Setting up a LAVA pipeline instance¶
Initial considerations¶
- The default setup of the LAVA packages and codebase is for the current dispatcher and Deploying Distributed Instances.
- A LAVA pipeline instance can have existing remote worker support alongside but uses a completely different mechanism to identify remote workers and run jobs on pipeline devices.
- If both systems are enabled, devices can support both pipeline and current JSON submissions. It is not yet possible to disable JSON submissions. If there is no relevant configuration for a device other than pipeline support, a JSON submission would be accepted but would stay in Submitted state until cancelled. See Changes for existing remote workers.
- The default setup provides both mechanisms, the only step required
to allow pipeline submissions to devices connected to
http://localhost
is to have pipeline devices available. - Distributed deployments need changes on each worker, see Changes for existing remote workers.
- Helpers will be developed in due course but currently, pipeline setup is principally a manual task for admins.
- If only pipeline devices are to be supported, the dispatchers
running
lava-slave
do not need to have thelava-server
package installed. Each dispatcher does need to be able to connect to the ZMQ port specified in thelava-master
configuration of the instance (which is then the only machine related to that instance which haslava-server
installed). Thelava-server
package on the master should be installed as a single master instance of LAVA. - The ZMQ protocol incorporates buffering at each end such that
either the
lava-master
or thelava-slave
service can be restarted at any time without affecting currently running jobs or requiring any changes or restarts at the other end of the connection. There are no other connections required between the slave and the master and the outgoing request from the slave is initiated by the slave, so it should be possible for the slave to be behind a local firewall, as long as the relevant ports are open for outgoing traffic. i.e. the slave pulls from the master, the master cannot push to the slave.
Detailed changes¶
The pipeline design designates the machine running Django and PostgreSQL
as the lava-master
and all other machines connected to that master
which will actually be running the jobs are termed lava-slave
machines.
If this slave has no devices which will be used by the current
dispatcher, only by the pipeline, just install lava-dispatcher
:
$ sudo apt install lava-dispatcher
Change the init script for
lava-slave
(/etc/init.d/lava-slave
) to point at the relevantlava-master
.Change the port numbers, if required, to match those in use on the
lava-master
.Restart
lava-slave
once the changes are complete:$ sudo service lava-slave restart
The administrator of the master will then be able to allocate pipeline devices to this slave.
Note
For security reasons, the slave does not declare the devices connected to it to the master. The slave actually needs no knowledge of what is connected or where. All this information is stored solely in the database of the master. Once this data is entered by the admin of the master, the slave then needs to connect and the admin can then select that slave for the relevant devices. Once selected, the slave can immediately start running pipeline jobs on those devices.
The administrator of the master will require the following information about the devices attached to each slave:
- Confirmation that a suitable template already exists, for each device i.e. Adding support for a device of a known type
- A completed and tested device dictionary for each device.
This information contains specific information about the local network setup of the slave and will be transmitted between the master and the slave in clear text over ZMQ. Any encryption would need to be arranged separately between the slave and the master. Information typically involves the hostname of the PDU, the port number of the device on that PDU and the port number of the serial connection for that device. The slave is responsible for ensuring that these ports are only visible to that slave. There is no need for any connections to be visible to the master.
Changes for existing remote workers¶
On an existing remote worker, a lava-master
daemon will already be
running on localhost (doing nothing). Once the migration to the
pipeline is complete, the lava-server
package can be removed
from all workers, so the above information relates to this endpoint. In
the meantime, remote workers should have lava-master
disabled on
localhost once the slave has been directed at the real master as above.
Disabling lava-master on workers¶
Warning
Only do this on the remote worker but make sure it is done on all remote workers before submitting pipeline jobs which would need the devices on those workers.
If a new worker does not need to run jobs using the current dispatcher,
i.e. if all devices on this worker are exclusive, then
lava-server
does not need to be installed and there is no lava-master
daemon to disable.
For existing workers, pipeline jobs will be likely be mixed with JSON
jobs. This leads to lava-server
being installed on the workers (solely
to manage the JSON jobs). On such workers, lava-master
should be
disabled once lava-slave
has been reconfigured:
$ sudo invoke-rc.d lava-master stop
$ sudo update-rc.d lava-master remove
$ sudo chmod a -x /etc/init.d/lava-master
$ sudo service lava-master status
lava-master: unrecognized service
Removing the executable bits stops the lava-master being re-enabled when the packages are updated.