We have been working on PoC of a CI environment design which will:
· Provide fast build or rebuild of environments for testing.
· Enable multiple independent concurrent builds
· Be available on-demand through automation or individual request.
· Be capable of fully utilising all available hardware
· Flexible enough to be used to build super-realistic development environments.
We intend to contribute and maintain our work within the Apache repos. However, we are currently building the POC, figuring out the requirements (and quirks) of the individual pieces, before pushing something concrete for to the community to review.
We envision that Trillian would cater for a number of use cases:
1. CloudStack community integration testing of master against multiple deployment scenarios (using ASF infra)
2. CloudStack community integration testing of PRs against multiple deployment scenarios (using ASF infra)
3. Organisations/individuals running the full suites of tests available in Marvin against any physical environment they have.
4. Organisations/individuals deploying and running the full suites of tests available in Marvin against virtualised infrastructures which can be deployed by Marvin.
As we intend Trillian to test multiple environments concurrently, we use nested virtualization on ESXi hosts (our testing has shown that this is the only hypervisor which can support the nested virtualisation of all other hypervisors with reasonable performance). We use Ansible to deploy and configure all aspects of the build as this will greatly lower the barrier to entry for independent testers.
We use CloudStack to provision the management server and virtualised (nested) hosts on the physical hosts. We are creating Ansible playbooks and roles which can:
1. Create guest instances using Rene’s Ansible 2.0 CloudStack modules - a Marvin VM, a Mgmt Server (CentOS or Ubuntu), any number of compute hosts (KVM, vSphere or XenServer. Hyper-V later)
2. Configure hosts (inc. installing the relevant CloudStack agent where required)
3. Install required ACS packages on management server
4. Configure a zone (including adding the compute hosts) via Marvin.
5. Run the required Marvin tests.
6. Return the results
We may need to propose enhancements to Marvin in order to sync the configuration of hosts with the configuration used by Marvin.
Using virtualised test environments, we can have multiple test scenarios running concurrently. To do this we have found that it is necessary to create pools or ranges of VLANs and IP addresses and allocate them to environments. So for any given physical environment which will be used for testing in, we take the total range(s) of IPs and VLANs available and carve them into non-overlapping chunks suitable for concurrent use as mgmt, public and guest networks. These are stored in a MariaDB database. When a range is being used in a testing environment, that range is marked as ‘inuse’ in the database. When creating a test environment, Trillian looks in the database for the next available VLAN range, the next available public IP range and so on. The returned values are used to populate a Marvin cfg file which in turn will be used to both build the environment and when running the Marvin testing. When the virtualised infra is cleaned up, the database will be updated to reflect that the used ranges are available again.
This initiative has only recently been started, and as stated earlier we are currently figuring out the requirements (and quirks) of the individual pieces and looking for the most suitable wrapper to glue it all together.
Also I have found that Marvin requires a little work to make the output more meaningful/readable (especially in the case of errors and exceptions) and to make it a little more intelligent about the tests it can/can’t run based on the chosen infrastructure components. I have also found unreachable or very slow ISO and template paths hardcoded into Marvin or individual tests.
We plan to enhance tests to address these issues and also reduce runtimes where possible.
We will continue to our work on the PoC and share the results.
We’d like to get some community agreement around the goals and use cases which we are planning to satisfy before we get too far down the road of our PoC. So please share your thoughts.