Varnish exporter for Prometheus
Go Shell
Latest commit 87700bf Dec 27, 2016 @filippog filippog committed with Recommend recording rules for backend metrics (#13)
Closes #12

Build Status

Varnish exporter for Prometheus

Grafana example

Scrapes the varnishstat -j JSON output on each Prometheus collect and exposes all reported metrics. Metrics with multiple backends or varnish defined identifiers (e.g. VBE.*.happy SMA.*.c_bytes LCK.*.creat) and other metrics with similar structure (e.g. MAIN.fetch_*) are combined under a single metric name with distinguishable labels. Vanish naming conventions are preserved as much as possible to be familiar to Varnish users when building queries, while at the same time trying to following Prometheus conventions like lower casing and using _ separators.

Handles runtime Varnish changes like adding new backends via vlc reload. Removed backends are reported by varnishstat until Varnish is restarted.

Advanced users can use -n -N, they are passed to varnishstat.

Tested to work against Varnish 4.1.0, 4.0.3 and 3.0.5. Missing category groupings in 3.x like MAIN. are detected and added automatically for label names to be consistent across versions, assuming of course that the Varnish project does not remove/change the stats.

I won't make any backwards compatibility promises at this point. Your built queries can break on new versions if metric names or labels are refined. If you find bugs or have feature requests feel free to create issues or send PRs.

Installing and running

You can find the latest binary releases for linux, darwin, windows, freebsd, openbsd and netbsd from the github releases page.

See prometheus_varnish_exporter -h for available options.

To test that varnishstat is found on the host machine and to preview all exported metrics run

prometheus_varnish_exporter -test

Grafana examples

Here are some of the queries that I used to create Grafana dashboards with Prometheus data exported by this exporter. I am fairly new to Prometheus so these examples might be not be optimal, but will get you started. If you are monitoring multiple Varnish instances, you need make the queries more detailed with Prometheus label selectors.

I'd be interested in Grafana dahsboard .json exports or Prometheus queries you make with this exporter. Perhaps we could make a wiki page of examples and premade generic dashboards.

// Frontend requests

// Frontend requests

// Network bytes frontend
irate(varnish_main_s_resp_hdrbytes[5m]) + irate(varnish_main_s_resp_bodybytes[5m])

// Network bytes per backend
sum by (backend) keep_common (irate(varnish_backend_beresp_hdrbytes[5m]) + irate(varnish_backend_beresp_bodybytes[5m]))

// Free memory (malloc allocator)

Varnish 4 and VCL UUIDs

Starting with version 1.2 backend and server labels are always set. For backend-related metrics and Varnish 4 the server tag will be set to the VCL UUIDs for that backend. Note that there might be multiple VCLs loaded at the same time and the server tag might not be meaningful in that case.

To aggregate all loaded VCLs into per-backend metric the following Prometheus recording rules are recommended:

backend:varnish_backend_bereq_bodybytes:sum = sum(varnish_backend_bereq_bodybytes) without (server)
backend:varnish_backend_bereq_hdrbytes:sum = sum(varnish_backend_bereq_hdrbytes) without (server)
backend:varnish_backend_beresp_bodybytes:sum = sum(varnish_backend_beresp_bodybytes) without (server)
backend:varnish_backend_beresp_hdrbytes:sum = sum(varnish_backend_beresp_hdrbytes) without (server)
backend:varnish_backend_conn:sum = sum(varnish_backend_conn) without (server)
backend:varnish_backend_happy:sum = sum(varnish_backend_happy) without (server)
backend:varnish_backend_pipe_hdrbytes:sum = sum(varnish_backend_pipe) without (server)
backend:varnish_backend_pipe_in:sum = sum(varnish_backend_pipe_in) without (server)
backend:varnish_backend_pipe_out:sum = sum(varnish_backend_pipe_out) without (server)
backend:varnish_backend_req:sum = sum(varnish_backend_req) without (server)


Use go build or ./ <version> for cross compilation.

Docker notes

Building a docker image would not be useful for this application. It operates by executing the varnishstat tool. This executable is of course not available inside the exporters container. We could install and run Varnish inside the same container, but this would make little sense and be harder to manage.

I do not know of a way to get the varnish stats over the network (--net="host" could be used to talk to it). If you have more info on this, please open an issue.