Principles for Modeling CouchDB Documents

问题:

I have a question that I've been trying to answer for some time now but can't figure out:

How do you design, or divide up, CouchDB documents?

Take a Blog Post for example.

The semi "relational" way to do it would be to create a few objects:

• Post
• User
• Comment
• Tag
• Snippet

This makes a great deal of sense. But I am trying to use couchdb (for all the reasons that it's great) to model the same thing and it's been extremely difficult.

Most of the blog posts out there give you an easy example of how to do this. They basically divide it up the same way, but say you can add 'arbitrary' properties to each document, which is definitely nice. So you'd have something like this in CouchDB:

• Post (with tags and snippets "pseudo" models in the doc)
• Comment
• User

Some people would even say you could throw the Comment and User in there, so you'd have this:

post {
    id: 123412804910820
    title: "My Post"
    body: "Lots of Content"
    html: "<p>Lots of Content</p>"
    author: {
        name: "Lance"
        age: "23"
    }
    tags: ["sample", "post"]
    comments {
        comment {
            id: 93930414809
            body: "Interesting Post"
        } 
        comment {
            id: 19018301989
            body: "I agree"
        }
    }
}

That looks very nice and is easy to understand. I also understand how you could write views that extracted just the Comments from all your Post documents, to get them into Comment models, same with Users and Tags.

But then I think, "why not just put my whole site into a single document?":

site {
    domain: "www.blog.com"
    owner: "me"
    pages {
        page {
            title: "Blog"
            posts {
                post {
                    id: 123412804910820
                    title: "My Post"
                    body: "Lots of Content"
                    html: "<p>Lots of Content</p>"
                    author: {
                        name: "Lance"
                        age: "23"
                    }
                    tags: ["sample", "post"]
                    comments {
                        comment {
                            id: 93930414809
                            body: "Interesting Post"
                        } 
                        comment {
                            id: 19018301989
                            body: "I agree"
                        }
                    }
                }
                post {
                    id: 18091890192984
                    title: "Second Post"
                    ...
                }
            }
        }
    }
}

You could easily make views to find what you wanted with that.

Then the question I have is, how do you determine when to divide the document into smaller documents, or when to make "RELATIONS" between the documents?

I think it would be much more "Object Oriented", and easier to map to Value Objects, if it were divided like so:

posts {
    post {
        id: 123412804910820
        title: "My Post"
        body: "Lots of Content"
        html: "<p>Lots of Content</p>"
        author_id: "Lance1231"
        tags: ["sample", "post"]
    }
}
authors {
    author {
        id: "Lance1231"
        name: "Lance"
        age: "23"
    }
}
comments {
    comment {
        id: "comment1"
        body: "Interesting Post"
        post_id: 123412804910820
    } 
    comment {
        id: "comment2"
        body: "I agree"
        post_id: 123412804910820
    }
}

... but then it starts looking more like a Relational Database. And often times I inherit something that looks like the "whole-site-in-a-document", so it's more difficult to model it with relations.

I've read lots of things about how/when to use Relational Databases vs. Document Databases, so that's not the main issue here. I'm more just wondering, what's a good rule/principle to apply when modeling data in CouchDB.

Another example is with XML files/data. Some XML data has nesting 10+ levels deep, and I would like to visualize that using the same client (Ajax on Rails for instance, or Flex) that I would to render JSON from ActiveRecord, CouchRest, or any other Object Relational Mapper. Sometimes I get huge XML files that are the entire site structure, like the one below, and I'd need to map it to Value Objects to use in my Rails app so I don't have to write another way of serializing/deserializing data:

<pages>
    <page>
        <subPages>
            <subPage>
                <images>
                    <image>
                        <url/>
                    </image>
                </images>
            </subPage>
        </subPages>
    </page>
</pages>

So the general CouchDB questions are:

1. What rules/principles do you use to divide up your documents (relationships, etc)?
2. Is it okay to put the entire site into one document?
3. If so, how do you handle serializing/deserializing documents with arbitrary depths levels (like the large json example above, or the xml example)?
4. Or do you not turn them into VOs, do you just decide "these ones are too nested to Object-Relational Map, so I'll just access them using raw XML/JSON methods"?

Thanks a lot for your help, the issue of how to divide up your data with CouchDB has been difficult for me to say "this is how I should do it from now on". I hope to get there soon.

I have studied the following sites/projects.

1. Hierarchical Data in CouchDB
2. CouchDB Wiki
3. Sofa - CouchDB App
4. CouchDB The Definitive Guide
5. PeepCode CouchDB Screencast
6. CouchRest
7. CouchDB README

...but they still haven't answered this question.

回答1:

There have been some great answers to this already, but I wanted to add some more recent CouchDB features to the mix of options for working with the original situation described by viatropos.

The key point at which to split up documents is where there might be conflicts (as mentioned earlier). You should never keep massively "tangled" documents together in a single document as you'll get a single revision path for completely unrelated updates (comment addition adding a revision to the entire site document for instance). Managing the relationships or connections between various, smaller documents can be confusing at first, but CouchDB provides several options for combining disparate pieces into single responses.

The first big one is view collation. When you emit key/value pairs into the results of a map/reduce query, the keys are sorted based on UTF-8 collation ("a" comes before "b"). You can also output complex keys from your map/reduce as JSON arrays: ["a", "b", "c"]. Doing that would allow you to include a "tree" of sorts built out of array keys. Using your example above, we can output the post_id, then the type of thing we're referencing, then its ID (if needed). If we then output the id of the referenced document into an object in the value that's returned we can use the 'include_docs' query param to include those documents in the map/reduce output:

{"rows":[
  {"key":["123412804910820", "post"], "value":null},
  {"key":["123412804910820", "author", "Lance1231"], "value":{"_id":"Lance1231"}},
  {"key":["123412804910820", "comment", "comment1"], "value":{"_id":"comment1"}},
  {"key":["123412804910820", "comment", "comment2"], "value":{"_id":"comment2"}}
]}

Requesting that same view with '?include_docs=true' will add a 'doc' key that will either use the '_id' referenced in the 'value' object or if that isn't present in the 'value' object, it will use the '_id' of the document from which the row was emitted (in this case the 'post' document). Please note, these results would include an 'id' field referencing the source document from which the emit was made. I left it out for space and readability.

We can then use the 'start_key' and 'end_key' parameters to filter the results down to a single post's data:

?start_key=["123412804910820"]&end_key=["123412804910820", {}, {}]

Or even specifically extract the list for a certain type:

?start_key=["123412804910820", "comment"]&end_key=["123412804910820", "comment", {}]

These query param combinations are possible because an empty object ("{}") is always at the bottom of the collation and null or "" are always at the top.

The second helpful addition from CouchDB in these situations is the _list function. This would allow you to run the above results through a templating system of some kind (if you want HTML, XML, CSV or whatever back), or output a unified JSON structure if you want to be able to request an entire post's content (including author and comment data) with a single request and returned as a single JSON document that matches what your client-side/UI code needs. Doing that would allow you to request the post's unified output document this way:

/db/_design/app/_list/posts/unified??start_key=["123412804910820"]&end_key=["123412804910820", {}, {}]&include_docs=true

Your _list function (in this case named "unified") would take the results of the view map/reduce (in this case named "posts") and run them through a JavaScript function that would send back the HTTP response in the content type you need (JSON, HTML, etc).

Combining these things, you can split up your documents at whatever level you find useful and "safe" for updates, conflicts, and replication, and then put them back together as needed when they're requested.

Hope that helps.

回答2:

The book says, if I recall correctly, to denormalize until "it hurts", while keeping in mind the frequency with which your documents might be updated.

1. What rules/principles do you use to divide up your documents (relationships, etc)?

As a rule of thumb, I include all data that is needed to display a page regarding the item in question. In other words, everything you would print on a real-world piece of paper that you would hand to somebody. E.g. a stock quote document would include the name of the company, the exchange, the currency, in addition to the numbers; a contract document would include the names and addresses of the counterparties, all information on dates and signatories. But stock quotes from distinct dates would form separate documents, separate contracts would form separate documents.

2. Is it okay to put the entire site into one document?

No, that would be silly, because:

• you would have to read and write the whole site (the document) on each update, and that is very inefficient;
• you would not benefit from any view caching.

回答3:

I know this is an old question, but I came across it trying to figure out the best approach to this exact same problem. Christopher Lenz wrote a nice blog post about methods of modeling "joins" in CouchDB. One of my take-aways was: "The only way to allow non-conflicting addition of related data is by putting that related data into separate documents." So, for simplicity sake you'd want to lean towards "denormalization". But you'll hit a natural barrier due to conflicting writes in certain circumstances.

In your example of Posts and Comments, if a single post and all of its comments lived in one document, then two people trying to post a comment at the same time (i.e. against the same revision of the document) would cause a conflict. This would get even worse in your "whole site in a single document" scenario.

So I think the rule of thumb would be "denormalize until it hurts", but the point where it will "hurt" is where you have a high likelihood of multiple edits being posted against the same revision of a document.

回答4:

I think Jake's response nails one of the most important aspects of working with CouchDB that may help you make the scoping decision: conflicts.

In the case where you have comments as an array property of the post itself, and you just have a 'post' DB with a bunch of huge 'post' documents in it, as Jake and others correctly pointed out you could imagine a scenario on a really popular blog post where two users submit edits to the post document simultaneously, resulting in a collision and a version conflict for that document.

ASIDE: As this article points out, also consider that each time you are requesting/updating that doc you have to get/set the document in its entirety, so passing around a massive documents that either represent the entire site or a post with a lot of comments on it can become a problem you would want to avoid.

In the case where posts are modeled separately from comments and two people submit a comment on a story, those simply become two "comment" documents in that DB, with no issue of conflict; just two PUT operations to add two new comments to the "comment" db.

Then to write the views that give you back the comments for a post, you would pass in the postID and then emit all the comments that reference that parent post ID, sorted in some logical ordering. Maybe you even pass in something like [postID,byUsername] as the key to the 'comments' view to indicate the parent post and how you want the results sorted or something along those lines.

MongoDB handles documents a bit differently, allowing indexes to be built on sub-elements of a document, so you might see the same question on the MongoDB mailing list and someone saying "just make the comments a property of the parent post".

Because of the write locking and single-master nature of Mongo, the conflicting revision issue of two people adding comments wouldn't spring up there and the query-ability of the content, as mentioned, isn't effected too poorly because of sub-indexes.

That being said, if your sub-elements in either DB are going to be huge (say 10s of thousands of comments) I believe it is the recommendation of both camps to make those separate elements; I have certainly seen that to be the case with Mongo as there are some upper bound limits on how big a document and its subelements can be.