RESTful URL design for search

I'm looking for a reasonable way to represent searches as a RESTful URLs.

The setup: I have two models, Cars and Garages, where Cars can be in Garages. So my urls look like:

/car/xxxx
  xxx == car id
  returns car with given id

/garage/yyy
  yyy = garage id
  returns garage with given id

A Car can exist on its own (hence the /car), or it can exist in a garage. What's the right way to represent, say, all the cars in a given garage? Something like:

/garage/yyy/cars     ?

How about the union of cars in garage yyy and zzz?

What's the right way to represent a search for cars with certain attributes? Say: show me all blue sedans with 4 doors :

/car/search?color=blue&type=sedan&doors=4

or should it be /cars instead?

The use of "search" seems inappropriate there - what's a better way / term? Should it just be:

/cars/?color=blue&type=sedan&doors=4

Should the search parameters be part of the PATHINFO or QUERYSTRING?

In short, I'm looking for guidance for cross-model REST url design, and for search.

[Update] I like Justin's answer, but he doesn't cover the multi-field search case:

/cars/color:blue/type:sedan/doors:4

or something like that. How do we go from

/cars/color/blue

to the multiple field case?

For the searching, use querystrings. This is perfectly RESTful:

/cars?color=blue&type=sedan&doors=4

An advantage to regular querystrings is that they are standard and widely understood and that they can be generated from form-get.

The RESTful pretty URL design is about displaying a resource based on a structure (directory-like structure, date: articles/2005/5/13, object and it's attributes,..), the slash / indicates hierarchical structure, use the -id instead.

Hierarchical structure

I would personaly prefer:

/garage-id/cars/car-id
/cars/car-id   #for cars not in garages

If a user removes the /car-id part, it brings the cars preview - intuitive. User exactly knows where in the tree he is, what is he looking at. He knows from the first look, that garages and cars are in relation. /car-id also denotes that it belongs together unlike /car/id.

Searching

The searchquery is OK as it is, there is only your preference, what should be taken into account. The funny part comes when joining searches (see below).

/cars?color=blue;type=sedan   #most prefered by me
/cars;color-blue+doors-4+type-sedan   #looks good when using car-id
/cars?color=blue&doors=4&type=sedan   #also possible, but & blends in with text

Or basically anything what isn't a slash as explained above.
The formula: /cars[?;]color[=-:]blue[,;+&], though I wouldn't use the & sign as it is unrecognizable from the text at first glance if that's your thing.

> ** Did you know that passing JSON object in URI is RESTful? **

Lists of options

/cars?color=black,blue,red;doors=3,5;type=sedan   #most prefered by me
/cars?color:black:blue:red;doors:3:5;type:sedan
/cars?color(black,blue,red);doors(3,5);type(sedan)   #does not look bad at all
/cars?color:(black,blue,red);doors:(3,5);type:sedan   #little difference

possible features?

Negate search strings (!)
To search any cars, but not black and red:
?color=!black,!red
color:(!black,!red)

Joined searches
Search red or blue or black cars with 3 doors in garages id 1..20 or 101..103 or 999 but not 5 /garage[id=1-20,101-103,999,!5]/cars[color=red,blue,black;doors=3]
You can then construct more complex search queries. (Look at CSS3 attribute matching for the idea of matching substrings. E.g. searching users containing "bar" user*=bar.)

Conclusion

Anyway, this might be the most important part for you, because you can do it however you like after all, just keep in mind that RESTful URI represents a structure which is easily understood e.g. directory-like /directory/file, /collection/node/item, dates /articles/{year}/{month}/{day}.. And when you omit any of last segments, you immediately know what you get.

So.., all these characters are allowed unencoded:

• unreserved: a-zA-Z0-9_.-~
  Typically allowed both encoded and not, both uses are then equivalent.
• special characters: $-_.+!*'(),
• reserved: ;/?:@=&
  May be used unencoded for the purpose they represent, otherwise they must be encoded.
• unsafe: <>"#%{}|^~[]`
  Why unsafe and why should rather be encoded: RFC 1738 see 2.2

Also see RFC 1738#page-20 for more character classes.

RFC 3986 see 2.2
Despite of what I previously said, here is a common distinction of delimeters, meaning that some "are" more important than others.

• generic delimeters: :/?#[]@
• sub-delimeters: !$&'()*+,;=

More reading:
Hierarchy: see 2.3, see 1.2.3
url path parameter syntax
CSS3 attribute matching
IBM: RESTful Web services - The basics
Note: RFC 1738 was updated by RFC 3986

Although having the parameters in the path has some advantages, there are, IMO, some outweighing factors.

• Not all characters needed for a search query are permitted in a URL. Most punctuation and Unicode characters would need to be URL encoded as a query string parameter. I'm wrestling with the same problem. I would like to use XPath in the URL, but not all XPath syntax is compatible with a URI path. So for simple paths, /cars/doors/driver/lock/combination would be appropriate to locate the 'combination' element in the driver's door XML document. But /car/doors[id='driver' and lock/combination='1234'] is not so friendly.

• There is a difference between filtering a resource based on one of its attributes and specifying a resource.

  For example, since

/cars/colors returns a list of all colors for all cars (the resource returned is a collection of color objects)

/cars/colors/red,blue,green would return a list of color objects that are red, blue or green, not a collection of cars.

To return cars, the path would be

/cars?color=red,blue,green or /cars/search?color=red,blue,green

• Parameters in the path are more difficult to read because name/value pairs are not isolated from the rest of the path, which is not name/value pairs.

One last comment. I prefer /garages/yyy/cars (always plural) to /garage/yyy/cars (perhaps it was a typo in the original answer) because it avoids changing the path between singular and plural. For words with an added 's', the change is not so bad, but changing /person/yyy/friends to /people/yyy seems cumbersome.

To expand on Peter's answer - you could make Search a first-class resource:

POST    /searches          # create a new search
GET     /searches          # list all searches (admin)
GET     /searches/{id}     # show the results of a previously-run search
DELETE  /searches/{id}     # delete a search (admin)

The Search resource would have fields for color, make model, garaged status, etc and could be specified in XML, JSON, or any other format. Like the Car and Garage resource, you could restrict access to Searches based on authentication. Users who frequently run the same Searches can store them in their profiles so that they don't need to be re-created. The URLs will be short enough that in many cases they can be easily traded via email. These stored Searches can be the basis of custom RSS feeds, and so on.

There are many possibilities for using Searches when you think of them as resources.

The idea is explained in more detail in this Railscast.

Justin's answer is probably the way to go, although in some applications it might make sense to consider a particular search as a resource in its own right, such as if you want to support named saved searches:

/search/{searchQuery}

or

/search/{savedSearchName}

I use two approaches to implement searches.

1. Simplest case, to query associated elements, and for navigation.

    /cars?q.garage.id.eq=1
   

This means, query cars that have garage ID equal to 1.

It is also possible to create more complex searches:

    /cars?q.garage.street.eq=FirstStreet&q.color.ne=red&offset=300&max=100

Cars in all garages in FirstStreet that are not red (3rd page, 100 elements per page).

2. Complex queries are considered as regular resources that are created and can be recovered.

    POST /searches  => Create
    GET  /searches/1  => Recover search
    GET  /searches/1?offset=300&max=100  => pagination in search
   

The POST body for search creation is as follows:

    {  
       "$class":"test.Car",
       "$q":{
          "$eq" : { "color" : "red" },
          "garage" : {
             "$ne" : { "street" : "FirstStreet" }
          }
       }
    }

It is based in Grails (criteria DSL): http://grails.org/doc/2.4.3/ref/Domain%20Classes/createCriteria.html

This is not REST. You cannot define URIs for resources inside your API. Resource navigation must be hypertext-driven. It's fine if you want pretty URIs and heavy amounts of coupling, but just do not call it REST, because it directly violates the constraints of RESTful architecture.

See this article by the inventor of REST.

In addition i would also suggest:

/cars/search/all{?color,model,year}
/cars/search/by-parameters{?color,model,year}
/cars/search/by-vendor{?vendor}

Here, Search is considered as a child resource of Cars resource.

There are a lot of good options for your case here. Still you should considering using the POST body.

The query string is perfect for your example, but if you have something more complicated, e.g. an arbitrary long list of items or boolean conditionals, you might want to define the post as a document, that the client sends over POST.

This allows a more flexible description of the search, as well as avoids the Server URL length limit.

RESTful does not recommend using verbs in URL's /cars/search is not restful. The right way to filter/search/paginate your API's is through Query Parameters. However there might be cases when you have to break the norm. For example, if you are searching across multiple resources, then you have to use something like /search?q=query

You can go through http://saipraveenblog.wordpress.com/2014/09/29/rest-api-best-practices/ to understand the best practices for designing RESTful API's

Though I like Justin's response, I feel it more accurately represents a filter rather than a search. What if I want to know about cars with names that start with cam?

The way I see it, you could build it into the way you handle specific resources:
/cars/cam*

Or, you could simply add it into the filter:
/cars/doors/4/name/cam*/colors/red,blue,green

Personally, I prefer the latter, however I am by no means an expert on REST (having first heard of it only 2 or so weeks ago...)

My advice would be this:

/garages
  Returns list of garages (think JSON array here)
/garages/yyy
  Returns specific garage
/garage/yyy/cars
  Returns list of cars in garage
/garages/cars
  Returns list of all cars in all garages (may not be practical of course)
/cars
  Returns list of all cars
/cars/xxx
  Returns specific car
/cars/colors
  Returns lists of all posible colors for cars
/cars/colors/red,blue,green
  Returns list of cars of the specific colors (yes commas are allowed :) )

Edit:

/cars/colors/red,blue,green/doors/2
  Returns list of all red,blue, and green cars with 2 doors.
/cars/type/hatchback,coupe/colors/red,blue,green/
  Same idea as the above but a lil more intuitive.
/cars/colors/red,blue,green/doors/two-door,four-door
  All cars that are red, blue, green and have either two or four doors.

Hopefully that gives you the idea. Essentially your Rest API should be easily discoverable and should enable you to browse through your data. Another advantage with using URLs and not query strings is that you are able to take advantage of the native caching mechanisms that exist on the web server for HTTP traffic.

Here's a link to a page describing the evils of query strings in REST: http://web.archive.org/web/20070815111413/http://rest.blueoxen.net/cgi-bin/wiki.pl?QueryStringsConsideredHarmful

I used Google's cache because the normal page wasn't working for me here's that link as well: http://rest.blueoxen.net/cgi-bin/wiki.pl?QueryStringsConsideredHarmful