The stability, availability and performance of the infrastructure over the years is a testament to the quality of the processes and people involved in the rollout and update of middleware products.
However, no process is perfect and it is only natural for people to move on - even particularly dedicated and talented people.
Moreover, there is an ongoing effort to further harmonise European infrastructures, in the context of the European Open Science Cloud.
The technology underwriting the infrastructure itself is rapidly evolving, with all manner of cloud, container, and serverless platforms being brought to bear.
This will mean more (and more complex) deployment scenarios, unforeseen integration issues, etc.
We have been thinking for some time about the issues that this presents, and how to address them systematically.
This describes how we address one small piece of the problem: trusted configurations of the Unified Middleware Distribution.
I will focus on the use of Ansible in this discussion, but the issues relate to pretty much any tool.
Ansible in the UMD world
We1 used to have a unique configuration management tool: YAIM23.
YAIM seems to have been born in a fit of exasperation4 resulting from the complexity, in the time before configuration management tools.
There was no Puppet, no Ansible, no Salt, no Chef… at most one could have hoped for a proto-Quattor.
The positive aspect of this monoculture was that there was not much room for confusion: learn how YAIM works, and it will do most of what you need it to do.
Sure, there may be parts of the system which it cannot be held responsible for, but in the wide and massively distributed grid world of the ’90s and first decade of the 21st century, it worked.
Not bad for a bunch of shell scripts.
YAIM actually served two purposes: configuration management and deployment.
Configuration data was written into the tool itself, in the form of a series of configuration files read as input.
This desired state was then achieved by YAIM itself by executing the “bunch of shell scripts” alluded to before.
This made it a bit difficult to separate configuration management from deployment.
Site configuration management
Things have changed.
Configuration management tools have grown to maturity and are now critically important to the success of any software-driven enterprise.
We have entire stacks to choose from now, which solve problems all the way from development through testing, deployment and monitoring in production.
One the one hand, this is great for sites – more choice is better, right?
If the team at a site has gained experience in a particular language (Go, Ruby, Python), they would naturally levitate to a particular stack or ecosystem which is written in that language (Hashi stack, Chef/Puppet, Ansible/SaltStack).
However, it wasn’t up to the sites.
The middleware products evolved independently, so they each ended up choosing their own preferred configuration management tools.
This means that a site now needs to adopt the stack that the product mandates if they are to deliver that component properly.
If they wanted to stick with their configuration management tool of choice, for whatever reason, it might mean that the product team couldn’t support them in debugging configuration issues, which may have a severe impact on the state of the service in production.
Are we really proposing a world in which all site admins need to know a plethora of configuration management stacks?
That doesn’t sound feasible. It doesn’t have to be that way.
The benefits of many tools
Let’s consider a world with just the two most widely used configuration management tools right now5.
It is difficult to say why Ansible or Puppet have the following they have in certain environments.
There are design and ecosystem considerations which suit different scenarios better in each case.
Both can be used to achieve continuous, correct deployment.
So the question is does it matter to the site admin whether a product expresses a preference for either ?
I hazard that it should not.
UMD deployment should be a conservative force6 :
End states should not depend on the path taken to get there
Indeed, having more than one tool can bring certain benefits to the ecosystem.
Developing configuration and deployment scenarios for a particular middleware component with two different tools can be thought of as a form of Pair Programming.
There are a few potential benefits from pair programming, to quote the Agile Alliance:
increased code quality: “programming out loud” leads to clearer articulation of the complexities and hidden details in coding tasks, reducing the risk of error or going down blind alleys
better diffusion of knowledge among the team, in particular when a developer unfamiliar with a component is pairing with one who knows it much better
better transfer of skills, as junior developers pick up micro-techniques or broader skills from more experienced team members
Pair programming allows better collaboration and quality code review.
If both tools should achieve the same state for a given middleware product, we should be able to review each others’ work.
One may counter that we are using different languages and paradigms – how can we review each others’ work in such a situation?
The answer lies in what is truly important: patterns instead of specific implementations.
Why are there four experiments at the LHC?
Another reason to rejoice that there is more than one tool is the same reason that there is more than one experiment at the LHC7: cross-validation8.
There are always biases and assumptions in deployment and configuration scenarios – these make their way into the code for deployment.
They implicitly exclude certain use cases or scenarios.
Cross-validating deployments with different tools tends to surface these assumptions and force us to confront them.
A good goal would be to achieve consistent deployments from a given state, regardless of the means to achieve it.
A healthy ecosystem
In any ecosystem, a monoculture is a sign that things are heading for collapse.
Reliance on a single tool and tribal knowledge around it is not a good sign, but healthy ‘inter-breeding’ of ideas from slightly different ways of doing things can probably lead to better health of the UMD ecosystem and whatever proceeds it.
Being forced to accept different tools, we can teach patterns and skills rather than tools.
This approach has the double benefit being useful to world beyond our borders in commerce and industry, while simultaneously making it easier for us as a community to attract talent from that world.
In summary: having Ansible in the UMD world can bring significant benefits, as long as it’s done in a way that supports and contributes to that world.
Let’s take a look at how that may come to be.
UMD in the Ansible world
Ansible is a general purpose automation tool – but don’t take it from me, take it from the horse’s mouth9:
Ansible is a radically simple IT automation engine
It is so simple that it can be really tempting to just solve problems quickly and be done with it.
However, this same power leads to massive divergence in the way in which problems are solved, making it difficult to trust that other peoples’ work will work for you.
As an example, consider that you are a site administrator responsible for deploying and maintaining a CREAM CE.
You decide to use Ansible - excellent choice10!
Typically, it’s best to re-use existing code, so you look for an Ansible role for configuring CREAM.
Perhaps you even find one on Galaxy11, but you don’t know anything else about it.
Immediately, some questions come to your mind:
Will it respect my local setup ?
Does it do the network configuration ?
Who maintains this?
Is it even correct?
It is impossible to answer these questions without digging into the code itself and judging for yourself, but we have admitted that we don’t want site administrators to have intimate knowledge of arbitrary configuration management systems.
We need some objective measure of quality which the role can be measured against.
The EGI Style Guide
The first step in developing trusted configurations for UMD with Ansible is therefore to develop a style guide.
I first announced this in a previous blog post and expanded on it in a subsequent on.
The guide helps those re-using roles trust those developing them, by expressing a few opinions on:
Usage of Ansible syntax in roles
Testing role scenarios, testing tools
Role release and publication
Collaborating with code
It builds upon several tools for checking the quality of code, as well as the Ansible guide itself and includes:
A better Ansible role skeleton: When creating new roles, one typically uses ansible-galaxy init <role name>.
The default skeleton has several important bits ‘missing’ which are necessary for engendering re-use and trust.
With ansible-galaxy init --role-skeleton=ansible-style-guide/egi-galaxy-skeleton <role name> you get:
Issue and PR templates, contributing guide, links to EGI support structures
This style guide therefore provides important context and objective measures of quality to the development of Ansible roles intended for use in EGI.
One role, many scenarios
The underlying platform is changing.
We are building sites on top of local clouds, putting services in virtual machines and containers, virtualising networks, building DMZs, etc.
This implies that there are several scenarios in which a particular product can be deployed – scenarios which should be accounted for and tested if possible.
The product development team should not be obliged to consider the specifics of any given scenario, as long as the product provides a means for proper configuration in that scenario.
A collaboration is necessary between the operators at the sites and the product development teams to ensure that the various production environments are considered and can be mocked up during testing.
Testing and Trusting
So, is it feasible to apply traditional Test-Driven Development12 to our infrastructure components?
If it were possible to write a specification of the desired state of a service after deployment, this would allow us to write infrastructure tests which that state would have to pass – and we could then write the code that allows that test to pass.
There are two widely-used toolkits for writing these specifications:
These tests are the basis for a transaction on the basis of trust.
As long as both dev and ops agree on the correctness and validity of the test, they have a common goal to work towards and can meet each other there.
What is needed is an easy tool to create the environment on which the service will then be provisioned and subsequently tested.
Molecule provides a general-purpose mock and testing framework for Ansible roles.
It allows the developer to define many deployment scenarios and test against them.
The easiest is to provision and test in is Docker, but molecule has plugins that can provision OpenStack or bare-metal scenarios, from given starting points.
I covered this in depth in a previous blog post, so I won’t go further into detail on it here.
Better use of existing infrastructure
Expecting individual product managers to achieve this high quality of production by themselves is a big ask.
In fact there is a lot of infrastructure within the EGI ecosystem as well as in the wider world which can make the maintainer and developer’s experience far better.
Issues such as13discoverability, re-usability and trustworthiness matter to operations:
discoverability: can I find what someone else has made?
re-usability: is what someone else has made relevant in my case? Is it immediately usable as-is?
trustworthiness: will the thing that someone else has made do what it says without collateral damage?
Luckily there are pieces of infrastructure created to address these issues, both for the software as well as the artefacts that are created.
Improving use of Ansible Galaxy
In the Puppet world there is Puppet forge; in the Ansible world there is Ansible Galaxy - a place to share the roles that one has created.
The EGI Ansible Style Guide does a bit of extra work to improve usage of this piece of infrastructure:
a good metadata schema is used so that roles posted to Galaxy will pass the initial linting test
that metadata is populated with sane defaults for the EGI environment, such that when searching for EGI things in Galaxy, they will be found
it ensures that roles created with the skeleton it proposes are properly tested
upon completion of the test, Galaxy is notified with a webhook, ensuring that the description and metadata are kept up to date.
Redistributing build artefacts for immediate re-use
Furthermore, the testing environment produces validated artefacts – container or VM images – which can in principle be immediately published to an artefact repository for re-use.
Part of the build process after successful validation is a push of the built container image to EGI’s Quay, or parts of EGI infrastructure such as AppDB.
Raise Vulnerabilities in Staging
Having an executable artefact in a repository before deploying it to a production environment is a very valuable thing.
These images are the thing which would get run at the site and having them in a safe environment before they get to the sites allows one to perform all sorts of other tests on them, apart from verifying their correctness for use.
One of the things that Quay brings is the application of then Clair vulnerability scanner.
It’s not a big leap to expect Secant14 to be applied to these artefacts too.
The result would be a pro-active pipeline which is more aware of vulnerabilities being introduced, rather than a reactive one which has to do extensive monitoring of the products in the wild.
With these new patterns in place, taking better advantage of collaboration, automation and discovery infrastructure, we can support many more deployment scenarios.
These may include:
small sites with few staff, deploying in known scenarios
remote, unmanned deployments
deployment on different platforms hitherto considered too difficult to support.
However, for this to become a reality, we need product teams and infrastructure engineers to collaborate, across several boundaries.
This collaboration has to happen at the level of peer review of contributions, pull requests, infrastructure specifications and documentation – not on the code of the product itself, but the pipeline for delivering that product in a viable state to the production environment.
The germs of such collaboration are already well in place, to varying degrees.
A significant amount of work has been done to bring the usage of Ansible to a EGI in a consistent manner, such that configuration and deployment scenarios developed for UMD by infrastructure engineers may be trusted by site operators.
UMD configurations should be put through the same rigorous testing as UMD products.
Furthermore, having more than one tool to achieve production states is a good thing, as long as there is a community of practice in EGI around those tools.
Such a community of practice is expressed in the EGI Ansible Style Guide, which includes a compliance profile to help developers adhere to it.
These contributions enable those developing infrastructure components to do so smoothly and collaboratively, and enable those operating infrastructure components to do so with confidence.
References and Footnotes
It is difficult to know who will associate with this “we” I refer to here. I intend this to cover the community of site administrators, middleware developers, members of co-ordinating projects and institutes which has contributed to and still uses components of the EGI ecosystem. ↩
Aiftimiei, C. (2013, May 28). EMI YAIM CORE V. 5.1.1. Zenodo. http://doi.org/10.5281/zenodo.6824 ↩
It seems that YAIM itself is not published anywhere. If anyone has the citation for it, with a DOI, please let me know. ↩
I mean that in the physical way - that the work done (i.e. final state) is independent of the path https://en.wikipedia.org/wiki/Conservative_force ↩
For the sake of clarity, I’m referring to the four experiments at the Large Hadron Collider (LHC) - ALICE, ATLAS, CMS and LHCb. The first and third have been explicitly designed to complement each other, whilst the other two have often provided cross-validation information on signals, with different emphasis. ↩
See e.g.: Voss, H. (2013). Classification. In Data Analysis in High Energy Physics (eds O. Behnke, K. Kröninger, G. Schott and T. Schörner‐Sadenius). doi:10.1002/9783527653416.ch5↩