The ultimate making app for a shipping multi-service system is actually a one-machine monolith with a UI.
If your team is experiencing the most common pains from working in a large SOA environment, the productivity payback will be enormous.
It’s important for me to take a second to remind you that there’s much more to this world than geekery. Please keep working for change all around you, including, especially, outside the monitor.
Stay safe, stay strong, stay angry, stay kind.
Black Lives Matter.
We’ve talked a lot about the idea of having a shipping app for our customers and a making app for us. We can use the same source base to make multiple binaries. We target customer needs with one of those, and we target developer needs with the rest.
The economics of this approach are straightforward: As long as a making app’s benefit — improved productivity on the shipping app — is less than its cost — the time spent working on something we don’t ship — we get a net gain in productivity.
"A one-machine monolith with a UI" is a lot to unpack.
Let’s talk about what it is, what production costs it mitigates, and, especially, how we can approach it in a stepwise change-harvesting fashion in real life.
A "one-machine monolith with a UI" is, on the outside, just a desktop app, same as your IDE, your word processor, and so on. And on the inside? Instead of connecting to the fifty remote services and databases of your shipping app, it contains them.
Now, be clear, we’re not talking about simulating those services. We’re talking about literally embedding their code inside the single binary that is the making app.
This sounds harder to do than it is, we’ll get to that, but before we go there, why might we want it, what is the possible benefit of a beast like this? Much of what we’d get is best expressed as "negative cost", it’s the waste being eliminated from our current approach.
Said another way, using the shipping environment for our myriad development purposes creates a bad fit between hand and tool. That bad fit costs us. The desktop-monolith fits much better, so it costs us much less.
Negative cost: we sidestep provisioning costs for a large swath of our work.
Fifty services & databases is usually fifty cloud machines. That means buying virtual hardware, and it also means having deployment and ops specialists. It means time.
If there’s one consistent feature of every enterprise SOA org I’ve worked with, it’s the ferocious commitment to having the fewest possible environments provisioned.
I don’t know the numbers, but they must be impressive, cuz I’ve seen teams struggle for months getting permission out of white-knuckled management. I swear they’d sooner buy every developer a new car than provision a new environment just for dev.
Negative cost: Nothing I can do as a developer in my one-box monolith can injure any other team’s work.
I can’t bring down the environment. I can’t spew garbage into the databases. I can’t fill the logs with an infinite loop I let slip into the code.
Correspondingly, no other team’s foolishness can keep me from working on the system. Those clowns who run the calendar service can’t shut me down cuz they changed the URL. A gal with a backhoe can’t keep me from programming by severing the backbone.
Positive benefit: the very most difficult outage causes — implicit state and flow connections between services — are easily found and tested for in a one-box making app.
Services constantly make decisions based on state fields in data. Consider adding a new role for one of your B2B flows. Add a new database record, yeah? Add the record, nobody’s using it, now add your logic based around it. Except, odds are good someone is using it.
Services say things like if(role is X), but they also say things like if(role is not X). At the time they were written, they had closure over the available roles. You just added a role, tho, and you broke that closure with unpredictable results.
The only way to find that is to find it by meticulously probing complex UI scenarios across multiple services in a scarce resource, a provisioned environment.
Unless. Unless all the services are right here. Unless you don’t have to use the UI to test a scenario.
Positive benefit: This thing is fast.
You’re eliminating latency, outages, SSO, VPN, passwords, you’re eliminating every common cloud-ish tax that you normally pay when you work in the cloud. The UI has to be functional, it doesn’t have to be pretty or branded.
Careful: a one-box monolith couldn’t support your actual customer base.
(Or, at least, I hope you dint use kubernetes to support a single user with a 100 transactions an hour.) But it doesn’t have to. It only has to support the developer at the box.
People tend to associate unresponsiveness of tools merely with the simple cost in time. In fact, unresponsive tools carry two much heavier costs: 1) multi-tasking, and 2) batching.
Nearly everyone thinks they’re good at multi-tasking. When the tool is slow, they start another task while it runs. Here’s the thing, tho: nearly everyone is actually bad at multi-tasking. Encouraging people to multi-task is begging them to lose focus & forget details.
And when a tool is less responsive, I use it less often. Instead, I use it in batches. And these batches inevitably violate the limits of human mental bandwidth. Changes A, B, C, and D, all evaluated at once, must be evaluated in a far larger context than when they were made.
So, you see, at least in theory, the potential merit of this approach is quite high.
This makes us turn now to practice, and specifically, the part of the practice that is "how would we get there and don’t tell me it’ll take 3 years and 30 million dollars?"
1) The crudest possible form of this, a first pass: provision a single cloud machine with a lot of disk and memory, and put every service binary and database on it.
Essentially, the cost is that one cloud machine and a whole lot of YAML-jiggling.
If you’re dockering, you can even skip the cloud machine. You might have to beef up dev hardware, but that’s chump-change compared to provisioning. Just run the vm on your own box. It’ll be slow, but not as slow as hitting a whole proviioned environment.
2) Now roll a custom dev UI on your desktop that connects to that monster you just made.
We already talked about the kind of things you can do with such a UI, so I won’t say more about that here.
3) Now take one service that’s running in your monster and embed it in your new UI app.
You can easily write that UI app so that it can be a UI for you and a service endpoint. I know you haven’t done that, but I have, and in most modern frameworks, it’s easy.
So your UI is hitting your VM for all the other services, but it’s hitting itself for that one service. This will improve the UI’s performance, but it will also enable you to make rapid changes in that one service without bouncing a whole docker instance.
4) Now, instead of having the UI-embedded service running as an endpoint, just call it’s logic directly.
It’s embedded now, right, it’s source code is the UI’s source code. No need for a transport layer, you’re already there.
This will have required you to separate transport from business logic, of course. (In some environments, a controller is already directly callable by a method. In some environments, not.) But there are compelling reasons to do this even without wanting a separate making app.
5) At this point, you can also make the UI start doing wicked things to the dataset that the embedded service is working with.
This opens a huge range of testing capability, and dramatically increases the safety of making and studying the results of change.
Wanna wipe your embedded service’s database? Bam. Wanna inject a "golden master DB" as the start? Bam. You’re all in the same code base, you’ve eliminated the protective layer of the transport, you can do any of these things.
6) Pick another service, and embed that service in your making-UI app, too.
Same stuff: the second one will go much faster than the first, cuz you’ll have seen most of the mistakes you can make by then.
The only stopping points are when a candidate service is written in a language or a service infrastructure that is incompatible with the service you’re working on. That is a real problem, of course, and it does happen. But in enterprise SOA environments, it’s far less common.
This is all a very high-level conversation. There are lots of details, and lots of variants, depending on what you’re working with now.
But in real life, it comes down to a handful of insights.
The shipping app is fit to the customer’s needs, not the developer’s needs. Developing the shipping app using only the shipping app costs us directly and indirectly, in both lump-sum and tax-like ways.
We can develop the shipping app using a making app that fits a developer’s needs far more closely. We can do it for less money than the costs we pay for trying to do it using the shipping app.
Our codebase is source not binary, and we can use one source to make many different binary images, including, in particular, a shipping app and any number of handing making apps.
The heart of our app is not in fact http transport, but the business logic it implements. Transport mechanisms are stable, cheap, not written by us, and not a major source of our defects, tho using them in development is a major source of our cost.
The code works for us, we don’t work for the code.
Truly, the sky is the limit. You can do absolutely ingenious things, and make astonishing productivity leaps, just by thinking about how to dual-purpose your source code, arranging it in one way to suit your customer’s needs, and another way to suit yours, both at the same time.
How to Participate
If you love the GeePaw Podcast, consider a monthly donation to help keep the content flowing. You can also subscribe to get weekly posts sent straight to your inbox. And to get more involved in the conversation, jump into the Camerata and start talking to other like-minded Change-Harvesters today.