Re:Death by a Thousand Microservices
by hliyan on Tuesday September 12, 2023 @03:07AM
(#37477691)
I wrote this 8 years ago: Microservices vs. "air-gapped" modules https://www.linkedin.com/pulse/maintainable-software-archite.... You can achieve the same end as a microservice using a module, by simply having lint rules that prevent you from importing other application level modules. This way the only possible comms interface is to pass events with a payload of primitive typed parameters.
[ Parent
| on HN ]
An interesting idea. Sort of a good half way. But one issue I see is that you can still have a shared data store. That means you can accidentally (or intentionally) use the database to pass back-channel messages —- have one module store data and another read it.
That could lead to hard to find bugs. Did this ever come up for you?
Re:Death by a Thousand Microservices
by foobarian on Tuesday September 12, 2023 @09:04AM
(#37480536)
I think systems like this are prone to buckling under product pressure. If it takes a few minutes to override a lint rule it's bound to happen sooner or later. With a microservice system the amount of effort is genuinely higher to add cross-dependencies.
The best "microservice" based system I've ever worked on was full ten years before the term had been invented. It had multiple services that read and wrote to the same database, but each used a common dynamically linked library that was the common interface to access that database (no naked SQL was allowed -- you had to #include <foo> and then do Foo* foo = Foo::findById("xxx"); etc. Worked remarkably well. The above common library was very carefully maintained and any addition had to be reviewed and approved. The approach seemed to work well.
That's the services model that Facebook uses and parts of Amazon use. A single data store surrounded by a bunch of small services with a prescribed data access layer.
That's really more of a monolith though. One of the key tenants of microservices is data separation -- each service has its own data store.
I think if we also "air-gap" (loosely) the dependencies, we get a typical monorepo in, say, JS or Golang? That is, a module in a monorepo is a special case of your airgapped modules?
Not exactly sure what you mean, but in my model, the modules all get built into a single runtime artifact/executable. In the case of monorepos as I understand them, the "mono"ness is in the version control system, but each code base builds to its own artifact or executable, which have to communicate with each other over a network interface.