Sanic prometheus metrics

After googling for a while I didn't find a library that would enable some prometheus metrics for Sanic-based apps, so I had to write one. It makes adding monitoring to your Sanic app super easy, just add one line to your code (ok, two if you count import :) and point Prometheus to a newly appeared /metrics endpoint.

Installation and enabling monitoring

Installatio:

pip install prometheus-sanic

Easy-peasy:

from sanic import Sanic
from prometheus_sanic import monitor

app = Sanic()
...

if __name__ == "__main__":
  monitor(app).expose_endpoint() # adds /metrics endpoint to your Sanic server
  app.run(host="0.0.0.0", port=8000)

Actually, there're two ways to run monitoring:

The one you've seen above, monitor(app).expose_endpoint(). It just adds a new route to your Sanic app, exposing /metrics endpoint on the same host and port your Sanic server runs. It might be useful if you run your app in a container and you do not want to expose different ports for metrics and everything else. You can customize the /metrics endpoint by passing the metrics_path keyword argument: monitor(app, metrics_path='/my_metrics_path').expose_endpoint().
monitor(app).start_server(addr=..., port=...). Runs a HTTP server on given address and port and exposes /metrics endpoint on it. This might be useful if you want to restrict access to your /metrics endpoint using some firewall rules

Versions compatibility

☑︎ use 0.1.0 for Sanic <= 0.4.1
☑︎ use 0.1.3 for Sanic >= 0.5.0
☑︎ use >= 0.1.4 if you need multiprocessing support
☑︎ use 0.1.6 if you have to use promtheus-client <= 0.4.2
☑︎ use 0.1.8 with prometheus-client >= 0.5.0
☑︎ use 0.2.0 with prometheus-client >= 0.7.1 and Sanic >= 18.12

Exposed metrics

At the moment prometheus-sanic provides four metrics:

sanic_request_count - total number of requests (labels: method, endpoint, status) [counter]
sanic_request_latency_sec - request latency in seconds (labels: method, endpoint) [histogram]
sanic_mem_rss_bytes - resident memory used by the process (in bytes) [gauge]
sanic_mem_rss_perc - a percent of total physical memory used by the process running Sanic [gauge]

Labels

method: a HTTP method (i.e. GET/POST/DELETE/etc)
endpoint: just a string, a name identifying a point handling a group of requests. By default it's just the first element of the relative path of the URL being called (i.e. for http://myhost/a/b/c you'll end up having /a as your endpoint). It is quite configurable, in fact it's up you what's gonna get to the endpoint label (see help(prometheus_sanic.monitor) for more details)
http_status: a HTTP status code

Multiprocess mode

Sanic allows to launch multiple worker processes to utilise parallelisation, which is great but makes metrics collection much trickier (read more) and introduces some limitations.

In order to collect metrics from multiple workers, create a directory and point a prometheus_multiproc_dir environment variable to it. Make sure the directory is empty before you launch your service:

% rm -rf /path/to/your/directory/*
% env prometheus_multiproc_dir=/path/to/your/directory python your_sanic_app.py

Unfortunately you can not use monitor(app).start_server(addr=..., port=...) in multiprocess mode as it exposes a prometheus endpoint from a newly created process.

Configuration

Best you can do is:

% ipython
In [1]: from prometheus_sanic import monitor
In [2]: help(monitor)

Prometheus quering examples:

Average latency over last 30 minutes:

rate(sanic_request_latency_sec_sum{endpoint='/your-endpoint'}[30m]) /
rate(sanic_request_latency_sec_count{endpoint='/your-endpoint'}[30m])

95th percentile of request latency:

histogram_quantile(0.95, sum(rate(sanic_request_latency_sec_bucket[5m])) by (le))

Physical memory usage percent over last 10 minutes:
```
rate(sanic_mem_rss_perc[10m])
```

swistakm/prometheus-sanic