Welcome to Tensor’s documentation!¶

Tensor is a modular gateway and event router for Riemann, built using the Twisted framework.

Contents:

Getting started¶

Installation¶

Tensor can be installed from PyPi with pip

$ pip install tensor

This will also install Twisted, protobuf and PyYAML

Or you can use the .deb package. Let the latest release from https://github.com/calston/tensor/releases/latest

$ aptitude install python-twisted python-protobuf python-yaml
$ wget https://github.com/calston/tensor/releases/download/0.2.0/tensor_0.2.0_amd64.deb
$ dpkg -i tensor_0.2.0_amd64.deb

This also gives you an init script and default config in /etc/tensor/

Creating a configuration file¶

Tensor uses a simple YAML configuration file

The first basic requirements are the Riemann server and port (defaults to localhost:5555) and the queue interval:

server: localhost
port: 5555
interval: 1.0
proto: udp

Tensors checks are Python classes (called sources) which are instantiated with the configuration block which defines them. Rather than being one-shot scripts, a source object remains in memory with its own timer which adds events to a queue. The interval defined above is the rate at which these events are rolled up into a message and sent to Riemann.

It is important then that interval is set to a value appropriate to how frequently you want to see them in Riemann, as well as the rate at which they collect metrics from the system. All interval attributes are floating point in seconds, this means you can check (and send to Riemann) at rates well below 1 second.

You can configure multiple outputs which receive a copy of every message for example

outputs:
    - output: tensor.outputs.riemann.RiemannUDP
      server: localhost
      port: 5555

If you enable multiple outputs then the server, port and proto options will go un-used and the default Riemann TCP transport won’t start.

You can configure as many outputs as you like, or create your own.

To configure the basic CPU usage source add it to the sources list in the config file

sources:
    - service: cpu
      source: tensor.sources.linux.basic.CPU
      interval: 1.0
      warning: {
        cpu: "> 0.5"
      }
      critical: {
        cpu: "> 0.8"
      }

This will measure the CPU from /proc/stat every second, with a warning state if the value exceeds 50%, and a critical state if it exceeds 80%

The service attribute can be any string you like, populating the service field in Riemann. The logical expression to raise the state of the event is (eg. critical) is assigned to a key which matches the service name.

Sources may return more than one metric, in which case it will add a prefix to the service. The state expression must correspond to that as well.

For example, the Ping check returns both latency and packet loss:

service: googledns
source: tensor.sources.network.Ping
interval: 60.0
destination: 8.8.8.8
critical: {
    googledns.latency: "> 100",
    googledns.loss: "> 0"
}

This will ping 8.8.8.8 every 60 seconds and raise a critical alert for the latency metric if it exceeds 100ms, and the packet loss metric if there is any at all.

critical and warning matches can also be a regular expression for sources which output keys for different devices and metrics:

service: network
source: tensor.sources.linux.basic.Network
...
critical: {
    network.\w+.tx_packets: "> 1000",
}

Starting Tensor¶

To start Tensor, simply use twistd to run the service and pass a config file:

twistd -n tensor -c tensor.yml

Sources¶

Introduction¶

Sources are Python objects which subclass tensor.objects.Source. They are constructed with a dictionary parsed from the YAML configuration block which defines them, and as such can read any attributes from that either optional or mandatory.

Since sources are constructed at startup time they can retain any required state, for example the last metric value to report rates of change or for any other purpose. However since a Tensor process might be running many checks a source should not use an excessive amount of memory.

The source configuration option is passed a string representing an object in much the same way as you would import it in a python module. The final class name is split from this string. For example specifying:

source: tensor.sources.network.Ping

is equivalent to:

from tensor.sources.network import Ping

Writing your own sources¶

A source class must subclass tensor.objects.Source and also implement the interface tensor.interfaces.ITensorSource

The source must have a get method which returns a tensor.objects.Event object. The Source parent class provides a helper method createEvent which performs the metric level checking (evaluating the simple logical statement in the configuration), sets the correct service name and handles prefixing service names.

A “Hello world” source:

from zope.interface import implements

from tensor.interfaces import ITensorSource
from tensor.objects import Source

class HelloWorld(Source):
    implements(ITensorSource)

    def get(self):
        return self.createEvent('ok', 'Hello world!', 0)

To hold some state, you can re-implement the __init__ method, as long as the arguments remain the same.

Extending the above example to create a simple flip-flop metric event:

from zope.interface import implements

from tensor.interfaces import ITensorSource
from tensor.objects import Source

class HelloWorld(Source):
    implements(ITensorSource)

    def __init__(self, *a):
        Source.__init__(self, *a)
        self.bit = False

    def get(self):
        self.bit = not self.bit
        return self.createEvent('ok', 'Hello world!', self.bit and 0.0 or 1.0)

You could then place this in a Python module like hello.py and as long as it’s in the Python path for Tensor it can be used as a source with hello.HelloWorld

A list of events can also be returned but be careful of overwhelming the output buffer, and if you need to produce lots of metrics it may be worthwhile to return nothing from get and call self.queueBack as needed.

Handling asynchronous tasks¶

Since Tensor is written using the Twisted asynchronous framework, sources can (and in most cases must) make full use of it to implement network checks, or execute other processes.

The simplest example of a source which executes an external process is the ProcessCount check:

from zope.interface import implements

from twisted.internet import defer

from tensor.interfaces import ITensorSource
from tensor.objects import Source
from tensor.utils import fork

class ProcessCount(Source):
    implements(ITensorSource)

    @defer.inlineCallbacks
    def get(self):
        out, err, code = yield fork('/bin/ps', args=('-e',))

        count = len(out.strip('\n').split('\n'))

        defer.returnValue(
            self.createEvent('ok', 'Process count %s' % (count), count)
        )

For more information please read the Twisted documentation at https://twistedmatrix.com/trac/wiki/Documentation

The tensor.utils.fork() method returns a deferred which can timeout after a specified time.

Thinking outside the box¶

Historically monitoring systems are poorly architected, and terribly inflexible. To demonstrate how Tensor offers a different concept to the boring status quo it’s interesting to note that there is nothing preventing you from starting a listening service directly within a source which processes and relays events to Riemann implementing some protocol.

Here is an example of a source which listens for TCP connections to port 8000, accepting any number on a line and passing that to the event queue:

from twisted.internet.protocol import Factory
from twisted.protocols.basic import LineReceiver
from twisted.internet import reactor

from zope.interface import implements

from tensor.interfaces import ITensorSource
from tensor.objects import Source

class Numbers(LineReceiver):
    def __init__(self, source):
        self.source = source

    def lineReceived(self, line):
        """
        Send any numbers received back to the Tensor queue
        """
        print repr(line)
        try:
            num = float(line)
            self.source.queueBack(
                self.source.createEvent('ok', 'Number: %s' % num, num)
            )
        except:
            pass

class NumbersFactory(Factory):
    def __init__(self, source):
        self.source = source

    def buildProtocol(self, addr):
        return Numbers(self.source)

class NumberProxy(Source):
    implements(ITensorSource)

    def startTimer(self):
        # Override starting the source timer, we don't need it
        f = NumbersFactory(self)
        reactor.listenTCP(8000, f)

    def get(self):
        # Implement the get method, but we can ignore it
        pass

Outputs¶

Introduction¶

Outputs are Python objects which subclass tensor.objects.Output. They are constructed with a dictionary parsed from the YAML configuration block which defines them, and as such can read any attributes from that either optional or mandatory.

Since outputs are constructed at startup time they can retain any required state. A copy of the queue is passed to all :method:`tensor.objects.Output.eventsReceived` calls which happen at each queue interval config setting as the queue is emptied. This list of tensor.objects.Event objects must not be altered by the output.

The output configuration option is passed a string representing an object the same way as sources configurations are

output: tensor.sources.network.Ping

Writing your own outputs¶

An output clas should subclass tensor.objects.Output.

The output can implement a createClient method which starts the output in whatever way necessary and can be a deferred. The output must also have a eventsReceived method which takes a list of tensor.objects.Event objects and process them accordingly, it can also be a deferred.

An example logging source:

from twisted.internet import reactor, defer
from twisted.python import log

from tensor.objects import Output

class Logger(Output):
    def eventsReceived(self, events):
        log.msg("Events dequeued: %s" % len(events))

If you save this as test.py the basic configuration you need is simply

outputs:
    - output: tensor.outputs.riemann.RiemannUDP
      server: localhost
      port: 5555

    - output: test.Logger

You should now see how many events are exiting in the Tensor log file

2014-10-24 15:35:27+0200 [-] Starting protocol <tensor.protocol.riemann.RiemannUDP object at 0x7f3b5ae15810>
2014-10-24 15:35:28+0200 [-] Events dequeued: 7
2014-10-24 15:35:29+0200 [-] Events dequeued: 2
2014-10-24 15:35:30+0200 [-] Events dequeued: 3

Example configurations¶

Replacing Munin¶

The first step is to create a TRIG stack (Tensor Riemann InfluxDB Grafana).

Step 1: Install Riemann¶

$ wget http://aphyr.com/riemann/riemann_0.2.6_all.deb
$ aptitude install openjdk-7-jre
$ dpkg -i riemann_0.2.6_all.deb

Step 2: Install InfluxDB¶

$ wget http://s3.amazonaws.com/influxdb/influxdb_latest_amd64.deb
$ sudo dpkg -i influxdb_latest_amd64.deb

Start InfluxDB, then quickly change the root/root default password because it also defaults to listening on all interfaces and apparently this is not important enough for them to fix.

Create a riemann and grafana database, and some users for them

$ curl -X POST 'http://localhost:8086/db?u=root&p=root' \
  -d '{"name": "riemann"}'
$ curl -X POST 'http://localhost:8086/db?u=root&p=root' \
  -d '{"name": "grafana"}'
$ curl -X POST 'http://localhost:8086/db/riemann/users?u=root&p=root' \
  -d '{"name": "riemann", "password": "riemann"}'
$ curl -X POST 'http://localhost:8086/db/grafana/users?u=root&p=root' \
  -d '{"name": "grafana", "password": "grafana"}'

Step 3: Install Grafana¶

$ aptitude install nginx
$ mkdir /var/www
$ cd /var/www
$ wget http://grafanarel.s3.amazonaws.com/grafana-1.8.1.tar.gz
$ tar -zxf grafana-1.8.1.tar.gz
$ mv grafana-1.8.1 grafana

Now we must create an nginx configuration in /etc/nginx/sites-enabled.

You can use something like this

server {
    listen 80;
    server_name <your hostname>;
    access_log /var/log/nginx/grafana-access.log;
    error_log /var/log/nginx/grafana-error.log;

    location / {
        alias /var/www/grafana/;
        index index.html;
        try_files $uri $uri/ /index.html;
    }
}

Next we need a configuration file for grafana. Open /var/www/grafana/config.js and use the following configuration

define(['settings'],
function (Settings) {
  return new Settings({
    datasources: {
      influxdb: {
        type: 'influxdb',
        url: "http://<your hostname>:8086/db/riemann",
        username: 'riemann',
        password: 'riemann',
      },
      grafana: {
        type: 'influxdb',
        url: "http://<your hostname>:8086/db/grafana",
        username: 'grafana',
        password: 'grafana',
        grafanaDB: true
      },
    },
    search: {
      max_results: 20
    },
    default_route: '/dashboard/file/default.json',
    unsaved_changes_warning: true,
    playlist_timespan: "1m",
    admin: {
      password: ''
    },
    window_title_prefix: 'Grafana - ',
    plugins: {
      panels: [],
      dependencies: [],
    }
  });
});

Step 4: Glue things together¶

Lets start by configuring Riemann to talk to InfluxDB. This is the full /etc/riemann/riemann.config file.

; -*- mode: clojure; -*-
; vim: filetype=clojure
(require 'capacitor.core)
(require 'capacitor.async)
(require 'clojure.core.async)

(defn make-async-influxdb-client [opts]
    (let [client (capacitor.core/make-client opts)
          events-in (capacitor.async/make-chan)
          resp-out (capacitor.async/make-chan)]
        (capacitor.async/run! events-in resp-out client 100 10000)
        (fn [series payload]
            (let [p (merge payload {
                    :series series
                    :time   (* 1000 (:time payload)) ;; s → ms
                })]
                (clojure.core.async/put! events-in p)))))

(def influx (make-async-influxdb-client {
        :host     "localhost"
        :port     8086
        :username "riemann"
        :password "riemann"
        :db       "riemann"
    }))

(logging/init {:file "/var/log/riemann/riemann.log"})

; Listen on the local interface over TCP (5555), UDP (5555), and websockets
; (5556)
(let [host "0.0.0.0"]
  (tcp-server {:host host})
  (udp-server {:host host})
  (ws-server  {:host host}))

(periodically-expire 60)

(let [index (index)]
  (streams
        index

        (fn [event]
            (let [series (format "%s.%s" (:host event) (:service event))]
                (influx series {
                    :time  (:time event)
                    :value (:metric event)
                })))))

You’re pretty much done at this point, and should see the metrics from the Riemann server process if you open up Grafana and look through the query builder.

Step 5: Using Tensor to retrieve stats from munin-node¶

First of all, install Tensor

$ pip install tensor

Next create /etc/tensor and a tensor.yml file in that directory.

The tensor.yml config file should look like this

ttl: 60.0
interval: 1.0

outputs:
    - output: tensor.outputs.riemann.RiemannTCP
      port: 5555
      server: <riemann server>

# Sources
sources:
    - service: mymunin
      source: tensor.sources.munin.MuninNode
      interval: 60.0
      ttl: 120.0
      critical: {
        mymunin.system.load.load: "> 2"
      }

This configures Tensor to connect to the munin-node on the local machine and retrieve all configured plugin values. You can create critical alert levels by setting the dot separated prefix for the service name and munin plugin.

You can now start Tensor

$ twistd -n tensor -c /etc/tensor/tensor.yml
2014-10-22 13:30:38+0200 [-] Log opened.
2014-10-22 13:30:38+0200 [-] twistd 14.0.2 (/home/colin/riemann-tensor/ve/bin/python 2.7.6) starting up.
2014-10-22 13:30:38+0200 [-] reactor class: twisted.internet.epollreactor.EPollReactor.
2014-10-22 13:30:38+0200 [-] Starting factory <tensor.protocol.riemann.RiemannClientFactory instance at 0x7faeec021b90>

This pretty much indiciates everything is alright, or else we’d see quickly see some errors.

Next we will add some graphs to Grafana

Step 6: Creating graphs in Grafana¶

Click on the green row tag on the left, and delete all but the last row. This will leave you with an empty graph.

Click the title of the graph, then click Edit.

In the edit screen the Metrics tab will be open already. Now we can add our munin metrics. If you start typing in the series field you should see your hosts and metrics autocomplete.

Many Munin metrics are counter types which are usually converted to a rate by the RRD aggregation on Munin Graph.

Handily the tensor.sources.munin.MuninNode source takes care of this by caching the metric between run intervals when that type is used.

If we wanted to graph our network interface all we need to do is make it a slightly better unit by multiplying the Byte/sec metric by 8, since Grafana provides a bit/sec legend format.

To do this start by clicking the gear icon on the metric query, then select Raw query mode.

Use the following query

select value * 8 from "<your hostname>.munin.network.if_eth0.down" where $timeFilter group by time($interval) order asc

And chose an alias of “RX”. Do the same for if_eth0.up and alias that “TX”. You should end up with something like this

Click on General to edit the title, and then on Axes & Grid change the Format to bps. Under Display Styles you can stack the data or play around with the look of the graph. Click Back to dashboard and you should end up with something as follows

API Documentation:

tensor¶

tensor.interfaces¶

tensor.objects¶

class tensor.objects.Event(state, service, description, metric, ttl, tags=, []hostname=None, aggregation=None, evtime=None)[source]¶

Bases: object

Tensor Event object

All sources pass these to the queue, which form a proxy object to create protobuf Event objects

Parameters:

state – Some sort of string < 255 chars describing the state
service – The service name for this event
description – A description for the event, ie. “My house is on fire!”
metric – int or float metric for this event
tags – List of tag strings
hostname – Hostname for the event (defaults to system fqdn)
aggregation – Aggregation function
evtime – Event timestamp override

class tensor.objects.Output(config, tensor)[source]¶

Bases: object

Output parent class

Outputs can inherit this object which provides a construct for a working output

Parameters:	config – Dictionary config for this queue (usually read from the yaml configuration) tensor – A TensorService object for interacting with the queue manager

createClient()[source]¶: Deferred which sets up the output

eventsReceived()[source]¶

Receives a list of events and processes them

Arguments: events – list of tensor.objects.Event

stop()[source]¶: Called when the service shuts down

class tensor.objects.Source(config, queueBack, tensor)[source]¶

Bases: object

Source parent class

Sources can inherit this object which provides a number of utility methods.

Parameters:	config – Dictionary config for this queue (usually read from the yaml configuration) queueBack – A callback method to recieve a list of Event objects tensor – A TensorService object for interacting with the queue manager

createEvent(state, description, metric, prefix=None, hostname=None, aggregation=None, evtime=None)[source]¶: Creates an Event object from the Source configuration

startTimer()[source]¶: Starts the timer for this source

tick(*args, **kwargs)[source]¶

Called for every timer tick. Calls self.get which can be a deferred and passes that result back to the queueBack method

Returns a deferred

tensor.service¶

class tensor.service.TensorService(config)[source]¶

Bases: twisted.application.service.Service

Tensor service

Runs timers, configures sources and and manages the queue

sendEvent(events)[source]¶: Callback that all event sources call when they have a new event or list of events

setupOutputs(config)[source]¶: Setup output processors

setupSources(config)[source]¶: Sets up source objects from the given config

tensor.utils¶

class tensor.utils.BodyReceiver(finished)[source]¶

Bases: twisted.internet.protocol.Protocol

Simple buffering consumer for body objects

class tensor.utils.ProcessProtocol(deferred, timeout)[source]¶

Bases: twisted.internet.protocol.ProcessProtocol

ProcessProtocol which supports timeouts

class tensor.utils.Resolver[source]¶

Bases: object

Helper class for DNS resolution

tensor.utils.fork(executable, args=(), env={}, path=None, timeout=3600)[source]¶

Provides a deferred wrapper function with a timeout function

Parameters:	executable (str.) – Executable args (tupple.) – Tupple of arguments env (dict.) – Environment dictionary timeout (int.) – Kill the child process if timeout is exceeded

tensor.protocol¶

tensor.protocol.riemann¶

class tensor.protocol.riemann.RiemannClientFactory[source]¶

Bases: twisted.internet.protocol.ReconnectingClientFactory

A reconnecting client factory which creates RiemannProtocol instances

class tensor.protocol.riemann.RiemannProtocol[source]¶

Bases: twisted.protocols.basic.Int32StringReceiver, tensor.protocol.riemann.RiemannProtobufMixin

Riemann protobuf protocol

class tensor.protocol.riemann.RiemannUDP(host, port)[source]¶

Bases: twisted.internet.protocol.DatagramProtocol, tensor.protocol.riemann.RiemannProtobufMixin

UDP datagram protocol for Riemann

tensor.protocol.sflow¶

tensor.protocol.sflow.server¶

class tensor.protocol.sflow.server.DatagramReceiver[source]¶

Bases: twisted.internet.protocol.DatagramProtocol

DatagramReceiver for sFlow packets

tensor.protocol.sflow.protocol¶

tensor.sources¶

tensor.sources.network¶

class tensor.sources.network.HTTP(config, queueBack, tensor)[source]¶

(service name).latency:
	Ping latency
(service name).loss:
	Packet loss

(service name).active:
	Active connections at this time
(service name).accepts:
	Accepted connections
(service name).handled:
	Handled connections
(service name).requests:
	Total client requests
(service name).reading:
	Reading requests
(service name).writing:
	Writing responses
(service name).waiting:
	Waiting connections

(service name).total_bytes:
	Bytes total for all requests
(service name).total_requests:
	Total request count
(service name).stats.(code).(requests\|bytes):
	Metrics by status code
(service name).user-agent.(agent).(requests\|bytes):
	Metrics by user agent
(service name).client.(ip).(requests\|bytes):
	Metrics by client IP
(service name).request.(request path).(requests\|bytes):
	Metrics by request path

(service_name).(queue).ready:
	Ready messages for queue
(service_name).(queue).unack:
	Unacknowledged messages for queue
(service_name).(queue).ready_rate:
	Ready rate of change per second
(service_name).(queue).unack_rate:
	Unacknowledge rate of change per second

Welcome to Tensor’s documentation!¶

Getting started¶

Installation¶

Creating a configuration file¶

Starting Tensor¶

Sources¶

Introduction¶

Writing your own sources¶

Handling asynchronous tasks¶

Thinking outside the box¶

Outputs¶

Introduction¶

Writing your own outputs¶

Example configurations¶

Replacing Munin¶

Step 1: Install Riemann¶

Step 2: Install InfluxDB¶

Step 3: Install Grafana¶

Step 4: Glue things together¶

Step 5: Using Tensor to retrieve stats from munin-node¶

Step 6: Creating graphs in Grafana¶

tensor¶

tensor.interfaces¶

tensor.objects¶

tensor.service¶

tensor.utils¶

tensor.protocol¶

tensor.protocol.riemann¶

tensor.protocol.sflow¶

tensor.protocol.sflow.server¶

tensor.protocol.sflow.protocol¶

tensor.sources¶

tensor.sources.network¶

tensor.sources.nginx¶

tensor.sources.munin¶

tensor.sources.riak¶

tensor.sources.rabbitmq¶

tensor.sources.linux¶

tensor.sources.linux.basic¶

tensor.sources.linux.process¶

tensor.sources.media¶

tensor.sources.media.libav¶

tensor.sources.sflow¶

tensor.sources.snmp¶

tensor.outputs¶

tensor.outputs.riemann¶

Indices and tables¶