Cross posting from: https://forums.aws.amazon.com/thread.jspa?messageID=766424
Hey,
Trying to apply this policy to a core instance group:
{
"Constraints": {
"MinCapacity": 0,
"MaxCapacity": 2
},
"Rules": [
{
"Name": "ScaleUp",
"Action": {
"Market": "ON_DEMAND",
"SimpleScalingPolicyConfiguration": {
"AdjustmentType": "EXACT_CAPACITY",
"ScalingAdjustment": 5,
"CoolDown": 300
}
},
"Trigger": {
"CloudWatchAlarmDefinition": {
"ComparisonOperator": "GREATER_THAN",
"MetricName": "AppsPending",
"Threshold": 0,
"Period": 300
}
}
},
{
"Name": "ScaleDown",
"Action": {
"Market": "ON_DEMAND",
"SimpleScalingPolicyConfiguration": {
"AdjustmentType": "EXACT_CAPACITY",
"ScalingAdjustment": 0,
"CoolDown": 300
}
},
"Trigger": {
"CloudWatchAlarmDefinition": {
"ComparisonOperator": "LESS_THAN_OR_EQUAL",
"MetricName": "AppsRunning",
"Threshold": 0,
"Period": 300
}
}
}
]
}
But I'm getting this error:
An error occurred (ValidationException) when calling the
PutAutoScalingPolicy operation: Auto Scaling constraint parameter
minCapacity should be at least 1 for Core Instance Group.
I'm no expert in EMR but from the docs I thought this would be possible (I can create a master only cluster manually in the UI, why does this difference exist?). The master node is running a job on a cron schedule, when that kicks in it generates the job and then the AutoScaling fires up the core instances to process it, downscaling when the job is done.
Any suggestions?
Thanks, Alex
PS. To clarify the functional requirements, I'm trying to run a zeppelin dashboard service on master, have it kick off a batch job every 24h which will need a few nodes and then downscale back to 0 nodes the rest of the time. Happy to consider other suggestions to achieve this if I've got the wrong end of the stick.
It's true that you can start a single-node, master-only cluster without any core nodes, but this is a special kind of "cluster" that runs everything on the master. It is not possible to transition from a multi-node cluster to a single-node cluster or vice versa. Because of this, the core instance group has a minimum of 1 instance, even when using autoscaling.
Single node cluster is not scalable. You need to have at least one core nodes along with the master node. So while applying scaling policy minimum number of core nodes should be 1.
Please find the screenshot from AWS document:
Please refer to link for more details:
https://docs.aws.amazon.com/emr/latest/ManagementGuide/emr-scale-on-demand.html
Related
AWS ECS cluster services do not start new tasks.
Already checked:
ECS EC2 instances are registered, active, full CPU and memory available, ECS agent is connected.
there are no events in ECS service "Events" tab, nothing about registering, starting, stopping, no errors, it's just empty.
Registered EC2 instances are set up correctly, in other cluster the same AMI is working perfect.
Task definition is correct, it used to work a day before and since then no changes happened.
Checked Service role contains all relevant policies
Querying ECS with AWS CLI aws ecs describe-services --services my-service --cluster my-cluster yields that deployment rollout is constantly IN_PROGRESS and stays like this.
Full response with configuration is here (I've substituted real names and IDs):
{
"serviceArn": "arn:aws:ecs:eu-central-1:my-account-id:service/my-cluster/my-service",
"serviceName": "my-service",
"clusterArn": "arn:aws:ecs:eu-central-1:my-account-id:cluster/my-cluster",
"loadBalancers": [
{
"targetGroupArn": "arn:aws:elasticloadbalancing:eu-central-1:my-account-id:targetgroup/my-service-lb/load-balancer-id",
"containerName": "my-service",
"containerPort": 8065
}
],
"serviceRegistries": [
{
"registryArn": "arn:aws:servicediscovery:eu-central-1:my-account-id:service/srv-srv_id",
"containerName": "my-service",
"containerPort": 8065
}
],
"status": "ACTIVE",
"desiredCount": 1,
"runningCount": 0,
"pendingCount": 0,
"launchType": "EC2",
"taskDefinition": "arn:aws:ecs:eu-central-1:my-account-id:task-definition/my-service:76",
"deploymentConfiguration": {
"deploymentCircuitBreaker": {
"enable": false,
"rollback": false
},
"maximumPercent": 200,
"minimumHealthyPercent": 100
},
"deployments": [
{
"id": "ecs-svc/deployment_id",
"status": "PRIMARY",
"taskDefinition": "arn:aws:ecs:eu-central-1:my-account-id:task-definition/my-service:76",
"desiredCount": 1,
"pendingCount": 0,
"runningCount": 0,
"failedTasks": 0,
"createdAt": "2022-06-28T09:15:08.241000+02:00",
"updatedAt": "2022-06-28T09:15:08.241000+02:00",
"launchType": "EC2",
"rolloutState": "IN_PROGRESS",
"rolloutStateReason": "ECS deployment ecs-svc/deployment_id in progress."
}
],
"roleArn": "arn:aws:iam::my-account-id:role/aws-service-role/ecs.amazonaws.com/AWSServiceRoleForECS",
"events": [],
"createdAt": "2022-06-28T09:15:08.241000+02:00",
"placementConstraints": [],
"placementStrategy": [
{
"type": "spread",
"field": "attribute:ecs.availability-zone"
}
],
"healthCheckGracePeriodSeconds": 120,
"schedulingStrategy": "REPLICA",
"createdBy": "arn:aws:iam::my-account-id:role/my-role",
"enableECSManagedTags": false,
"propagateTags": "NONE",
"enableExecuteCommand": false
}
The ECS service and service discovery entry is created using Terraform, and the service definition is
resource "aws_service_discovery_service" "ecs_discovery_service" {
name = var.service_name
dns_config {
namespace_id = var.service_discovery_hosted_zone_id
dns_records {
ttl = 10
type = "SRV"
}
}
health_check_custom_config {
failure_threshold = 1
}
}
resource "aws_ecs_service" "ecs_service" {
name = var.service_name
cluster = var.ecs_cluster_id
task_definition = var.task_definition_arn
desired_count = var.desired_count
deployment_minimum_healthy_percent = 100
deployment_maximum_percent = 200
health_check_grace_period_seconds = var.health_check_grace_period_seconds
target_group_arn = aws_lb_target_group.target_group.arn
container_name = var.service_name
container_port = var.service_container_port
ordered_placement_strategy {
type = "spread"
field = "attribute:ecs.availability-zone"
}
service_registries {
registry_arn = aws_service_discovery_service.ecs_discovery_service.arn
container_name = var.service_name
container_port = var.service_container_port
}
}
This code used to work pretty fine, and without any changes in infrastructure, after destroying and applying the infrastructure code, ECS does not start any new tasks.
I could narrow problem to the service discovery, as if I remove the service_registries section, the tasks are started as normal.
Removing the service discovery solves the issue, however it's not the proper solution and I don't understand what is the reason of the problem.
Again, the Service Role has the permissions for the service discovery.
"servicediscovery:DeregisterInstance",
"servicediscovery:Get*",
"servicediscovery:List*",
"servicediscovery:RegisterInstance",
"servicediscovery:UpdateInstanceCustomHealthStatus"
I can't find any ways to trace this strange behaviour and want to ask you guys for help:
could you give me any hints what / where I could check. I've checked multiple troubleshooting guides, however all of them rely on events in ECS service and I don't have any there, anything else I had in mind is checked.
maybe you know what could be the problem that the service discovery blocks the ECS to start new tasks? I thought ECS adds a SRV record to the registry when it starts the container and the container is healthy, however I could not see that any containers have been started at all.
I would be very thankful for any hints and let me know if you need any details.
Have a nice day and best regards.
I would like to forward the logs of select services running on my EKS cluster to CloudWatch for cluster-independent storage and better observability.
Following the quickstart outlined at https://docs.aws.amazon.com/AmazonCloudWatch/latest/monitoring/Container-Insights-setup-EKS-quickstart.html I've managed to get the logs forwarded via Fluent Bit service, but that has also generated 170 Container Insights metrics channels. Not only are those metrics not required, but they also appear to cost a fair bit.
How can I disable the collection of cluster metrics such as cpu / memory / network / etc, and only keep forwarding container logs to CloudWatch? I'm having a very hard time finding any documentation on this.
I think I figured it out - the cloudwatch-agent daemonset from quickstart guide is what's sending the metrics, but it's not required for log forwarding. All the objects with names related to cloudwatch-agent in quickstart yaml file are not required for log forwarding.
As suggested by Toms Mikoss, you need to delete the metrics object in your configuration file. This file is the one that you pass to the agent when starting
This applies to "on-premises" "linux" installations. I havent tested this on windows, nor EC2 but I imagine it will be similar. The AWS Documentation here says that you can also distribute the configuration via SSM, but again, I imagine the answer here is still applicable.
Example of file with metrics:
{
"agent": {
"metrics_collection_interval": 60,
"run_as_user": "root"
},
"logs": {
"logs_collected": {
"files": {
"collect_list": [
{
"file_path": "/var/log/nginx.log",
"log_group_name": "nginx",
"log_stream_name": "{hostname}"
}
]
}
}
},
"metrics": {
"metrics_collected": {
"cpu": {
"measurement": [
"cpu_usage_idle",
"cpu_usage_iowait"
],
"metrics_collection_interval": 60,
"totalcpu": true
}
}
}
}
Example of file without metrics:
{
"agent": {
"metrics_collection_interval": 60,
"run_as_user": "root"
},
"logs": {
"logs_collected": {
"files": {
"collect_list": [
{
"file_path": "/var/log/nginx.log",
"log_group_name": "nginx",
"log_stream_name": "{hostname}"
}
]
}
}
}
}
For reference, the command to start for linux on-premises servers:
sudo /opt/aws/amazon-cloudwatch-agent/bin/amazon-cloudwatch-agent-ctl -a fetch-config \
-m onPremise -s -c file:configuration-file-path
More details in the AWS Documentation here
I’m trying to get an AWS Auto Scaling Group to replace ‘unhealthy’ instances, but I can’t get it to work.
From the console, I’ve created a Launch Configuration and, from there, an Auto Scaling Group with an Application Load Balancer. I've kept all settings regarding the target group and listeners the same as the default settings. I’ve selected ‘ELB’ as an additional health check type for the Auto Scaling Group. I’ve consciously misconfigured the Launch Configuration to result in ‘broken’ instances -- there is no web server to listen to the port configured in the listener.
The Auto Scaling Group seems to be configured correctly and is definitely aware of the load balancer. However, it thinks the instance it has spun up is healthy.
// output of aws autoscaling describe-auto-scaling-groups:
{
"AutoScalingGroups": [
{
"AutoScalingGroupName": "MyAutoScalingGroup",
"AutoScalingGroupARN": "arn:aws:autoscaling:eu-west-1:<accountId>:autoScalingGroup:3edc728f-0831-46b9-bbcc-16691adc8f44:autoScalingGroupName/MyAutoScalingGroup",
"LaunchConfigurationName": "MyLaunchConfiguration",
"MinSize": 1,
"MaxSize": 3,
"DesiredCapacity": 1,
"DefaultCooldown": 300,
"AvailabilityZones": [
"eu-west-1b",
"eu-west-1c",
"eu-west-1a"
],
"LoadBalancerNames": [],
"TargetGroupARNs": [
"arn:aws:elasticloadbalancing:eu-west-1:<accountId>:targetgroup/MyAutoScalingGroup-1/1e36c863abaeb6ff"
],
"HealthCheckType": "ELB",
"HealthCheckGracePeriod": 300,
"Instances": [
{
"InstanceId": "i-0b589d33100e4e515",
// ...
"LifecycleState": "InService",
"HealthStatus": "Healthy",
// ...
}
],
// ...
}
]
}
The load balancer, however, is very much aware that the instance is unhealthy:
// output of aws elbv2 describe-target-health:
{
"TargetHealthDescriptions": [
{
"Target": {
"Id": "i-0b589d33100e4e515",
"Port": 80
},
"HealthCheckPort": "80",
"TargetHealth": {
"State": "unhealthy",
"Reason": "Target.Timeout",
"Description": "Request timed out"
}
}
]
}
Did I just misunderstand the documentation? If not, what else is needed to be done to get the Auto Scaling Group to understand that this instance is not healthy and refresh it?
To be clear, when instances are marked unhealthy manually (i.e. using aws autoscaling set-instance-health), they are refreshed as is expected.
Explanation
If you have consciously misconfigured the instance from the start and the ELB Health Check has never passed, then the Auto Scaling Group does not acknowledge yet that your ELB/Target Group is up and running. See this page of the documentation.
After at least one registered instance passes the health checks, it enters the InService state.
And
If no registered instances pass the health checks (for example, due to a misconfigured health check), ... Amazon EC2 Auto Scaling doesn't terminate and replace the instances.
I configured from scratch and arrived at the same behavior as what you described. To verify that this is indeed the root cause, check the Target Group status in the ASG. It is probably in Added state instead of InService.
[cloudshell-user#ip-10-0-xx-xx ~]$ aws autoscaling describe-load-balancer-target-groups --auto-scaling-group-name test-asg
{
"LoadBalancerTargetGroups": [
{
"LoadBalancerTargetGroupARN": "arn:aws:elasticloadbalancing:us-east-1:xxx:targetgroup/asg-test-1/abc",
"State": "Added"
}
Resolution
To achieve the desired behavior, what I did was
Run a simple web service on port 80. Ensure Security Group is open for the ELB to talk to EC2.
Wait until the ELB status is healthy. Ensure server is returning 200. You may need to create an empty index.html just to pass the health check.
Wait until the target group status has become InService in the ASG.
For example, for Step 3:
[cloudshell-user#ip-10-0-xx-xx ~]$ aws autoscaling describe-load-balancer-target-groups --auto-scaling-group-name test-asg
{
"LoadBalancerTargetGroups": [
{
"LoadBalancerTargetGroupARN": "arn:aws:elasticloadbalancing:us-east-1:xxx:targetgroup/test-asg-1-alb/abcdef",
"State": "InService"
}
]
}
Now that it is in service, turn off the web server and wait. Check often, though, as once ASG detects it is unhealthy it will terminate.
[cloudshell-user#ip-10-0-xx-xx ~]$ aws autoscaling describe-auto-scaling-groups
{
"AutoScalingGroups": [
{
"AutoScalingGroupName": "test-asg",
"AutoScalingGroupARN": "arn:aws:autoscaling:us-east-1:xxx:autoScalingGroup:abc-def-ghi:autoScalingGroupName/test-asg",
...
"LoadBalancerNames": [],
"TargetGroupARNs": [
"arn:aws:elasticloadbalancing:us-east-1:xxx:targetgroup/test-asg-1-alb/abc"
],
"HealthCheckType": "ELB",
"HealthCheckGracePeriod": 300,
"Instances": [
{
"InstanceId": "i-04bed6ef3b2000326",
"InstanceType": "t2.micro",
"AvailabilityZone": "us-east-1b",
"LifecycleState": "Terminating",
"HealthStatus": "Unhealthy",
"LaunchTemplate": {
"LaunchTemplateId": "lt-0452c90319362cbc5",
"LaunchTemplateName": "test-template",
"Version": "1"
},
...
},
...
]
}
i have a SpringBoot application which is showing helath of all the servers in react charts today. we have some applications(servers) deployed to GCP using Kubernetes. i would like to pull and show health of the servers, number of pods, cpu utilization etc in my spring boot application. i have searched all GKE related REST apis in documentation, how ever i found REST urls at https://container.googleapis.com. but, none of them are seems to help me. please help me find the set of REST api's to fetch the above said heath statistics.
You can follow the documentation
You will find all info you need like cpu utilization and other useful metrics
The "metric type" strings in this table must be prefixed with actions.googleapis.com/
Metric type: instance/cpu/utilization:
Fractional utilization of allocated CPU on this instance. Values are typically numbers between 0.0 and 1.0 (but some machine types allow bursting above 1.0). Charts display the values as a percentage between 0% and 100% (or more). This metric is reported by the hypervisor for the VM and can differ from agent.googleapis.com/cpu/utilization, which is reported from inside the VM. Sampled every 60 seconds. After sampling, data is not visible for up to 240 seconds.
instance_name: The name of the VM instance
Creating the GET request
#Raj: This is not the url for the get request, check this tutorial, you want to format your get request the following way (change parameters depending on your own values):
curl -X GET -H "Authorization: Bearer $TOKEN"\
"https://monitoring.googleapis.com/v3/projects/{{YOUR_PROJECT}}/timeSeries/?filter=metric.type+%3D+%22compute.googleapis.com%2Finstance%2Fcpu%2Futilization%22&\
interval.endTime=2017-01-30T21%3A45%3A00.000000Z\
&interval.startTime=2017-01-30T21%3A43%3A00.000000Z"
{
"timeSeries": [
{
"metric": {
"labels": {
"instance_name": "evan-test"
},
"type": "compute.googleapis.com/instance/cpu/utilization"
},
"resource": {
"type": "gce_instance",
"labels": {
"instance_id": "743374153023006726",
"zone": "us-east1-d",
"project_id": "evan-testing"
}
},
"metricKind": "GAUGE",
"valueType": "DOUBLE",
"points": [
{
"interval": {
"startTime": "2017-01-30T21:44:01.763Z",
"endTime": "2017-01-30T21:44:01.763Z"
},
"value": {
"doubleValue": 0.00097060417263416339
}
},
{
"interval": {
"startTime": "2017-01-30T21:43:01.763Z",
"endTime": "2017-01-30T21:43:01.763Z"
},
"value": {
"doubleValue": 0.00085122420706227329
}
}
]
},
...
]
Here is my goal: create on-demand Hadoop clusters (number of nodes to specify on-the-fly) using EMR5.3.0 or EMR 5.4.0 with Spark 2.1.0 through AWS CLI while storing the input and output data in S3 without worrying about managing a 24*7 cluster or HDFS for data storage.
Here are my challenges / questions:
a) I can do the above using 'aws create-cluster' commands and specify number of nodes? Is that correct? For example, if I specify the parameter
--instance-count 10
this will create one master node and 9 core nodes?
b) If I use 'aws create-cluster', can I add more worker nodes (I guess it's called task nodes) on the fly to speed up the job, using CLI?
c) If I install Anaconda and other software on the cluster (i.e. Master) and then save the Master and all slave nodes as AMI, can I still launch on-demand Hadoop cluster from these AMIs with a different number of nodes which I can specify on-the-fly with AWS CLI?
Thank you. Appreciate your feedback.
Using Autoscaling on AWS EMR, you can scale out and scale in nodes on a cluster. Scale out action can be triggered using Cloudwatch metrics(YARNMemoryAvailablePercentage and ContainerPendingRatio). Sample Policy below
"AutoScalingPolicy":
{
"Constraints":
{
"MinCapacity": 10,
"MaxCapacity": 50
},
"Rules":
[
{"Name": "Compute-scale-up",
"Description": "Scale out based on ContainerPending Mterics",
"Action":
{
"SimpleScalingPolicyConfiguration":
{"AdjustmentType": "CHANGE_IN_CAPACITY",
"ScalingAdjustment": 1,
"CoolDown":0}
},
"Trigger":
{"CloudWatchAlarmDefinition":
{"AlarmNamePrefix": "compute-scale-up",
"ComparisonOperator": "GREATER_THAN_OR_EQUAL",
"EvaluationPeriods": 3,
"MetricName": "ContainerPending",
"Namespace": "AWS/ElasticMapReduce",
"Period": 300,
"Statistic": "AVERAGE",
"Threshold": 10,
"Unit": "COUNT",
"Dimensions":
[
{"Key": "JobFlowId",
"Value": "${emr:cluster_id}"}
]
}
}
},
{"Name": "Compute-scale-down",
"Description": "Scale in",
"Action":
{
"SimpleScalingPolicyConfiguration":
{"AdjustmentType": "CHANGE_IN_CAPACITY",
"ScalingAdjustment": -1,
"CoolDown":300}
},
"Trigger":
{"CloudWatchAlarmDefinition":
{"AlarmNamePrefix": "compute-scale-down",
"ComparisonOperator": "GREATER_THAN_OR_EQUAL",
"EvaluationPeriods": 3,
"MetricName": "MemoryAvailableMB",
"Namespace": "AWS/ElasticMapReduce",
"Period": 300,
"Statistic": "AVERAGE",
"Threshold": 24000,
"Unit": "COUNT",
"Dimensions":
[
{"Key": "JobFlowId",
"Value": "${emr:cluster_id}"}
]
}
}
}
]
}
You can refer this blog for more details
https://aws.amazon.com/blogs/big-data/dynamically-scale-applications-on-amazon-emr-with-auto-scaling/
a) I can do the above using 'aws create-cluster' commands and specify number of nodes? Is that correct? For example, if I specify the parameter...
Yes.
If I use 'aws create-cluster', can I add more worker nodes (I guess it's called task nodes) on the fly to speed up the job, using CLI?
Since your goal is to add on-demand instances on the fly, I would suggest you to look after reserved or spot instances (based on your use-case/cost).
We use spot instances with 50% of bid price and use the terminate policy as 'after the completion of job'.
If I install Anaconda and other software on the cluster (i.e. Master) and then save the Master and all slave nodes as AMI, can I still launch on-demand Hadoop cluster from these AMIs with a different number of nodes which I can specify on-the-fly with AWS CLI?
Yes you can .