I have realized that the open source projects I am most proud of are a few that have existed for years now, increasing in popularity, with very little maintenance required. Including traject and bento_search. While community aspects matter for open source sustainability, the task gets so much easier when the code requires less effort to keep alive, for maintainers and utilizers. Using these projects as examples, can we as developers identify what makes code “inexpensive” to use and maintain over the long haul with little “churn”, and how to do that?
Rough transcript (really the script I wrote for myself)
Hi, I’m Jonathan Rochkind, and this is “Code that Lasts: Sustainable and Usable Open Source Code”
So, who am I? I have been developing open source library software since 2006, mainly in ruby and Rails.
Over that time, I have participated in a variety open source projects meant to be used by multiple institutions, and I’ve often seen us having challenges with long-term maintenance sustainability and usability of our software. This includes in projects I have been instrumental in creating myself, we’ve all been there!
We’re used to thinking of this problem in terms of needing more maintainers.
But let’s first think more about what the situation looks like, before we assume what causes it. In addition to features or changes people want not getting done, it also can look like, for instance:
Being stuck using out-of-date dependencies like old, even end-of-lifed, versions of Rails or ruby.
A reduction in software “polish” over time.
What do I mean by “polish”?
Engineer Richard Schneeman writes: [quote] “When we say something is “polished” it means that it is free from sharp edges, even the small ones. I view polished software to be ones that are mostly free from frustration. They do what you expect them to and are consistent.”
I have noticed that software can start out very well polished, but over time lose that polish.
This usually goes along with decreasing “cohesion” in software over time, a feeling like that different parts of the software start to no longer tell the developer a consistent story together.
While there can be an element of truth in needing more maintainers in some cases – zero maintainers is obviously too few — there are also ways that increasing the number of committers or maintainers can result in diminishing returns and additional challenges.
One of the theses of Fred Brooks famous 1975 book “The Mythical Man-Month” is sometimes called ”Brooks Law”: “under certain conditions, an incremental person when added to a project makes the project take more, not less time.”
Why? One of the main reasons Brooks discusses is the the additional time taken for communication and coordination between more people – with every person you add, the number of connections between people goes up combinatorily.
That may explain the phenomenon we sometimes see with so-called “Design by committee” where “too many cooks in the kitchen” can produce inconsistency or excessive complexity.
Cohesion and polish require a unified design vision— that’s not incompatible with increasing numbers of maintainers, but it does make it more challenging because it takes more time to get everyone on the same page, and iterate while maintaining a unifying vision. (There’s also more to be said here about the difference between just a bunch of committers committing PR’s, and the maintainers role of maintaining historical context and design vision for how all the parts fit together.)
Instead of assuming adding more committers or maintainers is the solution, can there instead be ways to reduce the amount of maintenance required?
I started thinking about this when I noticed a couple projects of mine which had become more widely successful than I had any right to expect, considering how little maintainance was being put into them.
Bento_search is a toolkit for searching different external search engines in a consistent way. It’s especially but not exclusively for displaying multiple search results in “bento box” style, which is what Tito Sierra from NCSU first called these little side by side search results.
I wrote bento_search for use at a former job in 2012. 55% of all commits to the project were made in 2012. 95% of all commits in 2016 or earlier. (I gave it a bit of attention for a contracting project in 2016).
But bento_search has never gotten a lot of maintenance, I don’t use it anymore myself. It’s not in wide use, but I found it kind of amazing, when I saw people giving me credit in conference presentations for the gem (thanks!), when I didn’t even know they were using it and I hadn’t been paying it any attention at all! It’s still used by a handful of institutions for whom it just works with little attention from maintainers. (The screenshot from Cornell University Libraries)
Traject is a Marc-to-Solr indexing tool written in ruby (or, more generally, can be a general purpose extract-transform-load tool), that I wrote with Bill Dueber from the University of Michigan in 2013.
We hoped it would catch on in the Blacklight community, but for the first couple years, it’s uptake was slow.
However, since then, it has come to be pretty popular in Blacklight and Samvera communities, and a few other library technologist uses. You can see the spikes of commit activity in the graph for a 2.0 release in 2015 and a 3.0 release in 2018 – but for the most part at other times, nobody has really been spending much time on maintaining traject. Every once in a while a community member submits a minor Pull Request, and it’s usually me who reviews it. Me and Bill remain the only maintainers.
And yet traject just keeps plugging along, picking up adoption and working well for adopters.
So, this made me start thinking, based on what I’ve seen in my career, what are some of the things that might make open source projects both low-maintenance and successful in their adoption and ease-of-use for developers?
One thing both of these projects did was take backwards compatibility very seriously.
The first step of step there is following “semantic versioning” a set of rules whose main point is that releases can’t include backwards incompatible changes unless they are a new major version, like going from 1.x to 2.0.
This is important, but it’s not alone enough to minimize backwards incompatible changes that add maintenance burden to the ecosystem. If the real goal is preventing the pain of backwards incompatibility, we also need to limit the number of major version releases, and limit the number and scope of backwards breaking changes in each major release!
The Bento_search gem has only had one major release, it’s never had a 2.0 release, and it’s still backwards compatible to it’s initial release.
Traject is on a 3.X release after 8 years, but the major releases of traject have had extremely few backwards breaking changes, most people could upgrade through major versions changing very little or most often nothing in their projects.
So OK, sure, everyone wants to minimize backwards incompatibility, but that’s easy to say, how do you DO it? Well, it helps to have less code overall, that changes less often overall all – ok, again, great, but how do you do THAT?
Parsimony is a word in general English that means “The quality of economy or frugality in the use of resources.”
In terms of software architecture, it means having as few as possible moving parts inside your code: fewer classes, types, components, entities, whatever: Or most fundamentally, I like to think of it in terms of minimizing the concepts in the mental model a programmer needs to grasp how the code works and what parts do what.
The goal of architecture design is, what is the smallest possible architecture we can create to make [quote] “simple things simple and complex things possible”, as computer scientist Alan Kay described the goal of software design.
We can see this in bento_search has very few internal architectural concepts.
The main thing bento_search does is provide a standard API for querying a search engine and representing results of a search. These are consistent across different searche engines,, with common metadata vocabulary for what results look like. This makes search engines interchangeable to calling code. And then it includes half a dozen or so search engine implementations for services I needed or wanted to evaluate when I wrote it.
This search engine API at the ruby level can be used all by itself even without the next part, the actual “bento style” which is a built-in support for displaying search engine results in a boxes on a page of your choice in a Rails app, way to, writing very little boilerplate code.
Traject has an architecture which basically has just three parts at the top.
There is a reader which sends objects into the pipeline.
There are some indexing rules which are transformation steps from source object to build an output Hash object.
And then a writer which which translates the Hash object to write to some store, such as Solr.
The reader, transformation steps, and writer are all independent and uncaring about each other, and can be mixed and matched.
That’s MOST of traject right there. It seems simple and obvious once you have it, but it can take a lot of work to end up with what’s simple and obvious in retrospect!
When designing code I’m often reminded of the apocryphal quote: “I would have written a shorter letter, but I did not have the time”
And, to be fair, there’s a lot of complexity within that “indexing rules” step in traject, but it’s design was approached the same way. We have use cases about supporting configuration settings in a file or on command line; or about allowing re-usable custom transformation logic – what’s the simplest possible architecture we can come up with to support those cases.
OK, again, that sounds nice, but how do you do it? I don’t have a paint by numbers, but I can say that for both these projects I took some time – a few weeks even – at the beginning to work out these architectures, lots of diagraming, some prototyping I was prepared to throw out, and in some cases “Documentation-driven design” where I wrote some docs for code I hadn’t written yet. For traject it was invaluable to have Bill Dueber at University of Michigan also interested in spending some design time up front, bouncing ideas back and forth with – to actually intentionally go through an architectural design phase before the implementation.
Figuring out a good parsimonious architecture takes domain knowledge: What things your “industry” – other potential institutions — are going to want to do in this area, and specifically what developers are going to want to do with your tool.
We’re maybe used to thinking of “use cases” in terms of end-users, but it can be useful at the architectural design stage, to formalize this in terms of developer use cases. What is a developer going to want to do, how can I come up with a small number of software pieces she can use to assemble together to do those things.
When we said “make simple things simple and complex things possible”, we can say domain analysis and use cases is identifying what things we’re going to put in either or neither of those categories.
The “simple thing” for bento_search , for instance is just “do a simple keyword search in a search engine, and display results, without having the calling code need to know anything about the specifics of that search engine.”
Another way to get a head-start on solid domain knowledge is to start with another tool you have experience with, that you want to create a replacement for. Before Traject, I and other users used a tool written in Java called SolrMarc — I knew how we had used it, and where we had had roadblocks or things that we found harder or more complicated than we’d like, so I knew my goals were to make those things simpler.
We’re used to hearing arguments about avoiding rewrites, but like most things in software engineering, there can be pitfalls on either either extreme.
I was amused to notice, Fred Brooks in the previously mentioned Mythical Man Month makes some arguments in both directions.
Brooks famously warns about a “second-system effect”, the [quote] “tendency of small, elegant, and successful systems to be succeeded by over-engineered, bloated systems, due to inflated expectations and overconfidence” – one reason to be cautious of a rewrite.
But Brooks in the very same book ALSO writes [quote] “In most projects, the first system built is barely usable….Hence plan to throw one away; you will, anyhow.”
It’s up to us figure out when we’re in which case. I personally think an application is more likely to be bitten by the “second-system effect” danger of a rewrite, while a shared re-usable library is more likely to benefit from a rewrite (in part because a reusable library is harder to change in place without disruption!).
We could sum up a lot of different princples as variations of “Keep it small”.
Both traject and bento_search are tools that developers can use to build something. Bento_search just puts search results in a box on a page; the developer is responsible for the page and an overall app.
Yes, this means that you have to be a ruby developer to use it. Does this limit it’s audience? While we might aspire to make tools that even not-really-developers can just use out of the box, my experience has been that our open source attempts at shrinkwrapped “solutions” often end up still needing development expertise to successfully deploy. Keeping our tools simple and small and not trying to supply a complete app can actually leave more time for these developers to focus on meeting local needs, instead of fighting with a complicated frameworks that doesn’t do quite what they need.
It also means we can limit interactions with any external dependencies. Traject was developed for use with a Blacklight project, but traject code does not refer to Blacklight or even Rails at all, which means new releases of Blacklight or Rails can’t possibly break traject.
Bento_search , by doing one thing and not caring about the details of it’s host application, has kept working from Rails 3.2 all the way up to current Rails 6.1 with pretty much no changes needed except to the test suite setup.
Sometimes when people try to have lots of small tools working together, it can turn into a nightmare where you get a pile of cascading software breakages every time one piece changes. Keeping assumptions and couplings down is what lets us avoid this maintenance nightmare.
And another way of keeping it small is don’t be afraid to say “no” to features when you can’t figure out how to fit them in without serious harm to the parsimony of your architecture. Your domain knowledge is what lets you take an educated guess as to what features are core to your audience and need to be accomodated, and which are edge cases and can be fulfilled by extension points, or sometimes not at all.
By extension points we mean we prefer opportunities for developer-users to write their own code which works with your tools, rather than trying to build less commonly needed features in as configurable features.
As an example, Traject does include some built-in logic, but one of it’s extension point use cases is making sure it’s simple to add whatever transformation logic a developer-user wants, and have it look just as “built-in” as what came with traject. And since traject makes it easy to write your own reader or writer, it’s built-in readers and writers don’t need to include every possible feature –we plan for developers writing their own if they need something else.
Looking at bento_search, it makes it easy to write your own search engine_adapter — that will be useable interchangeably with the built-in ones. Also, bento_search provides a standard way to add custom search arguments specific to a particular adapter – these won’t be directly interchangeable with other adapters, but they are provided for in the architecture, and won’t break in future bento_search releases – it’s another form of extension point.
These extension points are the second half of “simple things simple, complex things possible.” – the complex things possible. Planning for them is part of understanding your developer use-cases, and designing an architecture that can easily handle them. Ideally, it takes no extra layers of abstraction to handle them, you are using the exact architectural join points the out-of-the-box code is using, just supplying custom components.
So here’s an example of how these things worked out in practice with traject, pretty well I think.
Stanford ended up writing a package of extensions to traject called TrajectPlus, to take care of some features they needed that traject didn’t provide. Commit history suggests it was written in 2017, which was Traject 2.0 days.
I can’t recall, but I’d guess they approached me with change requests to traject at that time and I put them off because I couldn’t figure out how to fit them in parsimoniously, or didn’t have time to figure it out.
But the fact that they were *able* to extend traject in this way I consider a validation of traject’s architecture, that they could make it do what they needed, without much coordination with me, and use it in many projects (I think beyond just Stanford).
Much of the 3.0 release of traject was “back-port”ing some features that TrajectPlus had implemented, including out-of-the-box support for XML sources. But I didn’t always do them with the same implementation or API as TrajectPlus – this is another example of being able to use a second go at it to figure out how to do something even more parsimoniously, sometimes figuring out small changes to traject’s architecture to support flexibility in the right dimensions.
When Traject 3.0 came out – the TrajectPlus users didn’t necessarily want to retrofit all their code to the new traject way of doing it. But TrajectPlus could still be used with traject 3.0 with few or possibly no changes, doing things the old way, they weren’t forced to upgrade to the new way. This is a huge win for traject’s backwards compat – everyone was able to do what they needed to do, even taking separate paths, with relatively minimized maintenance work.
As I think about these things philosophically, one of my takeaways is that software engineering is still a craft – and software design is serious thing to be studied and engaged in. Especially for shared libraries rather than local apps, it’s not always to be dismissed as so-called “bike-shedding”.
It’s worth it to take time to think about design, self-reflectively and with your peers, instead of just rushing to put our fires or deliver features, it will reduce maintenance costs and increase values over the long-term.
And I want to just briefly plug “kithe”, a project of mine which tries to be guided by these design goals to create a small focused toolkit for building Digital Collections applications in Rails.
I could easily talk about all of this this another twenty minutes, but that’s our time! I’m always happy to talk more, find me on slack or IRC or email.
This last slide has some sources mentioned in the talk. Thanks for your time!