The following topics are discussed:

Any kind of metric can be generated and inserted in Scope with custom plugins. Metrics generated through a plugin are displayed in the user interface alongside the standard metrics found in Weave Scope.

Custom Metrics With Plugins

Official Plugins

Official Weave Scope plugins can be found at Weaveworks Plugins.

  • IOWait: is a Go plugin that uses iostat to provide host-level CPU IO wait or idle metrics.

  • HTTP Statistics: is a Python plugin that uses bcc to track multiple metrics about HTTP per process. It does this without any application-level instrumentation requirements and with a negligible performance toll. This plugin is a work in progress, and currently implements the following (for more information read the plugin documentation):

    • Number of HTTP requests per seconds.
    • Number of HTTP responses code per second (per code).

Note: The HTTP Statistics plugin requires a recent kernel version with ebpf support and it will not compile on dlite or on boot2docker hosts.

  • Traffic Control: This plugin allows you to modify latency and packet loss for a specific container via controls from the container’s detailed view in the Scope user interface.

  • Volume Count: This plugin (written in Python) requests the number of mounted volumes for each container, and provides a container-level count.

Note:Installed and running plugins are shown in the list of PLUGINS in the bottom right of the Scope UI.

How Plugins Communicate with Scope

In this section how the different components of a plugin communicate with Scope are described. You can also find practical examples of how plugins work in the Weaveworks Plugins repositories.

Plugin IDs

Each plugin must have a unique ID and this ID must not change during the plugin’s lifetime. Scope probes retrieve the plugin’s ID from the plugin’s socket filename. For example, if a socket is named my-plugin.sock, the scope probe deduces the ID as my-plugin. IDs may contain only alphanumeric sequences that are optionally separated by a dash.

Registering Plugins

All plugins listen for HTTP connections on a UNIX socket in the /var/run/scope/plugins directory. The Scope probe recursively scans that directory every 5 seconds and looks for any added or removed sockets.

If you want to run permissions or store any other information with the socket, you can also put the plugin UNIX socket into a sub-directory.

When a new plugin is detected, the Scope probe begins requesting reports from it via GET /report. It is therefore important that every plugin implements the report interface. Implementing the report interface also means handling specific requests.

All plugin endpoints are expected to respond within 500ms, and must respond using the JSON format.

Reporter Interface

When a Scope probe discovers a new plugin UNIX socket, it begins to periodically make a GET request to the /report endpoint. The report data structure returned from this is merged into the probe’s report and sent to the app. An example of the report structure can be viewed at the /api/report endpoint of any Scope app.

In addition to any data about the topology nodes, the report returned from the plugin must include some metadata about the plugin itself.

For example:

{
  ...,
  "Plugins": [
    {
      "id":          "plugin-id",
      "label":       "Human Friendly Name",
      "description": "Plugin's brief description",
      "interfaces":  ["reporter"],
      "api_version": "1",
    }
  ]
}

Note: The Plugins section includes exactly one plugin description that displays in the UI. The other plugin fields are:

  • id - checks for duplicate plugins. It is required.
  • label - a human readable label displayed in the UI. It is required.
  • description - displayed in the UI. It is required.
  • interfaces - a list of interfaces that the plugin supports. It is required, and must contain at least ["reporter"].
  • api_version - ensure both the plugin and the scope probe can speak to each other. It is required, and must match the probe’s value.

Controller Interface

Plugins may also implement the controller interface. Implementing the controller interface means that the plugin can react to HTTP POST control requests sent by the app. The plugin receives them only for the controls it exposed in its reports. All such requests come to the /control endpoint.

Add the “controller” string to the interfaces field in the plugin specification.

Control

The POST requests contain a JSON-encoded body with the following:

{
  "AppID": "some ID of an app",
  "NodeID": "an ID of the node that had the control activated",
  "Control": "the name of the activated control"
}

The body of the response should also be JSON-encoded data. In most cases, the body is an empty JSON object (so, “{}” after serialization). If an error occurs when handling the control, then the plugin sends a response with an error field set, for example:

{
  "error": "An error message here"
}

Sometimes the control activation can make the control obsolete, and so the plugin may want to hide it (for example, control for stopping the container should be hidden after the container is stopped). For this to work, the plugin sends a shortcut report by filling the ShortcutReport field in the response, like so:

{
  "ShortcutReport": { body of the report here }
}

How to Expose Controls

Each topology in the report (be it host, pod, endpoint and so on) contains a set of available controls that a node in the topology may want to show. The following (rather artificial) example shows a topology with two controls (ctrl-one and ctrl-two) and two nodes, each with a different control defined:

{
  "Host": {
    "controls": {
      "ctrl-one": {
        "id": "ctrl-one",
        "human": "Ctrl One",
        "icon": "far fa-futbol",
        "rank": 1
      },
      "ctrl-two": {
        "id": "ctrl-two",
        "human": "Ctrl Two",
        "icon": "fa fa-beer",
        "rank": 2
      }
    },
    "nodes": {
      "host1": {
        "latestControls": {
          "ctrl-one": {
            "timestamp": "2016-07-20T15:51:05Z01:00",
            "value": {
              "dead": false
            }
          }
        }
      },
      "host2": {
        "latestControls": {
          "ctrl-two": {
            "timestamp": "2016-07-20T15:51:05Z01:00",
            "value": {
              "dead": false
            }
          }
        }
      }
    }
  }
}

When control “ctrl-one” is activated, the plugin receives a request as follows:

{
  "AppID": "some ID of an app",
  "NodeID": "host1",
  "Control": "ctrl-one"
}

A short note about the “icon” field of the topology control - the value for it can be taken from Font Awesome Cheatsheet

Naming Nodes

Often the controller plugin may want to add controls to already existing nodes (for example add controls for network traffic management to nodes representing the running Docker container). To achieve that, it is important to make sure that the node ID in the plugin’s report matches the ID of the node created by the probe. The ID is a semicolon-separated list of strings.

For containers, images, hosts and others, the ID is usually formatted as ${name};<${tag}>. The ${name} variable is usually a name of a thing the node represents, like an ID of the Docker container or the hostname. The ${tag} denotes the type of the node.

There is a fixed set of tags used by the probe:

  • host
  • container
  • container_image
  • pod
  • service
  • deployment
  • replica_set

These are examples of “tagged” node names:

  • The Docker container with full ID 2299a2ca59dfd821f367e689d5869c4e568272c2305701761888e1d79d7a6f51: 2299a2ca59dfd821f367e689d5869c4e568272c2305701761888e1d79d7a6f51;<container>
  • The Docker image with name docker.io/alpine: docker.io/alpine;<container_image>
  • The host with name example.com: example.com;<host>

The fixed set of tags listed above is not a complete set of names a node can have though. For example, nodes representing processes have IDs formatted as ${host};${pid}. The easiest way to discover how the nodes are named are:

  1. Read the code in report/id.go.
  2. Browse the Weave Scope GUI, select some node and search for an id key in the nodeDetails array in the address bar.
    • For example in the http://localhost:4040/#!/state/{"controlPipe":null,"nodeDetails":[{"id":"example.com;<host>","label":"example.com","topologyId":"hosts"}],… URL, you can find the example.com;<host> which is an ID of the node representing the host.
  3. Mentally substitute the <SLASH> with /. This can appear in Docker image names, so docker.io/alpine in the address bar will be docker.io<SLASH>alpine.

A Guide to Developing Plugins

This section explains how to develop a simple plugin in Go. The code used here is a simplified version of the Scope IOWait plugin.

Setting up the Structure

As stated in the previous section, plugins need to be put into the /var/run/scope/plugins socket directory to be able to communicate with Scope. The best practice is to put the socket into a sub-directory and name it with the plugin ID (for example, /var/run/scope/plugins/plugins-id/plugins-id.sock). This is useful because the plugin can set more restrictive permissions to avoid unauthorized access as well as store other information along with the socket if needed.

Example of a helper function for setting up the socket:

func setupSocket(socketPath string) (net.Listener, error) {
	os.RemoveAll(filepath.Dir(socketPath))
	if err := os.MkdirAll(filepath.Dir(socketPath), 0700); err != nil {
		return nil, fmt.Errorf("failed to create directory %q: %v", filepath.Dir(socketPath), err)
	}
	listener, err := net.Listen("unix", socketPath)
	if err != nil {
		return nil, fmt.Errorf("failed to listen on %q: %v", socketPath, err)
	}

	log.Printf("Listening on: unix://%s", socketPath)
	return listener, nil
}

Because Scope detects running plugins by looking into the /var/run/scope/plugins directory, plugins should remove their socket and the directory (if created) when they exit. The side effect of not doing that is that the Scope UI will show that a plugin is running but that it is not reachable.

To remove the socket, and the directory, you can use the following helper function:

func setupSignals(socketPath string) {
	interrupt := make(chan os.Signal, 1)
	signal.Notify(interrupt, os.Interrupt, syscall.SIGTERM)
	go func() {
		<-interrupt
		os.RemoveAll(filepath.Dir(socketPath))
		os.Exit(0)
	}()
}

Also add the following to the main function the following:

defer func() {
		os.RemoveAll(filepath.Dir(socketPath))
	}()

This ensures that when the plugin terminates because of an error or an interrupt command, the /var/run/scope/plugins/plugins-id directory will be removed.

A bare minimum boilerplate can be the following:

package main

import (
	syscall
)

func main() {
	const socketPath = "/var/run/scope/plugins/my-plugin/my-plugin.sock"
	setupSignals(socketPath)

	listener, err := setupSocket(socketPath)

	plugin := &Plugin{}
	http.HandleFunc("/report", plugin.Report)

	defer func() {
		listener.Close()
		os.RemoveAll(filepath.Dir(socketPath))
	}()
}

Defining the Reporter Interface

As stated in the How Plugins Communicate with Scope section, the reporter interface is mandatory. Implementing the reporter interface means handling GET /report requests.

The following code snippet is sufficient to implement it:

// Plugin groups the methods a plugin needs
type Plugin struct {
	lock       sync.Mutex
}

type report struct {
	Plugins []pluginSpec
}

func (p *Plugin) makeReport() (*report, error) {
	rpt := &report{
		Plugins: []pluginSpec{
			{
				ID:          "plugin-id",
				Label:       "Plugin Name",
				Description: "Plugin short description",
				Interfaces:  []string{"reporter"},
				APIVersion:  "1",
			},
		},
	}
	return rpt, nil
}

// Report is called by scope when a new report is needed. It is part of the
// "reporter" interface, which all plugins must implement.
func (p *Plugin) Report(w http.ResponseWriter, r *http.Request) {
	p.lock.Lock()
	defer p.lock.Unlock()
	log.Println(r.URL.String())
	rpt, err := p.makeReport()
	if err != nil {
		log.Printf("error: %v", err)
		http.Error(w, err.Error(), http.StatusInternalServerError)
		return
	}
	raw, err := json.Marshal(*rpt)
	if err != nil {
		log.Printf("error: %v", err)
		http.Error(w, err.Error(), http.StatusInternalServerError)
		return
	}
	w.WriteHeader(http.StatusOK)
	w.Write(raw)
}

Report Data structures

A report can contain many types of information. If you go back to the Reporter Interface section, you will see the top-level Plugins attribute. Along with that, a report may contain multiple topologies. An example of a report containing a few topologies is the following:

{
  "Host": {},
  "Container": {},
  "Process": {},
  ...,
  "Plugins": [...,]
}

Topologies

A topology consists of a list of nodes along with controls and templates described below. These are the available topologies:

  • Endpoint nodes are (address, port) tuples on each host.
  • Process nodes are processes on each host.
  • Container nodes represent all Docker containers on hosts running probes.
  • Pods nodes represent all Kubernetes pods running on hosts running probes.
  • Service nodes represent all Kubernetes services running on hosts running probes.
  • Deployment nodes represent all Kubernetes deployments running on hosts running probes.
  • ReplicaSet nodes represent all Kubernetes ReplicaSets running on hosts running probes.
  • ContainerImage nodes represent all Docker container images on hosts running probes.
  • Host nodes are physical hosts that run probes.
  • ECSTask nodes represent AWS ECS tasks.
  • ECSService nodes represent AWS ECS services.
  • Overlay nodes are active peers in any software-defined network that’s overlaid on the infrastructure.

The topology structure consists of the following attributes:

  • nodes - is the list of the nodes that compose the topology.
  • controls - contains the list of IDs of the active controls at a particular time.
  • metadata_templates - contains the templates used to render data into the Scope UI.
  • table_templates - contains the templates used to render tables into the Scope UI.
  • metric_templates - contains the templates used to render metrics into the Scope UI.

Note: These attribute are not required. But a topology with no nodes does not have any information to render. metadata_templates, as well as table_templates, are needed to know how to render the information carried by nodes in the Scope UI.

Nodes

A Node contains information about a specific element of a topology. For example, the Host topology will contain nodes describing all the hosts in it. The same applies for containers and the Container topology, pods and the Pod topology and so on. Nodes are represented as follows:

{
  "Host": {
    "nodes" : {
      "hostID;<host>": {...}
    }
  },
  "Container": {
    "nodes" : {
      "containerID;<container>": {...}
    }
  },
  ...
}

Nodes are stored in a dictionary. The ID of nodes is different depending on the node topology. For instance, nodes representing hosts or containers have the format ID;<type>, where ID is the alphanumeric identifier of the nodes and type is the literal string host or container respectively. A node representing an endpoint could look like pc-4026531969;127.0.0.1;36238.

A node contains all the information about the represented object (e.g. host, container, pod, etc.). In particular, a node may contain:

  • latest - an id-value map containing the latest values. Each id has only one value.
  • latestControls - the latest available controls.
  • metrics - the collection of metrics to display in the UI. Each metric has multiple timestamped values.
  • sets - a string->set-of-strings map, for example a list of local networks.
  • counters - a string->int map.

Controls

Controls describe interfaces that expose actions that the user can perform on different objects (e.g. host, container, etc.). Controls are an element of nodes. In this way, each control in a node is attached to it and performs an action on the object described by the node itself. Below is an example of how controls are represented in the JSON report. In the report, the attribute latest_controls contains all the controls exposed by scope and/or plugins, but only those alive will be listed in the attribute controls.

"controls": {
	"timestamp": "2016-11-17T08:53:04.567890059Z",
	"controls": [
		"switchToIOWait"
	]
},
"latestControls": {
	"switchToIOWait": {
		"timestamp": "2016-11-17T08:53:03.171438309Z",
		"value": {
			"dead": false
		}
	},
	"switchToIdle": {
		"timestamp": "2016-11-17T08:53:03.171438309Z",
		"value": {
			"dead": true
		}
	}
}
  • timestamp specifies when the control was exposed.
  • value is an object containing the control value. At the moment, only the state is available.
  • dead is a boolean to know the state (active, dead) of a control. It is useful to show controls only when they are in a usable state.

Metadata

All metadata entries are placed within nodes in the section named latest. This section contains the latest values to display and consists of timestamp and value. Both should be written as JSON strings (with double quotes). Scope uses metadata_templates to display this data. To pair metadata with its template, it is necessary to use the metadata-template-id as a key to identify that particular piece of data. Example:

"metadata_templates": {
      "metadata-templates-id": {
        "id": "metadata-templates-id",
        "label": "Human-readable description",
        "priority": 1.6,
        "from": "latest"
      }
}

"latest": {
	"metadata-templates-id": {
		"timestamp": "2016-11-17T08:53:02.189193735Z",
		"value": "42"
	}
}

Metadata Templates

A metadata template describes a kind of metadata that is present in zero, one or several nodes of the topology and specifies how to display such metadata in Scope. Metadata templates are not placed within nodes but in the metadata_templates section of the JSON file.

"metadata_templates": {
      "traffic-control-pktloss": {
        "id": "traffic-control-pktloss",
        "label": "Human-readable description",
        "dataType": "number",
        "priority": 13.5,
        "from": "latest"
      },
      "another-plugins-id": {...}
}
  • id is a string identifying the particular metadata template (here traffic-control-pktloss) and is also used as a key to the template value.
  • label contains the label used by the Scope UI.
  • dataType specifies the type of data, and determines how the value is displayed. Possible values for this attribute are: “number”, “ip”, “datetime”, “link” and “” for strings.
  • priority is a floating point value used to decide the display ordering (lower values are displayed before higher ones). If omitted, the UI displays it last.
  • from indicates where to look for the metadata. The possible values are:
    • latest
    • sets
    • counters

Table Templates

Table Templates describe a table and also how to identify the metadata templates that belong to the table.

"table_templates": {
      "table-template-id": {
        "id": "table-template-id",
        "label": "Human-readable description",
        "prefix": "table-id-"
      },
      "another-table-template-id": {...}
}
  • table-template-identifier and id identify a particular table template.
  • label contains the label used by the Scope UI.
  • prefix is used to identify which metadata templates belong to the table.

To display data in a table, define a table template and prepend the table prefix to all of the metadata templates that identify the data you want to put into the table.

Multicolumn table

In order to display tabular data with multiple columns, additional fields are needed.

"table_templates": {
      "table-template-id": {
        "id": "table-template-id",
        "label": "Human-readable description",
        "prefix": "table-id-",
        "type": "multicolumn-table"
        "columns": [
          {
            "id": "table-column-id-1",
            "label": "Label 1",
            "dataType": ""
          },
          {...}
        ]
      },
      "another-table-template-id": {...}
}
  • type specifies the table type. Defaults to ‘property-list’ if absent.
  • columns defines the table’s columns
    • id the column id
    • label the column header
    • dataType: specifies the column type. Can be “” (for a string), “number”, “link”, or “ip”. Defaults to string if absent.

Values are placed in their respective columns and rows by specially formatting the entries in latest (prefix{unique row id}__column-id). _Note: that’s 3 underscores.

"latest": {
	"table-id-{unique-row-id}___table-column-id-1": {
		"timestamp": "2017-05-05T08:53:23.183293735Z",
		"value": "X"
	}...
}

Metrics

Metrics are a particular kind of data that can be plotted on the UI as a graph. Scope uses metric_templates to display graph data in Scope. To pair a metric with its template, use the metric-template-id as the key for identifying a particular metric. Metrics can be used to display CPU and memory usage, HTTP requests rate, I/O operations, etc. The following is an example of a report with a metric preceded by its metric template:

"metric_templates": {
      "metric-template-id": {
        "id": "metric-id",
        "label": "Human-readable description",
        "format": "percent",
        "priority": 1.6
      },
}

"metrics": {
	"metric-template-id": {
		"samples": [
			{
				"date": "2016-11-17T08:53:03.171424664Z",
				"value": 98.24
			},
			{
				"date": "2016-11-17T08:53:04.171789887Z",
				"value": 80.11
			}
		],
		"min": 0,
		"max": 100
	}
}
  • samples is the list of the samples for this report.
  • min is the minimum value possible.
  • max is the maximum value possible.

Metric Templates

Metric Templates describe a particular metric. The following is an example of metric template:

"metric_templates": {
      "metric-template-id": {
        "id": "metric-id",
        "label": "Human-readable description",
        "format": "percent",
        "priority": 1.6
      },
      "another-metric-template-id": {...}
}
  • metric-template-id and id identify a specific metric template.
  • label contains the label used by Scope UI.
  • format describes how the metrics are formatted and can be:
    • percent the metric value is a percentage.
    • filesize the metric value is a file size (e.g. memory usage), it is displayed with the suffix KB, MB, GB, etc.
    • integer the metric value is an integer.
  • priority is a floating point value used to decide the display ordering (lower values are displayed before higher ones).

Time Window

The Window attribute is a time window, expressed as a duration and it defines the period from which data in the report is considered valid. The default window is 15 seconds. You may change the window value using the option -app.window <SECONDS> when launching scope. However, using values smaller than 15 seconds increases the chance of information not being correctly displayed.

See Also