Canvas 1
New
Box Model Store
Line MT
Line FE
Line DP
Line DI
Box Model Serving
Box Performance Monitoring
Box Model Training
Box Model Store
Box Feature Engineering
Box Data Preperation
Box Data Ingestion
MD-Stage
M-Deployment-Outer
Model
Deployment
M-Deployment-Inner
MV-Stage
M-Validation-Outer
Model
Validation
M-Validation-Inner
MT-Stage
M-Training-Outer
Model
Training
M-Training-Inner
FE-Stage
F-Engineering-Outer
Feature
Engineering
F-Engineering-Inner
DP-Stage
D-Analysis-Outer
Data
processing
D-Analysis-Inner
DI-Stage
D-Ingestion-Outer
Data
Ingestion
D-Ingestion-Inner
DPS
Data
Transformation
Platform
DPS
Data
Transformation
Platform
DPS
Data
Transformation
Platform
Central OSP
Central
Object
Storage
Platform
Ingestion Service
Ingestion
Service
Ingestion Service
Ingestion
Service
Ingestion Service
Ingestion
Service
Data Source C
Data
Source
C
Data Source B
Data
Source
B
Data Source A
Data
Source
A
Dedicated pipeline for
independent and
concurrent processing
In-Memory data processing engine
(Spark Cluster)
Unstructured Data
(HDFS)
Streaming Data Aggregation
(Apache Flume)
Feature
Store
Data
Warehouse
Typically structured data
(Tanzu Greenplum)
In-Memory key-value store
(Redis)
Model
Training
Cluster
Light-weight downstream layer
serving data from data warehouse
(Container)
Monster VMs (24TB)
Bring Data close to Training cluster
(vSAN or Pure with NVME over
Fabric)
Host
Host
Host
GPU
GPU
GPU
GPU
Network Accessible Pool of GPUs
(VMware Bitfusion)
Multiple GPUs per host
(Horovod for distributed
scheduling)
Persistent memory
Project Capitola
Clusters of CPU
(NVIDIA vGPU | SR-IOV)
Kubernetes cluster
(TKGS)
Model Store
S3 API compatible Storage
(Minio)
MS-Stage
M-Serving-Outer
Model
Serving
M-Serving-Inner
Model
Serving
Platform
Deploy models as microservices
(Seldon Core / KFserving)
Central Model Repository
(Seldon)
Kubernetes Deployment
(TKGS)
Uses service mesh
(Istio)
Box Model Serving
App/website
Apps Tooling
K8S Tooling
MM-Stage
M-Monitoring-Outer
Model
Monitoring
M-Monitoring-Inner
Downstream Monitoring
(Grafana / Prometheus)
Upstream Monitoring
Extract, Transform, and Load
(ETL) data into Data Warehouse
Compute into feature
Host
GPU
GPU
GPU
GPU
GPU
GPU
GPU
GPU
GPU
GPU
GPU
GPU
Serving
Infrastructure
Optimization
(OctoML)