Skip to main content

Your submission was sent successfully! Close

Thank you for signing up for our newsletter!
In these regular emails you will find the latest updates from Canonical and upcoming events where you can meet our team.Close

Thank you for contacting us. A member of our team will be in touch shortly. Close

Monitoring

An important requirement of a production-ready system is to provide a way for the admins to monitor and actively observe the operation of the cluster. Charmed Spark delivers a comprehensive observability stack which includes:

  • business analytical monitoring via the deployment of Spark History Server, that provides a UI that allows Spark users (such as Data Scientist and Data Engineers) to analyze the logs of their jobs at a more business level, allowing to break down a Spark Job into its different parts and stages
  • cluster administration monitoring via integration with the Canonical Observability Stack (COS) for aggregating raw application metrics,
    and setting up alerts and dashboards based on resource utilization.

In the following we will describe the architecture and how the different components connect among themselves, providing the monitoring capabilities of the solution.

Before diving in, it is worth providing a brief reminder of the structure of a Spark Job.

Spark Job structure

A Spark Job is generally composed by two main types of processes:

  • Spark Driver, that executes the main program, interacts with the Cluster Manager to request computational resources (Executors), and control the workflow by providing instructions to the Executors on which operations/tasks to be executed. The Spark driver generally holds a small fraction of the data, generally resulting on the aggregation/reduction operation done by the Executors.

  • Spark Executors, that hold the raw data in memory and do the actual processing.
    The Executors generally communicate among themselves to shuffle data whenever needed, and they provide the horizontally scalable structure that enables analysis to be distributed and scaled out to massive datasets.

Spark Driver and Spark Executors are allocated by a Cluster Manager that in the Charmed Spark solution is represented by the K8s control plane. Therefore, in Charmed Spark, driver and executors run as Kubernetes pods, belonging to a given namespace. Every pod needs to be run using a service account that needs to be set up with the correct permissions. Charmed Spark tooling, i.e. the spark-client snap and the spark8t Python libraries, make sure that service account are created correctly and configured appropriately. The driver pod must be running with a service account that is able to create pods, services and configmaps in order to correctly spawn executors. Moreover, Charmed Spark service accounts also store Spark configuration centrally in Kubernetes as secrets, that must be readable/writable depending on their scope. Please refer to the explanations about the Charmed Spark hierachical configuration for more information.

Metrics

Spark Jobs are ephemeral processes that may be long-lived in some cases but also very short-lived on others. Spark Jobs may be launch occasionally, or multiple Spark Jobs may be running at the same times.

In this sense, Spark is very different from traditional database management systems (such as Kafka, MongoDB, PostgreSQL, etc), that are generally unique, with stable endpoints and up and running at all times.

Such services generally expose a stable endpoint where relevant metrics (such as memory consumption, errors, GC time, etc) can be scraped at regular intervals by a metric collector component, such as the Prometheus bundled in COS. However, because of the intrinsic characteristics of Spark, such an approach is not directly suitable and feasible for Spark.

Instead of being passively scraped and metrics being pulled directly by Prometheus, Spark Jobs running with Charmed Spark actively push the metrics into Prometheus Pushgateway, that is a sink component able to temporarily store the metrics, and exposing a stable, reliable and long-lived endpoint to be scraped by Prometheus. Metrics pushed into Prometheus Pushgateway are organized into groups and characterized by a series of labels.

Therefore, Spark Driver and Spark Executors need to be able to export metrics and push them to the Prometheus Pushgateway, in appropriate groups, so that they can be temporarily stored, exposed and then scraped by Prometheus. Charmed Spark comes with the spark-metrics exporter already packaged in the Charmed Spark OCI image, that is used as default base for both Driver and Executors pods. This exporter is designed to push metrics to the prometheus-pushgateway-k8s charm, which is integrated with the Canonical Observability Stack.

Once the metrics are ingested by Prometheus, they are then exposed to the user through Grafana where they can be visualized in custom dashboards. Custom alerting rules can also be defined and efficiently managed by AlertManager.

Logs

Logs of driver and executors are stored on the pod local filesystem by default, however they can also be forwarded to the Loki via the LOKI_URL environment variable. This variable is used by Pebble, to configure log-forwarding to remote Loki server.

This feature is available from Charmed Spark version 3.4.2/edge and requires a working Loki. Hence, we recommend using it with COS.

Charmed Spark also allows you to store these logs into S3 or Azure storage, so that they can be re-read and visualized using the Spark History Server, allowing users to monitor and visualize the information and metrics about the job execution.

Last updated 9 days ago. Help improve this document in the forum.