Podcast2021 MLOps Year in ReviewClaimServing models is the most important topic for 67% of the 1,000 people who completed the community questionnaire.1:43
Threads / Serving and deployment / 2022
Serving and deployment in 2022
37 sessions
Podcast2021 MLOps Year in Review
PodcastOn Structuring an ML Platform 1 Pizza TeamClaimNeoway's machine learning platform team provides internal services for feature engineering, data pipelines, model deployment, and monitoring.2:26
MeetupJust Build It! Tips for Making ML Engineering and MLOps RealClaimAndy McMahon says teams should bridge the gap between machine learning prototypes and production piece by piece instead of trying to deploy everything at once.12:00
Reading groupFederated Learning: Machine Learning on the EdgePushed backThe speaker disputed the idea that federated learning is not production-ready by describing existing large-scale deployments in products.40:02
PodcastCalibration for ML at EtsyClaimErica Greene said Etsy checked that calibration preserved ranking performance by verifying that offline AUC stayed the same.32:08
PodcastPlatform Thinking: A Lemonade Case StudyPushed backAutomatic model testing was not considered sufficient to make the team comfortable deploying a model, so people still had to take responsibility for testing.41:17
PodcastThe Journey from Data Scientist to MLOps EngineerClaimAle Solano says continuous integration and continuous delivery allowed the team to train and deploy models and see experiment results quickly.23:46
PodcastLessons from Studying FAANG ML SystemsClaimThe five fundamental ML platform components are a feature store, workflow orchestration, a model registry, model serving, and model monitoring.6:04
Reading groupMachine Learning Engineering in ActionClaimBen Wilson says his book summarizes more than a decade of mistakes and lessons from building machine learning and statistical inference solutions for multiple companies.1:18
PodcastMLOps at StripeClaimEmmanuel Ameisen says model deployment should be treated as a separate iteration cycle that also needs to be shortened and automated.16:03
MeetupApplications of Data ScienceClaimPallet deployed the classifier as a Dockerized Flask API on Google Cloud Run and calls it from the Pallet backend.23:50
PodcastDon't Listen Unless You Are Going to Do ML in ProductionClaimbanana.dev handles scaling, cost, and latency so customers can focus on model quality.4:18
PodcastML Platform Tradeoffs and Wondering Why to Use ThemClaimJavier Mansilla's team builds shared foundations for machine learning so individual teams do not have to reinvent monitoring, deployment, data safety, and related operational capabilities.7:48
PodcastBringing Audio ML Models into ProductionClaimValerio Velardo recommends that a small startup first build a minimal prototype and then shift to productization, infrastructure, experiment tracking, model registries, automated evaluation, and deployment.23:07
PodcastReal-time Model Inference in a Video Streaming EnvironmentClaimRunway's central engineering challenge is performing model inference on video quickly enough to race a user's playhead.17:03
PodcastGPU For Machine LearningClaimRun:AI helps organizations share GPU resources across data scientists, centralize compute resources across teams, and deploy multiple models efficiently.9:44
Scaling Real-time Machine Learning at ChimeClaimReal-time inference is appropriate when prediction latency is low, someone or something is blocked on the prediction, and information available at the moment is necessary.5:08
PodcastMLOps CritiquesPushed backMatthijs Brouns rejects the idea that model deployment is simply an already solved problem, arguing that important gaps remain in MLOps.17:08
MeetupBuilding a Movie Recommendation System on Tecton with SnowflakeClaimTecton takes data from its different sources, transforms it into machine-learning features, and provides access to those features for training and serving.4:31
MeetupThe Post Modern StackClaimJacopo Tagliabue says the pipeline starts with raw data in a warehouse and ends with a cloud endpoint that serves predictions.22:05
MeetupFLOps with Scaleout's Open-core PlatformClaimScaleout Systems expects future enterprise data centers to have fewer servers distributed across many locations instead of thousands of servers in one location.6:31
PodcastWhy and When to Use Kubeflow for MLOpsClaimKubeflow provides exploration, pipeline orchestration, and serving tools on an open-source platform.6:50
PodcastMLflow vs Kubeflow 2022Pushed backGeorge Pearse argues that Kubeflow's complexity makes it less suitable for small shops, while Byron Allen says Kubeflow has a valid role through managed services and in larger organizations.45:30
MeetupDevOps, Security, and Observability in MLPushed backLuke Marsden says Kubeflow's built-in metadata server is not very good.23:27
PodcastWhy You Need More Than AirflowClaimFlight uses signals to pause workflows for human approval or automated checks before progressing through deployment stages.39:11
PodcastTurning Redis into a Composable, ML Data PlatformPushed backSamuel Partee argued that Redis can be competitive with dedicated vector databases because vector search can be added to an existing Redis stack without a new dependency or service.27:38
PodcastMLflow Pipelines: Opinionated ML Pipelines in MLflowPushed backThe speakers dispute the idea that every model should be served in real time, because whether real-time serving is needed depends on the requirements and use case.15:45
PodcastML Platforms, Where to Start?ClaimThe AI platform initially prioritized simplifying the notebook experience, deploying models into production, and supporting model inference.18:19
PodcastMLOps at DoorDashClaimThe DoorDash machine learning platform aims to cover the applied machine learning lifecycle end to end and at scale, from feature engineering through model serving.4:23
PodcastBringing DevOps Agility to MLPushed backLuis Ceze clarified that his statement that MLOps should not exist was about treating deployed machine learning models like ordinary code, not denying the need for a specialized model-creation workflow.9:08
PodcastDatabricks Model Serving V2ClaimRafael Pierre says Databricks Model Serving V2 provides serverless real-time endpoints as a platform service, reducing the need to manage serving infrastructure.25:53
MeetupApplying DevOps Practices in Data and ML EngineeringPushed backAntoni Ivanov says Versatile Data Kit is focused on batch jobs rather than streaming, while a question asks about serving features with a sub-200-millisecond response time.28:46
PodcastWhat is Data / ML Like on League?ClaimIan Schweer says League's end-of-game data can provide a complete record of a game without requiring many services and databases to be queried to reconstruct it.29:43
MeetupMLOps in Practice: Common Challenges and Lessons LearnedClaimThe speakers group MLOps principles into versioning, testing, monitoring, automation, deployment and reproducibility.8:59
MeetupBuilding an Open Source MLOps Stack with ZenML Part 2ClaimZenML 0.20 changed the architecture from a client-driven model to a client-server model.9:36
PodcastMachine Learning Operations: What Is It and Why Do We Need It?ClaimNiklas Kühl works as a managing consultant in data science for IBM Consulting and leads the Applied AI in Services Lab at Karlsruhe Institute of Technology.0:00
MeetupVertex AI WorkshopClaimA custom prediction container must provide an HTTP server, a health-check endpoint, and a prediction endpoint.38:39