Patterns for handling batch operations in REST web services?

Question

A simple RESTful pattern for batches is to make use of a collection resource. For example, to delete several messages at once.

DELETE /mail?&id=0&id=1&id=2

It's a little more complicated to batch update partial resources, or resource attributes. That is, update each markedAsRead attribute. Basically, instead of treating the attribute as part of each resource, you treat it as a bucket into which to put resources. One example was already posted. I adjusted it a little.

POST /mail?markAsRead=true
POSTDATA: ids=[0,1,2]

Basically, you are updating the list of mail marked as read.

You can also use this for assigning several items to the same category.

POST /mail?category=junk
POSTDATA: ids=[0,1,2]

It's obviously much more complicated to do iTunes-style batch partial updates (e.g., artist+albumTitle but not trackTitle). The bucket analogy starts to break down.

POST /mail?markAsRead=true&category=junk
POSTDATA: ids=[0,1,2]

In the long run, it's much easier to update a single partial resource, or resource attributes. Just make use of a subresource.

POST /mail/0/markAsRead
POSTDATA: true

Alternatively, you could use parameterized resources. This is less common in REST patterns, but is allowed in the URI and HTTP specs. A semicolon divides horizontally related parameters within a resource.

Update several attributes, several resources:

POST /mail/0;1;2/markAsRead;category
POSTDATA: markAsRead=true,category=junk

Update several resources, just one attribute:

POST /mail/0;1;2/markAsRead
POSTDATA: true

Update several attributes, just one resource:

POST /mail/0/markAsRead;category
POSTDATA: markAsRead=true,category=junk

The RESTful creativity abounds.

Not at all -- I think the REST equivalent is (or at least one solution is) almost exactly that -- a specialized interface designed accommodate an operation required by the client.

I'm reminded of a pattern mentioned in Crane and Pascarello's book Ajax in Action (an excellent book, by the way -- highly recommended) in which they illustrate implementing a CommandQueue sort of object whose job it is to queue up requests into batches and then post them to the server periodically.

The object, if I remember correctly, essentially just held an array of "commands" -- e.g., to extend your example, each one a record containing a "markAsRead" command, a "messageId" and maybe a reference to a callback/handler function -- and then according to some schedule, or on some user action, the command object would be serialized and posted to the server, and the client would handle the consequent post-processing.

I don't happen to have the details handy, but it sounds like a command queue of this sort would be one way to handle your problem; it'd reduce the overall chattiness substantially, and it'd abstract the server-side interface in a way you might find more flexible down the road.

---

Update: Aha! I've found a snip from that very book online, complete with code samples (although I still suggest picking up the actual book!). Have a look here, beginning with section 5.5.3:

This is easy to code but can result in a lot of very small bits of traffic to the server, which is inefficient and potentially confusing. If we want to control our traffic, we can capture these updates and queue them locally and then send them to the server in batches at our leisure. A simple update queue implemented in JavaScript is shown in listing 5.13. [...]

The queue maintains two arrays. queued is a numerically indexed array, to which new updates are appended. sent is an associative array, containing those updates that have been sent to the server but that are awaiting a reply.

Here are two pertinent functions -- one responsible for adding commands to the queue (addCommand), and one responsible for serializing and then sending them to the server (fireRequest):

CommandQueue.prototype.addCommand = function(command)
{ 
    if (this.isCommand(command))
    {
        this.queue.append(command,true);
    }
}

CommandQueue.prototype.fireRequest = function()
{
    if (this.queued.length == 0)
    { 
        return; 
    }

    var data="data=";

    for (var i = 0; i < this.queued.length; i++)
    { 
        var cmd = this.queued[i]; 
        if (this.isCommand(cmd))
        {
            data += cmd.toRequestString(); 
            this.sent[cmd.id] = cmd;

            // ... and then send the contents of data in a POST request
        }
    }
}

That ought to get you going. Good luck!

While I think @Alex is along the right path, conceptually I think it should be the reverse of what is suggested.

The URL is in effect "the resources we are targeting" hence:

    [GET] mail/1

means get the record from mail with id 1 and

    [PATCH] mail/1 data: mail[markAsRead]=true

means patch the mail record with id 1. The querystring is a "filter", filtering the data returned from the URL.

    [GET] mail?markAsRead=true

So here we are requesting all the mail already marked as read. So to [PATCH] to this path would be saying "patch the records already marked as true"... which isn't what we are trying to achieve.

So a batch method, following this thinking should be:

    [PATCH] mail/?id=1,2,3 <the records we are targeting> data: mail[markAsRead]=true

of course I'm not saying this is true REST (which doesnt permit batch record manipulation), rather it follows the logic already existing and in use by REST.

Your language, "It seems very wasteful...", to me indicates an attempt at premature optimization. Unless it can be shown that sending the entire representation of objects is a major performance hit (we're talking unacceptable to users as > 150ms) then there's no point in attempting to create a new non-standard API behaviour. Remember, the simpler the API the easier it is to use.

For deletes send the following as the server doesn't need to know anything about the state of the object before the delete occurs.

DELETE /emails
POSTDATA: [{id:1},{id:2}]

The next thought is that if an application is running into performance issues regarding the bulk update of objects then consideration into breaking each object up into multiple objects should be given. That way the JSON payload is a fraction of the size.

As an example when sending a response to update the "read" and "archived" statuses of two separate emails you would have to send the following:

PUT /emails
POSTDATA: [
            {
              id:1,
              to:"[email protected]",
              from:"[email protected]",
              subject:"Try this recipe!",
              text:"1LB Pork Sausage, 1 Onion, 1T Black Pepper, 1t Salt, 1t Mustard Powder",
              read:true,
              archived:true,
              importance:2,
              labels:["Someone","Mustard"]
            },
            {
              id:2,
              to:"[email protected]",
              from:"[email protected]",
              subject:"Try this recipe (With Fix)",
              text:"1LB Pork Sausage, 1 Onion, 1T Black Pepper, 1t Salt, 1T Mustard Powder, 1t Garlic Powder",
              read:true,
              archived:false,
              importance:1,
              labels:["Someone","Mustard"]
            }
            ]

I would split out the mutable components of the email (read, archived, importance, labels) into a separate object as the others (to, from, subject, text) would never be updated.

PUT /email-statuses
POSTDATA: [
            {id:15,read:true,archived:true,importance:2,labels:["Someone","Mustard"]},
            {id:27,read:true,archived:false,importance:1,labels:["Someone","Mustard"]}
          ]

Another approach to take is to leverage the use of a PATCH. To explicitly indicate which properties you are intending to update and that all others should be ignored.

PATCH /emails
POSTDATA: [
            {
              id:1,
              read:true,
              archived:true
            },
            {
              id:2,
              read:true,
              archived:false
            }
          ]

People state that PATCH should be implemented by providing an array of changes containing: action (CRUD), path (URL), and value change. This may be considered a standard implementation but if you look at the entirety of a REST API it is a non-intuitive one-off. Also, the above implementation is how GitHub has implemented PATCH.

To sum it up, it is possible to adhere to RESTful principles with batch actions and still have acceptable performance.

The google drive API has a really interesting system to solve this problem (see here).

What they do is basically grouping different requests in one Content-Type: multipart/mixed request, with each individual complete request separated by some defined delimiter. Headers and query parameter of the batch request are inherited to the individual requests (i.e. Authorization: Bearer some_token) unless they are overridden in the individual request.

---

Example: (taken from their docs)

Request:

POST https://www.googleapis.com/batch

Accept-Encoding: gzip
User-Agent: Google-HTTP-Java-Client/1.20.0 (gzip)
Content-Type: multipart/mixed; boundary=END_OF_PART
Content-Length: 963

--END_OF_PART
Content-Length: 337
Content-Type: application/http
content-id: 1
content-transfer-encoding: binary


POST https://www.googleapis.com/drive/v3/files/fileId/permissions?fields=id
Authorization: Bearer authorization_token
Content-Length: 70
Content-Type: application/json; charset=UTF-8


{
  "emailAddress":"[email protected]",
  "role":"writer",
  "type":"user"
}
--END_OF_PART
Content-Length: 353
Content-Type: application/http
content-id: 2
content-transfer-encoding: binary


POST https://www.googleapis.com/drive/v3/files/fileId/permissions?fields=id&sendNotificationEmail=false
Authorization: Bearer authorization_token
Content-Length: 58
Content-Type: application/json; charset=UTF-8


{
  "domain":"appsrocks.com",
   "role":"reader",
   "type":"domain"
}
--END_OF_PART--

Response:

HTTP/1.1 200 OK
Alt-Svc: quic=":443"; p="1"; ma=604800
Server: GSE
Alternate-Protocol: 443:quic,p=1
X-Frame-Options: SAMEORIGIN
Content-Encoding: gzip
X-XSS-Protection: 1; mode=block
Content-Type: multipart/mixed; boundary=batch_6VIxXCQbJoQ_AATxy_GgFUk
Transfer-Encoding: chunked
X-Content-Type-Options: nosniff
Date: Fri, 13 Nov 2015 19:28:59 GMT
Cache-Control: private, max-age=0
Vary: X-Origin
Vary: Origin
Expires: Fri, 13 Nov 2015 19:28:59 GMT

--batch_6VIxXCQbJoQ_AATxy_GgFUk
Content-Type: application/http
Content-ID: response-1


HTTP/1.1 200 OK
Content-Type: application/json; charset=UTF-8
Date: Fri, 13 Nov 2015 19:28:59 GMT
Expires: Fri, 13 Nov 2015 19:28:59 GMT
Cache-Control: private, max-age=0
Content-Length: 35


{
 "id": "12218244892818058021i"
}


--batch_6VIxXCQbJoQ_AATxy_GgFUk
Content-Type: application/http
Content-ID: response-2


HTTP/1.1 200 OK
Content-Type: application/json; charset=UTF-8
Date: Fri, 13 Nov 2015 19:28:59 GMT
Expires: Fri, 13 Nov 2015 19:28:59 GMT
Cache-Control: private, max-age=0
Content-Length: 35


{
 "id": "04109509152946699072k"
}


--batch_6VIxXCQbJoQ_AATxy_GgFUk--

From my point of view I think Facebook has the best implementation.

A single HTTP request is made with a batch parameter and one for a token.

In batch a json is sent. which contains a collection of "requests". Each request has a method property (get / post / put / delete / etc ...), and a relative_url property (uri of the endpoint), additionally the post and put methods allow a "body" property where the fields to be updated are sent .

more info at: Facebook batch API

I would be tempted in an operation like the one in your example to write a range parser.

It's not a lot of bother to make a parser that can read "messageIds=1-3,7-9,11,12-15". It would certainly increase efficiency for blanket operations covering all messages and is more scalable.