ReleaseEngineering/Applications/Slavealloc
Application Description
Slavealloc is a client-server application. The client is runslave.py. Communication is via a very basic HTTP request to http://slavealloc.build.mozilla.org/gettac/$slavename, where the response is expected to be a buildbot.tac file suitable for use to start buildslave.
The slavealloc server is a implemented as a small Twisted application (source) which serves the tac generator, a JSON REST interface, and a client-side JavaScript interface.
The same source code also implements a command-line frontend to the REST interface.
Requirements
The server depends on
- MySQL database
- Python
- Twisted
- SQLAlchemy
External Resources
The slavealloc server uses the following external resources:
- two MySQL databases (production and staging)
- VM host for the slaveallocator VM
Security
The slave allocator hands out low-security slave passwords in the .tac files, which are stored in cleartext in the database. It does not do any sort of authentication either for read or modify operations, and relies on the Build VLAN firewalls to prevent external access.
Monitoring
The slavealloc host has the basic host monitoring from nagios (ping, filesystems, etc.), plus an HTTP GET to /api/pools, just to make sure the daemon is still responding.
Deployment
The slave allocator server is deployed on a single host, slavealloc.build.mozilla.org.
Server Setup
IT installed RHEL6 along with MySQL client libraries, and set up the proper firewall rules to allow database access.
As root, virtualenv-1.5.2 was installed into the system Python library. The following system packages were installed via yum:
- http://download.fedora.redhat.com/pub/epel/6/i386/epel-release-6-5.noarch.rpm (for EPEL packages; use --nogpgcheck)
- hg
- MySQL-python
- nginx (from EPEL repo)
- nscd (and enable it)
Nginx
Nginx frontends for both staging and production instances. The virtualhosts files are available in hg.
Note that on the x86_64 system slavealloc is currently installed on, the following must be added to the http section of nginx.conf:
# required to use virtualhosts - http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=547722 server_names_hash_bucket_size 33;
Twisted Daemon
The 'slavealloc' user runs the twisted daemon on this host. The user account is locked and accessed only via su from root.
The daemon is installed in a virtualenv at /tools/slavealloc-$rev, using the pre-checked python packages on the puppet server. Note that --no-site-packages is not used here, because we need access to the (binary) MySQL-python package which is installed systemwide:
cd /tools virtualenv slavealloc-8fe4dbc09d03 ./slavealloc/bin/pip install -e hg+http://hg.mozilla.org/build/tools@8fe4dbc09d03#egg=tools \ --no-index --find-links=http://staging-puppet.build.mozilla.org/staging/python-packages/ ln -s slavealloc-8fe4dbc09d03 slavealloc ln -s slavealloc-8fe4dbc09d03 slavealloc-staging
There is a make-slavealloc-virtualenv.sh script available in /tools to make this process automatic.
Once this was set up, the 'slavealloc dbinit' command was used to initialize the database.
The production and staging tac files are in /build/slavealloc. Staging runs on port 1079, and production runs on 1080. The files are similar to those in hg, and include the commented pool_recycle line, with timeout 400, to automatically expire connections before the MySQL server itself does. If this still occurs, we can look to the example of the Buildbot source for a better solution.
Startup is done via initscripts.
Slave Side
All slaves run runslave.py during startup. This file is distributed via puppet. The larger slave-startup process is described in ReleaseEngineering/Buildslave Startup Process.
Backups
The slavealloc server has
11 4 * * * slavealloc /tools/slavealloc/bin/slavealloc dbdump -D mysql://mumblemumblemumble > /builds/slavealloc/production-1080/dbdump.pkl
in /etc/cron.d/slavealloc-bkup, as a basic protection against someone accidentally running dbinit or doing something equally catastrophic.
Staging
As described above, http://staging-slavealloc.build.mozilla.org/ points to a staging implementation of the slave allocator. This implementation has its own database runs from a distinct daemon, although it is served from the same nginx instance.
NOTE: all slaves are configured use the production allocator. Allocations from the staging allocator will need to be simulated by hand (runslave.py has an command-line option to set the allocator URL). This is done to allow us, as a group, to move slaves between staging and production using a single slave allocator.
The prod-db-to-staging.sh script will copy the production db to staging for use when staging new changes.
Development
First, install build/tools in a virtualenv:
cd tools virtualenv sandbox sandbox/bin/pip install -e .
Then you can run the slavealloc daemon locally from the root of the tools repository with a simple:
sandbox/bin/twistd -noy lib/python/slavealloc/contrib/slavealloc-combined.tac
Note that due to what I believe to be a bug in pip, you may need to explicitly install Twisted to get the twistd executable installed:
pip install -U twisted
To set up a fresh database, use
sandbox/bin/slavealloc dbinit -D sqlite:///slavealloc.db
This configuration will use SQLite to access {{{slavealloc.db}}} in the current directory. You can hack on the static web content while the daemon is running.