Skip to content

Latest commit

 

History

History
138 lines (91 loc) · 6.15 KB

README.md

File metadata and controls

138 lines (91 loc) · 6.15 KB
Raw

test-all codecov

prometheus-push-client

Push metrics from your periodic long-running jobs to existing Prometheus/VictoriaMetrics monitoring system.

Currently supports pushes directly to VictoriaMetrics:

For pure Prometheus setups, several options are supported:

  • to pushgateway or prom-aggregation-gateway in OpenMetrics format via HTTP. Please read corresponding docs about appropriate use cases and limitations;
  • to StatsD or statsd-exporter in StatsD format via UDP. Prometheus and StatsD metric types are not fully compatible, so currenly all metrics become StatsD gauges, but rate, increase, histogram_quantile and other PromQL functions produce same results as if types never changed.

Install it via pip:

pip install prometheus-push-client

Metrics

This library uses prometheus-client metric implementation, but adds some minor tweaks.

Separate registry

New metric constructors use separate PUSH_REGISTRY as a default, not to interfere with other metrics already defined and monitored in existing projects.

Default labelvalues

With regular prometheus_client, defaults may be defined for either none or all the labels (with labelvalues), but that's not enough. Moreover labelvalues sometimes doesn't work as expected.

We probably want to define some defaults, like hostname, or more importantly, if we use VictoriaMetrics cluster, VictoriaMetrics_AccountID=<tenant_id> (try 0 as a default) label must always be set, and metrics without it will be ignored.

Following example shows how to use defaults, and how to override them if necessary.

import prometheus_push_client as ppc


counter1 = ppc.Counter(
    name="c1",
    labelnames=["VictoriaMetrics_AccountID", "host", "event_type"],
    default_labelvalues={
        "VictoriaMetrics_AccountID": 0,
        "host": socket.gethostname(),
    }
)


# regular usage
counter1.labels(event_type="login").inc()

# overriding defaults
counter1.labels(host="non-default", event_type="login").inc()
# same effect as above: defaults are applied in `labelvalues`
# order for "missing" labels in the beginning
counter1.labels("non-default", "login").inc()

Metrics with no labels are initialized at creation time. This can have unpleasant side-effect: if we initialize lots of metrics not used in currently running job, batch clients will have to push their non-changing values in every synchronization session.

To avoid that we'll have to properly isolate each task's metrics, which can be impossible or rather tricky, or we can create metrics with default, non-changing labels (like hostname). Such metrics will be initialized on first use (inc), and we'll be pushing only those we actually utilized.

Clients

Batch clients

Batch clients spawn synchronization jobs "in background" (meaning in a thread or asyncio task) to periodically send all metrics from ppc.PUSH_REGISTRY to the destination.

Clients will attempt to stop gracefully, synchronizing registry "one last time" after job exits or crashes. Sometimes this may mess up sampling, but the worst case I could artifically create looks like this:

graceful push effect

Best way to use them is via decorators / context managers. These clients are intended to be used with long running, but finite tasks, which could be spawned anywhere, therefor not easily accessible by the scraper. If that's not the case -- just use "passive mode" w/ the scraper instead.

def influx_udp_async(host, port, period=15.0):
def influx_udp_thread(host, port, period=15.0):
def statsd_udp_async(host, port, period=15.0):
def statsd_udp_thread(host, port, period=15.0):
def influx_http_async(url, verb="POST", period=15.0):
def influx_http_thread(url, verb="POST", period=15.0):
def openmetrics_http_async(url, verb="POST", period=15.0):
def openmetrics_http_thread(url, verb="POST", period=15.0):

Usage example:

import prometheus_push_client as ppc


req_hist = ppc.Histogram(
    name="external_requests",
    namespace="acme"
    subsystem="job123",
    unit="seconds",
    labelnames=["service"]
)


@ppc.influx_udp_async("victoria.acme.inc.net", 9876, period=15)
async def main(urls):
    # the job ...
    req_hist.labels(gethostname(url)).observe(response.elapsed)

# OR

async def main(urls):
    async with ppc.influx_udp_async("victoria.acme.inc.net", 9876, period=15):
        # the job ...
        req_hist.labels(gethostname(url)).observe(response.elapsed)

Please read about mandatory job tag within url while using pushgateway.

Streaming clients

If for some reason every metric change needs to be synced, UDP streaming clients are implemented in this library.

def influx_udp_aiostream(host, port):
def influx_udp_stream(host, port):
def statsd_udp_aiostream(host, port):
def statsd_udp_stream(host, port):

Usage is completely identical to batch clients' decorators / context managers.

⚠️ Histogram and Summary .time() decorator doesn't work in this mode atm, because it can't be monkey-patched easily.

Transports

Main goal is not to interrupt measured jobs with errors from monitoring code. Therefor all transports will attempt to catch all network errors, logging error info and corresponding tracebacks to stdout.