Friday, November 20, 2015

Initial experiences with the Prometheus monitoring system

I've been looking for a while for a monitoring system written in Go, self-contained and easy to deploy. I think I finally found what I was looking for in Prometheus, a monitoring system open-sourced by SoundCloud and started there by ex-Googlers who took their inspiration from Google's Borgmon system.

Prometheus is a pull system, where the monitoring server pulls data from its clients by hitting a special HTTP handler exposed by each client ("/metrics" by default) and retrieving a list of metrics from that handler. The output of /metrics is plain text, which makes it fairly easily parseable by humans as well, and also helps in troubleshooting.

Here's a subset of the OS-level metrics that are exposed by a client running the node_exporter Prometheus binary (and available when you hit http://client_ip_or_name:9100/metrics):

# HELP node_cpu Seconds the cpus spent in each mode.
# TYPE node_cpu counter
node_cpu{cpu="cpu0",mode="guest"} 0
node_cpu{cpu="cpu0",mode="idle"} 2803.93
node_cpu{cpu="cpu0",mode="iowait"} 31.38
node_cpu{cpu="cpu0",mode="irq"} 0
node_cpu{cpu="cpu0",mode="nice"} 2.26
node_cpu{cpu="cpu0",mode="softirq"} 0.23
node_cpu{cpu="cpu0",mode="steal"} 21.16
node_cpu{cpu="cpu0",mode="system"} 25.84
node_cpu{cpu="cpu0",mode="user"} 79.94
# HELP node_disk_io_now The number of I/Os currently in progress.
# TYPE node_disk_io_now gauge
node_disk_io_now{device="xvda"} 0
# HELP node_disk_io_time_ms Milliseconds spent doing I/Os.
# TYPE node_disk_io_time_ms counter
node_disk_io_time_ms{device="xvda"} 44608
# HELP node_disk_io_time_weighted The weighted # of milliseconds spent doing I/Os. See
# TYPE node_disk_io_time_weighted counter
node_disk_io_time_weighted{device="xvda"} 959264

There are many such "exporters" available for Prometheus, exposing metrics in the format expected by the Prometheus server from systems such as Apache, MySQL, PostgreSQL, HAProxy and many others (see a list here).

What drew me to Prometheus though was the fact that it allows for easy instrumentation of code by providing client libraries for many languages: Go, Java/Scala, Python, Ruby and others. 

One of the main advantages of Prometheus over alternative systems such as Graphite is the rich query language that it provides. You can associate labels (which are arbitrary key/value pairs) with any metrics, and you are then able to query the system by label. I'll show examples in this post. Here's a more in-depth comparison between Prometheus and Graphite.

Installation (on Ubuntu 14.04)

I put together an ansible role that is loosely based on Brian Brazil's demo_prometheus_ansible repo.

Check out my ansible-prometheus repo for this ansible role, which installs Prometheus, node_exporter and PromDash (a ruby-based dashboard builder). For people not familiar with ansible, most of the installation commands are in the install.yml task file. Here is the sequence of installation actions, in broad strokes.

For the Prometheus server:
  • download prometheus-0.16.1.linux-amd64.tar.gz from
  • extract tar.gz into /opt/prometheus/dist and link /opt/prometheus/prometheus-server to /opt/prometheus/dist/prometheus-0.16.1.linux-amd64
  • create Prometheus configuration file from ansible template and drop it in /etc/prometheus/prometheus.yml (more on the config file later)
  • create Prometheus default command-line options file from ansible template and drop it in /etc/default/prometheus
  • create Upstart script for Prometheus in /etc/init/prometheus.conf:
# Run prometheus

start on startup

chdir /opt/prometheus/prometheus-server

./prometheus -config.file /etc/prometheus/prometheus.yml
end script

For node_exporter:
  • download node_exporter-0.12.0rc1.linux-amd64.tar.gz from
  • extract tar.gz into /opt/prometheus/dist and move node_exporter binary to /opt/prometheus/bin/node_exporter
  • create Upstart script for Prometheus in /etc/init/prometheus_node_exporter.conf:
# Run prometheus node_exporter

start on startup

end script

For PromDash:
  • git clone from
  • follow instructions in the Prometheus tutorial from Digital Ocean (can't stop myself from repeating that D.O. publishes the best technical tutorials out there!)
Here is a minimal Prometheus configuration file (/etc/prometheus/prometheus.yml):

  scrape_interval: 30s
  evaluation_interval: 5s

  - job_name: 'prometheus'
      - targets:
  - job_name: 'node'
      - targets:

The configuration file format for Prometheus is well documented in the official docs. My example shows that the Prometheus server itself is monitored (or "scraped" in Prometheus parlance) on port 9090, and that OS metrics are also scraped from 5 clients which are running the node_exporter binary on port 9100, including the Prometheus server.

At this point, you can start Prometheus and node_exporter on your Prometheus server via Upstart:

# start prometheus
# start prometheus_node_exporter

Then you should be able to hit to see the metrics exposed by node_exporter, and more importantly to see the default Web console included in the Prometheus server. A demo page available from Robust Perception can be examined here.

Note that Prometheus also provides default Web consoles for node_exporter OS-level metrics. They are available at (the ansible-prometheus role installs nginx and redirects to the previous URL). The node consoles show CPU, Disk I/O and Memory graphs and also network traffic metrics for each client running node_exporter. 

Working with the MySQL exporter

I installed the mysqld_exporter binary on my Prometheus server box.

# cd /opt/prometheus/dist
# git clone
# cd mysqld_exporter
# make

Then I created a wrapper script I called

# cat

export DATA_SOURCE_NAME=“dbuser:dbpassword@tcp(dbserver:3306)/dbname”; ./mysqld_exporter

Two important notes here:

1) Note the somewhat awkward format for the DATA_SOURCE_NAME environment variable. I tried many other formats but only this one worked for me. The wrapper's script main purpose is to define this variable properly. With some of my other tries, I got this error message:

INFO[0089] Error scraping global state: Default addr for network 'dbserver:3306' unknown  file=mysqld_exporter.go line=697

You could also define this variable in ~/.bashrc but in that case it may clash with other  Prometheus exporters (the one for PostgreSQL for example) which also need to define this variable.

2) Note that the dbuser specified in the DATA_SOURCE_NAME variable needs to have either SUPER or REPLICATION CLIENT permissions to the MySQL server you need to monitor. I ran a SQL statement of this form:


I created an Upstart init script I called /etc/init/prometheus_mysqld_exporter.conf:

# cat /etc/init/prometheus_mysqld_exporter.conf
# Run prometheus mysqld exporter

start on startup

chdir /opt/prometheus/dist/mysqld_exporter

end script

I modified the Prometheus server configuration file (/etc/prometheus/prometheus.yml) and added a scrape job for the MySQL metrics:

  - job_name: 'mysql'
    honor_labels: true
      - targets:

I restarted the Prometheus server:

# stop prometheus
# start prometheus

Then I started up mysqld_exporter via Upstart:

# start prometheus_mysqld_exporter

If everything goes well, the metrics scraped from MySQL will be available at

Here are some of the available metrics:

# HELP mysql_global_status_innodb_data_reads Generic metric from SHOW GLOBAL STATUS.
# TYPE mysql_global_status_innodb_data_reads untyped
mysql_global_status_innodb_data_reads 12660
# HELP mysql_global_status_innodb_data_writes Generic metric from SHOW GLOBAL STATUS.
# TYPE mysql_global_status_innodb_data_writes untyped
mysql_global_status_innodb_data_writes 528790
# HELP mysql_global_status_innodb_data_written Generic metric from SHOW GLOBAL STATUS.
# TYPE mysql_global_status_innodb_data_written untyped
mysql_global_status_innodb_data_written 9.879318016e+09
# HELP mysql_global_status_innodb_dblwr_pages_written Generic metric from SHOW GLOBAL STATUS.
# TYPE mysql_global_status_innodb_dblwr_pages_written untyped
mysql_global_status_innodb_dblwr_pages_written 285184
# HELP mysql_global_status_innodb_row_ops_total Total number of MySQL InnoDB row operations.
# TYPE mysql_global_status_innodb_row_ops_total counter
mysql_global_status_innodb_row_ops_total{operation="deleted"} 14580
mysql_global_status_innodb_row_ops_total{operation="inserted"} 847656
mysql_global_status_innodb_row_ops_total{operation="read"} 8.1021419e+07
mysql_global_status_innodb_row_ops_total{operation="updated"} 35305

Most of the metrics exposed by mysqld_exporter are of type Counter, which means they always increase. A meaningful number to graph then is not their absolute value, but their rate of change. For example, for the mysql_global_status_innodb_row_ops_total metric, the rate of change of reads for the last 5 minutes (reads/sec) can be expressed as:


This is also an example of a Prometheus query which filters by a specific label (in this case {operation="read"})

A good way to get a feel for the metrics available to the Prometheus server is to go to the Web console and graphing tool available at You can copy and paste the ine above in the Expression edit box and click execute. You should see something like this graph in the Graph tab:

It's important to familiarize yourself with the 4 types of metrics handled by Prometheus: Counter, Gauge, Histogram and Summary. 

Working with the Postgres exporter

Although not an official Prometheus package, the Postgres exporter has worked just fine for me. 

I installed the postgres_exporter binary on my Prometheus server box.

# cd /opt/prometheus/dist
# git clone
# cd postgres_exporter
# make

Then I created a wrapper script I called

# cat

export DATA_SOURCE_NAME="postgres://dbuser:dbpassword@dbserver/dbname"; ./postgres_exporter

Note that the format for DATA_SOURCE_NAME is a bit different from the MySQL format.

I created an Upstart init script I called /etc/init/prometheus_postgres_exporter.conf:

# cat /etc/init/prometheus_postgres_exporter.conf
# Run prometheus postgres exporter

start on startup

chdir /opt/prometheus/dist/postgres_exporter

end script

I modified the Prometheus server configuration file (/etc/prometheus/prometheus.yml) and added a scrape job for the Postgres metrics:

  - job_name: 'postgres'
    honor_labels: true
      - targets:

I restarted the Prometheus server:

# stop prometheus
# start prometheus

Then I started up postgres_exporter via Upstart:

# start prometheus_postgres_exporter

If everything goes well, the metrics scraped from Postgres will be available at

Here are some of the available metrics:

# HELP pg_stat_database_tup_fetched Number of rows fetched by queries in this database
# TYPE pg_stat_database_tup_fetched counter
pg_stat_database_tup_fetched{datid="1",datname="template1"} 7.730469e+06
pg_stat_database_tup_fetched{datid="12998",datname="template0"} 0
pg_stat_database_tup_fetched{datid="13003",datname="postgres"} 7.74208e+06
pg_stat_database_tup_fetched{datid="16740",datname="mydb"} 2.18194538e+08
# HELP pg_stat_database_tup_inserted Number of rows inserted by queries in this database
# TYPE pg_stat_database_tup_inserted counter
pg_stat_database_tup_inserted{datid="1",datname="template1"} 0
pg_stat_database_tup_inserted{datid="12998",datname="template0"} 0
pg_stat_database_tup_inserted{datid="13003",datname="postgres"} 0
pg_stat_database_tup_inserted{datid="16740",datname="mydb"} 3.5467483e+07
# HELP pg_stat_database_tup_returned Number of rows returned by queries in this database
# TYPE pg_stat_database_tup_returned counter
pg_stat_database_tup_returned{datid="1",datname="template1"} 6.41976558e+08
pg_stat_database_tup_returned{datid="12998",datname="template0"} 0
pg_stat_database_tup_returned{datid="13003",datname="postgres"} 6.42022129e+08
pg_stat_database_tup_returned{datid="16740",datname="mydb"} 7.114057378094e+12
# HELP pg_stat_database_tup_updated Number of rows updated by queries in this database
# TYPE pg_stat_database_tup_updated counter
pg_stat_database_tup_updated{datid="1",datname="template1"} 1
pg_stat_database_tup_updated{datid="12998",datname="template0"} 0
pg_stat_database_tup_updated{datid="13003",datname="postgres"} 1
pg_stat_database_tup_updated{datid="16740",datname="mydb"} 4351

These metrics are also of type Counter, so to generate meaningful graphs for them, you need to plot their rates. For example, to see the rate of rows returned per second from the database called mydb, you would plot this expression:


The Prometheus expression evaluator available at is again your friend. BTW, if you start typing pg_ in the expression field, you'll see a drop-down filled automatically with all the available metrics starting with pg_. Handy!

Working with the AWS CloudWatch exporter

This is one of the officially supported Prometheus exporters, used for graphing and alerting on AWS CloudWatch metrics. I installed it on the Prometheus server box. It's a java app, so it needs a JDK installed, and also maven for building the app.

# cd /opt/prometheus/dist
# git clone
# apt-get install maven2 openjdk-7-jdk
# cd cloudwatch_exporter
# mvn package

The cloudwatch_exporter app needs AWS credentials in order to connect to CloudWatch and read the metrics. Here's what I did:
  1. created an AWS IAM user called cloudwatch_ro and downloaded its access key and secret key
  2. created an AWS IAM custom policy called CloudWatchReadOnlyAccess-201511181031, which includes the default CloudWatchReadOnlyAccess policy (the custom policy is not stricly necessary, and you can use the default one, but I preferred to use a custom one because I may need to further edits to the policy file)
  3. attached the CloudWatchReadOnlyAccess-201511181031 policy to the cloudwatch_ro user
  4. created a file called ~/.aws/credentials with the contents:

The cloudwatch_exporter app also needs a json file containing the CloudWatch metrics we want it to retrieve from AWS. Here is an example of ELB-related metrics I specified in a file called cloudwatch.json:

  "region": "us-west-2",
  "metrics": [
    {"aws_namespace": "AWS/ELB", "aws_metric_name": "RequestCount",
     "aws_dimensions": ["AvailabilityZone", "LoadBalancerName"],
     "aws_dimension_select": {"LoadBalancerName": [“LB1”, “LB2”]},
     "aws_statistics": ["Sum"]},
    {"aws_namespace": "AWS/ELB", "aws_metric_name": "BackendConnectionErrors",
     "aws_dimensions": ["AvailabilityZone", "LoadBalancerName"],
     "aws_dimension_select": {"LoadBalancerName": [“LB1”, “LB2”]},
     "aws_statistics": ["Sum"]},
    {"aws_namespace": "AWS/ELB", "aws_metric_name": "HTTPCode_Backend_2XX",
     "aws_dimensions": ["AvailabilityZone", "LoadBalancerName"],
     "aws_dimension_select": {"LoadBalancerName": [“LB1”, “LB2”]},
     "aws_statistics": ["Sum"]},
    {"aws_namespace": "AWS/ELB", "aws_metric_name": "HTTPCode_Backend_4XX",
     "aws_dimensions": ["AvailabilityZone", "LoadBalancerName"],
     "aws_dimension_select": {"LoadBalancerName": [“LB1”, “LB2”]},
     "aws_statistics": ["Sum"]},
    {"aws_namespace": "AWS/ELB", "aws_metric_name": "HTTPCode_Backend_5XX",
     "aws_dimensions": ["AvailabilityZone", "LoadBalancerName"],
     "aws_dimension_select": {"LoadBalancerName": [“LB1”, “LB2”]},
     "aws_statistics": ["Sum"]},
    {"aws_namespace": "AWS/ELB", "aws_metric_name": "HTTPCode_ELB_4XX",
     "aws_dimensions": ["AvailabilityZone", "LoadBalancerName"],
     "aws_dimension_select": {"LoadBalancerName": [“LB1”, “LB2”]},
     "aws_statistics": ["Sum"]},
    {"aws_namespace": "AWS/ELB", "aws_metric_name": "HTTPCode_ELB_5XX",
     "aws_dimensions": ["AvailabilityZone", "LoadBalancerName"],
     "aws_dimension_select": {"LoadBalancerName": [“LB1”, “LB2”]},
     "aws_statistics": ["Sum"]},
    {"aws_namespace": "AWS/ELB", "aws_metric_name": "SurgeQueueLength",
     "aws_dimensions": ["AvailabilityZone", "LoadBalancerName"],
     "aws_dimension_select": {"LoadBalancerName": [“LB1”, “LB2”]},
     "aws_statistics": ["Maximum", "Sum"]},
    {"aws_namespace": "AWS/ELB", "aws_metric_name": "SpilloverCount",
     "aws_dimensions": ["AvailabilityZone", "LoadBalancerName"],
     "aws_dimension_select": {"LoadBalancerName": [“LB1”, “LB2”]},
     "aws_statistics": ["Sum"]},
    {"aws_namespace": "AWS/ELB", "aws_metric_name": "Latency",
     "aws_dimensions": ["AvailabilityZone", "LoadBalancerName"],
     "aws_dimension_select": {"LoadBalancerName": [“LB1”, “LB2”]},
     "aws_statistics": ["Average"]},

Note that you need to look up the exact syntax for each metric name, dimensions and preferred statistics in the AWS CloudWatch documentation. For ELB metrics, the documentation is here. The CloudWatch name corresponds to the cloudwatch_exporter JSON parameter aws_metric_name, dimensions corresponds to aws_dimensions, and preferred statistics corresponds to aws_statistics.

I modified the Prometheus server configuration file (/etc/prometheus/prometheus.yml) and added a scrape job for the CloudWatch metrics:

  - job_name: 'cloudwatch'
    honor_labels: true
      - targets:

I restarted the Prometheus server:

# stop prometheus
# start prometheus

I created an Upstart init script I called /etc/init/prometheus_cloudwatch_exporter.conf:

# cat /etc/init/prometheus_cloudwatch_exporter.conf
# Run prometheus cloudwatch exporter

start on startup

chdir /opt/prometheus/dist/cloudwatch_exporter

   /usr/bin/java -jar target/cloudwatch_exporter-0.2-SNAPSHOT-jar-with-dependencies.jar 9106 cloudwatch.json
end script

Then I started up cloudwatch_exporter via Upstart:

# start prometheus_cloudwatch_exporter

If everything goes well, the metrics scraped from CloudWatch will be available at

Here are some of the available metrics:

# HELP aws_elb_request_count_sum CloudWatch metric AWS/ELB RequestCount Dimensions: [AvailabilityZone, LoadBalancerName] Statistic: Sum Unit: Count
# TYPE aws_elb_request_count_sum gauge
aws_elb_request_count_sum{job="aws_elb",load_balancer_name=“LB1”,availability_zone="us-west-2a",} 1.0
aws_elb_request_count_sum{job="aws_elb",load_balancer_name=“LB1”,availability_zone="us-west-2c",} 1.0
aws_elb_request_count_sum{job="aws_elb",load_balancer_name=“LB2”,availability_zone="us-west-2c",} 2.0
aws_elb_request_count_sum{job="aws_elb",load_balancer_name=“LB2”,availability_zone="us-west-2a",} 12.0
# HELP aws_elb_httpcode_backend_2_xx_sum CloudWatch metric AWS/ELB HTTPCode_Backend_2XX Dimensions: [AvailabilityZone, LoadBalancerName] Statistic: Sum Unit: Count
# TYPE aws_elb_httpcode_backend_2_xx_sum gauge
aws_elb_httpcode_backend_2_xx_sum{job="aws_elb",load_balancer_name=“LB1”,availability_zone="us-west-2a",} 1.0
aws_elb_httpcode_backend_2_xx_sum{job="aws_elb",load_balancer_name=“LB1”,availability_zone="us-west-2c",} 1.0
aws_elb_httpcode_backend_2_xx_sum{job="aws_elb",load_balancer_name=“LB2”,availability_zone="us-west-2c",} 2.0
aws_elb_httpcode_backend_2_xx_sum{job="aws_elb",load_balancer_name=“LB2”,availability_zone="us-west-2a",} 12.0
# HELP aws_elb_latency_average CloudWatch metric AWS/ELB Latency Dimensions: [AvailabilityZone, LoadBalancerName] Statistic: Average Unit: Seconds
# TYPE aws_elb_latency_average gauge
aws_elb_latency_average{job="aws_elb",load_balancer_name=“LB1”,availability_zone="us-west-2a",} 0.5571935176849365
aws_elb_latency_average{job="aws_elb",load_balancer_name=“LB1”,availability_zone="us-west-2c",} 0.5089397430419922
aws_elb_latency_average{job="aws_elb",load_balancer_name=“LB2”,availability_zone="us-west-2c",} 0.035556912422180176
aws_elb_latency_average{job="aws_elb",load_balancer_name=“LB2”,availability_zone="us-west-2a",} 0.0031794110933939614

Note that there are 3 labels available to query the metrics above: job, load_balancer_name and availability_zone. 

If we specify something like aws_elb_request_count_sum{job="aws_elb"} in the expression evaluator at, we'll see 4 graphs, one for each load_balancer_name/availability_zone combination. 

To see only graphs related to a specific load balancer, say LB1, we can specify an expression of the form:
In this case, we'll see 2 graphs for LB1, one for each availability zone.

In order to see the request count across all availability zones for a specific load balancer, we need to apply the sum function: sum(aws_elb_request_count_sum{job="aws_elb",load_balancer_name="LB1"}) by (load_balancer_name) 
In this case, we'll see one graph with the request count across the 2 availability zones pertaining to LB1.

If we want to graph all load balancers but only show one graph per balancer, summing all availability zones for each balancer, we would use an expression like this: sum(aws_elb_request_count_sum{job="aws_elb"}) by (load_balancer_name)
So in this case we'll see 2 graphs, one for LB1 and one for LB2, with each graph summing the request count across the availability zones for LB1 and LB2 respectively.

Note that in all the expressions above, since the job label has the value "aws_elb" common to all metrics, it can be dropped from the queries because it doesn't produce any useful filtering.

For other AWS CloudWatch metrics, consult the Amazon CloudWatch Namespaces, Dimensions and Metrics Reference.

Instrumenting Go code with Prometheus

For me, the most interesting feature of Prometheus is that allows for easy instrumentation of the code. Instead of pushing metrics a la statsd and Graphite, a web app needs to implement a /metrics handler and use the Prometheus client library code to publish app-level metrics to that handler. The Prometheus server will then hit /metrics on the client and pull/scrape the metrics.

More specifics for Go code instrumentation

1) Declare and register Prometheus metrics in your code

I have the following 2 variables defined in an init.go file in a common package that gets imported in all of the webapp code:

var PrometheusHTTPRequestCount = prometheus.NewCounterVec(
        Namespace: "myapp",
        Name:      "http_request_count",
        Help:      "The number of HTTP requests.",
    []string{"method", "type", "endpoint"},

var PrometheusHTTPRequestLatency = prometheus.NewSummaryVec(
        Namespace: "myapp",
        Name:      "http_request_latency",
        Help:      "The latency of HTTP requests.",
    []string{"method", "type", "endpoint"},

Note that the first metric is a CounterVec, which in the Prometheus client_golang library specifies a Counter metric that can also get labels associated with it. The labels in my case are "method", "type" and "endpoint". The purpose of this metric is to measure the HTTP request count. Since it's a Counter, it will increase monotonically, so for graphing purposes we'll need to plot its rate and not its absolute value.

The second metric is a SummaryVec, which in the client_golang library specifies a Summary metric with labels. I have the same labels are for the CounterVec metric. The purpose of this metric is to measure the HTTP request latency. Because it's a Summary, it will provide the absolute measurement, the count, as well as quantiles for the measurements.

These 2 variables then get registered in the init function:

func init() {
    // Register Prometheus metric trackers

2) Let Prometheus handle the /metrics endpoint

The GitHub README for client_golang shows the simplest way of doing this:

http.Handle("/metrics", prometheus.Handler())
http.ListenAndServe(":8080", nil)

However, most of the Go webapp code will rely on some sort of web framework, so YMMV. In our case, I had to insert the prometheus.Handler function as a variable pretty deep in our framework code in order to associate it with the /metrics endpoint.

3) Modify Prometheus metrics in your code

The final step in getting Prometheus to instrument your code is to modify the Prometheus metrics you registered by incrementing Counter variables and taking measurements for Summary variables in the appropriate places in your app. In my case, I increment PrometheusHTTPRequestCount in every HTTP handler in my webapp by calling its Inc() method. I also measure the HTTP latency, i.e. the time it took for the handler code to execute, and call the Observe() method on the PrometheusHTTPRequestLatency variable.

The values I associate with the "method""type" and "endpoint" labels come from the endpoint URL associated with each instrumented handler. As an example, for an HTTP GET request to a URL such as, "method" is the HTTP method used in the request ("GET"), "type" is "customers", and "endpoint" is "/customers/find".

Here is the code I use for modifying the Prometheus metrics (R is an object/struct which represents the HTTP request):

    // Modify Prometheus metrics
    pkg, endpoint := common.SplitUrlForMonitoring(R.URL.Path)
    method := R.Method
    PrometheusHTTPRequestCount.WithLabelValues(method, pkg, endpoint).Inc()
    PrometheusHTTPRequestLatency.WithLabelValues(method, pkg, endpoint).Observe(float64(elapsed) / float64(time.Millisecond))

4) Retrieving your metrics

Assuming your web app runs on port 8080, you'll need to modify the Prometheus server configuration file and add a scrape job for app-level metrics. I have something similar to this in /etc/prometheus/prometheus.xml:

- job_name: 'myapp-api'
      - targets:
          group: 'production'
      - targets:
          group: 'test'

Note an extra label called "group" defined in the configuration file. It has the values "production" and "test" respectively, and allows for the filtering of Prometheus measurements by the environment of the monitored nodes.

Whenever the Prometheus configuration file gets modified, you need to restart the Prometheus server:

# stop prometheus
# start prometheus

At this point, the metrics scraped from the webapp servers will be available at

Here are some of the available metrics:

# HELP myapp_http_request_count The number of HTTP requests.
# TYPE myapp_http_request_count counter
myapp_http_request_count{endpoint="/merchant/register",method="GET",type="admin"} 2928
# HELP myapp_http_request_latency The latency of HTTP requests.
# TYPE myapp_http_request_latency summary
myapp_http_request_latency{endpoint="/merchant/register",method="GET",type="admin",quantile="0.5"} 31.284808
myapp_http_request_latency{endpoint="/merchant/register",method="GET",type="admin",quantile="0.9"} 33.353354
myapp_http_request_latency{endpoint="/merchant/register",method="GET",type="admin",quantile="0.99"} 33.353354
myapp_http_request_latency_sum{endpoint="/merchant/register",method="GET",type="admin"} 93606.57930099976

myapp_http_request_latency_count{endpoint="/merchant/register",method="GET",type="admin"} 2928

Note that myapp_http_request_count and myapp_http_request_latency_count show the same value for the method/type/endpoint combination in this example. You could argue that myapp_http_request_count is redundant in this case. There could be instances where you want to increment a counter without taking a measurement for the summary, so it's still useful to have both. 

Also note that myapp_http_request_latency, being a summary, computes 3 different quantiles: 0.5, 0.9 and 0.99 (so 50%, 90% and 99% of the measurements respectively fall under the given numbers for the latencies).

5) Graphing your metrics with PromDash

The PromDash tool provides an easy way to create dashboards with a look and feel similar to Graphite. PromDash is available at

First you need to define a server by clicking on the Servers link up top, then entering a name ("prometheus") and the URL of the Prometheus server ("").

Then click on Dashboards up top, and create a new directory, which offers a way to group dashboards. You can call it something like "myapp". Now you can create a dashboard (you also need to select the directory it belongs to). Once you are in the Dashboard create/edit screen, you'll see one empty graph with the default title "Title". 

When you hover over the header of the graph, you'll see other buttons available. You want to click on the 2nd button from the left, called Datasources, then click Add Expression. Note that the server field is already pre-filled. If you start typing myapp in the expression field, you should see the metrics exposed by your application (for example myapp_http_request_count and myapp_http_request_latency).

To properly graph a Counter-type metric, you need to plot its rate. Use this expression to show the HTTP request/second rate measured in the last minute for all the production endpoints in my webapp:


(the job and group values correspond to what we specified in /etc/prometheus/prometheus.xml)

If you want to show the HTTP request/second rate for test endpoints of "admin" type, use this expression:


If you want to show the HTTP request/second rate for a specific production endpoint, use an expression similar to this:


Once you enter the expression you want, close the Datasources form (it will save everything). Also change the title by clicking on the button called "Graph and Axis Settings". In that form, you can also specify that you want the plot lines stacked as opposed to regular lines.

 For latency metrics, you don't need to look at the rate. Instead, you can look at a specific quantile. Let's say you want to plot the 99% quantile for latencies observed in all production endpoint, for write operations (corresponding to HTTP methods which are not GET). Then you would use an expression like this:


As for the HTTP request/second graphs, you can refine the latency queries by specifying a type, an endpoint or both:


I hope you have enough information at this point to go wild with dashboards! Remember, who has the most dashboards wins!

Wrapping up

I wanted to write this blog post so I don't forget all the stuff that was involved in setting up and using Prometheus. It's a lot, but it's also not that bad once you get a hang for it. In particular, the Prometheus server itself is remarkably easy to set up and maintain, a refreshing change from other monitoring systems I've used before.

One thing I haven't touched on is the alerting mechanism used in Prometheus. I haven't looked at that yet, since I'm still using a combination of Pingdom, monit and Jenkins for my alerting. I'll tackle Prometheus alerting in another blog post.

I really like Prometheus so far and I hope you'll give it a try!

Thursday, November 05, 2015

Notes on testing in golang

I've been doing a lot of testing of APIs written in the Go programming language in the last few months. It's been FUN! Writing code in Golang is also very fun. Can't have good code quality without tests though, so I am going to focus on the testing part in this post.

Unit testing

Go is a "batteries included" type of language, just like Python, so naturally it comes with its own testing package, which provides support for automated execution of unit tests. Here's an excerpt from its documentation:
Package testing provides support for automated testing of Go packages. It is intended to be used in concert with the “go test” command, which automates execution of any function of the form
func TestXxx(*testing.T)
where Xxx can be any alphanumeric string (but the first letter must not be in [a-z]) and serves to identify the test routine.
The functionality offered by the testing package is fairly bare-bones though, so I've actually been using another package called testify which provides test suites and more friendly assertions.

Whether you're using testing or a 3rd party package such as testify, the Go way of writing unit tests is to include them in a file ending with _test.go in the same directory as your code under test. For example, if you have a file called customers.go which deals with customer management business logic, you would write unit tests for that code and put them in file called customers_test.go in the same directory as customers.go. Then, when you run the "go test" command in that same directory, your unit tests will be automatically run. In fact, "go test" will discover all tests in files named *_test.go and run them. You can find more details on Go unit testing in the Testing section of the "How to Write Go Code" article.

Integration testing

I'll give some examples of how I organize my integration tests. Let's take again the example of testing an API what deals with the management of customers. An integration test, by definition, will hit the API endpoint from the outside, via HTTP. This is in contrast with a unit test which will test the business logic of the API handler internally, and will live as I said above in the same package as the API code.

For my integration tests, I usually create a directory per set of endpoints that I want to test, something like core-api for example. In there I drop a file called main.go where I set some constants used throughout my tests:

package main

import (

const API_VERSION = "v2"
const API_HOST = ""
const API_PORT = 8000
const API_PROTO = "http"
const API_INIT_KEY = "some_init_key"
const API_SECRET_KEY = "some_secret_key"
const TEST_PHONE_NUMBER = "+15555550000"
const DEBUG = true

func init() {
        fmt.Printf("API_PROTO:%s; API_HOST:%s; API_PORT:%d\n", API_PROTO, API_HOST, API_PORT)

func main() {

For integration tests related to the customer API, I create a file called customer_test.go with the following boilerplate:

package main

import (


// Define the suite, and absorb the built-in basic suite
// functionality from testify - including a T() method which
// returns the current testing context
type CustomerTestSuite struct {
        apiURL          string
        testPhoneNumber string

// Set up variables used in all tests
// this method is called before each test
func (suite *CustomerTestSuite) SetupTest() {
        suite.apiURL = fmt.Sprintf("%s://%s:%d/%s/customers", API_PROTO, API_HOST, API_PORT, API_VERSION)
        suite.testPhoneNumber = TEST_PHONE_NUMBER

// Tear down variables used in all tests
// this method is called after each test
func (suite *CustomerTestSuite) TearDownTest() {

// In order for 'go test' to run this suite, we need to create
// a normal test function and pass our suite to suite.Run
func TestCustomerTestSuite(t *testing.T) {
        suite.Run(t, new(CustomerTestSuite))

By using the testify package, I am able to define a test suite, a struct I call CustomerTestSuite which contains a testify suite.Suite as an anonymous field. Golang uses composition over inheritance, and the effect of embedding a suite.Suite in my test suite is that I can define methods such as SetupTest and TearDownTest on my CustomerTestSuite. I do the common set up for all test functions in SetupTest (which is called before each test function is executed), and the common tear down for all test functions in TearDownTest (which is called after each test function is executed).

In the example above, I set some variables in SetupTest which I will use in every test function I'll define. Here is an example of a test function:

func (suite *CustomerTestSuite) TestCreateCustomerNewEmailNewPhone() {
        url := suite.apiURL
        random_email_addr := fmt.Sprintf("", common.RandomInt(1, 1000000))
        phone_num := suite.testPhoneNumber
        status_code, json_data := create_customer(url, phone_num, random_email_addr)

        customer_id := get_nested_item_property(json_data, "customer", "id")

        assert_success_response(suite.T(), status_code, json_data)
        assert.NotEmpty(suite.T(), customer_id, "customer id should not be empty")

The actual HTTP call to the backend API that I want to test happens inside the create_customer function, which I defined in a separate utils.go file:

func create_customer(url, phone_num, email_addr string) (int, map[string]interface{}) {
        fmt.Printf("Sending request to %s\n", url)

        payload := map[string]string{
                "phone_num":  phone_num,
                "email_addr": email_addr,
        ro := &grequests.RequestOptions{}
        ro.JSON = payload

        var resp *grequests.Response
        resp, _ = grequests.Post(url, ro)

        var json_data map[string]interface{}
        status_code := resp.StatusCode
        err := resp.JSON(&json_data)
        if err != nil {
                fmt.Println("Unable to coerce to JSON", err)
                return 0, nil

        return status_code, json_data

Notice that I use the grequests package, which is a Golang port of the Python Requests package. Using grequests allows me to encapsulate the HTTP request and response in a sane way, and to easily deal with JSON.

To go back to the TestCreateCustomerNewEmailNewPhone test function, once I get back the response from the API call to create a customer, I call another helper function called assert_success_response, which uses the assert package from testify in order to verify that the HTTP response code was 200 and that certain JSON parameters that we send back with every response (such as error_msg, error_code, req_id) are what we expect them to be:

func assert_success_response(testobj *testing.T, status_code int, json_data map[string]interface{}) {
        assert.Equal(testobj, 200, status_code, "HTTP status code should be 200")
        assert.Equal(testobj, 0.0, json_data["error_code"], "error_code should be 0")
        assert.Empty(testobj, json_data["error_msg"], "error_msg should be empty")
        assert.NotEmpty(testobj, json_data["req_id"], "req_id should not be empty")
        assert.Equal(testobj, true, json_data["success"], "success should be true")

To actually run the integration test, I run the usual 'go test' command inside the directory containing my test files.

This pattern has served me well in creating an ever-growing collection of integration tests against our API endpoints.

Test coverage

Part of Golang's "batteries included" series of tools is a test coverage tool. To use it, you first need to run 'go test' with various coverage options. Here is a shell script we use to produce our test coverage numbers:


# Run all of our go unit-like tests from each package

set -e
set -x

cp /dev/null $CTMP
cp /dev/null $CREAL
cp /dev/null $CMERGE

go test -v -coverprofile=$CTMP -covermode=count -parallel=9 ./auth
cat $CTMP > $CREAL

go test -v -coverprofile=$CTMP -covermode=count -parallel=9 ./customers
cat $CTMP |tail -n+2 >> $CREAL

# Finally run all the go integration tests

go test -v -coverprofile=$C -covermode=count -coverpkg=./auth,./customers ./all_test.go
cat $CTMP |tail -n+2 >> $CREAL

rm $CTMP

# Merge the coverage report from unit tests and integration tests

cd $GOPATH/src/core_api/
cat $CREAL | go run ../samples/mergecover/main.go >> $CMERGE

set +x

echo "You can run the following to view the full coverage report!::::"
echo "go tool cover -func=$CMERGE"
echo "You can run the following to generate the html coverage report!::::"
echo "go tool cover -html=$CMERGE -o coverage.html"

The first section of the bash script above runs 'go test' in covermode=count against every sub-package we have (auth, customers etc). It combines the coverprofile output files (CTMP) into a single file (CREAL).

The second section runs the integration tests by calling 'go test' in covermode=count, with coverpkg=[comma-separated list of our packages], against a file called all_test.go. This file starts an HTTP server exposing our APIs, then hits our APIs by calling 'go test' from within the integration test directory.

The coverage numbers from the unit tests and integration tests are then merged into the CMERGE file by running the mergecover tool.

At this point, you can generate an html file via go tool cover -html=$CMERGE -o coverage.html, then inspect coverage.html in a browser. Aim for more than 80% coverage for each package under test.