nused

> The general outlook is that, OT - while being simple, they have this classic TP2 problem that would make it harder to arrive at a consistent state eventually - without a help from a central server. The alternative is to have a complicated system, almost as complicated as CRDTs.

Well this characterization isn't quite right. With regard to TP2, this is a solved problem, with or without using a server. Most of the conversations around TP2 seems to miss the crucial point that TP2 is not needed for convergence. CRDT academic literature has largely been responsible for propagating the notion that TP2 is a must for every OT system - without TP2, all is lost -- this is a myth. I wrote about it extensively here [1].

This article [2] gives an analysis of 7 coeditors based on OT, including the likes of Google Docs, showing that they avoid having to design transformation functions that require TP2. Some of the systems are fully de-centralized [3], others require a server.

It is also worth making a finer distinction between systems that do need a central server -- there are "transformation servers" and "messaging severs". Google Docs, for example, demands that its servers perform some portion of the transformation logic. Many other solutions rely on using the server only for message transmission/serialization, thus the server side logic is significantly simpler.

[1] https://arxiv.org/pdf/1905.01302.pdf

[2] Achieving convergence in operational transformation: conditions, mechanisms, and systems.ACM CSCW (2014)

[3] A generic operation transformation scheme for consistency maintenance in real-time cooperative editing systems ACM GROUP (1997)

Enjoyed your demos. Quill.js delta fits nicely with OT (and we do mentioned in https://arxiv.org/pdf/1905.01517.pdf).

Here is a Quill.js coediting demo, without using a ShareDB backend: https://codepen.io/dnus/pen/OojaeN

Thanks for the mention. We've re-organized the material and updated the submissions by splitting into three parts. We wanted to extend the content but felt that in its original form the article was too dense. The new structure consists of three related but more self-contained articles to help articulate the key differences:

Overview: https://arxiv.org/abs/1905.01518

Systems perspective: https://arxiv.org/abs/1905.01517

Correctness/theoretical perspective: https://arxiv.org/abs/1905.01302

It appears you have specific comments and/or feedback? Either post them on HN or send them to the primary contact's email. Either way, we'll make an earnest effort to respond.

This discussion is of interest to folks both inside and outside of academia and the developer community brings valuable hands-on experiences to the table. So, let's keep it inclusive.

DOI 10.1145/2957276.2957300 is an attempt at using binary-tree variants to improve CRDT's searching performance bottleneck in converting external index position to internal object identifiers. We address these specific complexities in Section 4.2.1.

DOI 10.1145/2957276.2957300 doesn't deal with operation compression. If you believe it does, please point out where we might find it.

It is trivial state opinions like 'OT is complex', but much harder work to back it up with concrete evidence. I'd think folks will find the discussion more informative without the constant hand-waving.

This paper follows the same reasoning as in http://hal.inria.fr/inria-00071213/, so it invariably arrives at the same bogus conclusion "Most OT algorithms have been proved incorrect". Please feel free to explain if you feel otherwise, but quoting yet another publication doesn't really help its case here.

What did catch my eye in this paper is that the authors claimed to have used CoWord as a case study (and of course, claimed that the theorem prover showed CoWord doesn't satisfy TP2). I worked on CoWord, and it was never open-sourced so how the authors got hands on CoWord's transformation functions (which were inside a compiled C++ binary OT-engine) for evaluation is truly troubling. Care to explain?

No points were missed - the paper just showed why the points are wrong. I'd be happy to discuss them if you care to read it: there are three concrete correctness problems with CRDT (Logoot) in sections 4.2.4 - 4.2.6.

Re: the articles you cited: Molli et al.'s http://hal.inria.fr/inria-00071213/ (A) and their earlier paper: https://www.lri.fr/~mbl/ENS/CSCW/2013/papers/Imine-ECSCW03.p... (B), are unfortunately wrong about OT correctness, demonstrably so.

The formalism can obscure the underlying argument, but both A and B uses the same reasoning, which goes something like this:

1. Assumption: An OT system's transformation functions must always satisfy TP1 and TP2, independent of the OT algorithm or protocol, otherwise they are incorrect.

2. Proof: take any existing OT system, disregard the OT algorithm or protocol, just look at the OT transformation functions in isolation, and show that the functions do not satisfy TP1 and TP2.

3. Conclusion: All known OT systems are incorrect

The starting assumption of the proof is a false premise: OT's transformation functions do not always require TP2. The proof itself, of evaluating correctness on a subset of OT in isolation, then broadly generalize to OT as a whole, is a sleight of hand (and sloppy theory work). Hence, the conclusion is plainly wrong.

TP2 is only required under a very specific pre-condition, and meaning that avoiding this pre-condition means removing the need to preserve TP2 at the level of transformation functions. TP2/CP2-avoidance strategy is commonly used in many OT algorithms and protocols: GOT, JUPITER, G-WAVE/DOCS, COT, TIBOT etc., which are provably correct (see https://www.computer.org/csdl/trans/td/2016/03/07060680-abs..... Transformation functions capable of preserving TP2/CP2 for string-wise text editing (with verified correctness and without tombstones or other meta-data)are also available (see https://dl.acm.org/citation.cfm?doid=2998181.2998252)

What I also find really dubious is the verification method used in A and B: paper B starts with using a theorem prover to definitively prove that all previously known OT systems are wrong. Then the paper proposed new transformation functions that the same theorem prover showed to be definitively correct. Paper A, which came after paper B by the same authors, used another theorem prover to show again that all previously known OT functions to be wrong, definitively, including the ones that they proposed to be "definitely correct" from B. So A basically contradicts the results in B, and both were definitively "proven correct".

Enjoyed reading your thinking around what "Intention preservation" means for CoCalc. I think "Intention" was left open-ended since it was always meant to be defined in the context of a specific application, e.g. image vs text editing would like have different notions of user-intentions and what it means to preserve them.

TP2 has been made into this monster that seems to scare anyone starting to look at OT. You don't need to worry about TP2 with a suitable protocol with a server or fully peer-to-peer, and either solution are simple enough to implement. A fine point is that TP2 and intention-violation are distinct properties: you can easily achieve convergence by serializing the operations through a total order at every site (but your won't have intentions preserved), on the other hand, if you insert 'a' and I insert 'b' at the same location simultaneously, it's reasonable for both results 'ab' or 'ba' to be intention preserved, but without transformations we'd each see a different result (thus divergence).

We've documented plenty of CRDT's performance and correctness issues in that article. I think a bigger app design question I would ask is if you want to "lock-in" your apps internal data model with a consistency-maintenance scheme -- there is real value in separation of concerns (SoC) .

Yes, OT has a time complexity of O(1) for processing a local change.

1. I think you are confusing the cost of OT with the cost of implementing standard editor operations (i.e, insert/delete characters/strings/object).

Editors have existed before collaborative editing came along and folks have figured out how to build it to be performant using ropes, piece table or many other data structures -- optimized for their own set of needs (more discussions here https://news.ycombinator.com/item?id=15381886). OT is orthogonal to the choice of design/implementation of the underlying editor, but interacts with the editor through an abstract-data-type, which in the case of text editing is a character sequence. Lumping the cost of OT with the cost of editor operations is a common mistake I see.

2. Hashtables in RGA are used only for processing remote operations. Comparing local to remote processing cost is like comparing apples and oranges.

Elaborations from the article, on page 9: We should further highlight that the use of position-based operations does not imply the text sequence must be implemented as an array of characters. The positional reference to the text sequence could be (and has commonly been) implemented in numerous data structures, such as an array of characters, the linked-list structures, the buffer-gap structures, and piece tables, etc. [8,69].

.. and on page 11: In OT-based co-editors, for example, the choice of strategies (e.g. what native data structures or operation models to use) for implementing efficient document editing is completely left to application designers, and the support for real-time collaboration is layered above and interfaced with the editing application’s exposed abstract-data-type, which is, in the case of text editing, a sequence of characters [8, 21]

There are a few things related to TP2, aka Convergence Property 2 (CP2), that needs to be unpacked here.

Do your system always need to preserve TP2/CP2 to converge correctly? Contrary to what many folks and papers will claim, answer is categorically no.

There are two general approaches to arrive this desired outcome: (1) avoiding it or (2) preserving it. Let's look at the basics of what you need for each:

(1) Avoiding it - this can be done with a centralized server, as well as in a fully distributed, decentralized, peer-to-peer context. So I'm in agreement with josephg in that a central server will avoid TP2/CP2, but I'd like to amend that with a decentralized avoidance strategy is available and just as easy to adopt. To back this up, this paper [1] looks at TP2/CP2-avoidance systematically in representative protocols and algorithms: JUPITER, NICE, Google Wave, COT, TIBOT/TIBOT2. TIBOT/TIBOT2 are two fully decentralized algorithms/protocols that avoids TP2/CP2.

In addition, [1] also answers the more interesting question of why these systems were able to avoid TP2/CP2, by giving the general conditions for TP2/CP2-avoidance. You can apply the conditions to design a new protocol or algorithm that avoids TP2/CP2.

[1] https://www.dropbox.com/s/u4mhm4ol4qq24uz/TPDS2412938-POT.pd...

(2) Preserving it - By TP2/CP2 preservation, it usually refers to the task of writing transformation functions (e.g. between insert and delete, or delete and delete) that satisfies the aforementioned condition. This is a solved problem for String-wise operations, from 2016 [2]:

- You can take these String-wise transformation functions and use them with many of the existing OT algorithms in fully-decentralized context and satisfy all TP1 and TP2.

- The transformation functions are free of Tombstones (or its variants) or other schemes (DAG etc) to ensure TP2.

Happy to say more about it if someone has an interest.

[2] https://dl.acm.org/citation.cfm?doid=2998181.2998252.

You can implement correct and performant OT with both a centralized server or with a fully decentralized topology, using TP2/CP2-avoidance and/or with TP2/CP-preservation. I've built multiple production systems with all the varieties. I generally do prefer OT with TP2/CP2-avoidance on a server, but it's not because OT can't avoid TP2/CP2 without a server or preserve TP2/CP2, it is that there is not much real-world evidence showing clear benefits moving a collaborative editing system to decentralized topology, esp when most of the CRDT-based p2p proposals in papers that I've come across involves a server one way or another.

TP2/CP2 in OT is often stood up as a straw-man so folks can throw stones against its correctness, it is really a solved problem. The unfortunately outcome is that you see a ton of papers adopting this strategy to get published and overtime the conversation becomes really convoluted.

We discuss these ideas and try to dispel many of these confusions in detail in the arxiv manuscript. A good starting point would be Section 4.1.3: OT Solutions to the CP2 Issue and the FT Puzzle

<- This, concurs with our experience.

When we profiled our string OT implementation on a single 2-3 GHz CPU, with 2-3 CPI (Cycle Per Instruction)), it was able to handle about 10M transformations per second. On a half dozen of OT based system I've worked on, the OT implementation was never where we found the real engineering complexity or the performance bottleneck in working coeditors.

The significance of using string-wise operation model in OT is under appreciated and can't be emphasized enough for real world systems. The first string-wise OT transformation functions can be traced back to Achieving Convergence, Causality Preservation, and Intention Preservation in Real-Time Cooperative Editing Systems, 1998. We also found that skip, insert, and delete can improve compression of string edits.

I think you'll find the article supports your sentiment to an extent.

One of the key messages we tried to get across is that academic research for co-editing ought to go beyond theoretical analysis but rather drive research by pushing the envelope in new areas of application and/or features. OT's evolution has been symbiotic with new applications, while we haven't seen this in CRDT (for coediting) research. The reality is that you'll find new CRDT papers every year that still (a) make fairly broad claims of benefits are either theoretically dubious or not backed/validated by application/system implementations, (b) dwelling on technical issues that were resolve years ago. To move beyond this, we want to have put these issues in context (a general transformation approach) and dispel the most common fallacies in theoretical analysis.

Related to the 'visual-merging' feature you mentioned, when I worked on making MS Word collaborative as a research project, one of the UX research questions was how to allow users to visualize the different combined effects of updates to objects (e.g. if you changed the background of object X to red and I change it to yellow at the same time). We came up with a multi-versioning technique to display the potentially different combinations and help users select one that's the most desirable. It is definitely a interesting and challenging problem to consider for text documents.

I co-authored the paper and stand by the rigor of the evidence and arguments. If you believe there is a mistake, I would be happy to have that discussion to the extent that substantive arguments and/or evidences are presented.

In referring to "CRDT", the paper qualifies upfront that it is "CRDT for co-editors" that we focus specifically on (title and 2nd sentence in the abstract). CRDT counters, sets etc are not in scope.

Casual and blanket theoretical characterizations of OT and CRDT algorithms is exactly the kind of folly that the article is trying to address.

For example, it's a logical contradiction to begin the statement with "OT's complexity is _generally_ governed by editing history (H), i.e. O(H)" to conclude that "CRDTs have much better perf than OT", then adding a qualification that argument applies only for the specific off-line editing scenario (which btw is also not true).

It's also inaccurate to base CRDT's complexity on N - the document size, i.e. number of characters visible in the document. You need to include tombstones for the case of WOOT variants or use garbage collection for RGA (which then requires vector clocks). These nuances are described in sections 4.3.

OT's complexity is governed by O(c) where c is number of concurrent operations, period.

For off-line edits, the short answers is that OT's complexity is also not O(H), if H is the number of character edits, because you can easily apply compression. Now, the longer answer: there is pretty strong spatial locality in the distribution of edits over a document -- we don't sporadically and randomly add or delete characters around a document, but intuitively, most of the inserts will be adjacent characters (i.e. strings), and many of the deletes are over newly inserted strings. OT uses string-wise operations, hence it will compress n consecutive character insert ops down to a single string op. In addition you can compress delete edits over newly inserted content: i.e. [insert 'To be or not yoo be' at 0, delete 'y' at '14', delete 'o' at '14', insert 't' at '14'] => [insert 'To be or not to be']. This is basic idea behind "operation compression" which is what OT would use to support off-line, non-real-time, asynchronous editing, however you want to call it. There is a better example here (http://www3.ntu.edu.sg/home/czsun/projects/otfaq/#_Toc321146...).