What does edge-based and level-based mean? - kubernetes

What does "level-based" and "edge-based" mean in general?
I read "In other words, the system's behavior is level-based rather than edge-based" from kubernetes documentation:
https://github.com/GoogleCloudPlatform/kubernetes/blob/master/docs/api-conventions.md
with Google, I only find:
http://www.keil.com/forum/9423/edge-based-vs-level-based-interrupt/
Thank you.

It also has a more general definition (at least the way we tend to use it in the documentation). A piece of logic is "level based" if it only depends on the current state. A piece of logic is "edge-based" if it depends on history/transitions in addition to the current state.
"Level based" components are more resilient because if they crash, they can come back up and just look at the current state. "Edge-based" components must store the history they rely on (or depend on some other component that stores it), so that when they come back up they can look at the current state and the history. Also, if there is some kind of temporary network partition and an edge-based component misses some of the updates, then it will compute the wrong output.
However, "level based" components are usually less efficient, because they may need to scan a lot of state in order to compute an output, rather than just reading deltas.
Many components are a mixture of the two.
Simple example: You want to build a component that reports the number of pods in READY state. A level-based implementation would fetch all the pods from etcd (or the API server) and count. An edge-based implementation would do that once at startup, and then just watch for pods entering and exiting READY state.

I'd say they explained it pretty well on the site:
When a new version of an object is POSTed or PUT, the "spec" is updated and available immediately. Over time the system will work to bring the "status" into line with the "spec". The system will drive toward the most recent "spec" regardless of previous versions of that stanza. In other words, if a value is changed from 2 to 5 in one PUT and then back down to 3 in another PUT the system is not required to 'touch base' at 5 before changing the "status" to 3.
So from that statement, we know that "level base" means a PUT request does not need to be satisfied if the primary goal does not require it; it's free to skip PUT requests when seen fit.
This makes me assume that "edge based" systems require every PUT request to be satisfied, even if some requests could be skipped without altering the final result, since that would be the alternative to skipping requests.
I'm no RESTful developer (you can see by my account activity). I could not find any source of information for these things anywhere else, so I'm going based on the explanation they gave, which seems pretty straight forward.

The Kubernetes API doesn't store a history of all the changes made to an object. The controller that is responsible for that object cannot reliably observe each change; it may only observe the current state of the object.
The term "level" means the current state of the object, and the term "edge" means a transition to a new state. Controllers are "level-based" because they cannot reliably observe the transitions.
From the API conventions:
When a new version of an object is POSTed or PUT, the spec is updated and available immediately. Over time the system will work to bring the status into line with the spec. The system will drive toward the most recent spec regardless of previous versions of that stanza. For example, if a value is changed from 2 to 5 in one PUT and then back down to 3 in another PUT the system is not required to 'touch base' at 5 before changing the status to 3. In other words, the system's behavior is level-based rather than edge-based. This enables robust behavior in the presence of missed intermediate state changes.

Related

Data syncing with pouchdb-based systems client-side: is there a workaround to the 'deleted' flag?

I'm planning on using rxdb + hasura/postgresql in the backend. I'm reading this rxdb page for example, which off the bat requires sync-able entities to have a deleted flag.
Q1 (main question)
Is there ANY point at which I can finally hard-delete these entities? What conditions would have to be met - eg could I simply use "older than X months" and then force my app to only ever displays data for less than X months?
Is such a hard-delete, if possible, best carried out directly in the central db, since it will be the source of truth? Would there be any repercussions client-side that I'm not foreseeing/understanding?
I foresee the number of deleted's growing rapidly in my app and i don't want to have to store all this extra data forever.
Q2 (bonus / just curious)
What is the (algorithmic) basis for needing a 'deleted' flag? Is it that it's just faster to check a flag rather than to check for the omission of an object from, say, a very large list. I apologize if it's kind of a stupid question :(
Ultimately it comes down to a decision that's informed by your particular business/product with regards to how long you want to keep deleted entities in your system. For some applications it's important to always keep a history of deleted things or even individual revisions to records stored as a kind of ledger or history. You'll have to make a judgement call as to how long you want to keep your deleted entities.
I'd recommend that you also add a deleted_at column if you haven't already and then you could easily leverage something like Hasura's new Scheduled Triggers functionality to run a recurring job that fully deletes records older than whatever your threshold is.
You could also leverage Hasura's permissions system to ensure that rows that have been deleted aren't returned to the client. There is documentation and examples for ways to work with soft deletes and Hasura
For your second question it is definitely much faster to check for the deleted flag on records than to have to try and diff the entire dataset looking for things that are now missing.

Complicated job aggregate

I have a very complicated job process and it's not 100% clear to me where to handle what.
I don't want to have code, it just the question who is responsible for what.
Given is the following:
There is a root directory "C:\server"
Inside are two directories "ftp" and "backup"
Imagine the following process:
An external customer sends a file into the ftp directory.
An importer application get's the file and now the fun starts.
A job aggregate have to be created for this file.
The command "CreateJob(string file)" is fired.
?. The file have to be moved from ftp to backup. Inside the CommandHandler or inside the Aggregate or on JobCreated event?
StartJob(Guid jobId) get's called. A third folder have to be created "in-progress", File have to be copied from backup to in-progress. Who does it?
So it's unclear for me where Filesystem things have to be handled if the Aggregate can not work correctly without the correct filesystem.
Because my first approach was to do that inside an Infrastructure layer/lib which listen to the events from the job layer. But it seems not 100% correct?!
And top of this, what is with replaying?
You can't replay things/files that were moved, you have to somehow simulate that a customer sends the file to the ftp folder...
Thankful for answers
The file have to be moved from ftp to backup. Inside the CommandHandler or inside the Aggregate or on JobCreated event?
In situations like this, I move the file to the destination folder in the Application service that sends the command to the Aggregate (or that calls a command-like method on the Aggregate, it's the same) before the command is sent to the Aggregate. In this way, if there are some problems with the file-system (not enough permissions or space is not available etc) the command is not sent. These kind of problems should not reach our Aggregate. We most protect it from the infrastructure. In fact we should keep the Aggregate isolated from anything else; it must contain only pure business logic that is used to decide what events get generated.
Because my first approach was to do that inside an Infrastructure layer/lib which listen to the events from the job layer. But it seems not 100% correct?!
Indeed, this seems like over engineering to me. You must KISS.
StartJob(Guid jobId) get's called. A third folder have to be created "in-progress", File have to be copied from backup to in-progress. Who does it?
Whoever's calling the StartJob could do the moving, before the StartJob gets called. Again, keep the Aggregate pure. In this case it depends on your framework/domain details.
And top of this, what is with replaying? You can't replay things/files that where moved, you have to somehow simulate that a customer sends the file to the ftp folder...
The events are loaded from the event store and replayed in two situations:
Before every command gets sent to the Aggregate, the Aggregate Repository loads all the events from the event store then it applies every one of them to the Aggregate, probably calling some applyThisEvent(TheEvent) method on the Aggregate. So, this methods should be with no side effects (pure) otherwise you change the outside world again and again at every command execution and you don't want that.
The read-models (the projections, the query-models) that present data to the user listen to those events and update the database tables that hold the data that the users see. The events are sent to those read-models after they are generated and every time the read-models are being recreated. When you invent a new read-model, you must pass it all the events that were previous generated by the aggregates in order to build the correct/complete state on them. If your read-model's event listeners have side effects what do you think happens when you replay those long past events? The outside world is modified again and again and you don't want that! The read-models only interpret the events, they don't generate other events and they don't change the outside world.
There is a special third case when events reach another type of model, a Saga. A Saga must receive an event only once! This is the case that you thought to use in Because my first approach was to do that inside an Infrastructure layer/lib which listen to the events from the job layer. You could do this in your case but is not KISS.
I have a very complicated job process and it's not 100% clear to me where to handle what. I don't want to have code, it just the question who is responsible for what.
The usual answer is that the domain model -- aka the "aggregate" makes decisions, and saves them. Observing those decisions, some event handler induces side effects.
And top of this, what is with replaying? You can't replay things/files that where moved, you have to somehow simulate that a customer sends the file to the ftp folder...
You replay the events to the aggregate, so that it is restored to the state where it made the last decision. That's a separate concern from replaying the side effects -- which is part of the motivation for handling the side effects elsewhere.
Where possible, of course, you prefer to have the side effects be idempotent, so that a duplicated message doesn't create a problem. But notice that from the point of view of the model, it doesn't actually matter whether the side effect succeeds or not.

CQRS and Passing Data

Suppose I have an aggregate containing some data and when it reaches a certain state, I'd like to take all that state and pass it to some outside service. For argument and simplicity's sake, lets just say it is an aggregate that has a list and when all items in that list are checked off, I'd like to send the entire state to some outside service. Now when I'm handling the command for checking off the last item in the list, I'll know that I'm at the end but it doesn't seem correct to send it to the outside system from the processing of the command. So given this scenario what is the recommended approach if the outside system requires all of the state of the aggregate. Should the outside system build its own copy of the data based on the aggregate events or is there some better approach?
Should the outside system build its own copy of the data based on the aggregate events.
Probably not -- it's almost never a good idea to share the responsibility of rehydrating an aggregate from its history. The service that owns the object should be responsible for rehydration.
First key idea to understand is when in the flow the call to the outside service should happen.
First, the domain model processes the command arguments, computing the update to the event history, including the ChecklistCompleted event.
The application takes that history, and saves it to the book of record
The transaction completes successfully.
At this point, the application knows that the operation was successful, but the caller doesn't. So the usual answer is to be thinking of an asynchronous operation that will do the rest of the work.
Possibility one: the application takes the history that it just saved, and uses that history to create schedule a task to rehydrate a read-only copy of the aggregate state, and then send that state to the external service.
Possibility two: you ditch the copy of the history that you have now, and fire off an asynchronous task that has enough information to load its own copy of the history from the book of record.
There are at least three ways that you might do this. First, you could have the command schedule the task as before.
Second, you could have a event handler listening for ChecklistCompleted events in the book of record, and have that handler schedule the task.
Third, you could read the ChecklistCompleted event from the book of record, and publish a representation of that event to a shared bus, and let the handler in the external service call you back for a copy of the state.
I was under the impression that one bounded context should not reach out to get state from another bounded context but rather keep local copies of the data it needed.
From my experience, the key idea is that the services shouldn't block each other -- or more specifically, a call to service B should not block when service A is unavailable. Responding to events is fundamentally non blocking; does it really matter that we respond to an asynchronously delivered event by making an asynchronous blocking call?
What this buys you, however, is independent evolution of the two services - A broadcasts an event, B reacts to the event by calling A and asking for a representation of the aggregate that B understands, A -- being backwards compatible -- delivers the requested representation.
Compare this with requiring a new release of B every time the rehydration logic in A changes.
Udi Dahan raised a challenging idea - the notion that each piece of data belongs to a singe technical authority. "Raw business data" should not be replicated between services.
A service is the technical authority for a specific business capability.
Any piece of data or rule must be owned by only one service.
So in Udi's approach, you'd start to investigate why B has any responsibility for data owned by A, and from there determine how to align that responsibility and the data into a single service. (Part of the trick: the physical view of a service can span process boundaries; in other words, a process may be composed from components that belong to more than one service).
Jeppe Cramon series on microservices is nicely sourced, and touches on many of the points above.
You should never externalise your state. Reporting on that state is a function of the read side, as it produces reports and you'll need that data to call the service. The structure of your state is plastic, and you shouldn't have an external service that relies up that structure otherwise you'll have to update both in lockstep which is a bad thing.
There is a blog that puts forward a strong argument that the process manager is the correct place to put this type of feature (calling an external service), because that's the appropriate place for orchestrating events.

Can I use Time as globally unique event version?

I found time as the best value as event version.
I can merge perfectly independent events of different event sources on different servers whenever needed without being worry about read side event order synchronization. I know which event (from server 1) had happened before the other (from server 2) without the need for global sequential event id generator which makes all read sides to depend on it.
As long as the time is a globally ever sequential event version , different teams in companies can act as distributed event sources or event readers And everyone can always relay on the contract.
The world's simplest notification from a write side to subscribed read sides followed by a query pulling the recent changes from the underlying write side can simplify everything.
Are there any side effects I'm not aware of ?
Time is indeed increasing and you get a deterministic number, however event versioning is not only serves the purpose of preventing conflicts. We always say that when we commit a new event to the event store, we send the new event version there as well and it must match the expected version on the event store side, which must be the previous version plus exactly one. If there will be a thousand or three millions of ticks between two events - I do not really care, this does not give me the information I need. And if I have missed one event on the go is critical to know. So I would not use anything else than incremental counter, with events versioned per aggregate/stream.

Can watchman send why a file changed?

Is watchman capable of posting to the configured command, why it's sending a file to that command?
For example:
a file is new to a folder would possibly be a FILE_CREATE flag;
a file that is deleted would send to the command the FILE_DELETE flag;
a file that's modified would send a FILE_MOD flag etc.
Perhaps even when a folder gets deleted (and therefore the files thereunder) would send a FOLDER_DELETE parameter naming the folder, as well as a FILE_DELETE to the files thereunder / FOLDER_DELETE to the folders thereunder
Is there such a thing?
No, it can't do that. The reasons why are pretty fundamental to its design.
The TL;DR is that it is a lot more complicated than you might think for a client to correctly process those individual events and in almost all cases you don't really want them.
Most file watching systems are abstractions that simply translate from the system specific notification information into some common form. They don't deal, either very well or at all, with the notification queue being overflown and don't provide their clients with a way to reliably respond to that situation.
In addition to this, the filesystem can be subject to many and varied changes in a very short amount of time, and from multiple concurrent threads or processes. This makes this area extremely prone to TOCTOU issues that are difficult to manage. For example, creating and writing to a file typically results in a series of notifications about the file and its containing directory. If the file is removed immediately after this sequence (perhaps it was an intermediate file in a build step), by the time you see the notifications about the file creation there is a good chance that it has already been deleted.
Watchman takes the input stream of notifications and feeds it into its internal model of the filesystem: an ordered list of observed files. Each time a notification is received watchman treats it as a signal that it should go and look at the file that was reported as changed and then move the entry for that file to the most recent end of the ordered list.
When you ask Watchman for information about the filesystem it is possible or even likely that there may be pending notifications still due from the kernel. To minimize TOCTOU and ensure that its state is current, watchman generates a synchronization cookie and waits for that notification to be visible before it responds to your query.
The combination of the two things above mean that watchman result data has two important properties:
You are guaranteed to have have observed all notifications that happened before your query
You receive the most recent information for any given file only once in your query results (the change results are coalesced together)
Let's talk about the overflow case. If your system is unable to keep up with the rate at which files are changing (eg: you have a big project and are very quickly creating and deleting files and the system is heavily loaded), the OS can't fit all of the pending notifications in the buffer resources allocated to the watches. When that happens, it blows those buffers and sends an overflow signal. What that means is that the client of the watching API has missed some number of events and is no longer in synchronization with the state of the filesystem. If that client is maintains state about the filesystem it is no longer valid.
Watchman addresses this situation by re-examining the watched tree and synthetically marking all of the files as being changed. This causes the next query from the client to see everything in the tree. We call this a fresh instance result set because it is the same view you'd get when you are querying for the first time. We set a flag in the result so that the client knows that this has happened and can take appropriate steps to repair its own state. You can configure this behavior through query parameters.
In these fresh instance result sets, we don't know whether any given file really changed or not (it's possible that it changed in such a way that we can't detect via lstat) and even if we can see that its metadata changed, we don't know the cause of that change.
There can be multiple events that contribute to why a given file appears in the results delivered by watchman. We don't them record them individually because we can't track them with unbounded history; imagine a file that is incrementally being written once every second all day long. Do we keep 86400 change entries for it per day on hand and deliver those to our clients? What if there are hundreds of thousands of files like this? We'd have to truncate that data, and at that point the loss in the data reduces how well you can reason about it.
At the end of all of this, it is very rare for a client to do much more than try to read a file or look at its metadata, and generally speaking, they want to do that only when the file has stopped changing. For this use case, watchman-wait, watchman-make and trigger all have the concept of a settle period that causes the change notifications to be delayed in delivery until after the filesystem has stopped changing.