banshee/services-collection/services/haproxy/https/stats

HAProxy statistics

Returns, for each frontend, backend and server of an HAProxy load balancer, its state, its current sessions against their limit, and its errors per second, read from its statistics page in CSV over HTTP. Five watches are offered, each one set on the probe.

Versions

Generic: the CSV of the statistics page is the same from HAProxy 1.5 to 3.x, new columns being added at its end; each column is found by its name in the first line. The Data Plane API is not read here.

How it works

One GET to the address of the statistics page, ;csv added when the address does not carry it (the form the HAProxy management guide gives); nothing else is sent. The first line names the columns:

# pxname,svname,qcur,qmax,scur,smax,slim,stot,bin,bout,dreq,dresp,ereq,econ,eresp,wretr,wredis,status,weight,act,bck,...,type,...
www,FRONTEND,,,12,80,2000,5120,...,OPEN,...,0,...
www,web1,0,0,5,40,,2400,...,UP,...,2,...
www,BACKEND,0,0,5,40,200,2400,...,UP,...,1,...

An instance is <pxname>/<svname>: www/FRONTEND, www/web1, www/BACKEND. The listening sockets (type 3) are left out.
status is the first word of the column status, in capitals: OPEN for a frontend that accepts, UP, DOWN, MAINT, DRAIN, NOLB; a server without health check reads NO CHECK. During a change the column reads UP 1/3 (going down) or DOWN 1/2 (coming up): the state it still has, UP or DOWN, is returned.
sessions is scur; sessions-percent is scur against slim, only when a limit is set (maxconn of the frontend, the backend or the server).
errors is the rate of ereq + econ + eresp (request errors, connection errors, response errors) between this run and the previous one, per second; it comes from the second run on.
A load balancer that refuses the connection, or does not answer in the delay of protocol:http:timeout, is critical: "HAProxy down". A code 401, 403 or 404 stops the collector misconfigured; any other code but 200 unreachable. A page that is not the CSV of the statistics stops it misconfigured.

Prerequisites

A statistics page served by HAProxy, open to the poller's address, for instance:

frontend stats
    bind :8404
    stats enable
    stats uri /stats
    stats auth monitoring:<a password>
    acl poller src 192.0.2.10
    http-request deny unless poller

192.0.2.10 stands for the poller's address. stats auth is optional; without stats admin, the page only reads.

Its full address in the variable haproxy:url, reachable from the poller.
With stats auth, the account in protocol:http:user and protocol:http:password, served on https://: an account never goes to http:// (put the page behind bind :8404 ssl crt …).

Variables

Name	Label	Says	Default
`haproxy:url`	Statistics page address	the address of the statistics page; the collector adds `;csv`	—

Outputs

Name	Label	Unit	Per instance
`status`	State	—	yes, per frontend, backend or server
`sessions`	Current sessions	—	yes
`sessions-percent`	Sessions against their limit	%	yes
`errors`	Errors	errors/s	yes

Watches

Name	Title	Kind	Defaults	Checked by default
`down`	Down	state to avoid, on `status`	critical on `DOWN`	yes
`maintenance`	In maintenance or draining	state to avoid, on `status`	warning on `MAINT`, `DRAIN`, `NOLB`	no
`saturation`	Sessions near their limit	above a threshold, on `sessions-percent`	warning above 80 %, critical above 95 %	yes
`errors`	Errors	above a threshold, on `errors`	warning above 1 error/s, for 3 runs	no
`missing`	Frontend, backend or server missing	expected instances, on `status`	no name given: it raises nothing until names are set on the probe	no

A backend whose servers are all down reads DOWN itself: "Down" raises on it and on each of its servers. A server taken out on purpose reads MAINT, which "Down" does not raise; leave it out by the selection of the probe, or check "In maintenance or draining".

Limits

With several processes (nbproc, before HAProxy 2.5), the page shows only the process that answers.
A reload sets the counters back to zero: the error rate is skipped for one run.
An address that already carries ;csv is used as is.
Not checked against an HAProxy: the columns are those the management guide of HAProxy lists for the CSV statistics (docs.haproxy.org).