ALBERTSONS - 使用 DATABRICKS SERVERLESS 提供近乎实时的模型服务.pdf

1、2024 Databricks Inc.All rights reserved1NEAR REALNEAR REAL-TIME INFENRECE TIME INFENRECE WITH DATABRICKS WITH DATABRICKS SERVERLESS COMPUTESERVERLESS COMPUTEAmit Adiraju,Diana AdamAmit Adiraju,Diana Adam2024 Databricks Inc.All rights reserved2Amit AdirajuSr.ML EngineerDiana AdamDiana AdamSr.Director

2、 Data ScienceSr.Director Data Science2024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedFORECASTING AT ALBERTSONSFORECASTING AT ALBERTSONS32024 Databricks Inc.All rights reserved2024 Databricks Inc.All rights reservedA future looking view into the estimated performance of

3、our products is critical for the short-and long-term strategy planning of What,Where,When and for How Much to sell.Forecasting needs to be done at the relevant granularity to support these goals and have a high degree of accuracy.Business users need the ability to simulate in near real timethe impac

4、t of different strategies on the forecast in based on various objectives to pick the best one.FORECASTING USE CASE REQUIREMENTSFORECASTING USE CASE REQUIREMENTS4 4Total Number of Forecasting Models124,000Algorithms UsedFB Prophet,Moving Average,Temporal Fusion Transformer and GPVARRegistryMlflow(mod

5、el registry,as well as experiment tracking)Inference&ServingRun simulations for different inputs in less than 15 seconds per request(1400 individual forecast models per request)GeographyBusiness StrategyProduct HierarchyForecasting Keys2024 Databricks Inc.All rights reserved2024 Databricks Inc.All r

6、ights reserved1.On Demand Batch Inference Job with group of models.2.One REST end point per model.1.User Request-FAST API on AKS-Trigger Job Cluster for Inference.2.Deploy one prophet model,per REST endpoint,and query group of endpoints based on users request.1.Cluster Startup time=3 mins+model infe