I have an event-sourced CQRS architecture hosted in kubernetes. I have a single writer (the "denormalizer service") that listens to events and writes the denormalized views to a datastore. These views are then served by a separate view service. When the denormalizer image is updated through a deployment with new projections it replays all events from the beginning and writes the new views to a different datastore.
So I need 2 instances of denormalizer, one with the old code, and another replaying the events through the new code. When the new code is finished replaying I need to:
1) signal to the view service to switch to the newly written datastore and then,
2) bring down the old denormalizer deployment as it is no longer required.
Problem is, (to my limited knowledge) kubernetes seems ill-equipped to deal with this scenario.
Any idea how I would do something like this?
I don't know the specifics of your system, but two solutions come to my mind.
Using readinessProbe
You can define a readinessProbe for your writer service. Make it report the service is ready, when the rewrite is done. Then, the rolling updater will know when to shut down the old version of the writer and start serving traffic to the new one. The only thing you'd need to do more is notify the viewer to switch to new data source. This could be done by the writer calling some API on the viewer service.
Using separate process
You can create a special process that will execute the procedure you described using Kubernetes API. It is more work than the first solution, but gives you more control over the whole process. It would observe your repository if there are new versions of writers, if yes, it would start a new service, wait for it to be ready, kill the old writer and notify the viewer.
I've been looking at using service fabric as a platform for a new solution that we are building and I am getting hung up on data / stage management. I really like the concept of reliable services and the actor model and as we have started to prototype out some things it seems be working well.
With that beings said I am getting hung up on state management and how I would use it in a 'real' project. I am also a little concerned with how the data feels like a black box that I can't interrogate or manipulate directly if needed. A couple scenarios I've thought about are:
How would I share state between two developers on a project? I have an Actor and as long as I am debugging the actor my state is maintained, replicated, etc. However when I shut it down the state is all lost. More importantly someone else on my team would need to set up the same data as I do, this is fine for transactional data - but certain 'master' data should just be constant.
Likewise I am curious about how I would migrate data changes between environments. We periodically pull production data down form our SQL Azure instance today to keep our test environment fresh, we also push changes up from time to time depending on the requirements of the release.
I have looked at the backup and restore process, but it feels cumbersome, especially in the development scenario. Asking someone to (or scripting the) restore on every partition of every stateful service seems like quite a bit of work.
I think that the answer to both of these questions is that I can use the stateful services, but I need to rely on an external data store for anything that I want to retain. The service would check for state when it was activated and use the stateful service almost as a write-through cache. I'm not suggesting that this needs to be a uniform design choice, more on a service by service basis - depending on the service needs.
Does that sound right, am I overthinking this, missing something, etc?
Thanks
Joe
If you want to share Actor state between developers, you can use a shared cluster. (in Azure or on-prem). Make sure you always do upgrade-style deployments, so state will survive. State is persisted if you configure the Actor to do so.
You can migrate data by doing a backup of all replica's of your service and restoring them on a different cluster. (have the service running and trigger data-loss). It's cumbersome yes, but at this time it's the only way. (or store state externally)
Note that state is safe in the cluster, it's stored on disk and replicated. There's no need to have an external store, provided you do regular state backups and keep them outside the cluster. Stateful services can be more than just caches.
First of all, this is a question regarding my thesis for school. I have done some research about this, it seems like a problem that hasn't been tackled yet (might not be that common).
Before jumping right into the problem, I'll give a brief example of my use case.
I have multiple namespaces containing microservices depending on a state X. To manage this the microservices are put in a namespace named after the state. (so namespaces state_A, state_B, ...)
Important to know is that each microservice needs this state at startup of the service. It will download necessary files, ... according to the state. When launching it with state A version 1, it is very likely that the state gets updated every month. When this happens, it is important to let all the microservices that depend on state A upgrade whatever necessary (databases, in-memory state, ...).
My current approach for this problem is simply using events, the microservices that need updates when the state changes can subscribe on the event and migrate/upgrade accordingly. The only problem I'm facing is that while the service is upgrading, it should still work. So somehow I should duplicate the service first, let the duplicate upgrade and when the upgrade is successful, shut down the original. Because of this the used orchestration service would have to be able to create duplicates (including duplicating the state).
My question is, are there already solutions for my problem (and if yes, which ones)? I have looked into Netflix Conductor (which seemed promising with its workflows and events), Amazon SWF, Marathon and Kubernetes, but none of them covers my problem.
Best of all the existing solution should not be bound to a specific platform (Azure, GCE, ...).
For uninterrupted upgrade you should use clusters of nodes providing your service and perform a rolling update, which takes out a single node at a time, upgrading it, leaving the rest of the nodes for continued servicing. I recommend looking at the concept of virtual services (e.g. in kubernetes) and rolling updates.
For inducing state I would recommend looking into container initialization mechanisms. For example in docker you can use entrypoint scripts or in kubernetes there is the concept of init containers. You should note though that today there is a trend to decouple services and state, meaning the state is kept in a DB that is separate from the service deployment, allowing to view the service as a stateless component that can be replaced without losing state (given the interfacing between the service and required state did not change). This is good in scenarios where the service changes more frequently and the DB design less frequently.
Another note - I am not sure that representing state in a namespace is a good idea. Typically a namespace is a static construct for organization (of code, services, etc.) that aims for stability.
Let's say we have a User, Wallet REST microservices and an API gateway that glues things together. When Bob registers on our website, our API gateway needs to create a user through the User microservice and a wallet through the Wallet microservice.
Now here are a few scenarios where things could go wrong:
User Bob creation fails: that's OK, we just return an error message to the Bob. We're using SQL transactions so no one ever saw Bob in the system. Everything's good :)
User Bob is created but before our Wallet can be created, our API gateway hard crashes. We now have a User with no wallet (inconsistent data).
User Bob is created and as we are creating the Wallet, the HTTP connection drops. The wallet creation might have succeeded or it might have not.
What solutions are available to prevent this kind of data inconsistency from happening? Are there patterns that allow transactions to span multiple REST requests? I've read the Wikipedia page on Two-phase commit which seems to touch on this issue but I'm not sure how to apply it in practice. This Atomic Distributed Transactions: a RESTful design paper also seems interesting although I haven't read it yet.
Alternatively, I know REST might just not be suited for this use case. Would perhaps the correct way to handle this situation to drop REST entirely and use a different communication protocol like a message queue system? Or should I enforce consistency in my application code (for example, by having a background job that detects inconsistencies and fixes them or by having a "state" attribute on my User model with "creating", "created" values, etc.)?
What doesn't make sense:
distributed transactions with REST services. REST services by definition are stateless, so they should not be participants in a transactional boundary that spans more than one service. Your user registration use case scenario makes sense, but the design with REST microservices to create User and Wallet data is not good.
What will give you headaches:
EJBs with distributed transactions. It's one of those things that work in theory but not in practice. Right now I'm trying to make a distributed transaction work for remote EJBs across JBoss EAP 6.3 instances. We've been talking to RedHat support for weeks, and it didn't work yet.
Two-phase commit solutions in general. I think the 2PC protocol is a great algorithm (many years ago I implemented it in C with RPC). It requires comprehensive fail recovery mechanisms, with retries, state repository, etc. All the complexity is hidden within the transaction framework (ex.: JBoss Arjuna). However, 2PC is not fail proof. There are situations the transaction simply can't complete. Then you need to identify and fix database inconsistencies manually. It may happen once in a million transactions if you're lucky, but it may happen once in every 100 transactions depending on your platform and scenario.
Sagas (Compensating transactions). There's the implementation overhead of creating the compensating operations, and the coordination mechanism to activate compensation at the end. But compensation is not fail proof either. You may still end up with inconsistencies (= some headache).
What's probably the best alternative:
Eventual consistency. Neither ACID-like distributed transactions nor compensating transactions are fail proof, and both may lead to inconsistencies. Eventual consistency is often better than "occasional inconsistency". There are different design solutions, such as:
You may create a more robust solution using asynchronous communication. In your scenario, when Bob registers, the API gateway could send a message to a NewUser queue, and right-away reply to the user saying "You'll receive an email to confirm the account creation." A queue consumer service could process the message, perform the database changes in a single transaction, and send the email to Bob to notify the account creation.
The User microservice creates the user record and a wallet record in the same database. In this case, the wallet store in the User microservice is a replica of the master wallet store only visible to the Wallet microservice. There's a data synchronization mechanism that is trigger-based or kicks in periodically to send data changes (e.g., new wallets) from the replica to the master, and vice-versa.
But what if you need synchronous responses?
Remodel the microservices. If the solution with the queue doesn't work because the service consumer needs a response right away, then I'd rather remodel the User and Wallet functionality to be collocated in the same service (or at least in the same VM to avoid distributed transactions). Yes, it's a step farther from microservices and closer to a monolith, but will save you from some headache.
This is a classic question I was asked during an interview recently How to call multiple web services and still preserve some kind of error handling in the middle of the task. Today, in high performance computing, we avoid two phase commits. I read a paper many years ago about what was called the "Starbuck model" for transactions: Think about the process of ordering, paying, preparing and receiving the coffee you order at Starbuck... I oversimplify things but a two phase commit model would suggest that the whole process would be a single wrapping transaction for all the steps involved until you receive your coffee. However, with this model, all employees would wait and stop working until you get your coffee. You see the picture ?
Instead, the "Starbuck model" is more productive by following the "best effort" model and compensating for errors in the process. First, they make sure that you pay! Then, there are message queues with your order attached to the cup. If something goes wrong in the process, like you did not get your coffee, it is not what you ordered, etc, we enter into the compensation process and we make sure you get what you want or refund you, This is the most efficient model for increased productivity.
Sometimes, starbuck is wasting a coffee but the overall process is efficient. There are other tricks to think when you build your web services like designing them in a way that they can be called any number of times and still provide the same end result. So, my recommendation is:
Don't be too fine when defining your web services (I am not convinced about the micro-service hype happening these days: too many risks of going too far);
Async increases performance so prefer being async, send notifications by email whenever possible.
Build more intelligent services to make them "recallable" any number of times, processing with an uid or taskid that will follow the order bottom-top until the end, validating business rules in each step;
Use message queues (JMS or others) and divert to error handling processors that will apply operations to "rollback" by applying opposite operations, by the way, working with async order will require some sort of queue to validate the current state of the process, so consider that;
In last resort, (since it may not happen often), put it in a queue for manual processing of errors.
Let's go back with the initial problem that was posted. Create an account and create a wallet and make sure everything was done.
Let's say a web service is called to orchestrate the whole operation.
Pseudo code of the web service would look like this:
Call Account creation microservice, pass it some information and a some unique task id 1.1 Account creation microservice will first check if that account was already created. A task id is associated with the account's record. The microservice detects that the account does not exist so it creates it and stores the task id. NOTE: this service can be called 2000 times, it will always perform the same result. The service answers with a "receipt that contains minimal information to perform an undo operation if required".
Call Wallet creation, giving it the account ID and task id. Let's say a condition is not valid and the wallet creation cannot be performed. The call returns with an error but nothing was created.
The orchestrator is informed of the error. It knows it needs to abort the Account creation but it will not do it itself. It will ask the wallet service to do it by passing its "minimal undo receipt" received at the end of step 1.
The Account service reads the undo receipt and knows how to undo the operation; the undo receipt may even include information about another microservice it could have called itself to do part of the job. In this situation, the undo receipt could contain the Account ID and possibly some extra information required to perform the opposite operation. In our case, to simplify things, let's say is simply delete the account using its account id.
Now, let's say the web service never received the success or failure (in this case) that the Account creation's undo was performed. It will simply call the Account's undo service again. And this service should normaly never fail because its goal is for the account to no longer exist. So it checks if it exists and sees nothing can be done to undo it. So it returns that the operation is a success.
The web service returns to the user that the account could not be created.
This is a synchronous example. We could have managed it in a different way and put the case into a message queue targeted to the help desk if we don't want the system to completly recover the error". I've seen this being performed in a company where not enough hooks could be provided to the back end system to correct situations. The help desk received messages containing what was performed successfully and had enough information to fix things just like our undo receipt could be used for in a fully automated way.
I have performed a search and the microsoft web site has a pattern description for this approach. It is called the compensating transaction pattern:
Compensating transaction pattern
All distributed systems have trouble with transactional consistency. The best way to do this is like you said, have a two-phase commit. Have the wallet and the user be created in a pending state. After it is created, make a separate call to activate the user.
This last call should be safely repeatable (in case your connection drops).
This will necessitate that the last call know about both tables (so that it can be done in a single JDBC transaction).
Alternatively, you might want to think about why you are so worried about a user without a wallet. Do you believe this will cause a problem? If so, maybe having those as separate rest calls are a bad idea. If a user shouldn't exist without a wallet, then you should probably add the wallet to the user (in the original POST call to create the user).
IMHO one of the key aspects of microservices architecture is that the transaction is confined to the individual microservice (Single responsibility principle).
In the current example, the User creation would be an own transaction. User creation would push a USER_CREATED event into an event queue. Wallet service would subscribe to the USER_CREATED event and do the Wallet creation.
If my wallet was just another bunch of records in the same sql database as the user then I would probably place the user and wallet creation code in the same service and handle that using the normal database transaction facilities.
It sounds to me you are asking about what happens when the wallet creation code requires you touch another other system or systems? Id say it all depends on how complex and or risky the creation process is.
If it's just a matter of touching another reliable datastore (say one that can't participate in your sql transactions), then depending on the overall system parameters, I might be willing to risk the vanishingly small chance that second write won't happen. I might do nothing, but raise an exception and deal with the inconsistent data via a compensating transaction or even some ad-hoc method. As I always tell my developers: "if this sort of thing is happening in the app, it won't go unnoticed".
As the complexity and risk of wallet creation increases you must take steps to ameliorate the risks involved. Let's say some of the steps require calling multiple partner apis.
At this point you might introduce a message queue along with the notion of partially constructed users and/or wallets.
A simple and effective strategy for making sure your entities eventually get constructed properly is to have the jobs retry until they succeed, but a lot depends on the use cases for your application.
I would also think long and hard about why I had a failure prone step in my provisioning process.
One simple Solution is you create user using the User Service and use a messaging bus where user service emits its events , and Wallet Service registers on the messaging bus, listens on User Created event and create Wallet for the User. In the mean time , if user goes on Wallet UI to see his Wallet, check if user was just created and show your wallet creation is in progress, please check in some time
What solutions are available to prevent this kind of data inconsistency from happening?
Traditionally, distributed transaction managers are used. A few years ago in the Java EE world you might have created these services as EJBs which were deployed to different nodes and your API gateway would have made remote calls to those EJBs. The application server (if configured correctly) automatically ensures, using two phase commit, that the transaction is either committed or rolled back on each node, so that consistency is guaranteed. But that requires that all the services be deployed on the same type of application server (so that they are compatible) and in reality only ever worked with services deployed by a single company.
Are there patterns that allow transactions to span multiple REST requests?
For SOAP (ok, not REST), there is the WS-AT specification but no service that I have ever had to integrate has support that. For REST, JBoss has something in the pipeline. Otherwise, the "pattern" is to either find a product which you can plug into your architecture, or build your own solution (not recommended).
I have published such a product for Java EE: https://github.com/maxant/genericconnector
According to the paper you reference, there is also the Try-Cancel/Confirm pattern and associated Product from Atomikos.
BPEL Engines handle consistency between remotely deployed services using compensation.
Alternatively, I know REST might just not be suited for this use case. Would perhaps the correct way to handle this situation to drop REST entirely and use a different communication protocol like a message queue system?
There are many ways of "binding" non-transactional resources into a transaction:
As you suggest, you could use a transactional message queue, but it will be asynchronous, so if you depend on the response it becomes messy.
You could write the fact that you need to call the back end services into your database, and then call the back end services using a batch. Again, async, so can get messy.
You could use a business process engine as your API gateway to orchestrate the back end microservices.
You could use remote EJB, as mentioned at the start, since that supports distributed transactions out of the box.
Or should I enforce consistency in my application code (for example, by having a background job that detects inconsistencies and fixes them or by having a "state" attribute on my User model with "creating", "created" values, etc.)?
Playing devils advocate: why build something like that, when there are products which do that for you (see above), and probably do it better than you can, because they are tried and tested?
In micro-services world the communication between services should be either through rest client or messaging queue. There can be two ways to handle the transactions across services depending on how are you communicating between the services. I will personally prefer message driven architecture so that a long transaction should be a non blocking operation for a user.
Lets take you example to explain it :
Create user BOB with event CREATE USER and push the message to a message bus.
Wallet service subscribed to this event can create a wallet corresponding to the user.
The one thing which you have to take care is to select a robust reliable message backbone which can persists the state in case of failure. You can use kafka or rabbitmq for messaging backbone. There will be a delay in execution because of eventual consistency but that can be easily updated through socket notification. A notifications service/task manager framework can be a service which update the state of the transactions through asynchronous mechanism like sockets and can help UI to update show the proper progress.
Personally I like the idea of Micro Services, modules defined by the use cases, but as your question mentions, they have adaptation problems for the classical businesses like banks, insurance, telecom, etc...
Distributed transactions, as many mentioned, is not a good choice, people now going more for eventually consistent systems but I am not sure this will work for banks, insurance, etc....
I wrote a blog about my proposed solution, may be this can help you....
https://mehmetsalgar.wordpress.com/2016/11/05/micro-services-fan-out-transaction-problems-and-solutions-with-spring-bootjboss-and-netflix-eureka/
Eventual consistency is the key here.
One of the services is chosen to become primary handler of the event.
This service will handle the original event with single commit.
Primary handler will take responsibility for asynchronously communicating the secondary effects to other services.
The primary handler will do the orchestration of other services calls.
The commander is in charge of the distributed transaction and takes control. It knows the instruction to be executed and will coordinate executing them. In most scenarios there will just be two instructions, but it can handle multiple instructions.
The commander takes responsibility of guaranteeing the execution of all instructions, and that means retires.
When the commander tries to effect the remote update and doesn’t get a response, it has no retry.
This way the system can be configured to be less prone to failure and it heals itself.
As we have retries we have idempotence.
Idempotence is the property of being able to do something twice such a way that the end results be the same as if it had been done once only.
We need idempotence at the remote service or data source so that, in the case where it receives the instruction more than once, it only processes it once.
Eventual consistency
This solves most of distributed transaction challenges, however we need to consider couple of points here.
Every failed transaction will be followed by a retry, the amount of attempted retries depends on the context.
Consistency is eventual i.e., while the system is out of consistent state during a retry, for example if a customer has ordered a book, and made a payment and then updates the stock quantity. If the stock update operations fail and assuming that was the last stock available, the book will still be available till the retry operation for the stock updating has succeeded. After the retry is successful your system will be consistent.
Why not use API Management (APIM) platform that supports scripting/programming? So, you will be able to build composite service in the APIM without disturbing micro services. I have designed using APIGEE for this purpose.
Our org is planning on basing parts of our business model on the premise of recurring workflows in CRM 2011. However, we sometimes run into an issue with a backed up workflow queue, or for some reason need to restart the server (update rollups, etc.), or in some other way find we have to restart the CRM's async service.
What would happen to any workflows in the "waiting" phase in this scenario?
I see the workflow in the AsyncOperationBase table with the "waiting" statuscode; when the service comes back online, does it look at this table and resume accordingly?
In the above scenario, what what would happen if the service was stopped, and in the interim, the workflow reached its PostponeUntil date? Does the service look at all non-complete future and backdated workflows and decide what to with each? Or does the workflow just fail altogether?
Any fails in the process would obviously be a deal breaker for this element of the CRM system, and we'd have to develop an external component to handle recurring items.
I'd expect there to be some documentation on this, but I can only guess that the WaitSubscription class has something to do with this topic, but it's for the most part undocumented.
For now, we've decided to go with an external service to manage this, due to the seemingly black box nature of the async process. Tracing the calls to the database server does show a lot of calls to the AsyncOperationBase table, which tends to make me believe the service always checks to see if a job is being done, but in the absence of extreme testing, for now it's more secure to use a separate service for this requirement.