You are on page 1of 35

Fact-based Monitoring

puppetconf 2014
Alexis Lê-Quôc @alq
Alexis Lê-Quôc, @alq
CTO at Datadog
Poll: Monitoring makes me…

happy cry
proud want to hide
Puppet brings Automation to
Systems Management
Improve
Monitoring
the way Puppet has
improved
Systems Management
“The good old days”

• Your “CMDB” was Excel

• SSH in and hack away

• Little time for anything else


Then Puppet came…

• Expressive rules that capture expected result

• Using facts and classifiers, a.k.a. metadata to figure out where to


apply changes

• That freed up a lot of our time*

* on a per-machine basis
“Puppet brings immunity of configuration to change in
infrastructure”

–Me (just now)


I have seen this before…
“[SQL brings] immunity of application to change in storage
structure and access strategy”

–C.J. Date (1977)

http://www.cs.berkeley.edu/~brewer/cs262/SystemR.pdf
SQL
• 1974 IBM introduces System R and its Structured Query Language

• Expressive rules that capture expected result

• Using facts and predicates, a.k.a. metadata to figure out what data
to get

• That freed up a lot of development time


SQL

• From a time-consuming, imperative mess (“how”)

• … to expressive data queries (“what”)

SQL query

SELECT (desired facts)



FROM (existing facts)

WHERE (matching criteria)
Puppet

• From a time-consuming, imperative mess (“how”)

• … to expressive configuration queries (“what”)

puppet apply

CHANGE (desired facts)



FROM (existing puppet facts)

WHERE (matching puppet classes)
Is there a pattern?
“Break free from ever more complex naming conventions for
hostnames as a means of identity. Use a very rich set of meta
data provided by each machine to address them.”

–MCollective overview
MCollective

• From a time-consuming, imperative mess (“how”)

• … to expressive orchestration queries (“what”)

mco rpc service restart service=nginx\


-F webpool=A

EXEC (desired actions)



FROM (existing puppet facts)

WHERE (matching puppet classes)
Back to monitoring

• Monitoring is to behavior what Puppet is to configuration

• Monitoring is to behavior what MCollective is to orchestration


Monitoring

• From a time-consuming, imperative mess (“how”)

• … to expressive monitoring queries (“what”)

Monitoring query

MONITOR (desired behavior)



FROM (existing heartbeats/metrics)

WHERE (matching puppet facts)
Examples
• “All provisioned web servers in the production environment,
datacenter ABC must respond to queries within 200ms”

• “All PostgreSQL servers must have a postgres: bgwriter process


running”

• “At least one ActiveMQ server is up to support mcollective"

• Never mention a hostname


Hosts are not the center of the
monitoring universe.

Facts are!
Hosts are just places where facts occur.
The proof is in the pudding…
Hosts at the center of the universe
a.k.a. the Wrong Way
“Its fairly straightforward, so hopefully you find things easy to
understand…”

–Nagios Core 4 manual on monitoring clusters


Host-centric: Monitor a DNS cluster
check_command
check_service_cluster!"DNS Cluster"!0!1!
$SERVICESTATEID:host1:DNS Service$,$SERVICESTATEID:host2:DNS
Service$,$SERVICESTATEID:host3:DNS Service$

Where do host1, host2, host3 come from?


Host-centric: can’t use facts directly
• “Host groups solve this problem”. No, they don’t.

• Combinatorial explosion, e.g. trivially

• 4 data centers (us-1, us-2, eu, apac)

• 5 classes (web, db, cache, appserver, hadoop)

• 3 environments (test, staging, prod)

• => up to 119 materialized host groups


Nagios-bashing?

• No!

• Same fatal flaw with all host-centric monitoring tools

• Host-centric monitoring forces an extra, expensive step:

• replicate fact-based conditionals in host-centric templates


“Please note that this module is not for the faint of heart. Even I
(the author) have my head hurt each time I have to make
modifications to it…”

–puppet-nagios author
Facts at the center of the universe
a.k.a. the Right Way
"De Revolutionibus manuscript p9b" by Nicolas Copernicus - www.bj.uj.edu.pl. Licensed under Public domain via Wikimedia Commons - http://commons.wikimedia.org/wiki/File:De_Revolutionibus_manuscript_p9b.jpg#mediaviewer/
File:De_Revolutionibus_manuscript_p9b.jpga
Earlier Examples

• “All provisioned web servers in the production environment,


datacenter ABC must respond to queries within 200ms”

• “All PostgreSQL servers must have a postgres: bgwriter process


running”

• “At least one ActiveMQ server is up to support mcollective"


In Sensu (heartbeats)
• “All PostgreSQL servers must have a postgres: bgwriter process
running”

class postgres::monitoring::sensu {
sensu::subscription { 'postgres': }
}

• Monitoring using a fact-based query

• Is node of class “postgres” and subscribed to “postgres” or not?

• If so, it will execute the postgres check


In Datadog (metrics)
• “All provisioned web servers in the production environment,
datacenter ABC must respond to queries within 200ms”
$ puppet module install datadog-datadog_agent

class {
‘datadog_agent’:
api_key => …,
tags => [$environment],
fact_to_tags => [“datacenter”]
}
include datadog_agent::integrations::nginx
In Datadog (metrics)
• Monitoring using a fact-based query

• Puppet facts directly reused

max(nginx.request.latency{production,datacenter:ABC}) < 200


What to take away
Fact-based monitoring

1. Hosts are not at the center of the monitoring universe

2. Expressive monitoring uses queries

3. Monitoring queries should use Puppet facts


Thank you!

You might also like