But too many are eager to jump into distributed computing without understanding what they are bring into their development workflow and debugging scenarios.
Absolutely. More people need to understand the binding hierarchy:
- "early binding": function A calls function B. A specific implementation is selected at compile+link time.
- "late binding": function A calls function B. The available implementations are linked in at build time, but the specific implementation is selected at runtime.
From this point on, code can select implementations that did not even exist at the time function A was written:
- early binding + dynamic linking: a specific function name is selected at compile+link time, and the runtime linker picks an implementation in a fairly deterministic manner
- late binding + dynamic linking: an implementation is selected at runtime in an extremely flexible way, but still within the same process
- (D)COM/CORBA: an implementation of an object is found .. somewhere. This may be in a different thread, process, or system. The system provides transparent marshalling.
- microservices: a function call involves marshalling an HTTP request to a piece of software potentially written by a different team and hosted in a datacenter somewhere on a different software lifecycle.
At each stage, your ability to predict and control what happens at the time of making a function call goes down. Beyond in-process functions, your ability to get proper backtraces and breakpoint code is impaired.
You see this under various guises in the web world - "event-driven", "microservices", "dependency injection", "module pattern". It's a very easy thing to see the appeal of, and seems to check a lot of "good architecture" boxes. There are a lot of upsides too - scaling, encapsulation, testability, modular updates.
Unfortunately, it also incurs a very high and non-obvious cost - that it's much more difficult to properly trace events through the system. Reasoning through any of these decoupled patterns frequently takes specialized constructs - additional debugging views, logging, or special instances with known state.
It is for this reason that I argue that lower-hierarchy bindings should be viewed with skepticism - if you _cannot_ manage to solve a problem with tight coupling, then resort to a looser coupling. Introduce a loose coupling when there is measurable downside to maintaining a tighter coupling. Even then, choose the next step down the heirarchy (i.e. a new file, class, or module rather than a new service or pubsub system).
Here, as everywhere, it is a tradeoff about how understandable versus flexible you build a system. I think it is very easy to lean towards flexibility to the detriment of progress.
Microservices put function calls behind a network call. This adds uncertainty to the call - but you could argue this is a good thing.
In the actor model, as implemented in Erlang, actors are implemented almost as isolated processes over a network. You can't accidentally share memory, you can't bind state through a function call - you have to send a message, and await a response.
And yet this model has led to extremely reliable systems, despite being extremely similar to service oriented architecture in many ways.
Why? Because putting things behind a network can, counter intuitively, lead to more resilient systems.
I don't think you would get the same benefits as Erlang unless you either actually write in Erlang or replicate the whole fault tolerant culture and ecosystem that Erlang has created to deal with the fact that it is designed around unreliable networks. And while I haven't worked with multi-node BEAM, I bet single-node is still more reliable than multi-node. Removing a source of errors is still less errors.
If your argument is that we should in fact run everything on BEAM or equivalently powerful platforms, I'm all on board. My current project is on Elixir/Phoenix.
I think the idea is to move the general SaaS industry from the local monolith optimum to the better global distributed optimum that Erlang currently inhabits. Or rather, beyond the Erlang optimum insofar as we want the benefits of the Erlang operation model without restricting ourselves to the Erlang developer/package ecosystem. So yeah, the broader "micro service" culture hasn't yet caught up to Erlang because industry-wide culture changes don't happen over night, especially considering the constraints involved (compatibility with existing software ecosystems). This doesn't mean that the current state of the art of microservices is right for every application or even most applications, but it doesn't mean that they're fundamentally unworkable either.
Erlang without a network and distribution is going to be more resilient than Erlang with a network.
If you're talking about the challenges of distributed computing impacting the design of Erlang, then I agree. Erlang has a wonderful design for certain use cases. I'm not sure Erlang can replace all uses of microservices, however, because from what I understand and recall, Erlang is a fully connected network. The communication overhead of Erlang will be much greater than that of a microservice architecture that has a more deliberate design.
I doubt that this is supposed to be true. Erlang is based on the idea of isolated processes and transactions - it's fundamental to Armstrong's thesis, which means being on a network shouldn't change how your code is built or designed.
Maybe it ends up being true, because you add more actual failures (but not more failure cases). That's fine. In Erlang that's the case. I wouldn't call that resiliency though, the resiliency is the same - uptime, sure, could be lower.
What about in other languages that don't model systems this way? Where mutable state can be shared? Where exceptions can crop up, and you don't know how to roll back state because it's been mutated?
In a system where you have a network boundary, and if you follow Microservice architecture you're given patterns to deal with many others.
It's not a silver bullet. Just splitting code across a network boundary won't magically make it better. But isolating state is a powerful way to improve resiliency if you leverage it properly (, which is what Microservice architecture intends).
You could also use immutable values and all sorts of other things to help get that isolation of state. There's lots of ways to write resilient software.
At minimum, you need to start dealing with things like service discovery and accounting for all of the edge cases where one part of your system is up while another part is down, or how to deal with all of the transitional states without losing work.
> Why? Because putting things behind a network can, counter intuitively, lead to more resilient systems
If you're creating resiliency to a set of problems that you've created by going to a distributed system, it's not necessarily a net win.
I really don't get this phobia. You already have to deal with that everywhere, don't you? I mean, you run a database. You run a web app calling your backend. You run mobile clients calling your backend. You call with services. The distributed system is more often than not already there. Why are we fooling ourselves into believing that just because you choose to bundle everything in a mega-executable that you're not running a distributed system?
If anything,explicitly acknowledging that you already run a distributed system frames the problem ina way that you are forced to face failure modes you opt to ignore.
These are not strictly a "program on network vs program not on network" issue. "Accidentally sharing memory" can be solved by language design. It is correct that a system that is supposedly designed as inter-actor communication is not ideal when designed and written in a "function call-like" manner, but erlang only partially solves this phenomena by enforcing a type of architecture, which locks away the mentioned bad practice.
> ...despite being extremely similar to service oriented architecture in many ways. Why? Because putting things behind a network can, counter intuitively, lead to more resilient systems.
This is a strange logic. Resilient system is usually achieved by putting a service/module/functionality/whatever in a network of replications, meanwhile service oriented architecture talks about loosening the coupling between different computers that acts differently.
I do agree on microservice != turn function calls into distributed computing problems.
It MAY happen to systems eagerly designed with microservice architecture without a proper prior architectural validations, but it is not always the case.
For sure - I am definitely not trying to say that there is "one true approach".
It gets much more interesting when you don’t call functions at all - you post messages. You have no idea which systems are going to handle them or where... and at that point, microservices are freeing.
All this focus on function call binding just seems so... small, compared to what distributed microservice architectures are actually for.
Monoliths aren’t merely monolithic in terms of having a monolithic set of addressable functionality; they are also monolithic in terms of how they access resources, how they take dependencies, how they are built, tested and deployed, how they employ scarce hardware, and how they crash.
Microservices help solve problems that linkers fundamentally struggle with. Things like different parts of code wanting to use different versions of a dependency. Things like different parts of the codebase wanting to use different linking strategies.
Adding in network hops is a cost, true. But monoliths have costs too: resource contention; build and deploy times; version locking
Also, not all monolithic architectures are equal. If you’re talking about a monolithic web app with a big RDBMS behind it, that is likely going to have very different problems than a monolithic job-processing app with a big queue-based backend.
In any case, too many rush for micro-services with the intent reason to use the network as a package boundary.
Have you ever tried to debug spaghetti RPC calls across the network?
I sadly have.
I think we can all agree that straw man architectures don’t work. Everyone should employ the true Scotsman architecture.
What was originally a package, gets its own process and REST endpoint, sorry nowadays it should be gRPC, the network boilerplate gets wrapped in nice function calls, and gets used everywhere just like the original monolith code.
Just like almost no one does REST as it was originality intended, most microservices end up reflecting the monolith with an additional layer of unwanted complexity.
Good programming practices to refactor monoliths never get touched upon, as otherwise the sale would lose its appeal.
It is essentially impossible without the (sadly rare) design pattern that I gave at https://news.ycombinator.com/item?id=26016854.
If I am ever unfortunate enough to work on a microservices architecture again, I'll see if I can get it used.
The other advantage of the network boundary is that you can use different languages / technologies for each of your modules / services.
Don't get me wrong, sometimes it's worth it (I particularly like Spark's facilities for distributed statistical modelling), but I really don't get (and have never gotten) why you would want to inflict that pain upon yourself if you don't have to.
I’ve been developing for more years than dime if you have lived, and the best thing I’ve heard in years was that Google interviews were requiring developers to understand the overhead of requests.
In addition, they should require understanding of design complexity of asynchronous queues, needing and suffering from management overhead of dead letter, scaling by sharding queues if it makes more sense vs decentralizing and having to have non-transactionality unless it’s absolutely needed.
But not just Google- everyone. Thanks, Mr. Fowler for bringing this into the open.
Adding network is not a limitation. And frankly, I don't understand why you say things like understanding network. Like reliability is taken care of, routing is taken care of. The remaining problems of unboundedness and causal ordering are taken care of (by various frameworks and protocols).
For dlq management, you can simply use a persistent dead letter queue. I mean it's a good thing to have dlq because failures will always happen. About which order to procese queue etc. These are trivial questions.
You say things as if you have been doing software development for ages, but you're missing out on some very simple things.
Each of them introduces a new, rare failure mode; because there are now many rare failure modes, you have frequent-yet-inexplicable errors.
Dead letter queues back up and run out of space; network splits still happen;
And secondly, if you do end up with q distributed systems, remember how many independently failing components there are because thag directly translates to complexity.
On both these counts I agree. Microservices is no silver bullet. Network partitions and failure happen almost every day where I work. But most people are not dealing with that level of problems, partly because of cloud providers.
Same kind of problems will be found on a single machine also. Like you'd need some sort of write ahead log, checkpointing, maybe optimize your kernel for faster boot up, heap size and gc rate.
All of these problems do happen, but most people don't need to think about it.
You'll also likely use multiple databases (caching in e. g. Redis) and a job queue for longer tasks.
You'll also probably already have multiple instances talking to the databases, as well as multiple workers processing jobs.
Pretending that the monolith is a single thing is sneakily misleading. It's already a distributed system
Making a system to be reliable is really really hard and take many resources, which seldom companies pursuit.
Requests can fail in a host of ways that a call simply cannot, the complexity is massively greater than a method call.
I realized this one day when I was drawing some nice sequence diagrams and presenting it to a senior and he said "But who's ensuring the sequence?". You'll never ask this question in a single threaded system.
Having said that, these things are unavoidable. The expectations from a system are too great to not have distributed systems in picture.
Monoliths are so hard to deploy. It's even more problematic when you have code optimized for both sync cpu intensive stuff and async io in the same service. Figuring out the optimal fleet size is also harder.
I'd love to hear some ways to address this issue and also not to have microservice bloat.
Getting engineers who don't intuitively understand or maybe even care how to avoid coupling in monoliths to work on a distributed application can result in all the same class of problems plus chains of network calls and all the extra overhead you should be avoiding.
It seems like you tell people to respect the boundaries, and if that fails you can make the wall difficult to climb. The group of people that respect the boundaries whether virtual or not, will continue to respect the boundaries. The others will spend huge amounts of effort and energy getting really good at finding gaps in the wall and/or really good at climbing.
If you take a look at dependency management at open source software, you'll see a mostly unified procedure, that scales to an "entirety of mankind" sized team without working too badly for single developers, so it can handle your team size too.
That problem that the "microservices bring better architectures" people are trying to solve isn't open by any measure. It was patently solved, decades ago, with stuff that work much better than microservices, in a way that is known to a large chunk of the developers and openly published all over the internet for anybody that wants to read about.
Microservices still have their use. It's just that "it makes people write better code" isn't true.
I've often wondered if this is a pattern sitting underneath our noses. I.e., Starting with a monolith with strong boundaries, and giving architects/developers a way to more gracefully break apart the monolith. Today it feels very manual, but it doesn't need to be.
What if we had frameworks that more gracefully scaled from monoliths to distributed systems? If we baked something like GRPC into the system from the beginning, we could more gracefully break the monolith apart. And the "seams" would be more apparent inside the monolith because the GRPC-style calls would be explicit.
(Please don't get too hung up on GRPC, I'm thinking it could be any number of methods; it's more about the pattern than the tooling).
The advantages to this style would be:
* Seeing the explicit boundaries, or potential boundaries, sooner.
* Faster refactoring: it's MUCH easier to refactor a monolith than refactor a distributed architecture.
* Simulating network overhead. For production, the intra-boundary calls would just feel like function calls, but in a develop or testing environment, could you simulate network conditions: lag, failures, etc.
I'm wondering if anything like this exists today?
If you've only got a basic IPC system (say, Unix domain sockets), then you could stream a standard seriaization format across them (MessagePack, Protobuf, etc.).
To your idea of gracefully moving to network-distributed system: If nothing else, couldn't you just actually start with gRPC and connect to localhost?
Is there something I'm missing?
When you start with gRPC and connect to localhost, usually the worst that can happen with a RPC call is that the process crashes, and your RPC call eventually times out.
But other than that everything else seems to work as a local function call.
Now when you move the server into another computer, maybe it didn't crash, it was just a network hiccup and now you are getting a message back that the calling process is no longer waiting, or you do two asynchronous calls, but due to the network latency and packet distribution, they get processed out of order.
Or eventually one server is not enough for the load, and you decide to add another one, so you get some kind of load mechanism in place, but also need to take care for unprocessed messages that one of the nodes took responsibility over, and so forth.
There is a reason why there are so many CS books and papers on distributed systems.
Using them as mitigation for teams that don't understand how to write modular code, only escalates the problem, you move from spaghetti calls in process, to spaghetti RPC calls and having to handle network failures in the process.
(I side with the monolith, FWIW...I love Carl Hewitt's work and all, it just brings in a whole set of stuff a single actor doesn't need... I loved the comment on binding and RPC above, also the one in which an RPC call's failure modes were compared to the (smaller profile) method call's)
Most languages provide a way to separate the declaration of an interface from its implementation. And a common language makes it much easier to ensure that changes that break that interface are caught at compile time.
Let's say we have 3 different modules, all domains: sales, work, and materials. A customer places an order, someone on the factory floor needs to process it, and they need materials to do it. Materials know what work they are for, and work knows what order it's for (there's probably a better way to do this. This is just an example).
On the frontend, users want to see all the materials for a specific order. You could have a single query in the materials module that joins tables across domains. Is that ok? I guess in this instance the materials module wouldn't be importing from other modules. It does have to know about sales though.
Here's another contrived example. We have a certain material and want to know all the orders that used this material. Since we want orders, it makes sense to me to add this in the sales module. Again, you can perform joins to get the answer, and again this doesn't necessarily involve importing from other modules. Conceptually, though, it just doesn't feel right.
In your examples you need to add extra layers, just like you would do with the microservices.
There would be the DTOs that represent the actual data that gets across the models, the view models that package the data together as it makes sense for the views, the repository module that actually abstracts if the data is accessed via SQL, ORM, RPC or whatever.
You should look into something like:
"Domain-Driven Design: Tackling Complexity in the Heart of Software"
https://www.amazon.com/Eric-J-Evans/dp/0321125215
"Component Software: Beyond Object-Oriented Programming"
https://www.amazon.com/Component-Software-Beyond-Object-Orie...
Joining on the database layer is still adding a dependency between domains. The data models still need to come out of one domain. Dependencies add complexity. So joining is just like importing a module, but worse because it's hidden from the application.
If you really need a join or transaction, you need to think as if you had microservices. You'd need to denormalize data from one domain into another. Then the receiving domain can do whatever it wants.
Of course, you can always break these boundaries and add dependencies. But you end up with the complexity that comes with in, in the long run.
If I understand your example, the usual solution is to separate your business objects from your business logic, and add a data access layer between them.
In terms of dependencies, you would want the data access layer module to depend on the business object modules. And your business logic modules would depend on both the data access layer and business object modules. You may find that it is ok to group business objects from multiple domains into a single module.
Note that this somewhat mirrors the structure you might expect to see in a microservices architecture.
As an engineer, the thought process goes: I can use the same old tried and true patterns that will just get the job done. That would be safe and comfortable, but it won't add anything to my skillset/resume.
Or we could try out this sexy new tech that the internet is buzzing about, it will make my job more interesting, and better position me to move onto my next job. Or at least give me more options.
It's essentially the principal-agent problem. And by the way, I don't blame developers for taking this position.
I suppose if you work in an organization where everyone (including management) is very disciplined and respects the high level architecture then this isn't much of a benefit, but I've never had the pleasure of working in such an organization.
That said, I hear all the time about people making a mess with micro services, so I'm sure there's another side, and I haven't managed to figure out yet why these other experiences don't match my own. I've mostly seen micro service architecture as an improvement to my experiences with monoliths (in particular, it seems like it's really hard to do many small, frequent releases with monoliths because the releases inevitably involving collaborating across every team). Maybe I've just never been part of an organization that has done monoliths well.
This fallacy is the distributed computing bogeyman.
Just because you peel off a responsibility out of a monolith that does not mean you suddenly have a complex mess. This is a false premise. Think about it: one of the first things to be peeled off a monolith are expensive fire-and-forget background tasks, which more often than not are already idempotent.
Once these expensive background tasks are peeled off, you gets far more manageable and sane system which is far easier to reason about and develop and maintain and run.
Hell, one of the basic principles of microservices is that you should peel off responsibilities that are totally isolated and independent. Why should you be more concerned about the possibility of 10% of your system being down if the alternative is having 100% of your system down? More often than not you don't even bat an eye if you get your frontend calling your backend and half a dozen third-party services. Why should it bother you if your front-end calls two of your own services instead of just one?
I get the concerns about microservicea, but this irrational monolith-mania has no rational basis either.
One of the patterns that I have noted in the past 3 decades in the field is that operational complexity is far more accessible than conceptual complexity to the practitioners. Microservices shift complexity from conceptual to operational. With some rare exceptions, most MS designs I've seen were proposed by teams that were incapable of effective conceptual modeling of their domain.
That is the question being discussed.
Most of microservice implementation is within a single team. That is the real question being discussed.
I think it's not, but I would like to hear other opinions.
I guess the context is only implelied in your parent post.
The rise of the network API in the last 20 years has proven it's own benefits. Whether you are calling a monolith of a microservice, it's easier to upgrade the logic without recompiling and re-linking all dependencies.
For example, microservices tend to communicate via strings of bytes (e.g. containing HTTP requests with JSON payloads, or whatever). We could do a similar thing with `void` in C, or `byte[]` in Java, etc.
Languages which support 'separate compilation' only need to recompile modules which have changed; if all of our modules communicate using `void` then only the module containing our change will be recompiled.
It's also easy to share a `void` between modules written in different languages, e.g. using a memory-mapped file, a foreign function interface, etc.
It's also relatively* easy to hook such systems into a network; it introduces headaches regarding disconnection, packet loss, etc. but those are all standard problems with anything networked. The actual logic would work as-is (since all of the required parsing, validation, serialisation, etc. is already there, for shoehorning our data in and out of `void*`).
If these benefits were so clear, I would expect to see compiled applications moving in this direction. Yet instead I see the opposite: stronger, more elaborate contracts between modules (e.g. generic/parametric types, algebraic data types, recursive types, higher-kinded types, existential types, borrow checkers, linear types, dependent types, etc.)