org.apache.spark.ml.util.MLWriter def save(path : scala.Predef.String) } #UnifiedAnalytics #SparkAISummit 13

14 .Challenge: Spark PipelineModel Load Latency ● Zipped Spark Pipeline and Pipeline Model Type Spark Pipeline / Protobuf Load protobuf ﬁles were comparable sizes (up to 10s of MBs) GBDT Regression 21.22x GBDT Binary Classification 28.63x ● Spark Pipeline load latency was very high relative to custom Linear Regression 29.94x protobuf load latency Logistic Regression 43.97x ● Impacts online serving resource RF Binary Classification 8.05x agility and health monitoring RF Regression 12.16x #UnifiedAnalytics #SparkAISummit 14

15 .Tuning Load Latency: Part 1 Replaced sc.textfile with local metadata read ● DefaultParamsReadable.load uses sc.textfile ● Forming RDD of strings for small 1-line ﬁle was slower than simple load ● Replaced with java I/O for local ﬁle case, which was much faster ○ Updated loadMetadata method in mllib/src/main/scala/org/apache/spark/ml/util/ReadWrite.scala ● Big reduction in latency of metadata read #UnifiedAnalytics #SparkAISummit 15

16 .Tuning Load Latency: Part 2 Replaced sparkSession.read.parquet with ParquetUtil.read ● Spark distributed read/select for small Transformer data was very slow ● Replaced with direct parquet read/getRecord, which was much faster ○ Relevant to Transformers like LogisticRegression, StringIndexer, LinearRegression ● Signiﬁcant reduction in latency of Transformer data read #UnifiedAnalytics #SparkAISummit 16

17 .Tuning Load Latency: Part 3 Updated Tree Ensemble model data save and load to use Parquet directly ● Coalesced tree ensemble node and metadata weights DataFrames at save time to avoid writing large number of small ﬁles that are slow to read ● Loading Tree Ensemble models invoked a groupByKey,sortByKey ○ Spark distributed read/select/sort/collect was very slow ● Replaced with direct parquet read/getRecord, which was much faster ● Signiﬁcant reduction in latency of tree ensemble data read #UnifiedAnalytics #SparkAISummit 17

18 .Before and After: Tuned Pipeline Load Latency Pipeline Model Type Spark Pipeline / Tuned Spark Greatly improved MLLib load Protobuf Load Pipeline / Protobuf latency, while retaining Load current on-disk format! GBDT Regression 21.22x 2.05x GBDT Binary 28.63x 2.50x Classification Linear Regression 29.94x 2.03x Logistic Regression 43.97x 2.88x RF Binary Classification 8.05x 3.14x RF Regression 12.16x 3.01x #UnifiedAnalytics #SparkAISummit 18

19 .Challenge: SparkContext Cleaner Performance ● Michelangelo online serving creates local SparkContext to handle load of any unoptimized Transformers ● Periodic context cleaner runs induced non-trivial latency in serving request responses ● Solution: Stopped SparkContext when models not actively being loaded. ○ Model load only happens at service startup or when new models are deployed into production online serving #UnifiedAnalytics #SparkAISummit 19

20 .Challenge: Serving APIs too slow for online serving ● Added OnlineTransformer trait to Transformers to be served online ○ Single & small list APIs which leverage low-level spark predict methods trait OnlineTransformer { def scoreInstances(instances: List[Map[String, Any]]): List[Map[String, Any]] def scoreInstance(instance: Map[String, Any]): Map[String, Any] } ○ Injected at Transformer load time, so pipeline models trained outside of Michelangelo can be served online by Michelangelo #UnifiedAnalytics #SparkAISummit 20

21 .Michelangelo Use of Spark MLlib Evolution Outcome ● Michelangelo is using updated Spark MLlib interface in production ○ Spark PipelineModel on-disk representation ○ Optimized Transformer loads to support online serving ○ OnlineTransformer trait to provide online serving APIs #UnifiedAnalytics #SparkAISummit 21

22 .Example Use Cases Enabled by Evolved MA MLlib ● Flexible Pipeline Model Deﬁnition ○ Model Pipeline including TFTransformer ● Flexible Use of Michelangelo ○ Train Model in Notebook, Serve Model in Michelangelo #UnifiedAnalytics #SparkAISummit 22

23 .Flexible Pipeline Model Deﬁnition ● Interoperability with non-Michelangelo components / pipelines ○ Cross framework, system, language support via Estimators / Transformers ● Allow customizability of PipelineModel, Estimators, Transformers while fully integrated into Michelangelo’s Training and Serving infrastructure ○ Combines Spark’s Data Processing with Training using custom libraries e.g. XGBoost, Tensorﬂow #UnifiedAnalytics #SparkAISummit 23

24 .Flexible Pipeline Deﬁnition Example: TFTransformer ● Serving TensorFlow Models with TFTransformer https://eng.uber.com/cota-v2/ ○ Spark Pipeline built from training contains both data processing transformers and TensorFlow transformations (TFTransformer) ○ P95 serving latency < 10ms ○ Combines the distributed computation of Spark and low-latency serving using CPUs and the acceleration of DL training using GPUs #UnifiedAnalytics #SparkAISummit 24

25 .Serving TF Models using TFTransformer #UnifiedAnalytics #SparkAISummit 25

26 .Flexible Use Example: Train in DSW, Serve in MA ● Decouple Michelangelo into functional components ● Consolidate custom data processing, feature engineering, model deﬁnition, train, and serve around notebook environments (DSW) #UnifiedAnalytics #SparkAISummit 26

27 .Experiment in DSW, Serve in Michelangelo #UnifiedAnalytics #SparkAISummit 27

28 .Key Learnings in Evolving Michelangelo ● Pipeline representation of models is powerful ○ Encodes all steps in operational modeling ○ Enforces consistency between training and serving ● Pipeline representation of models needs to be ﬂexible ○ Model pipeline can encapsulate complex stages ○ Complexity stems from diﬀering workﬂow and user needs #UnifiedAnalytics #SparkAISummit 28

29 .Conclusion ● Michelangelo updated use of Spark MLlib is working well in production ● Propose to open source our changes to Spark MLlib ○ Submitted Spark MLlib Online Serving SPIP ■ https://issues.apache.org/jira/browse/SPARK-26247 ○ Posted 2 patches ■ Patch to reduce spark pipeline load latency ■ Patch to add OnlineTransformer trait for online serving APIs #UnifiedAnalytics #SparkAISummit 29

0点赞

0收藏

2下载