Docker Monitoring Continued: Prometheus and Sysdig

I recently compared several docker monitoring tools and services. Since the article went live we have gotten feedback about additional tools that should be included in our survey. I would like to highlight two such tools; Prometheus and Sysdig cloud. Prometheus is a capable self-hosted solution which is easier to manage than sensu. Sysdig cloud on the other hand provides us with another hosted service much like Scout and Datadog. Collectively they help us add more choices to their respective classes. As before I will be using the following six criteria to evaluate Prometheus and Sysdig cloud: 1) ease of deployment, 2) level of detail of information presented, 3) level of aggregation of information from entire deployment, 4) ability to raise alerts from the data and 5) Ability to monitor non-docker resources 6) cost.


First lets take a look at Prometheus; it is a self-hosted set of tools which collectively provide metrics storage, aggregation, visualization and alerting. Most of the tools and services we have looked at so far have been push based, i.e. agents on the monitored servers talk to a central server (or set of servers) and send out their metrics. Prometheus on the other hand is a pull based server which expects monitored servers to provide a web interface from which it can scrape data. There are several exporters available for Prometheus which will capture metrics and then expose them over http for Prometheus to scrape. In addition there are libraries which can be used to create custom exporters. As we are concerned with monitoring docker containers we will use the container_exporter capture metrics. Use the command shown below to bring up the container-exporter docker container and browse to http://MONITORED_SERVER_IP:9104/metrics to see the metrics it has collected for you. You should launch exporters on all servers in your deployment. Keep track of the respective *MONITORED_SERVER_IP*s as we will be using them later in the configuration for Prometheus.

docker run -p 9104:9104 -v /sys/fs/cgroup:/cgroup -v /var/run/docker.sock:/var/run/docker.sock prom/container-exporter

Once we have got all our exporters running we are can launch Prometheus server. However, before we do we need to create a configuration file for Prometheus that tells the server where to scrape the metrics from. Create a file called prometheus.conf and then add the following text inside it.

  scrape_interval: 15s
  evaluation_interval: 15s
    monitor: exporter-metrics


- job_name: prometheus
  scrape_interval: 5s

    # These endpoints are scraped via HTTP.
    - targets: ['localhost:9090','MONITORED_SERVER_IP:9104']

In this file there are two sections, global and job(s). In the global section we set defaults for configuration properties such as data collection interval (scrape_interval). We can also add labels which will be appended to all metrics. In the jobs section we can define one or more jobs that each have a name, an optional override scraping interval as well as one or more targets from which to scrape metrics. We are adding two targets, one is the Prometheus server itself and the second is the container-exporter we setup earlier. If you setup more than one exporter your can setup additional targets to pull metrics from all of them. Note that the job name is available as a label on the metric hence you may want to setup separate jobs for your various types of servers. Now that we have a configuration file we can start a Prometheus server using the prom/prometheus docker image.

docker run -d --name prometheus-server -p 9090:9090 -v $PWD/prometheus.conf:/prometheus.conf prom/prometheus -config.file=/prometheus.conf

After launching the container, Prometheus server should be available in your browser on the port 9090 in a few moments. Select Graph from the top menu and select a metric from the drop down box to view its latest value. You can also write queries in the expression box which can find matching metrics. Queries take the form METRIC_NAME{LABEL_NAME=LABEL_VALUE, ...}. You can find more details of the query syntax here.

We are able to drill down into the data using queries to filter out data from specific server types (jobs) and containers. All metrics from containers are labeled with the image name, container name and the host on which the container is running. Since metric names do not encompass container or server name we are able to easily aggregate data across our deployment. For example we can filter for the container_memory_usage_bytes {image=\“ubuntu\“} to get information about the memory usage of all ubuntu containers in our deployment. Using the built in functions we can also aggregate the resulting set of of metrics. For example average_over_time(container_memory_usage_bytes {image=\“ubuntu\“}[5m]) will show the memory used by ubuntu containers, averaged over the last five minutes. Once you are happy with with a query you can click over to the Graph tab and see the variation of the metric over time.

Temporary graphs are great for ad-hoc investigations but you also need to have persistent graphs for dashboards. For this you can use the Prometheus Dashboard Builder. To launch Prometheus Dashboard Builder you need access to an SQL database which you can create using the official MySQL Docker image. The command to launch the MySQL container is shown below, note that you may select any value for database name, user name, user password and root password however keep track of these values as they will be needed later.

docker run -p 3306:3306 --name promdash-mysql      \
   -e MYSQL_DATABASE=<database-name>               \
   -e MYSQL_USER=<database-user>                   \
   -e MYSQL_PASSWORD=<user-password>               \
   -e MYSQL_ROOT_PASSWORD=<root-password>          \
   -d mysql

Once you have the database setup, use the rake installation inside the promdash container to initialize the database. You can then run the Dashboard builder by running the same container. The command to initialize the database and bring up the Prometheus Dashboard Builder are shown below.

# Initialize Database
docker run --rm -it --link promdash-mysql:db \
  -e DATABASE_URL=mysql2://<database-user>:<user-password>@db:3306/<database-name> prom/promdash ./bin/rake db:migrate

# Run Dashboard
docker run -d --link promdash-mysql:db -p 3000:3000 --name prometheus-dash \
 -e DATABASE_URL=mysql2://<database-user>:<user-password>@db:3306/<database-name> prom/promdash

Once your container is running you can browse to port 3000 and load up the dashboard builder UI. In the UI you need to click Servers in the top menu and New Server to add your Prometheus Server as a datasource for the dashboard builder. Add http://PROMETHEUS_SERVER_IP:9090 to the list of servers and hit Create Server.

Now click Dashboards in the top menu, here you can create Directories (Groups of Dashboards) and Dashboards. For example we created a directory for Web Nodes and one for Database Nodes and in each we create a dashboard as shown below.

Once you have created a dashboard you can add metrics by mousing over the title bar of a graph and selecting the data sources icon (Three Horizontal lines with an addition sign following them ). You can then select the server which you added earlier, and a query expression which you tested in the Prometheus Server UI. You can add multiple data sources into the same graph in order to see a comparative view.

You can add multiple graphs (each with possibly multiple data sources) by clicking the Add Graph button. In addition you may select the time range over which your dashboard displays data as well as a refresh interval for auto-loading data. The dashboard is not as polished as the ones from Scout and DataDog, for example there is no easy way to explore metrics or build a query in the dashboard view. Since the dashboard runs independently of the Prometheus server we can’t ‘pin’ graphs generated in the Prometheus server into a dashboard. Furthermore several times we noticed that the UI would not update based on selected data until we refreshed the page. However, despite its issues the dashboard is feature competitive with DataDog and because Prometheus is under heavy development, we expect the bugs to be resolved over time. In comparison to other self-hosted solutions Prometheus is a lot more user friendly than Sensu and allows you present metric data as graphs without using third party visualizations. It also is able to provide much better analytical capabilities than CAdvisor.

Prometheus also has the ability to apply alerting rules over the input data and displaying those on the UI. However, to be able to do something useful with alerts such send emails or notify pagerduty we need to run the the Alert Manager. To run the Alert Manager you first need to create a configuration file. Create a file called alertmanager.conf and add the following text into it:

notification_config {
    name: "ubuntu_notification"
    pagerduty_config {
        service_key: "<PAGER_DUTY_API_KEY>"
    email_config {
        email: "<TARGET_EMAIL_ADDRESS>"
    hipchat_config {
        auth_token: "<HIPCHAT_AUTH_TOKEN>"
        room_id: 123456
aggregation_rule {
    filter {
        name_re: "image"
        value_re: "ubuntu:14.04"
    repeat_rate_seconds: 300
    notification_config_name: "ubuntu_notification"

In this configuration we are creating a notification configuration called ubuntu_notification, which specifies that alerts must go to the PagerDuty, Email and HipChat. We need to specify the relevant API keys and/or access tokens for the HipChat and PagerDutyNotifications to work. We are also specifying that the alert configuration should only apply to alerts on metrics where the label image has the value ubuntu:14.04. We specify that a triggered alert should not retrigger for at least 300 seconds after the first alert is raised. We can bring up the Alert Manager using the docker image by volume mounting our configuration file into the container using the command shown below.

docker run -d -p 9093:9093 -v $PWD:/alertmanager prom/alertmanager -logtostderr -config.file=/alertmanager/alertmanager.conf

Once the container is running you should be able to point your browser to port 9093 and load up the Alarm Manger UI. You will be able to see all the alerts raised here, you can ‘silence’ them or delete them once the issue is resolved. In addition to setting up the Alert Manager we also need to create a few alerts. Add rule_file: \“/prometheus.rules\” in a new line into the global section of the prometheus.conf file you created earlier. This line tells Prometheus to look for alerting rules in the prometheus.rules file. We now need to create the rules file and load it into our server container. To do so create a file called prometheus.rules in the same directory where you created prometheus.conf. and add the following text to it:

ALERT HighMemoryAlert
  IF container_memory_usage_bytes{image="ubuntu:14.04"} > 1000000000
  FOR 1m
  WITH {}
  SUMMARY "High Memory usage for Ubuntu container"
  DESCRIPTION "High Memory usage for Ubuntu container on {{$labels.instance}} for container {{$}} (current value: {{$value}})"

In this configuration we are telling Prometheus to raise an alert called HighMemoryAlert if the container_memory_usage_bytes metric for containers using the Ubuntu:14.04 image goes above 1 GB for 1 minute. The summary and the description of the alerts is also specified in the rules file. Both of these fields can contain placeholders for label values which are replaced by Prometheus. For example our description will specify the server instance (IP) and the container name for metric raising the alert. After launching the Alert Manager and defining your Alert rules, you will need to re-run your Prometheus server with new parameters. The commands to do so are below:

# stop and remove current container
docker stop prometheus-server && docker rm prometheus-server

# start new container
docker run -d --name prometheus-server -p 9090:9090         \
  -v $PWD/prometheus.conf:/prometheus.conf                  \
  -v $PWD/prometheus.rules:/prometheus.rules                \
  prom/prometheus                                           \
  -config.file=/prometheus.conf                             \

Once the Prometheus Server is up again you can click Alerts in the top menu of the Prometheus Server UI to bring up a list of alerts and their statuses. If and when an alert is fired you will also be able to see it in the Alert Manager UI and any external service defined in the alertmanager.conf file.

Collectively the Prometheus tool-set’s feature set is on par with DataDog which has been our best rated Monitoring tool so far. Prometheus uses a very simple format for input data and can ingest from any web endpoint which presents the data. Therefore we can monitor more or less any resource with Prometheus, and there are already several libraries defined to monitor common resources. Where Prometheus is lacking is in level of polish and ease of deployment. The fact that all components are dockerized is a major plus however, we had to launch 4 different containers each with their own configuration files to support the Prometheus server. The project is also lacking detailed, comprehensive documentation for these various components. However, in caparison to self-hosted services such as CAdvisor and Sensu, Prometheus is a much better toolset. It is significantly easier setup than sensu and has the ability to provide visualization of metrics without third party tools. It is able has much more detailed metrics than CAdvisor and is also able to monitor non-docker resources. The choice of using pull based metric aggregation rather than push is less than ideal as you would have to restart your server when adding new data sources. This could get cumbersome in a dynamic environment such as cloud based deployments. Prometheus does offer the Push Gateway to bridge the disconnect. However, running yet another service will add to the complexity of the setup. For these reasons I still think DataDog is probably easier for most users, however, with some polish and better packaging Prometheus could be a very compelling alternative, and out of self-hosted solutions Prometheus is my pick.

Score Card:

  1. Easy of deployment: **
  2. Level of detail: *****
  3. Level of aggregation: *****
  4. Ability to raise alerts: ****
  5. Ability to monitor non-docker resources: Supported
  6. Cost: Free

Sysdig Cloud

Sysdig cloud is a hosted service that provides metrics storage, aggregation, visualization and alerting. To get started with sysdig sign up for a trial account at and complete the registration form. Once you complete the registration form and log in to the account, you will be asked to Setup your Environment and be given a curl command similar to the shown below. Your command will have your own secret key after the -s switch. You can run this command on the host running docker and which you need to monitor. Note that you should replace the [TAGS] place holder with tags to group your metrics. The tags are in the format TAG_NAME:VALUE so you may want to add a tag role:web or deployment:production. You may also use the containerized sysdig agent.

# Host install of sysdig agent
curl -s | sudo bash -s 12345678-1234-1234-1234-123456789abc [TAGS]

# Docker based sysdig agent
docker run --name sysdig-agent --privileged --net host \
  -e ACCESS_KEY=12345678-1234-1234-1234-123456789abc   \
  -e TAGS=os:rancher                                   \
  -v /var/run/docker.sock:/host/var/run/docker.sock    \
  -v /dev:/host/dev -v /proc:/host/proc:ro             \
  -v /boot:/host/boot:ro                               \
  -v /lib/modules:/host/lib/modules:ro                 \
  -v /usr:/host/usr:ro sysdig/agent

Even if you use docker you will still need to install Kernel headers in the host OS. This goes against Docker’s philosophy of isolated micro services. However, installing kernel headers is fairly benign. Installing the headers and getting sysdig running is trivial if you are using a mainstream kernel such us CentOS, Ubuntu or Debian. Even the Amazon’s custom kernels are supported however RancherOS’s custom kernel presented problems for sysdig as did the tinycore kernel. So be warned if you would like to use Sysdig cloud on non-mainstream kernels you may have to get your hands dirty with some system hacking.

After you run the agent you should see the Host in the Sysdig cloud console in the Explore tab. Once you launch docker containers on the host those will also be shown. You can see basic stats about the CPU usage, memory consumption, network usage. The metrics are aggregated for the host as well as broken down per container.

Screen Shot 2015-04-14 at 12.06.36
PMBy selecting one of the hosts or containers you can get a whole host of other metrics including everything provided by the docker stats API. Out of all the systems we have seen so far sysdig certainly has the most comprehensive set of metrics out of the box. You can also select from several pre-configured dashboards which present a graphical or tabular representation of your deployment.

Screen Shot 2015-04-16 at 11.26.53

You can see live metrics, by selecting Real-time Mode (Target Icon) or select a window of time over which to average values. Furthermore, you can also setup comparisons which will highlight the delta of current values and values at a point in the past. For example the table below shows values compared with those from ten minutes ago. If the CPU usage is significantly higher than 10 minutes ago you may be experiencing load spikes and need to scale out. The UI is at par with, if not better than DataDog for identifying and exploring trends in the data.Screen Shot
2015-04-19 at 4.59.09

In addition to exploring data on an ad-hoc basis you can also create persistent dashboards. Simply click the pin icon on any graph in the explore view and save it to a named dashboard. You can view all the dashboards and their associated graphs by clicking the Dashboards tab. You can also select the bell icon on any graph and create an alert from the data. The Sysdig cloud supports detailed alerting criteria and is again one of the best we have seen. The example below shows an alert which triggers if the count of containers labeled web falls below three on average for the last ten minutes. We are also segmenting the data by the region tag, so there will be a separate check for web nodes in North America and Europe. Lastly, we also specify a Name, description and Severity for the alerts. You can control where alerts go by going to Settings (Gear Icon) > Notifications and add email addresses or SNS Topics to send alerts too. Note all alerts go to all notification endpoints which may be problematic if you want to wake up different people for different alerts.Screen Shot 2015-04-19 at

I am very impressed with Sysdig cloud as it was trivially easy to setup, provides detailed metrics with great visualization tools for real-time and historical data. The requirement to install kernel headers on the host OS is troublesome though and lack of documentation and support for non-standard kernels could be problematic in some scenarios. The alerting system in the Sysdig cloud is among the best we have seen so far, however, the inability to target different email addresses for different alerts is problematic. In a larger team for example you would want to alert a different team for database issues vs web server issues. Lastly, since it is in beta the pricing for Sysdig cloud is not easily available. I have reached out to their sales team and will update this article if and when they get back to me. If sysdig is price competitive then Datadog has serious competition in the hosted service category.

Score Card:

  1. Easy of deployment: ***
  2. Level of detail: *****
  3. Level of aggregation: *****
  4. Ability to raise alerts: ****
  5. Ability to monitor non-docker resources: Supported
  6. Cost: Must Contact Support

To learn more about monitoring and managing Docker, please join us for our next Rancher online meetup. Usman is a server and infrastructure engineer, with experience in building large scale distributed services on top of various cloud platforms. You can read more of his work at, or follow him on twitter @usman_ismailor onGitHub. Update: The article has been updated for new prometheus.conf format. Thanks too Sebastian Gerard from @ruxit for help with the update. 12/08/2015 --Usman Ismail