Prometheus was developed at SoundCloud starting in 2012, when the company discovered that their existing metrics and monitoring solutions were not sufficient for their needs. Specifically, they identified needs that Prometheus was built to meet including: a multi-dimensional data model, operational simplicity, scalable data collection, and a powerful query language, all in a single tool. The project was open-source from the beginning, and began to be used by Boxever and Docker users as well, despite not being explicitly announced. Prometheus was inspired by the monitoring tool Borgmon used at Google. By 2013, Prometheus was introduced for production monitoring at SoundCloud. The official public announcement was made in January 2015. In May 2016, the Cloud Native Computing Foundation accepted Prometheus as its second incubated project, after Kubernetes. The blog post announcing this stated that the tool was in use at many companies including Digital Ocean, Ericsson, CoreOS, Weaveworks, Red Hat, and Google. Prometheus 1.0 was released in July 2016. Subsequent versions were released through 2016 and 2017, leading to Prometheus 2.0 in November 2017. In August 2018, the Cloud Native Computing Foundation announced that the Prometheus project had graduated.
Architecture
A typical monitoring platform with Prometheus is composed of multiple tools:
Multiple exporters that typically run on the monitored host to export local metrics.
Prometheus to centralize and store the metrics.
Alertmanager to trigger alerts based on those metrics.
PromQL is the query language used to create dashboards and alerts.
Data storage format
Prometheus data is stored in the form of metrics, with each metric having a name that is used for referencing and querying it. Each metric can be drilled down by an arbitrary number of key=value pairs. Labels can include information on the data source and other application-specific breakdown information such as the HTTP status code, query method, endpoint, etc. The ability to specify an arbitrary list of labels and to query based on these in real time is why Prometheus' data model is called multi-dimensional. Prometheus stores data locally on disk, which helps for fast data storage and fast querying. There is ability to store metrics in remote storage.
Data collection
Prometheus collects data in the form of time series. The time series are built through a pull model: the Prometheus server queries a list of data sources at a specific polling frequency. Each of the data sources serves the current values of the metrics for that data source at the endpoint queried by Prometheus. The Prometheus server then aggregates data across the data sources. Prometheus has a number of mechanisms to automatically discover resources that it should be using as data sources.
PromQL
Prometheus provides its own query language PromQL that lets users select and aggregate data. PromQL is specifically adjusted to work in convention with a Time-Series Database and therefore provides time-related query functionalities. Examples include the rate function, the instant vector and the range vector which can provide many samples for each queried time series. Prometheus has four clearly defined metric types around which the PromQL components revolve. The four types are
Gauge
Counter
Histogram
Summary
Alerts and monitoring
Configuration for alerts can be specified in Prometheus that specifies a condition that needs to be maintained for a specific duration in order for an alert to trigger. When alerts trigger, they are forwarded to Alertmanager service. Alertmanager can include logic to silence alerts and also to forward them to email, Slack, or notification services such as PagerDuty.
Dashboards
Prometheus is not intended as a dashboarding solution. Although it can be used to graph specific queries, it is not a full-fledged dashboarding solution and needs to be hooked up with Grafana to generate dashboards; this has been cited as a disadvantage due to the additional setup complexity.
Interoperability
Prometheus favors white-box monitoring. Applications are encouraged to publish internal metrics to be collected periodically by Prometheus. Some exporters and agents for various applications are available to provide metrics. Prometheus supports some monitoring and administration protocols to allow interoperability for transitioning: Graphite, StatsD, SNMP, JMX, and CollectD. Prometheus focuses on the availability of the platform and basic operations. The metrics are typically stored for few weeks. For long term storage, the metrics can be streamed to remote storage solutions.
Standardization into OpenMetrics
There is an effort to promote Prometheus exposition format into a standard known as OpenMetrics. Some products adopted the format: InfluxData's TICK suite, InfluxDB, Google Cloud Platform, and DataDog.
Usage
Prometheus was first used in-house at SoundCloud, where it was developed, for monitoring their systems. The Cloud Native Computing Foundation has a number of case studies of other companies using Prometheus. These include digital hosting service Digital Ocean, digital festival DreamHack, and email and contact migration service ShuttleCloud. Separately, Pandora Radio has mentioned using Prometheus to monitor its data pipeline. GitLab provides a Prometheus integration guide to export GitLab metrics to Prometheus and it is activated by default since version 9.0
