Looking at the example described in Couch DB Joins.
It discusses view collation and how you can have one document for your blog posts, and then each comment is a separate document in CouchDB. So for example, I could have "My Post" and 5 comments associated with "My Post" for a total of 6 documents. In their example, "myslug" is stored both in the post document, and each comment document, so that when I search CouchDB with the key "myslug" it returns all the documents.
Here's the problem/question. Let's say I want to search on the author in the comments and a post that also has a category of "news". How would this work exactly?
So for example:
function(doc) {
if (doc.type == "post") {
emit([doc._id, 0], doc);
} else if (doc.type == "comment") {
emit([doc.post, 1], doc);
}
}
That will load my blog post and comments based on this: ?startkey=["myslug"]
However, I want to do this, grab the comments by author bob, and the post that has the category news. For this example, bob has written three comments to the blog post with the category news. It seems as if CouchDB only allows me search on keys that exist in both documents, and not search on a key in one document, and a key in another that are "joined" together with the map function.
In other words, if post and comments are joined by a slug, how do I search on one field in one document and another field in another document that are joined by the id aka. slug?
In SQL it would be something like this:
SELECT * FROM comments JOIN doc.id ON doc.post WHERE author = bob AND category = news
I've been investigating couchdb for about a week so I'm hardly qualified to answer your question, but I think I've come to the conclusion it can't be done. View results need to be tied to one and only one document so the view can be updated. You are going to have to denormalize, at least if you don't want to do a grunt search. If anyone's come up with a clever way to do this I'd really like to know.
There are several ways that you can approximate a SQL join on CouchDB. I've just asked a similar question here: Why is CouchDB's reduce_limit enabled by default? (Is it better to approximate SQL JOINS in MapReduce views or List views?)
You can use MapReduce (not a good option)
You can use lists (This will iterate over a result set before emitting results, meaning you can 'combine' documents in a number of creative ways)
You can also apparently use 'collation', though I haven't figured this out yet (seems like I always get a count and can only use the feature with Reduce - if I'm on the right track)
Related
So I have a very specific question about the optimum way of storing & then fetching data from a MongoDB database and ill try my best to explain the use case:
I have a content publishing platform that I've built. On this platform, a user can say, write a story, and the story gets saved as a document in the 'stories' collection in the database in a structure such as this:
{
_id : s_12345,
title : This is a story,
...
}
now on the same platform, let's say, another user writes a 'news article', which gets saved as a document in a separate 'news' collection. But now, the interesting thing is, while writing this news article, the user could 'tag' a story of their choice so that this news article would show up in a 'related content' section when some user on the platform is viewing that particular story. So the data structure of this news article could be:
{
_id : n_12345,
title : This is a news article,
related_to_tag : s_12345 //id of the story
...
}
Now from my understanding as of now, there are 2 ways of doing this:
OPTION 1: when a user tries to view this story (s_12345), we make a get request to the server, fetch this particular story document from the 'stories' collection in the database, then cycle through ALL the documents in the 'news' collection and pick up all the documents that have the related_to_tag === s_12345, and then return the story document + all these related news documents to the client. However this operation seems pretty expensive to me, especially if I have, let's say, 10,000 news articles in the news collection.
OPTION 2: At the time of posting the 'news article' to the database, I also find the story (s_123456) in the stories collection, and write a reference to the news article in this story document itself, like so:
{
_id : s_12345,
title : This is a story,
related_content : n_12345
...
}
The second option seems better to me. Because then, when a user tries to get the story, I already know all the other news articles that are related to it, and simply have to run a mongoose populate function to populate these news articles. but it brings up other complications such as:
what happens when the author of the news article deletes it? that means that I will have to find the story document (s_12345), and delete the related_content reference (n_12345) as well.
Or maybe I could run a weekly cron-job that does this sort of cleanup.
Also what happens if, while I am doing this double write operation (write the news article to database + write a reference to the news article to the story document), the second operation fails for whatever reason. that would create data inconsistency.
Anyway, this is a question that I have been struggling with for quite some time now, hope I have explained my use case clearly.
Awaiting your responses!
Abrar
Let's use the classic example of blog context. In our domain we have the following scenarios: Users can write Posts. Posts must be cataloged at least in one Category. Posts can be described using Tags. Users can comment on Posts.
The four entities (Post, Category, Tag, Comment) are implemented as different aggregates because of I have not detected any rule for that an entity data should interfere in another. So, for each aggregate I will have one repository that represent it. Too, each aggregate reference others by his id.
Following CQRS, from this scenario I have deducted typical use cases that result on commands such as WriteNewPostCommand, PublishPostCommand, DeletePostCommand etc... along with their respective queries to get data from repositories. FindPostByIdQuery, FindTagByTagNameQuery, FindPostsByAuthorIdQuery etc...
Depending on which site of the app we are (backend or fronted) we will have queries more or less complex. So, if we are on the front page maybe we need build some widgets to get last comments, latest post of a category, etc... Queries that involve a simple Query object (few search criterias) and a QueryHandler very simple (a single repository as dependency on the handler class)
But in other places this queries can be more complex. In an admin panel we require to show in a table a relation that satisfy a complex search criteria. Might be interesting search posts by: author name (no id), categories names, tags name, publish date... Criterias that belongs to different aggregates and different repositories.
In addition, in our table of post we dont want to show the post along with author ID, or categories ID. We need to show all information (name user, avatar, category name, category icon etc).
My questions are:
At infrastructure layer, when we design repositories, the search methods (findAll, findById, findByCriterias...), should have return the corresponding entity referencing to all associations id's? I mean, If a have a method findPostById(uuid) or findPostByCustomFilter(filter), should return a post instance with a reference to all categories id it has, all tags id, and author id that it has? Or should my repo have some kind of method that populates a given post instance with the associations I want?
If I want to search posts created from 12/12/2014, written by John, and categorised on "News" and "Videos" categories and tags "sci-fi" and "adventure", and get the full details of each aggregate, how should create my Query and QueryHandler?
a) Create a Query with all my parameters (authorName, categoriesNames, TagsNames, if a want retrive User, Category, Tag association full detailed) and then his QueryHandler ensamble the different read models in a only one. Or...
b) Create different Queries (FindCategoryByName, FindTagByName, FindUserByName) and then my web controller calls them for later
call to FindPostQuery but now passing him the authorid, categoryid, tagid returned from the other queries?
The b) solution appear more clean but it seems me more expensive.
On the query side, there are no entities. You are free to populate your read models in any way suits your requirements best. Whatever data you need to display on (a part of) the screen, you put it in the read model. It's not the command side repositories that return these read models but specialized query side data access objects.
You mentioned "complex search criteria" -- I recommend you model it with a corresponding SearchCriteria object. This object would be technnology agnostic, but it would be passed to your Query side data access object that would know how to combine the criteria to build a lower level query for the specific data store it's targeted at.
With simple applications like this, it's easier to not get distracted by aggregates. Do event sourcing, subscribe to the events by one set of tables that is easy to query the way you want.
Another words, it sounds like you're main goal is to be able to query easily for the scenarios you describe. Start with that end goal. Now write your event handler to adjust your tables accordingly.
Start with events and the UI. Then everything else will fit easily. Google "Event Modeling" as it will help you formulate ideas sound what and how you want to build these style of applications.
I can see three problems in your approach and they need to be solved separately:
In CQRS the Queries are completely separate from the Commands. So, don't try to solve your queries with your Commands pipelines repositories. The point of CQRS is precisely to allow you to solve the commands and queries in very different ways, as they have very different requirements.
You mention DDD in the question title, but you don't mention your Bounded Contexts in the question itself. If you follow DDD, you'll most likely have more than one BC. For example, in your question, it could be that CategoryName and AuthorName belong to two different BCs, which are also different from the BC where the blog posts are. If that is the case and each BC properly owns its own data, the data that you want to search by and show in the UI will be stored potentially in different databases, therefore implementing a query in the DB with a join might not even be possible.
Searching and Reading data are two different concerns and can/should be solved differently. When you search, you get some search criteria (including sorting and paging) and the result is basically a list of IDs (authorIds, postIds, commentIds). When you Read data, you get one or more Ids and the result is one or more DTOs with all the required data properties. It is normal that you need to read data from multiple BCs to populate a single page, that's called UI composition.
So if we agree on these 3 points and especially focussing on point 3, I would suggest the following:
Figure out all the searches that you want to do and see if you can decompose them to simple searches by BC. For example, search blog posts by author name is a problem, because the author information could be in a different BC than the blog posts. So, why not implement a SearchAuthorByName in the Authors BC and then a SearchPostsByAuthorId in the Posts BC. You can do this from the Client itself or from the API. Doing it in the client gives the client a lot of flexibility because there are many ways a client can get an authorId (from a MyFavourites list, from a paginated list or from a search by name) and then get the posts by authorId is a separate operation. You can do the same by tags, categories and other things. The Post will have Ids, but not the extra details about those IDs.
Potentially, you might want more complicated searches. As long as the search criteria (including sorting fields) contain fields from a single BC, you can easily create a read model and execute the search there. Note that this is only for the search criteria. If the search result needs data from multiple BCs you can solve it with UI composition. But if the search criteria contain fields from multiple BCs, then you'll need some sort of Search engine capable of indexing data coming from multiple sources. This is especially evident if you want to do full-text search, search by categories, tags, etc. with large quantities of data. You will need to use some specialized service like Elastic Search and it won't belong to any of your existing BCs, it'll be like a supporting service.
From CQRS you will have a separeted Stack for Queries and Commands. Your query stack should represent a diferente module, namespace, dll or package at your project.
a) You will create one QueryModel and this query model will return whatever you need. If you are familiar with Entity Framework or NHibernate, you will create a Façade to hold this queries togheter, DbContext or Session.
b) You can create this separeted queries, but saying again, if you are familiar with any ORM your should return the set that represents the model, return every set as IQueryable and use LET (Linq Expression Trees) to make your Query stack more dynamic.
Using Entity Framework and C# for exemple:
public class QueryModelDatabase : DbContext, IQueryModelDatabase
{
public QueryModelDatabase() : base("dbname")
{
_products = base.Set<Product>();
_orders = base.Set<Order>();
}
private readonly DbSet<Order> _orders = null;
private readonly DbSet<Product> _products = null;
public IQueryable<Order> Orders
{
get { return this._orders.Include("Items").Include("Items.Product"); }
}
public IQueryable<Product> Products
{
get { return _products; }
}
}
Then you should do queries the way you need and return anything:
using (var db = new QueryModelDatabase())
{
var queryable = from o in db.Orders.Include(p => p.Items).Include("Details.Product")
where o.OrderId == orderId
select new OrderFoundViewModel
{
Id = o.OrderId,
State = o.State.ToString(),
Total = o.Total,
OrderDate = o.Date,
Details = o.Items
};
try
{
var o = queryable.First();
return o;
}
catch (InvalidOperationException)
{
return new OrderFoundViewModel();
}
}
There's a bunch of questions on here related to pagination using CouchDB, but none that quite fit what I'm wondering about.
Basically, I have a result set ranked by number of votes, and I want to page through the set in descending order.
Here's the map for reference.
function(doc) {
emit(doc.votes);
}
Now, the problem. I found out that startkey_docid doesn't work on it's own. You have to use it in combination with startkey. The thing is, for the query, I don't use a startkey parameter (I'm not looking to restrict the results, just get the most->least). I was thinking I could just use startkey={{doc.votes}}&startkey_docid={{doc._id}} instead, but the number of votes for a document could have changed by the time someone clicks the "Next Page" link.
The way to solve this seemed obvious: just set startkey=99999999 so that it will return all documents in the database and I can just use startkey_docid to start at the one where we left off last time. Oddly, when I do that, the startkey_docid stopped working and just allowed all results to be returned again. Apparently startkey needs to exactly equal the key on the document whose _id is used in startkey_docid.
What I'm asking is whether anyone knows a workaround for using startkey_docid to page when the actual startkey could have changed by the time you want to use it? Should my application just lookup the document by _id and immediately use the doc.votes value hoping it hasn't changed in the few milliseconds between requests? Even that doesn't seem very reliable.
EDIT: Ended up switching to Mongo for the speed, so this question turned out to be kinda moot.
I have never done something like this but I think I have some idea how to do it. What you can do is to take a snapshot of the ratings and refer to it in every page. You probably want your view not to consume to much space, so you should not map separate copies of the documents with votes not changed after taking the snapshot. So, you can do the following:
Add some history of ratings with timestamp to your document.
Map the ratings AND history like this.
In your app get the current time: start_time = Date.now() and query all pages.
Cleanup the history older then the oldest active sessions.
The problem is that if you emit [votes, date] and try to paginate you will never know how many document you have to fetch to get desired number per page. There can always be some older version which you will have to skip, and you will have make next get from DB. Thats why you can consider emitting: [date, votes], read the view always twice -- for start_time and current time, and merge and sort the result (like in merge-sort).
Ad.1:
{ ...,
votes: 12,
history: [
{date: 1357390271342, votes: 10},
{date: 1357390294682, votes: 11}
]
}
Ad.2:
function (doc) {
emit([{}, doc.votes], null);
doc.history && doc.history.forEach(function(h) {
emit([h.date, h.votes], null);
});
}
Ad.3:
?startkey=[start_time, votes]&limit=items_per_page_plus1
?startkey=[{}, votes]&limit=items_per_page_plus1
Merge lists, sort by votes in your app (on in a list function).
If you will have problems with using start_docid then you can emit [date, votes, id] and query with the ID explicitly. Even when this particular doc changes its votes it will still be available in the history.
Ad.4:
If you emit [date, votes] then you can just get outdated history width: ?startkey=[0]&endkey=[oldest_active_session_time]&inclusive_end=false and update them with update handler:
function(doc, req) {
if (!doc || !doc.history) return [null, 'Error'];
var history = new Array();
var oldest = +(req.query.date);
doc.history.forEach(function(h) {
if (h.date >= oldest)
history.push(h);
});
doc.history = history;
return [doc, 'OK'];
}
Note: I have not tested it, so it is expected not to run without modifications :)
As far as I know CouchDB uses b-tree shadowing to make updates and in principle is should be possible to access older revisions of the view. I am not into the CouchDB design, so it is just a guess and there seems not to be any (documented) API for this.
I can't figure out any simple solution by now, but there are options:
Replicate not-so-often your sorting list to small dedicated db so it will be much more stale than stale=ok
Modify your schema in a way that you'll be able to sort by some more stable data. Look at the banking/ledger example in CouchDb guide: http://guide.couchdb.org/draft/recipes.html#banking. Try to log every vote and reduce them hourly for example. As a bonus you'll get a history/trends :)
I'm kind of surprised this question has been left unanswered because the functionality of CouchDB Futon basically does this when you are paginating through the results of a map function. I opened up firebug to see what was happening in the javascript console as I paginated and saw that for every set of paginated results it is passing the startkey along with startkey_docid. So although the question is how do I paginate without including startkey, CouchDB specifies that the startkey is required and demonstrates how it can work. The endkey is not specified, so if there is only one result for the specified startkey, the next set of paginated results will also contain the next key of the sorted results that do not match the startkey.
So to clarify a bit, the answer to this problem is that as you are paginating and keeping track of the startkey_docid, you also need to capture the startkey of the same document that will be the start of the next set of results. When you are calling the paginated results use both the captured startkey and startkey_docid as couchdb requires. Leave endkey off so that the results will continue on to the next key of the sorted results.
The usecase scenario for wanting to be able to paginate without specifying a key is kind of odd. So let's say that the start docid of the next paginated result did change it's key value drastically from a 9 to a 3. And we are also assuming that there is only one instance of the docid existing in the map results, even though it could potentially appear multiple times (which I believe is why the startkey needs to be specified). As the user is clicking the next button, the user's paginated results will have now moved from looking at rank 9 to rank 3. But if you are including the startkey in addition to the startkey_docid, the paginated results would just start all over at the beginning of the rank 9 results which is a more logical progression than potentially jumping over a large set of results.
I created a simple view to return the blog title and time
function(doc) {
if ( doc.TITLE) emit(doc.TIME, doc.TITLE);
}
what is a simple way to display newest blog articles first (by default it is the other way around)?
Just apply descending sort order at view request e.g.
GET /dbname/_design/titles/_view/by_time?descending=True
And view output would be sorted in reversed way - newest blog articles will go first. Remember that startkey/endkey parameters will limit key range for this reversed order. More about view query parameters you could found in CouchDB wiki
I have three document types MainCategory, Category, SubCategory... each have a parentid which relates to the id of their parent document.
So I want to set up a view so that I can get a list of SubCategories which sit under the MainCategory (preferably just using a map function)... I haven't found a way to arrange the view so this is possible.
I currently have set up a view which gets the following output -
{"total_rows":16,"offset":0,"rows":[
{"id":"11098","key":["22056",0,"11098"],"value":"MainCat...."},
{"id":"11098","key":["22056",1,"11098"],"value":"Cat...."},
{"id":"33610","key":["22056",2,"null"],"value":"SubCat...."},
{"id":"33989","key":["22056",2,"null"],"value":"SubCat...."},
{"id":"11810","key":["22245",0,"11810"],"value":"MainCat...."},
{"id":"11810","key":["22245",1,"11810"],"value":"Cat...."},
{"id":"33106","key":["22245",2,"null"],"value":"SubCat...."},
{"id":"33321","key":["22245",2,"null"],"value":"SubCat...."},
{"id":"11098","key":["22479",0,"11098"],"value":"MainCat...."},
{"id":"11098","key":["22479",1,"11098"],"value":"Cat...."},
{"id":"11810","key":["22945",0,"11810"],"value":"MainCat...."},
{"id":"11810","key":["22945",1,"11810"],"value":"Cat...."},
{"id":"33123","key":["22945",2,"null"],"value":"SubCat...."},
{"id":"33453","key":["22945",2,"null"],"value":"SubCat...."},
{"id":"33667","key":["22945",2,"null"],"value":"SubCat...."},
{"id":"33987","key":["22945",2,"null"],"value":"SubCat...."}
]}
Which QueryString parameters would I use to get say the rows which have a key that starts with ["22945".... When all I have (at query time) is the id "11810" (at query time I don't have knowledge of the id "22945").
If any of that makes sense.
Thanks
The way you store your categories seems to be suboptimal for the query you try to perform on it.
MongoDB.org has a page on various strategies to implement tree-structures (they should apply to Couch and other doc dbs as well) - you should consider Array of Ancestors, where you always store the full path to your node. This makes updating/moving categories more difficult, but querying is easy and fast.