500 Nodes https://blog.openai.com/scaling-kubernetes-to-2500-nodes/ Agenda • What is the typical ML workflow and some of their shortcomings • Why DevOps? • Why Containers, Kubernetes, and Helm? • Intro Inception v3 and transfer learning • Automate repeatable ML experiments with containers • Deploy ML components to Kubernetes with Kubeflow • Scale and test ML experiments with Helm • Manage training jobs and pipelines with TF Serving • Rapid prototyping with self-service Jupyter notebook from JupyterHub Simplified ML Workflow/Pipeline What is DevOps? • “A cross-disciplinary community of practice dedicated to the

trainingType == 0): 21 ml = RS. ReinforcementSarsa ( startState ) 22 elif( trainingType == 1): 23 ml = RQ. ReinforcementQlearning ( startState ) 24 elif( trainingType == 2): 25 ml = SS. SpaceSaveSarsa == 3): 27 ml = SQ. SpaceSaveQlearning ( startState ) 28 elif( trainingType == 4): 29 ml = ODS.OneDSarsa(startState ) 30 else: 31 ml = SSSS. SpaceSaveSarsaSpread ( startState ) 38 32 ml.setup () time.time () 56 move = ml.makeMove(board) 57 end = time.time () 58 moveTime = end -start 59 board[move [1]] = board[move [1]] - move [0] 60 playstatistics .write("Move by ML , took " + str(move [0])

Applications  Evaluate in larger scope of real datasets and query workloads  Examples: Customer workload, ML pipelines (dataprep, training, eval) Experiments and Result Presentation 7 706.015 Introduction topics: compression, ML accuracy  “Real” Data Repositories  Wide selection of available datasets w/ different characteristics  UCI ML Repository: https://archive.ics.uci.edu/ml/index.php  Florida tamu.edu/  Google dataset search: https://datasetsearch.research.google.com/  Common Datasets in ML: ImageNet, Mnist, CIFAR, KDD, Criteo  Common Datasets in DM: Census, Taxi, Airlines, DBLP, benchmarks

从 Swift 到 机器器学习 CreateML - Swifter 通向 ML 的⾦金金钥匙 ? 王 巍 (onevcat) 2018.09.15, @Swift Conf. 后移动开发时代 Google Trends: “iOS Develop” WWDC 2013 转变⼀一般都会带来痛苦 如何评价 2017 年年初华为开始「清理理」34 岁以上的职员？ 程序员能纯靠技术渡过中年年危机吗？ 框架背后的特征提取 17KB in Demo Transfer Learning ✅ VisionFeaturePrint_Scene ? VisionFeaturePrint_Screen Core ML Community Tools visionFeaturePrint glmClassifier 开始定义 pipeline 输⼊入 299x299 保留留特性的同时限制处理理时间 VisionFeaturePrint 集成在 iOS 12 和 macOS 10.14 中 问题 No.1 Core ML - iOS 11 • CoreML 从 iOS 11 开始⽀支持 • ⽆无法读取 CreateML 创建的带有 VisionFeaturePrint 的模型 Core ML - iOS 11 使⽤用 TuriCreate 的其他模型进⾏行行特征提取 可以得到与 CreateML

Area = 16,30 $ ￥ € $ €￥ $ £ ￥ ￡ ￥ UDF in ClickHouse Concept, Develpoment, and Application in ML Systems Begin Content Area = 16,30 2 About CraiditX CraiditX 氪信, a finance AI startup since 2015 Contributor Begin Content Area = 16,30 4 OLAP in ML Systems Begin Content Area = 16,30 5 Begin Content Area = 16,30 6 Intensive Tasks in a ML System • Pre-analyzing the data • Extracting features relationship graphs • Generating reports • ... Begin Content Area = 16,30 7 Intensive Tasks in a ML System • Pre-analyzing the data = Finding the useful part of data + Summerizing data • Extracting

09/2018 TU Graz, Austria  BMK endowed chair for data management  Data management for data science (ML systems internals, end-to-end data science lifecycle)  2012-2018 IBM Research – Almaden, USA  Declarative Architecture of ML Systems (AMLS, SS) Data Integration and Large-Scale Analysis (DIA, WS) Master Bachelor Data management from user/application perspective Distributed Data Management ML system internals Topic selection needs time  pipeline model  Ex. Compressed Linear Algebra  Problem: Iterative ML algorithms + memory-bandwidth-bound operations  crucial to fit data in memory  automatic lossless

Third-party apps Manual QA trace: r trace: r trace: r trace: r CI Pipeline | CONFIDENTIAL 16 ML-assisted Context Rule Learning createProduct(…): Response { “productId”: “HDSN1890675”, “src”: is effort intensive Solution • ML-driven identification of candidate relationships • Supervised system to accept true positives • No code! | CONFIDENTIAL 17 ML-assisted Assertion Rule Learning Assertion creation/maintenance is effort intensive Solution • Comprehensive comparison of results • ML-driven identification of decision rules • Human review to accept the learned rules • No code! Test

Frontend Deep Learning Frameworks https://github.com/xilinx© Copyright 2018 Xilinx TVM as Unified ML Front End >> 6 Relay (and NNVM) Graph Parser XIR Compiler Quantizer Partitioner @relay.transform i.e. ZC104/Ultra96) https://github.com/Xilinx/ml-suite/blob/master/examples/caffe/Benchmark_README.md Two measurements we track: Latency & Throughput ˃ ML pipeline contains multiple stages, performance Performance results based on Xilinx own runtime pipeline available in github (https://github.com/Xilinx/ml-suite/blob/master/examples/deployment_modes/mp_classify.py) Streamlined multi-process pipeline using

(BELK) • Beats + Elasticsearch + Logstash + Kibana • 接⼊入层还会搭配Kafka • 重要企业级组件都在商业组件X-Pack中 • 安全、ML、SQL、监控、告警、Transform等 • 提供⼀一个开源免费的APM⽅方案 Kafka + EBLK 引⼊队列，解决丢数据问题 部署、维护复杂度较为复杂 Elasticsearch核⼼能⼒ 阿⾥里里云⽇日志服务 ElasticSearch ML (X-Pack) Splunk AI增强 - 预测 • 对周期性趋势性数据进⾏行行预测 阿⾥里里云⽇日志服务 ElasticSearch ML (X-Pack) Splunk AI增强 - 根因分析 • 关联事件上下⽂文与相关链路路指标，提供根因辅助 阿⾥里里云⽇日志服务 ElasticSearch ML (X-Pack) Splunk 使⽤用开源⽅方案时需要考虑

Advanced SIMD Algorithms in Pictures name : mismatch | size : 10000 | type : char | group : apple_ml | padding : min 400 Advanced SIMD Algorithms in Pictures name : copy not 0s | size ; 10000 | type : int | group : apple_ml | padding : min sd:copy 1l6o0 -ever:algori Advanced SIMD Algorithms in Pictures name : copy not 0s | size ; 10000 | type : short | group : apple_ml | padding : min sd:copy ever:algoricopy_Isparse_output ee:algo::copy_f 45o0 4000 35oo