Very comprehensive and clarifying. Coming from OOP, it helps me to understand sagas like this: - Decentralized Saga == Observer Pattern - Centralized Saga == Mediator Pattern
Very comprehensive presentation on Messaging and architecting Microservices. Certainly answered some of my questions and verifies certain decisions I have made
This is a great talk and really useful for building things at scale. Now before everyone starts building systems like this, it's probably worth defining how far an ACID / SQL database can get someone before they need to implement this shiny microservice architecture.
In terms of monolithic applications you need to weigh the pros and cons. - If scalability isn't going to be an issue, like you have few transactions, then the added complexity of microservices may not be worth it. - If you're developing a simple application. Microservices are needlessly complex. - If you want to launch quickly. Its better to get your proof of concept out there and validate your business idea (and make money on it). You can always iterate to microservices later. A good way to do this is with the API gateway Pattern which sits in between the consumer and your services. Which in a nutshell is where you route requests to the new microservice as you add them and route everything else to your old monolith
This was really a good presentation. Next, I've got to find more on failure scenarios in Saga & on creating, handling compensation processes and events.
Wonderful session ,gained some great insights about Transaction management in microservice world. But one question I have is like in CQRS for Get/View purposes Chris mentioned we will have one more no sql schema like mongo something with same data. This will include whole lot of complexity in maintaining/deployment/cost perceptive right. Any other alternatives or suggestion to solve those complextities?
Mh, I do not understand why a workflow engine is a wrong approach. I think, Chris is wrong here. First, Saga is definitely complicated! It is not easy. Count the required concepts and the corresponding implementation parts that in addition needs to be correct: transactional outbox, unique message id, de-duplicating of incoming messages/calls, compensation actions, reverse order of compensation invocations, retries of remote calls, dead letter queue, resume execution, visualization of a saga execution for debugging. Second, take the workflow engine as implementation detail und hide it behind the Saga Dsl of, e.g., the eventuate tram library. Then, the workflow engine is not too much for such a 'small' problem. Instead, it is an ideal solution. It supports all of the concepts from above and requires minimal implementation and configuration effort from developers.
Microservices come with soooo many complications but I have a hard time imagining going back to having everything under a single monolithic system, where I can easily fall into the temptation of coupling So it's a bit sad that the reason to stick with microservices has nothing to do with something the pattern actually offers but with how it forces us to save us from ourselves
If you think that microservices are not worth it, it means they are not worth it in your case :) and 95% of companies and products, don't need microservices.
You can force decoupling by putting code in separate repository as separate library/package. Also world is not divided into hot-dogs and not-hot-dogs (microservices and monoliths) - there is whole spectrum of modularity, dependency control, services width, runtime control (ie. supervisor/actor trees) etc. etc. As a side note please don't forget to prefix the whole complexity of this presentation with prelude "You can just do single line sql query or... ".
Hmm... The argument around 17:00->18:00. One has to ask: How can the order service fail? If it is unreachable of course... but if it just has to acknowledge that req. and add it to it's own internal pending state, ... and then when complete, do an async callback to the service making the request. Then the req. to the order service is basically just a handoff of responsibility. Injecting a message broker is the same ... and it has the exactly the same problem: What if the message broker is unreachable.?? I not sure I see the argument for a message broker here. PS: And the model presented at 27:38 could just as well have been done with sync REST invocations. The customer service simply put "credit reserved" to the order resource.
This is a great talk except for the bit "I know we all want to use Kafka" - nope. We definitely don't. Seriously though, this talk helped me solidify my vocabulary around this complex topic.
I really don't get the obsession with avoiding synchronous communications, when this whole time a customer is sitting at a loading screen, or waiting for a confirmation email. I don't see the same advice given when talking about sending requests to a DB over a network, or to an event queue over a network, or to third party APIs over a network, etc. This is just something consultants say, knowing it will take a ton of billing time to implement.
example on API Composition Pattern @31:06 makes little sense: why API Gateway would call Customer Service if customerId is already known as input param in findOrdersForCustomers(customerId)?
in case it needs some data from the customer service, like - maybe the user-preferences determine things about which orders should be hidden (e.g. maybe the customer chose to not display orders that are of the category "18+" (this is an example that came to mind but surely there are better ones...), or what currency they should be displayed in
FYI: eventuate.io is for Java and the talk is great! dasync.io solves same problems for .NET in a language-oriented way instead of providing a framework.
at first, I thought he is Vision(Paul Bettany), why the color of his face is so red :) I got it. it's a reflection issue of back scene light which it didn't rendered properly by OpenGL fragment shader
I came up with a solution where i use websockets to update UI when projections are updated. But it requires some "gost" styling to give a user an indication that particular entity is still in update process.
I didn't get this 42:16. Does he saying that the event store is responsible for publishing the events into the message broker (kafka) once the event is stored?
You are going to Dubai Desert Safari and while in the middle of desert, u r seeking water bottle. - creating a order then reserving the money is same Case. go with buttle at first place. Sir
It's actually "Semantic Lock". For example, Chris posted an article here -> www.er.crichardson.com/post/antipatterns/2019/07/09/developing-sagas-part-1.html
Data warehouse loading still has latency in 2021 under normal conditions and unless you specifically construct something that *streams* a replication of your transnational database into your analytical database. So, for sufficiently complex data or for big enough data --- it's a trade-off between fast queries that show older state, or slow queries that show current state in any case CQRS (in his presentation) demonstrates how to solve the transnational READS in a natural application flow that is split into micro-services, NOT JUST analytical aggregations over masses of data
I disagree fundamentally to some of Chris's claims. He rejects what IMO actually should be done, then elaborates on stupid solutions with all their problems ... Chris seems lost in the djungle.
Very comprehensive and clarifying.
Coming from OOP, it helps me to understand sagas like this:
- Decentralized Saga == Observer Pattern
- Centralized Saga == Mediator Pattern
One of the best presentations about distributed data, finally a very comprehensive CQRS explanation
Great presentation! I've gone back and reviewed it a couple of times and learned new things each time I view it.
Excellent talk. He knows how to talk in a way people understand unlike some techies who just try to show off..
Chris, you are a great speaker, Awesome video with wonderful explanations, brings loads of clarity on mistakes that we have been doing.
Such a great talk. Thank you Chris Richardson, I learned alot.
Very comprehensive presentation on Messaging and architecting Microservices. Certainly answered some of my questions and verifies certain decisions I have made
Excellent presentation and detailing about each concept. I love the most as one talk covers mostly design and technical design.
This is a great talk and really useful for building things at scale. Now before everyone starts building systems like this, it's probably worth defining how far an ACID / SQL database can get someone before they need to implement this shiny microservice architecture.
You can still use ACID / SQL databases in a microservice architecture, they just shouldn't span into different services.
In terms of monolithic applications you need to weigh the pros and cons.
- If scalability isn't going to be an issue, like you have few transactions, then the added complexity of microservices may not be worth it.
- If you're developing a simple application. Microservices are needlessly complex.
- If you want to launch quickly. Its better to get your proof of concept out there and validate your business idea (and make money on it).
You can always iterate to microservices later. A good way to do this is with the API gateway Pattern which sits in between the consumer and your services. Which in a nutshell is where you route requests to the new microservice as you add them and route everything else to your old monolith
Great presentation, it helped me a great deal. I think I'm beginning to understand why Temporal is such an amazing framework, too.
This was really a good presentation. Next, I've got to find more on failure scenarios in Saga & on creating, handling compensation processes and events.
This is great talk! I think we used have DTO or VM but haven't relate it to microservices and CQRS which is a great use!
Excellent talk!!
Great talk, very impressive, thanks
Wonderful session ,gained some great insights about Transaction management in microservice world. But one question I have is like in CQRS for Get/View purposes Chris mentioned we will have one more no sql schema like mongo something with same data. This will include whole lot of complexity in maintaining/deployment/cost perceptive right. Any other alternatives or suggestion to solve those complextities?
Mh, I do not understand why a workflow engine is a wrong approach. I think, Chris is wrong here. First, Saga is definitely complicated! It is not easy. Count the required concepts and the corresponding implementation parts that in addition needs to be correct: transactional outbox, unique message id, de-duplicating of incoming messages/calls, compensation actions, reverse order of compensation invocations, retries of remote calls, dead letter queue, resume execution, visualization of a saga execution for debugging.
Second, take the workflow engine as implementation detail und hide it behind the Saga Dsl of, e.g., the eventuate tram library. Then, the workflow engine is not too much for such a 'small' problem. Instead, it is an ideal solution. It supports all of the concepts from above and requires minimal implementation and configuration effort from developers.
Zeebe/Camunda for transactions., CouchDB for read model and db for mscs.
Awesome talk, very useful
Microservices come with soooo many complications but I have a hard time imagining going back to having everything under a single monolithic system, where I can easily fall into the temptation of coupling
So it's a bit sad that the reason to stick with microservices has nothing to do with something the pattern actually offers but with how it forces us to save us from ourselves
If you think that microservices are not worth it, it means they are not worth it in your case :) and 95% of companies and products, don't need microservices.
You can force decoupling by putting code in separate repository as separate library/package.
Also world is not divided into hot-dogs and not-hot-dogs (microservices and monoliths) - there is whole spectrum of modularity, dependency control, services width, runtime control (ie. supervisor/actor trees) etc. etc.
As a side note please don't forget to prefix the whole complexity of this presentation with prelude "You can just do single line sql query or... ".
When he was talking about semantic locking he said he had another talk that goes into it in detail 14:30. Where is that video?
I learned a lot. Great talk!!
Hmm... The argument around 17:00->18:00.
One has to ask: How can the order service fail?
If it is unreachable of course... but if it just has to acknowledge that req. and add it to it's own internal pending state, ... and then when complete, do an async callback to the service making the request. Then the req. to the order service is basically just a handoff of responsibility.
Injecting a message broker is the same ... and it has the exactly the same problem: What if the message broker is unreachable.??
I not sure I see the argument for a message broker here.
PS: And the model presented at 27:38 could just as well have been done with sync REST invocations. The customer service simply put "credit reserved" to the order resource.
A Postgres mention. w00t!
This is a great talk except for the bit "I know we all want to use Kafka" - nope. We definitely don't. Seriously though, this talk helped me solidify my vocabulary around this complex topic.
Excellent Session
I really don't get the obsession with avoiding synchronous communications, when this whole time a customer is sitting at a loading screen, or waiting for a confirmation email. I don't see the same advice given when talking about sending requests to a DB over a network, or to an event queue over a network, or to third party APIs over a network, etc. This is just something consultants say, knowing it will take a ton of billing time to implement.
Oh nice, I've just found out I've been doing the Transactional Outbox pattern. :P
example on API Composition Pattern @31:06 makes little sense: why API Gateway would call Customer Service if customerId is already known as input param in findOrdersForCustomers(customerId)?
in case it needs some data from the customer service, like - maybe the user-preferences determine things about which orders should be hidden (e.g. maybe the customer chose to not display orders that are of the category "18+" (this is an example that came to mind but surely there are better ones...), or what currency they should be displayed in
These problems are easily solved with DDD: well-defined Bounded Contexts (customers shouldn't have knowledge of orders) and Aggregates (sagas)
Super good
FYI: eventuate.io is for Java and the talk is great! dasync.io solves same problems for .NET in a language-oriented way instead of providing a framework.
hey, the link doesn't work. can you update with new one?
thanks. very useful!
Juval Lowy (iDesign and Righting Software) advocates decomposing by volatility, Chris here mentions functional decomposition. Which to choose ?
at first, I thought he is Vision(Paul Bettany), why the color of his face is so red :)
I got it. it's a reflection issue of back scene light which it didn't rendered properly by OpenGL fragment shader
I came up with a solution where i use websockets to update UI when projections are updated. But it requires some "gost" styling to give a user an indication that particular entity is still in update process.
What's the semantic lock... Or is it semantic logs? I didn't get that part ... It's at 12:08
I didn't get this 42:16. Does he saying that the event store is responsible for publishing the events into the message broker (kafka) once the event is stored?
Oh sagas! Like the Star War trilogy😮
Solution to transactions is Saga pattern. As he states @11:00. Then what is Process Managers for?
A saga in my mind has no state whereas a process manager is state machine based
FYI I tried to buy the book but got an error saying the coupon code has expired.
Is that other talk he mentions also available anywhere? Can't find it immediately: ruclips.net/video/kyNL7yCvQQc/видео.html
You are going to Dubai Desert Safari and while in the middle of desert, u r seeking water bottle.
- creating a order then reserving the money is same Case.
go with buttle at first place. Sir
If you're looking to stream changes from your database, be it PostgreSQL, MySQL or MongoDB check out Lapidus: github.com/JarvusInnovations/lapidus
Can't find anything on "symantec lock" for sagas, can anyone follow up?
semantic lock
It's actually "Semantic Lock". For example, Chris posted an article here -> www.er.crichardson.com/post/antipatterns/2019/07/09/developing-sagas-part-1.html
Good presentation close to enterprise system reality who worried about transaction management with MicroServices. Thanks.
1.5x
1.75x
Can't believe this CQRS rubbish presented here. Especially the example used. We solved this as an industry decades ago with the data warehouse.
Data warehouse loading still has latency in 2021 under normal conditions and unless you specifically construct something that *streams* a replication of your transnational database into your analytical database.
So, for sufficiently complex data or for big enough data ---
it's a trade-off between fast queries that show older state,
or slow queries that show current state
in any case CQRS (in his presentation) demonstrates how to solve the transnational READS in a natural application flow that is split into micro-services,
NOT JUST analytical aggregations over masses of data
You just didn't get it
Wow, only 0.01% of developers will ever find this useful ... and most of those probably would have been better off with a different approach too. LOL.
can i ask you of example of better approach please?
I think 0.01% is an accurate number if you include wordpress and js developers.
blm
i didnt know samuel l jackson was a dev
apache starts *stares motherfuckerly*
I disagree fundamentally to some of Chris's claims. He rejects what IMO actually should be done, then elaborates on stupid solutions with all their problems ... Chris seems lost in the djungle.
Why does the business logic of a business transaction "belong to the services" and not to the business transaction?