how i enriched prometheus metrics article cover

How I enriched prometheus metrics

by Bruno, on 05-May-2022

VMWare exporter metrics didn't have all the informations we wanted. This is how I enriched them.


In my current mission, in order to set up a metric-federation, we have to gather vmware metrics.

So we used the vmware_exporter by Daniel Pryor, and to not edit it, i wrote a small front made with fastapi.

Metrics are get by fastapi, parsed, enriched, reserialized, and scrapped by prometheus.

Metrics names need to be mapped, as they are unified with metrics from openstack and IBM cloud.

MapTable is the one below :

    "vmware_vm_power_state" : "power_state"
    "vmware_vm_cpu_usage_average"': "cpu_usage_percentage
    "vmware_vm_mem_consumed_average": "memory_used"
    "vmware_vm_memory_max": "memory_total"
    "vmware_vm_mem_usage_average": "memory_usage_percentage",
    "vmware_vm_num_cpu": "vcpus"

Step1 : Parsing vmware_exporter metrics

Exporter is hit from within a function named request_exporter().

The output of the exporter is one big str file that need to be parsed in order to be processed.

I used prometheus_client.parser for this.

from prometheus_client.parser import text_string_to_metric_families

data_from_exporter = request_exporter(target)
metric_families = text_string_to_metric_families(data_from_exporter)

The variable metric_families is a multilevel object containing metrics grouped by metric name.

It can be represented as a dict, where key is the metric_name and the value is a list containing all metrics as tuples.

Example :

    "vmware_vm_power_state": [
        ("vmware_vm_power_state", {"vm_name": "vm1", "label2": "value2"}, 1),
        ("vmware_vm_power_state", {"vm_name": "vm2", "label2": "value2"}, 0),

So I made a little algorithm to flatten the object as a list of dicts :

processed_exporter_data = []

for family in data_from exporter:
    for sample in family.samples:
        row dict = {}
        row dict["__name__"] = sample[0]
        for key, value in sample[1].items():
            row dict[key] = value
        row_dict["value"] = sample[2]
        processed exporter_data.append(row_dict)

I will have then a list of dicts, a dataframe can be created out of it.

    {"__name__": "vmware_vm_power_state", "vm_name": "vm1", "value": 1},
    {"__name__": "vmware_vm_power_state", "vm_name": "vm2", "value": 0},

Now let's play with dataframe ...

Step2: Playing with dataframes

First of all, i'm creating the dataframe :

import pandas

metrics_df = pandas.DataFrame(processed_exporter_data)

Then, I am purging the uneeded labels :


This way, I am using the dataframe to delete the 4 columns. inplace=Trueis used in order to overwrite existing dataframe and not create a new one.

Then i have to merge metric data with data from two other internal datasources (one for the hosts/vm referential, one for the business lines).

These are also represented as dataframes.

businesslines_referential_df = referential_df.merge(
    businesslines_df, how="left", left_on="ecosystem", right_index=True

This merge will enrich referential dataframe with businesslines data, from right to left, according to ecosystem name and use the index from the right dataframe as the join key.

Now it is time to merge this dataframe with the metrics dataframe.

metrics_df = metrics_df.merge(
    businesslines_referential_df, how="inner", left_on="vm_name", right_index=True

This merge is made with the inner method, so this way we only keep common columns, according to vm_name.

Now i'm replacing original metric name with the mapping we talked about earlier :

metrics_df.replace({"__name_": MAPPING_METRICS_NAME}, inplace=True)

Now "vmware_vm_power_state" is mapped to "power_state".

The result is similar to this :

__name__ vm_name ecosystem business_line value
power_state vm1 TESTECO BL1 1

Now to be scrapped by prometheus that data needs to be reserialized.

Step3: How i re-serialized data ?

Prometheus Metrics strings are pretty simple :

metric_name{label1="value1", label2="value2"} 0

Using list comprehension i am iterating over the dataframe to print a string containing in the begining the metric_name and in the end the metric_value.

return "\n".join(
        f"{row['__name__']}{generate_dict_label(row)} {row['value']}"
        for index, row in metrics df.iterrows()

The tricky thing to do is to generate labels stuff, because we cannot tweak output of a dict in python.

Dict in python look like {"key": "value"}, we have to generate a string that look like {label="value"}

So I wrote a little function that generate that string.

def generate_dict_label(row):

    labels = set(row.index) - {"__name__", "value"}

    return (
        + ",".join(
                for key, value in row.items()
                if key in labels
        + "}"

Finally, serialized data looks like :

power_state{vm_name="vm1", ecosystem="TESTECO", business_line="BL1"} 1

Now, metrics are scrapped by prometheus calling a small module made with fastapi, that will parse, enrich, and reserialize data, that comes from a prometheus vmware exporter.