AWSTemplateFormatVersion: "2010-09-09"
Description: "Amazon CloudWatch dashboard to visualize multimodel endpoint metrics"
Parameters:
  EndpointC5XL:
    Type: String
    Description: 'Name for the SageMaker endpoint deployed with ml.c5.xlarge instance. NOT the ARN.'
  EndpointC52XL:
    Type: String
    Description: 'Name for the SageMaker endpoint deployed with ml.c5.2xlarge instance. NOT the ARN.'

Resources:
  MMECloudWatchDashboard:
      Type: AWS::CloudWatch::Dashboard
      Properties:
        DashboardBody: !Sub '{
            "widgets": [
                {
                    "height": 6,
                    "width": 6,
                    "y": 12,
                    "x": 18,
                    "type": "metric",
                    "properties": {
                        "metrics": [
                            [ "/aws/sagemaker/Endpoints", "CPUUtilization", "EndpointName", "${EndpointC5XL}", "VariantName", "AllTraffic" ],
                            [ ".", "MemoryUtilization", ".", ".", ".", "." ],
                            [ ".", "DiskUtilization", ".", ".", ".", "." ],
                            [ ".", "LoadedModelCount", ".", ".", ".", ".", { "yAxis": "right" } ]
                        ],
                        "view": "timeSeries",
                        "stacked": false,
                        "region": "${AWS::Region}",
                        "title": "c5.xl-instance-metrics",
                        "period": 60,
                        "stat": "Average",
                        "yAxis": {
                            "right": {
                                "label": "",
                                "min": 30,
                                "max": 55
                            },
                            "left": {
                                "min": 0,
                                "max": 120
                            }
                        }
                    }
                },
                {
                    "height": 6,
                    "width": 9,
                    "y": 12,
                    "x": 9,
                    "type": "metric",
                    "properties": {
                        "metrics": [
                            [ "AWS/SageMaker", "InvocationsPerInstance", "EndpointName", "${EndpointC5XL}", "VariantName", "AllTraffic", { "stat": "Sum", "visible": false } ],
                            [ ".", "Invocations", ".", ".", ".", ".", { "stat": "Sum", "visible": false } ],
                            [ ".", "ModelLoadingTime", ".", ".", ".", "." ],
                            [ ".", "Invocation5XXErrors", ".", ".", ".", ".", { "stat": "Sum", "visible": false } ],
                            [ ".", "Invocation4XXErrors", ".", ".", ".", ".", { "stat": "Sum", "visible": false } ],
                            [ ".", "ModelCacheHit", ".", ".", ".", ".", { "stat": "Sum", "visible": false } ],
                            [ ".", "ModelDownloadingTime", ".", ".", ".", "." ],
                            [ ".", "ModelLoadingWaitTime", ".", ".", ".", "." ],
                            [ ".", "ModelLatency", ".", ".", ".", "." ],
                            [ ".", "OverheadLatency", ".", ".", ".", "." ],
                            [ ".", "ModelUnloadingTime", ".", ".", ".", "." ]
                        ],
                        "view": "timeSeries",
                        "stacked": true,
                        "title": "c5.xl-MME-latency-metrics",
                        "region": "${AWS::Region}",
                        "period": 60,
                        "stat": "Average"
                    }
                },
                {
                    "height": 3,
                    "width": 24,
                    "y": 9,
                    "x": 0,
                    "type": "metric",
                    "properties": {
                        "metrics": [
                            [ "AWS/SageMaker", "Invocation5XXErrors", "EndpointName", "${EndpointC5XL}", "VariantName", "AllTraffic", { "stat": "Sum", "label": "Invocation5XXErrors", "visible": false } ],
                            [ ".", "Invocation4XXErrors", ".", ".", ".", ".", { "stat": "Sum", "label": "Invocation4XXErrors", "visible": false } ],
                            [ ".", "Invocations", ".", ".", ".", ".", { "stat": "Sum", "label": "Invocations" } ],
                            [ ".", "InvocationsPerInstance", ".", ".", ".", ".", { "stat": "Sum", "label": "InvocationsPerInstance", "visible": false } ],
                            [ ".", "ModelLatency", ".", ".", ".", ".", { "label": "ModelLatency" } ],
                            [ ".", "OverheadLatency", ".", ".", ".", ".", { "label": "OverheadLatency" } ],
                            [ ".", "ModelLoadingWaitTime", ".", ".", ".", ".", { "label": "ModelLoadingWaitTime" } ],
                            [ ".", "ModelDownloadingTime", ".", ".", ".", ".", { "label": "ModelDownloadingTime", "visible": false } ],
                            [ ".", "ModelLoadingTime", ".", ".", ".", ".", { "label": "ModelLoadingTime" } ],
                            [ ".", "ModelUnloadingTime", ".", ".", ".", ".", { "label": "ModelUnloadingTime" } ],
                            [ "/aws/sagemaker/Endpoints", "LoadedModelCount", ".", ".", ".", "." ],
                            [ "AWS/SageMaker", "ModelCacheHit", ".", ".", ".", ".", { "label": "ModelCacheHit", "visible": false } ]
                        ],
                        "view": "singleValue",
                        "region": "${AWS::Region}",
                        "stat": "Average",
                        "period": 60,
                        "title": "c5.xl Latest values"
                    }
                },
                {
                    "height": 6,
                    "width": 9,
                    "y": 12,
                    "x": 0,
                    "type": "metric",
                    "properties": {
                        "metrics": [
                            [ "AWS/SageMaker", "InvocationsPerInstance", "EndpointName", "${EndpointC5XL}", "VariantName", "AllTraffic" ],
                            [ ".", "Invocations", ".", ".", ".", "." ],
                            [ ".", "Invocation5XXErrors", ".", ".", ".", ".", { "yAxis": "right" } ],
                            [ ".", "Invocation4XXErrors", ".", ".", ".", ".", { "yAxis": "right" } ],
                            [ ".", "ModelCacheHit", ".", ".", ".", "." ]
                        ],
                        "view": "timeSeries",
                        "stacked": false,
                        "region": "${AWS::Region}",
                        "stat": "Sum",
                        "period": 60,
                        "title": "c5.xl-Invocations-modelcache",
                        "yAxis": {
                            "right": {
                                "label": "Errors"
                            },
                            "left": {
                                "label": "Invocations"
                            }
                        }
                    }
                },
                {
                    "height": 6,
                    "width": 6,
                    "y": 3,
                    "x": 18,
                    "type": "metric",
                    "properties": {
                        "metrics": [
                            [ "/aws/sagemaker/Endpoints", "CPUUtilization", "EndpointName", "${EndpointC52XL}", "VariantName", "AllTraffic" ],
                            [ ".", "MemoryUtilization", ".", ".", ".", "." ],
                            [ ".", "DiskUtilization", ".", ".", ".", "." ],
                            [ ".", "LoadedModelCount", ".", ".", ".", ".", { "yAxis": "right" } ]
                        ],
                        "view": "timeSeries",
                        "stacked": false,
                        "region": "${AWS::Region}",
                        "title": "c5.2xl-instance-metrics",
                        "period": 60,
                        "stat": "Average",
                        "yAxis": {
                            "left": {
                                "min": 0,
                                "max": 120
                            },
                            "right": {
                                "min": 30,
                                "max": 55
                            }
                        }
                    }
                },
                {
                    "height": 6,
                    "width": 9,
                    "y": 3,
                    "x": 9,
                    "type": "metric",
                    "properties": {
                        "metrics": [
                            [ "AWS/SageMaker", "InvocationsPerInstance", "EndpointName", "${EndpointC52XL}", "VariantName", "AllTraffic", { "stat": "Sum", "visible": false } ],
                            [ ".", "Invocations", ".", ".", ".", ".", { "stat": "Sum", "visible": false } ],
                            [ ".", "ModelLoadingTime", ".", ".", ".", "." ],
                            [ ".", "Invocation5XXErrors", ".", ".", ".", ".", { "stat": "Sum", "visible": false } ],
                            [ ".", "Invocation4XXErrors", ".", ".", ".", ".", { "stat": "Sum", "visible": false } ],
                            [ ".", "ModelCacheHit", ".", ".", ".", ".", { "stat": "Sum", "visible": false } ],
                            [ ".", "ModelDownloadingTime", ".", ".", ".", "." ],
                            [ ".", "ModelLoadingWaitTime", ".", ".", ".", "." ],
                            [ ".", "ModelLatency", ".", ".", ".", "." ],
                            [ ".", "OverheadLatency", ".", ".", ".", "." ],
                            [ ".", "ModelUnloadingTime", ".", ".", ".", "." ]
                        ],
                        "view": "timeSeries",
                        "stacked": true,
                        "title": "c5.2xl-MME-latency-metrics",
                        "region": "${AWS::Region}",
                        "period": 60,
                        "stat": "Average"
                    }
                },
                {
                    "height": 3,
                    "width": 24,
                    "y": 0,
                    "x": 0,
                    "type": "metric",
                    "properties": {
                        "metrics": [
                            [ "AWS/SageMaker", "Invocation5XXErrors", "EndpointName", "${EndpointC52XL}", "VariantName", "AllTraffic", { "stat": "Sum", "label": "Invocation5XXErrors", "visible": false } ],
                            [ ".", "Invocation4XXErrors", ".", ".", ".", ".", { "stat": "Sum", "label": "Invocation4XXErrors", "visible": false } ],
                            [ ".", "Invocations", ".", ".", ".", ".", { "stat": "Sum", "label": "Invocations" } ],
                            [ ".", "InvocationsPerInstance", ".", ".", ".", ".", { "stat": "Sum", "label": "InvocationsPerInstance", "visible": false } ],
                            [ ".", "ModelLatency", ".", ".", ".", ".", { "label": "ModelLatency" } ],
                            [ ".", "OverheadLatency", ".", ".", ".", ".", { "label": "OverheadLatency" } ],
                            [ ".", "ModelLoadingWaitTime", ".", ".", ".", ".", { "label": "ModelLoadingWaitTime" } ],
                            [ ".", "ModelDownloadingTime", ".", ".", ".", ".", { "label": "ModelDownloadingTime", "visible": false } ],
                            [ ".", "ModelLoadingTime", ".", ".", ".", ".", { "label": "ModelLoadingTime" } ],
                            [ ".", "ModelUnloadingTime", ".", ".", ".", ".", { "label": "ModelUnloadingTime" } ],
                            [ "/aws/sagemaker/Endpoints", "LoadedModelCount", ".", ".", ".", "." ],
                            [ "AWS/SageMaker", "ModelCacheHit", ".", ".", ".", ".", { "label": "ModelCacheHit", "visible": false } ]
                        ],
                        "view": "singleValue",
                        "region": "${AWS::Region}",
                        "stat": "Average",
                        "period": 60,
                        "title": "c5.2xl Latest values"
                    }
                },
                {
                    "height": 6,
                    "width": 9,
                    "y": 3,
                    "x": 0,
                    "type": "metric",
                    "properties": {
                        "metrics": [
                            [ "AWS/SageMaker", "InvocationsPerInstance", "EndpointName", "${EndpointC52XL}", "VariantName", "AllTraffic" ],
                            [ ".", "Invocations", ".", ".", ".", "." ],
                            [ ".", "Invocation5XXErrors", ".", ".", ".", ".", { "yAxis": "right" } ],
                            [ ".", "Invocation4XXErrors", ".", ".", ".", ".", { "yAxis": "right" } ],
                            [ ".", "ModelCacheHit", ".", ".", ".", "." ]
                        ],
                        "view": "timeSeries",
                        "stacked": false,
                        "region": "${AWS::Region}",
                        "stat": "Sum",
                        "period": 60,
                        "title": "c5.2xl-Invocations-modelcache",
                        "setPeriodToTimeRange": true,
                        "legend": {
                            "position": "bottom"
                        },
                        "yAxis": {
                            "right": {
                                "label": "Errors"
                            },
                            "left": {
                                "label": "Invocations"
                            }
                        }
                    }
                }
            ]
        }
        '