Hobbit – Monitoring that works
description
Transcript of Hobbit – Monitoring that works
Hobbit – Monitoring that works
Hobbit MonitorLinuxforum 2007March 3rd 2007
Henrik Storner <[email protected]>
Agenda
Hobbit history What does “monitoring” mean? Demo / Screen shots Architecture: The Hobbit components
Hobbit server Network server monitoring Server monitoring with the clients
Setting it up: Quick tour of the configuration Custom checks Hobbit at the CSC Copenhagen data center Future
Hobbit history
In 2001, CSC's Managed Web Services division had no monitoring of websites, only box-monitoring with Unicenter/TNG.
Enter Big Brother – the UI is great, but BB is written in Korn shell and is slooooow!
bbgen toolkit (2002) eliminated some slow parts of BB, but kept the BB daemon.
Hobbit (2005) is mostly compatible with BB but has a completely different architecture.
Hobbit contains no BB parts.
What does “monitoring” mean?
Availability: Can I access the website? Performance: ... without getting bored ? Capacity: ... when we triple the number of users?
It is vital that you know which of these questions you must answer.
The main focus for Hobbit is availability(but there's a bit of the others thrown in)
Hobbit overview
A small demo
Hobbit architecture
1 Hobbit server holds all current data, i.e. the status of everything we monitor.
Usually, the Hobbit server also hosts the web interface, and runs tasks which take care of storing history- and trend-data.
1+ servers perform network tests, and reports the results to the Hobbit server (may be the same)
Clients collect data from each monitored server, and send it to the Hobbit server for analysis.
The Hobbit server
Stores current data in RAM only, and never does any slow disk I/O or process forking
Historical and trend data stored on disk Core hobbitd daemon feeds server tasks via an IPC
“channel”, using shared memory Server tasks handle e.g. client data analysis, trend
data updates, and alerting Some tasks can be distributed on multiple servers
(e.g. client data analysis) Extensible – you can write your own tasks, e.g. a
task that stores all measurements in a database.
Hobbit web interface
Overview pages are static HTML, rebuilt once a minute with the current status data.
Detailed status pages are dynamically generated “Critical Systems” view is dynamic Will probably switch to an all-dynamic setup The web UI is not particularly attractive or
flexible, so web designers are welcome! Some customization is possible by modifying
header- and footer-files
Network service monitoring
ping : Is the server alive ? Connect : Will it accept new connections ? Service : Is the network service running ? Application : Is it working ?
It is easy to check that the service is running because it uses a standard protocol (eg HTTP)
But the end-user only cares about the application!
Webserver says it's “200 OK”
Check the actual data returned !
And the devil's in the detail...
Standard network tests
ping FTP SSH SMTP POP IMAP(S) NNTP(S) LDAP(S) query HTTP(S) w/ content SSL certificate
AJP13 VNC clamd spamd cupsd rsync Oracle TNS listener
add your own service definition
Server monitoring – Hobbit clients
Usually a “Hobbit client” runs on the server. Clients are really dumb – they know how to
collect some data, but they know nothing about interpreting the data they collect.
Runs uptime, free, df, ps, netstat, mount, who ... Collects server-side statistics Scans server log files for new entries Collects data for directories and individual files The raw data is sent to the Hobbit server for
analysis.
Standard server tests
CPU load average System uptime System clock Memory usage Swap usage File system usage Process counts Network ports
File attributes File data Directory sizes Log file data
Data can be graphed
Setting it up
All configuration is kept on the Hobbit server All configuration files are text based Uses regular expressions a lot
bb-hosts (list hosts Hobbit knows about, defines network service tests and the web page layout)
hobbit-alerts.cfg (rules for sending alerts) hobbit-clients.cfg (rules for analyzing data from
clients) client-local.cfg (instructions for client data
collection)
bb-hosts
## Master configuration file for Hobbit#
group My hosts127.0.0.1 localhost # bbd http://localhost/
192.168.1.1 demohost # pop3 http://127.0.0.1/ smtp \
cont=Login;https://www/Login.php;Please.*userid
hobbit-alerts.cfg
HOST=demohost SERVICE=httpMAIL [email protected] TIME=W:0800:2200
SERVICE=httpMAIL [email protected] /usr/local/bin/smsalert +4512345678
hobbit-clients.cfg
HOST=*DISK / 80 90 EXHOST=backup.foo.com
HOST=db.foo.comDISK %/data/ IGNORE
HOST=%web[1-9]PROC apache MIN=4 MAX=20DIR /var/log/apache SIZE<100000 TRACK yellowFILE /var/www/index.html \
MD5=dd2cf7192db28919203eef126943b
client-local.cfg
# This file tell clients what file/log data to report
[linux]log:/var/log/messages
[web1]dir:/var/log/apachefile:/var/www/default.html:md5
Why server-side analysis?
Managing configuration files on each monitored server is impossible when you have 2000 clients.
Bulk configuration updates are much easier Configuration settings can apply to groups of hosts Adding new analysis tools only requires upgrading
the Hobbit server – not all of the clients (provided they already collect the necessary data, of course)
Having RAW data available is USEFUL. Only downside: Your Hobbit server must spend
some cpu time analyzing the raw client data
Custom checks
Custom checks normally check something, then send a red/yellow/green “status” message
A check can run locally on a host as part of the client installation
A check can run centrally and pull data from several hosts (eg. grabbing data with SNMP)
A check can run on the Hobbit server, using data that has already been collected (“combo-tests” or extra client-data analysis)
Numeric data can be tracked in graphs
Simple client-side check
#!/bin/bashCOLUMN=weather; COLOR=green
DEGREES=`/usr/local/bin/getweather temperature`if [ $DEGREES -ge 30 ]; then COLOR=red; fi
$BB $BBDISP \ “status $MACHINE.$COLUMN $COLOR `date` temperature=$DEGREES”
exit 0
Server-side checks
You can hook modules into all kinds of Hobbit data: Status messages, data collected from Hobbit clients and so on.
E.g. Hobbit clients run “who” to report who is logged on.
To monitor for a root login on all servers only takes is 62 lines of Perl (see hobbitd_rootlogin.pl in source)
Windows, SNMP and other stuff
Windows client: BBWinNote: Does not support central configuration
SNMP add-on: Devmon Both are OSS, available on Sourceforge.net Other add-ons available, e.g. for database
monitoring. Add-ons for Big Brother (available from
deadcat.net) can be used – but check licensing
Hobbit@CSC - Summary
The Copenhagen data center is the largest CSC data center in EMEA, globally in the top 5.
Hobbit/BBWin/BB clients on 90% of all servers. Hobbit is considered mission-critical.
Lots of network tests, especially for Web- and middleware systems (J2EE and LDAP)
Web application monitoring done through customer-built “monitoring” web pages
Hobbit@CSC – Multiple views
Multiple sets of web pages with Hobbit data: One set grouped by account manager, then by account:
Lets the account manager quickly see if his customers are running OK
One set grouped by sysadmin group, then by account: Lets the system administrators quickly see what servers need attention
One set for customers who want access to Hobbit The “Critical Systems” view is monitored 24x7
Hobbit@CSC - reports
Availability reports pre-generated for daily, weekly and monthly availability
Reports and detailed history available on-line for 3 months
Monthly reports available for 12 months Graphs clean-up automatically, provide data for
1½ years (1 day average)
Hobbit@CSC: Statistics
1 Web / 2 net serversSun E220R server450 Mhz Ultrasparc II1 GB RAM2x72 GB SCSI disk
1 RRD serverHP DL3803 Ghz Xeon1 GB RAM2x72 GB SCSI disk
3.800 hosts 28.000 statuses 9.500.000 updates/day
= 111 updates/second 3.100 network tests 40.000 webpages/day
27.000 RRD files = ~160.000 RRD datasets
8.500 RRD graphs/day
Hobbit@CSC :Load on main server
Future work
Load balancing of Hobbit tasks: 4.3.0 Graph updates and viewing History log storage Client data analysis Network checks
High availability ? Maybe not ... can be handled externally
Re-design the web UI – any volunteers ? Automated web checking of a full user session,
perhaps using Mozilla or Konqueror
The End
Questions ?