This document describes the architecture of Sourcegraph’s monitoring stack, and the technical decisions we have made to date and why.
To better understand our goals with Sourcegraph’s monitoring stack, please read monitoring pillars: long-term vision.
We use a custom declarative Go generator syntax for:
- Defining the services we monitor.
- Describing what those services do to site admins.
- Laying out dashboards in a uniform, consistent, and simple way.
- Generating the Prometheus alerting rules and Grafana dashboards.
- Generating documentation in the form of “possible solutions” for site admins to follow when alerts are firing.
This allows us to assert constraints and principles that we want to hold ourselves to, as described in our monitoring pillars, as well as provide integrations with our monitoring architecture.
To learn more about adding monitoring using the generator, see adding monitoring.
We use Grafana for:
- Displaying generated dashboards for our Prometheus metrics and alerts.
- Providing an interface to query Prometheus metrics and Jaeger traces.
sourcegraph/grafana image handles shipping Grafana and Sourcegraph monitoring dashboards. It bundles:
- Preconfigured Grafana, which displays data from Prometheus and Jaeger
- Dashboards generated by the monitoring generator.
For convenience, Grafana is served on
/-/debug/grafana on all Sourcegraph deployments via a reverse-proxy restricted to admins.
Edits through the web UI are prohibited, for two reasons:
- We do not want customers, or ourselves, to introduce dashboards or modifications through the Grafana web UI (see monitoring pillars: long-term-vision)
- Services served via reverse-proxy in this manner could be vulnerable to cross-site request forgery (CSRF), which is complicated to resolve (#6075) without just forbidding all POST requests.
- The monitoring generator creates provisioned dashboards that, once loaded, cannot be edited in Grafana (technical limitation in Grafana.)
We use Prometheus for:
- Collecting high-level, and low-cardinality, metrics from our services.
- Defining Sourcegraph alerts as both:
sourcegraph/prometheus image handles shipping Sourcegraph metrics and alerting. It bundles:
- Preconfigured Prometheus, which consumes metrics from Sourcegraph services.
- Alert and recording rules generated by the monitoring generator.
- Alertmanager, which handles alerts from Prometheus.
- prom-wrapper, which subscribes to updates in site configuration and propagates relevant settings to Alertmanager configuration.
alert_count metrics are special Prometheus recording rules that evaluate a single upper or lower bound, as defined in an Observable and generated by the monitoring generator. This metric is always either 0 if the threshold is not exceeded (or data does not exist, if configured), or 1 if the threshold is exceeded. This allows historical alert data to easily be consumed programmatically.
Learn more about the
alert_count metrics in the metrics guide.
alert_count: Why did we introduce
alert_count metrics? At the time, a customer needed the ability to monitor their Sourcegraph instances’ health and we did not have any Prometheus metrics, alert definitions, or dashboards that they could consume. We knew that we would need to over time solidify our entire monitoring stack, which created a situation where in the interim we needed something that they could pipe to their own internal custom metrics and alerting system to monitor Sourcegraph health - without being tied to our exact Prometheus metrics, etc. such that we could continue to improve and refine them over the coming versions of Sourcegraph. To solve this, we introduced this
alert_count metric which made it possible for them to query the firing alerts in Sourcegraph via the Prometheus API (we did not have alertmanager at the time either). There is some evidence that we may be able to use
ALERTS instead of
alert_count (tracking issue), but one problem there is that
ALERTS does not show non-firing alerts and that has proven very useful both for this customer and to have in Grafana in general to instill some confidence that we are in fact monitoring a pretty substantial number of things. We believe
alert_count can probably be replaced with
ALERTS, but more discussion / a plan is needed.
We use Alertmanager for:
- Providing data about currently active Sourcegraph alerts.
- Routing alerts to appropriate receivers and silencing them when desired, configured using site configuration.
Rationale for Alertmanager: An approach for notifiers using Grafana was considered, but had some issues outlined in #11832, so Alertmanager was selected as our notification provider.
The prom-wrapper is the entrypoint program for
sourcegraph/prometheus and it:
- Handles starting up Prometheus and Alertmanager
- Applies updates to site configuration by generating a diff and applying changes
- Most notably, this includes configuring notifiers and silences for Sourcegraph alerts
- Exposes endpoints for configuration issues, alerts summary statuses, and reverse-proxies Prometheus and Alertmanager
Rationale for an all-in-one Prometheus image with prom-wrapper: This allows us to avoid adding a new separate service (which must be handled in all our deployment types), thus simplifying the deployment story for our monitoring stack, while also improving the alert debugging workflow (for example, simply port-forward
svc/prometheus to get access to the entire alerting stack), with minimal disadvantages (for example, high-availability Prometheus and Alertmanager can still be configured via
ALERTMANAGER_ADDITIONAL_FLAGS, and Alertmanager can be disabled via
There is currently no process defining how custom additions should be made to our monitoring stack (for example, to accomodate Cloud-specific needs).
Notifiers for alerts are configured via the
deploy-sourcegraph-cloud frontend ConfigMap for
Alerts go to the #alerts channel, with critical alerts going to OpsGenie. Critical alerts are assigned responders in OpsGenie based on the
owner field of each alert - the mapping of
owner to engineering team can be found in the Sourcegraph Cloud site configuration.
We use Blackbox Exporter for:
- Monitoring external and internal http endpoints
- Configuring alerts based on http response codes to notify teams when sourcegraph is unreachable via an external URL as well as
sourcegraph-frontend-internalto assist in diagnosing the root cause of sourcegraph being unavailable.
Prometheus scrapes Blackbox Exporter every 30s, which will send a request to endpoints configured via the
deploy-sourcegraph-dotcom prometheus ConfigMap for
Alerts are configured separately via the
deploy-sourcegraph-dotcom prometheus ConfigMap for
Rationale for Blackbox Exporter: Site24x7 has been a source of flaky alerts, outlined in #10742 and more broadly in #11966. In an effort increase reliablity and broaden the scope of our monitoring, Blackbox Exporter was selected as it integrates well into our existing Prometheus and Alertmanager stack.
Cloudflare Analytics is used to extract useful data about the performance of our WAF, as well as the overall traffic distribution to our instances. Note that the retention of analytics data is relatively short due to the limits on our plan. See our Cloudflare developer guide for more details on how to use the analytics API.