Eventual Consistency in Real-time Web Apps(blog.johnryding.com)
blog.johnryding.com
Eventual Consistency in Real-time Web Apps
http://blog.johnryding.com/post/89055480988/eventual-consistency-in-real-time-web-apps
5 comments
The complexity depends on what constraints you put on yourself. Here's an example, suppose you have a UI panel with a number of checkboxes, radiobuttons, and sliders, and you want to synchronize the state across several users, without blocking (i.e. if a user changes state, he doesn't need to wait for server response to change change state again). Global state exists on a single server. Let's say your users start changing state furiously, what happens? ...well, the object state will live in several places (and slightly different for every client):
1) Local: In the view but before event dispatched to the model.
2) Local: In model but before it was synced to the server.
2a) Local: After model sync and in queue waiting to be sent to the server.
3) Remote: In transit to the server.
4) Remote: On the server (i.e. the one true state).
5) Remote: In transit to the clients.
6) Local: Received but before it was synced to the model.
7) Local: In model but before synced to the view.
8) In the initial state sync (e.g. when a new client joins)
So at any given point, every client state is inconsistent with every other client (maybe even significantly if the latency for one or more of the clients is bad) since pieces of the eventual state are spread across all these locations. If now every user stops changing state furiously, every client should eventually be consistent. It's not a hard problem but not completely trivial. For example, you will need conflict resolution rules on the server and in your client and traditional distributed issues like (temporary) network partitioning and latency need to be taken into account. Message ordering will probably matter.
>It sounds like the client is just trying to receive a stream of updates, which could have been made by a single writer
Not necessarily. You have the same synchronization issues whether the updates are coming from a single source, or from multiple sources.
1) Local: In the view but before event dispatched to the model.
2) Local: In model but before it was synced to the server.
2a) Local: After model sync and in queue waiting to be sent to the server.
3) Remote: In transit to the server.
4) Remote: On the server (i.e. the one true state).
5) Remote: In transit to the clients.
6) Local: Received but before it was synced to the model.
7) Local: In model but before synced to the view.
8) In the initial state sync (e.g. when a new client joins)
So at any given point, every client state is inconsistent with every other client (maybe even significantly if the latency for one or more of the clients is bad) since pieces of the eventual state are spread across all these locations. If now every user stops changing state furiously, every client should eventually be consistent. It's not a hard problem but not completely trivial. For example, you will need conflict resolution rules on the server and in your client and traditional distributed issues like (temporary) network partitioning and latency need to be taken into account. Message ordering will probably matter.
>It sounds like the client is just trying to receive a stream of updates, which could have been made by a single writer
Not necessarily. You have the same synchronization issues whether the updates are coming from a single source, or from multiple sources.
The EtherPad UI had a bunch of checkboxes and whatnot that were synchronized live, so I definitely acknowledge the problem. We're working on it at Meteor now too.
The way I look at it, as long as the server has the final say over the "official" value at any given time -- which seems to be the case here, and is true in all real apps I've seen up to the backend datastore's ability to provide it -- you don't really have a consistency or convergence problem. You just have a UI problem. It's easy for the client to track both the latest official value and what the user is doing, and tell the server about the latter. The only question is what to show the user while their action is still outstanding, and the answer may be different for different kinds of UI. You may want to "fake it till you make it" (Meteor's default, aka latency compensation) or show some kind of loading indicator.
What about conflicts? Well, depending on your app, your options for how the server handles operations from the client include: they sometimes fail; they always succeed via "last writer wins" (good for a checkbox, say); they always succeed via operational transforms; or there is application-level conflict resolution that ropes in the user (implemented on top of the foregoing).
There's a trend at the moment to view everything as a distributed system, but if you treat a client as a tool for viewing information and calling APIs, it seems a lot simpler.
The way I look at it, as long as the server has the final say over the "official" value at any given time -- which seems to be the case here, and is true in all real apps I've seen up to the backend datastore's ability to provide it -- you don't really have a consistency or convergence problem. You just have a UI problem. It's easy for the client to track both the latest official value and what the user is doing, and tell the server about the latter. The only question is what to show the user while their action is still outstanding, and the answer may be different for different kinds of UI. You may want to "fake it till you make it" (Meteor's default, aka latency compensation) or show some kind of loading indicator.
What about conflicts? Well, depending on your app, your options for how the server handles operations from the client include: they sometimes fail; they always succeed via "last writer wins" (good for a checkbox, say); they always succeed via operational transforms; or there is application-level conflict resolution that ropes in the user (implemented on top of the foregoing).
There's a trend at the moment to view everything as a distributed system, but if you treat a client as a tool for viewing information and calling APIs, it seems a lot simpler.
We already have this sort of eventual consistency in Derby, built on top of ShareJS. All documents (JSON objects) can be edited synchronously in the browser, and changes are automatically replicated to all other clients.
If two people make a change to the same documents at the same time, the changes are merged in a way that will never conflict. For example if we both simultaneously insert into a list embedded in some part of the data model, derby sends a "list insert at position X" operation, which is seamlessly transformed & merged into the document at all sites. The insertion position is updated automatically before merging using the same logic that makes collaborative text editing in etherpad work. Except you can do it over the JSON structures that power your web app.
If two people make a change to the same documents at the same time, the changes are merged in a way that will never conflict. For example if we both simultaneously insert into a list embedded in some part of the data model, derby sends a "list insert at position X" operation, which is seamlessly transformed & merged into the document at all sites. The insertion position is updated automatically before merging using the same logic that makes collaborative text editing in etherpad work. Except you can do it over the JSON structures that power your web app.
>There's a trend at the moment to view everything as a distributed system, but if you treat a client as a tool for viewing information and calling APIs, it seems a lot simpler.
Ok, I guess I must have misunderstood. If you're the guy whose building on-top some synchronization layer, then yeah, it's not really a problem. The sync layer is just another controller that manipulates your model or another consumer of your model changes.
If you're the guy implementing the synchronization layer, then you are thinking of distributed systems.
Ok, I guess I must have misunderstood. If you're the guy whose building on-top some synchronization layer, then yeah, it's not really a problem. The sync layer is just another controller that manipulates your model or another consumer of your model changes.
If you're the guy implementing the synchronization layer, then you are thinking of distributed systems.
It's not always as simple as a PUT landing. Imagine you're tracking something like a phone call in real time. You want to fetch the current state of the call and start tracking it, but it's changing relatively rapidly. You also don't want to have to push the full state of the call down the notification pipe on every update, because you're handling a lot of these.
Merging the incoming deltas with fetched API state becomes a bit tricky, because you have to account for changes that might be happening while your GET is in transit. Alternatively, it may be that your notifications are slower than the GET request, depending on network conditions. An option that works in both of these cases (and doesn't involve version tagging all of your assets) is the approach described in the blog post.
Merging the incoming deltas with fetched API state becomes a bit tricky, because you have to account for changes that might be happening while your GET is in transit. Alternatively, it may be that your notifications are slower than the GET request, depending on network conditions. An option that works in both of these cases (and doesn't involve version tagging all of your assets) is the approach described in the blog post.
That is certainly simpler.
I will say that our notifications come from a different service that is separate from the service the responds to HTTP requests, which leads to the process that is described in my post.
I will say that our notifications come from a different service that is separate from the service the responds to HTTP requests, which leads to the process that is described in my post.
Are there any "standard" models for treating a realtime web app as just another distributed database node (with, of course, extra security precautions and having to do server-side data re-validation)? I'm aware of CouchDB/PouchDB and Meteor's use of mini-Mongo client-side. Are there others?
This seems to be like the big problem of modern (web) app development. Even for not-very-realtime apps, the problem is essentially the same, and non trivial (single page web) apps I've seen end up hand-cooking some sort of sync protocol through a mixture of REST calls and maybe some websockets here and there.
I like the Couch/Pouch idea, but it's really only handy when CouchDB fits your data. You can stream all CouchDB updates to a RDBMS using the changes stream, for example when you want to easily create management information, but I don't see how you can avoid making the CouchDB bunch-of-documents-per-db concept central to the app.
I haven't looked into Meteor, but my suspicion is that due to Mongo's comparably document-oriented setup, it can't be all that different.
I like the Couch/Pouch idea, but it's really only handy when CouchDB fits your data. You can stream all CouchDB updates to a RDBMS using the changes stream, for example when you want to easily create management information, but I don't see how you can avoid making the CouchDB bunch-of-documents-per-db concept central to the app.
I haven't looked into Meteor, but my suspicion is that due to Mongo's comparably document-oriented setup, it can't be all that different.
When I had to do that for some apps, the main issues were
* while being offline some notifications get lost so one needs to get a sense of the general state of the central server
* I wanted to synchronize only part of the database to each client (data it has the right to see)
* pushes to the server could not be generic due to business logic (again, rights of the user)
Thus I started recording changes on the client, pushing them to the central server through functional endpoints and then pulling all changes from the server. To pull data from the server I found a paper[1] with an interesting way of encoding data so that the client can get a sense of the difference by getting a data structure of size proportional to the diff (no matter the whole dataset size). I developed a library for that, pretty limited for now [2].
[#2] http://conferences.sigcomm.org/sigcomm/2011/papers/sigcomm/p... [#1] https://github.com/3musket33rs/mathsync
Thus I started recording changes on the client, pushing them to the central server through functional endpoints and then pulling all changes from the server. To pull data from the server I found a paper[1] with an interesting way of encoding data so that the client can get a sense of the difference by getting a data structure of size proportional to the diff (no matter the whole dataset size). I developed a library for that, pretty limited for now [2].
[#2] http://conferences.sigcomm.org/sigcomm/2011/papers/sigcomm/p... [#1] https://github.com/3musket33rs/mathsync
Actors.
Actors says nothing about consistency; you have to do it all on your own. Basically a non solution in this area.
+1. Actors are simply a unit of concurrency and means of communication between those units. Their scope and division of labor, as well as, the associated communication protocols are left to developer. They provide no inherent consistency guarantees.
If you have a CRDT implementation you can use on both sides, you can perform operations on the client and merge with the server correctly.
That's exactly what I'm planning on doing, as that seems to be simultaneously correct and relatively easy to program. I'm surprised, though, that I haven't found a "nicely packaged" version of some CRDTs, a server DB, and a client-side DB. I suppose it is probably because everyone's use-cases are so different and that using CRDTs limit the data you can use, to some degree.
I'm just learning about CRDTs because of this thread.
Which CRDT things did you find, did you experiment with any, which are you using and why?
Which CRDT things did you find, did you experiment with any, which are you using and why?
(semi cross-posted from another comment)
Aral Balkan gave a talk on a CRDT called WOOT for text editing[1] that I found really helpful to get the general idea of the concept. (Really only the last 8 or 9 minutes of his talk.)
If you want more of the nitty-gritty on some of the different types of CRDT, there's fairly readable paper on the topic[2].
From there, you can start using Google Scholar to find the other papers that have been written.
[1] https://www.youtube.com/watch?v=NSTZ4mIv_wk [2] http://hal.upmc.fr/docs/00/55/55/88/PDF/techreport.pdf
Aral Balkan gave a talk on a CRDT called WOOT for text editing[1] that I found really helpful to get the general idea of the concept. (Really only the last 8 or 9 minutes of his talk.)
If you want more of the nitty-gritty on some of the different types of CRDT, there's fairly readable paper on the topic[2].
From there, you can start using Google Scholar to find the other papers that have been written.
[1] https://www.youtube.com/watch?v=NSTZ4mIv_wk [2] http://hal.upmc.fr/docs/00/55/55/88/PDF/techreport.pdf
I hadn't seen the merge / fill idea before. Worth reading to get more ideas on this subject.
OT but seems like a good time to ask, does anyone have any experience with Operational Transformations for dealing with syncing in web apps? I've started throwing the idea around for a web app I'm building and it seems like a really interesting pattern. The only real implementation I can find is in sharejs [1]. I work with python on the backend and I can't really find an implementation of it (though for my constrained use-case I can generate good enough code to do it myself).
It seems like a good approach but there's very little in the way of libraries implementing it, so I was wondering if it's somehow problematic.
[1] http://sharejs.org/
OT but seems like a good time to ask, does anyone have any experience with Operational Transformations for dealing with syncing in web apps? I've started throwing the idea around for a web app I'm building and it seems like a really interesting pattern. The only real implementation I can find is in sharejs [1]. I work with python on the backend and I can't really find an implementation of it (though for my constrained use-case I can generate good enough code to do it myself).
It seems like a good approach but there's very little in the way of libraries implementing it, so I was wondering if it's somehow problematic.
[1] http://sharejs.org/
In general, I've read that OT tends to be very complicated to get right, which is why only a few libraries implement it. If you can use what's called a CRDT, your life will be easier, but there are fewer data structures that are supported.
Aral Balkan gave a talk on OT vs a CRDT called WOOT for text editing[1] that I found really helpful.
If you want more of the nitty-gritty on some of the different types of CRDT, there's fairly readable paper on the topic[2].
[1] https://www.youtube.com/watch?v=NSTZ4mIv_wk [2] http://hal.upmc.fr/docs/00/55/55/88/PDF/techreport.pdf
Aral Balkan gave a talk on OT vs a CRDT called WOOT for text editing[1] that I found really helpful.
If you want more of the nitty-gritty on some of the different types of CRDT, there's fairly readable paper on the topic[2].
[1] https://www.youtube.com/watch?v=NSTZ4mIv_wk [2] http://hal.upmc.fr/docs/00/55/55/88/PDF/techreport.pdf
That's great information, thank you. The paper looks really interesting - going to take a little while to get through, might start with the video :)
I'd also read that it's hard to get right (I think that was in a comment on HN by one of the people who worked on wave/sharejs). In my case I've limited myself to two different flavours of small object with a limited set of fields (instead of a totally generalised system). Most of the algorithms I've seen focus on the text element of the syncing but I'm much more interested in the object / list / field part for my use case.
Sharejs actually have a generic system of handling the OT on arbitrary json structures, but I'm not sure how well it works.
I'd also read that it's hard to get right (I think that was in a comment on HN by one of the people who worked on wave/sharejs). In my case I've limited myself to two different flavours of small object with a limited set of fields (instead of a totally generalised system). Most of the algorithms I've seen focus on the text element of the syncing but I'm much more interested in the object / list / field part for my use case.
Sharejs actually have a generic system of handling the OT on arbitrary json structures, but I'm not sure how well it works.
I've not tried it, but you might also be interested in an approach called differential synchronization[1]. Basically, it's just a specific way of passing diffs around to guarantee convergence, though (as far as I can tell) it doesn't guarantee much else (as in do you lose data if a patch can't be applied?). The really nice bits, though, are that it is a simple algorithm and that it can be applied to anything that you have diff/patch algorithms for.
[1] https://www.youtube.com/watch?v=S2Hp_1jqpY8
[1] https://www.youtube.com/watch?v=S2Hp_1jqpY8
I already watched that video the other day! Didn't quite get to the end of it but I thought the general idea was interesting.
My use case is a little different, and one that OT seems to suit.
I have multiple users working on the same documents but they don't want/need to see each other's changes straight away. They perform a number of operations and then eventually save. When they save I'd like to propagate that that history to the other user(s) that haven't saved yet. It feels like OT works there - you have two threads of history and you need to rebase the uncommitted one on top of committed one.
Maybe I'm coming at the problem wrong, but it feels like something that could work quite well as a pattern in general.
My use case is a little different, and one that OT seems to suit.
I have multiple users working on the same documents but they don't want/need to see each other's changes straight away. They perform a number of operations and then eventually save. When they save I'd like to propagate that that history to the other user(s) that haven't saved yet. It feels like OT works there - you have two threads of history and you need to rebase the uncommitted one on top of committed one.
Maybe I'm coming at the problem wrong, but it feels like something that could work quite well as a pattern in general.
I'm looking at trying to build something using CRDT; I've had my eye on the field for a while and it seems like by far the best solution.
Yes, you give up generality. But in exchange you get a more natural system with fewer strange edge cases that has good performance and won't lose data under pressure.
Yes, you give up generality. But in exchange you get a more natural system with fewer strange edge cases that has good performance and won't lose data under pressure.
OT also won't lose data under pressure.
The biggest downsides of CRDTs are that your data ends up much bigger (although usually by a constant multiple), and nobody has figured out how to edit JSON documents using CRDTs yet.
Its not all that surprising that CRDTs will be larger - imagine a stream of operations happening to a document (think about them like tiny git commits, if that helps). You can end up with a lot of changes for a small document (tens of thousands of changes for a single page of text isn't unheard of). CRDTs in effect embed the history of operations into the document, while OT stores them separately. As such, with OT you can eventually migrate the operation log to secondary storage (or throw it away). You can never really do that with CRDTs.
The JSON problem is harder, and obviously very important for realtime web apps that do more than text editing. Decent CRDTs will be trivial to add to ShareJS when they're ready.
The biggest downsides of CRDTs are that your data ends up much bigger (although usually by a constant multiple), and nobody has figured out how to edit JSON documents using CRDTs yet.
Its not all that surprising that CRDTs will be larger - imagine a stream of operations happening to a document (think about them like tiny git commits, if that helps). You can end up with a lot of changes for a small document (tens of thousands of changes for a single page of text isn't unheard of). CRDTs in effect embed the history of operations into the document, while OT stores them separately. As such, with OT you can eventually migrate the operation log to secondary storage (or throw it away). You can never really do that with CRDTs.
The JSON problem is harder, and obviously very important for realtime web apps that do more than text editing. Decent CRDTs will be trivial to add to ShareJS when they're ready.
I've had a lot of success with embedding initial content for a page as JSON in a script tag with a specific ID and a type of "application/json". This way you can rely on the content being there and sidestep the whole timing issue completely. You can potentially speed up your loading time as well, if you have a bunch of things you'd usually fire additional requests off for.
I believe I got this idea from Backbone's documentation [1] originally, but I've seen it mentioned in other places as well.
[1]: http://backbonejs.org/#FAQ-bootstrap
I believe I got this idea from Backbone's documentation [1] originally, but I've seen it mentioned in other places as well.
[1]: http://backbonejs.org/#FAQ-bootstrap
The Errai J2EE framework for GWT has an 'Operational Transform Eventually Consistent" Web Socket bus implementation for those curious. (https://github.com/errai/errai/tree/master/errai-otec) Demo video: https://www.youtube.com/watch?v=ChCasRr0cZc
I don't see anything I recognize as eventual consistency in the traditional sense. It sounds like the client is just trying to receive a stream of updates, which could have been made by a single writer, and the only consistency issues are caused by the different overlapping mechanisms for getting updates.