Koji Infrastructure SOP

We are transitioning from two buildsystems, koji for Fedora and plague for EPEL, to just using koji. This page documents both.

Koji and plague are our buildsystems. They share some of the same machines to do their work.

Contact Information

Owner: Fedora Infrastructure Team

Contact: #fedora-admin, sysadmin-build group

Persons: mbonnet, dgilmore, f13, notting, mmcgrath

Location: Phoenix

Servers: koji.fedoraproject.org, buildsys.fedoraproject.org, xenbuilder[1-4] , hammer1, ppc[1-4]

Purpose: Build packages for Fedora.

Description

Users submit builds to koji.fedoraproject.org or buildsys.fedoraproject.org. From there it gets passed on to the builders.

At present plague and koji are unaware of each other. A result of this may be an overloaded builder. A easy fix for this is not clear at this time

Troubleshooting and Resolution

Restarting Koji

If for some reason koji needs to be restarted, make sure to restart the koji master first, then the builders. If the koji master has been down for a short enough time the builders do not need to be restarted.

service httpd restart
service kojira restart
service kojid restart

If postgres becomes interrupted in some way, koji will need to be restarted. As long as the koji master daemon gets restarted the builders should reconnect automatically. If the db server has been restarted and the builders don't seem to be building, restart their daemons as well.

kojid won't start or some builders won't connect

In the event that some items are able to connect to koji while some are not, please make sure that the database is not filled up on connections. This is common if koji crashes and the db connections aren't properly cleared. Upon restart many of the connections are full so koji cannot reconnect. Clearing old connections is easy, guess about how long it the new koji has been up and pick a number of minutes larger then that and kill those queries. From db3 as postgres run:

echo "select procpid from pg_stat_activity where usename='koji' and now() - query_start >= '00:40:00' order by query_start;" | psql koji | grep "^  " | xargs kill

OOM (Out of Memory) Issues

Out of memory issues occur from time to time on the build machines. There are a couple of options for correction. The first fix is to just restart the machine and hope it was a one time thing. If the problem continues please choose from one of the following options.

Increase Memory

The xen machines can have memory increased on their corresponding xen hosts. At present this is the table:

xen3	xenbuilder1
xen4	xenbuilder2
disabled	xenbuilder3
xen8	xenbuilder4

Edit /etc/xen/xenbuilder[1-4] and add more memory.

Decrease weight

Each builder has a weight as to how much work can be given to it. Presently the only way to alter weight is actually changing the database on db3:

$ sudo su - postgres
-bash-2.05b$ psql koji
koji=# select * from host limit 1;
id | user_id |          name          |  arches   | task_load | capacity | ready | enabled
----+---------+------------------------+-----------+-----------+----------+-------+---------
6 |     130 | ppc3.fedora.redhat.com | ppc ppc64 |       1.5 |        4 | t     | t
(1 row)
koji=# update host set capacity=2 where name='ppc3.fedora.redhat.com';

Simply update capacity to a lower number.

Disk Space Issues

The builders use a lot of temporary storage. Failed builds also get left on the builders, most should get cleaned but plague does not. The easiest thing to do is remove some older cache dirs.

Step one is to turn off both koji and plague:

/etc/init.d/plague-builder stop
/etc/init.d/kojid stop

Next check to see what file system is full:

df -h

If any one of the following directories is full, send an outage notification as outlined in: Infrastructure/OutageTemplate to the fedora-infrastructure-list and fedora-devel-list, then contact Mike McGrath

1. /mnt/koji
2. /mnt/ntap-fedora1/scratch
3. /pub/epel
4. /pub/fedora

Typically just / will be full. The next thing to do is determine if we have any extremely large builds left on the builder. Typical locations include /var/lib/mock and /mnt/build (/mnt/build actually is on the local filesystem):

du -sh /var/lib/mock/* /mnt/build/*

• /var/lib/mock/dist-f8-build-10443-1503: classic koji build
• /var/lib/mock/fedora-6-ppc-core-57cd31505683ef1afa533197e91608c5a2c52864: classic plague build

If nothing jumps out immediately, just start deleting files older than one week. Once enough space has been freed start koji and plague back up:

/etc/init.d/plague-builder start
/etc/init.d/kojid start