MongoDB/NoSQL: Keeping Document Change History

问题:

A fairly common requirement in database applications is to track changes to one or more specific entities in a database. I've heard this called row versioning, a log table or a history table (I'm sure there are other names for it). There are a number of ways to approach it in an RDBMS--you can write all changes from all source tables to a single table (more of a log) or have a separate history table for each source table. You also have the option to either manage the logging in application code or via database triggers.

I'm trying to think through what a solution to the same problem would look like in a NoSQL/document database (specifically MongoDB), and how it would be solved in a uniform way. Would it be as simple as creating version numbers for documents, and never overwriting them? Creating separate collections for "real" vs. "logged" documents? How would this affect querying and performance?

Anyway, is this a common scenario with NoSQL databases, and if so, is there a common solution?

回答1:

Good question, I was looking into this myself as well.

Create a new version on each change

I came across the Versioning module of the Mongoid driver for Ruby. I haven't used it myself, but from what I could find, it adds a version number to each document. Older versions are embedded in the document itself. The major drawback is that the entire document is duplicated on each change, which will result in a lot of duplicate content being stored when you're dealing with large documents. This approach is fine though when you're dealing with small-sized documents and/or don't update documents very often.

Only store changes in a new version

Another approach would be to store only the changed fields in a new version. Then you can 'flatten' your history to reconstruct any version of the document. This is rather complex though, as you need to track changes in your model and store updates and deletes in a way that your application can reconstruct the up-to-date document. This might be tricky, as you're dealing with structured documents rather than flat SQL tables.

Store changes within the document

Each field can also have an individual history. Reconstructing documents to a given version is much easier this way. In your application you don't have to explicitly track changes, but just create a new version of the property when you change its value. A document could look something like this:

{
  _id: "4c6b9456f61f000000007ba6"
  title: [
    { version: 1, value: "Hello world" },
    { version: 6, value: "Foo" }
  ],
  body: [
    { version: 1, value: "Is this thing on?" },
    { version: 2, value: "What should I write?" },
    { version: 6, value: "This is the new body" }
  ],
  tags: [
    { version: 1, value: [ "test", "trivial" ] },
    { version: 6, value: [ "foo", "test" ] }
  ],
  comments: [
    {
      author: "joe", // Unversioned field
      body: [
        { version: 3, value: "Something cool" }
      ]
    },
    {
      author: "xxx",
      body: [
        { version: 4, value: "Spam" },
        { version: 5, deleted: true }
      ]
    },
    {
      author: "jim",
      body: [
        { version: 7, value: "Not bad" },
        { version: 8, value: "Not bad at all" }
      ]
    }
  ]
}

Marking part of the document as deleted in a version is still somewhat awkward though. You could introduce a state field for parts that can be deleted/restored from your application:

{
  author: "xxx",
  body: [
    { version: 4, value: "Spam" }
  ],
  state: [
    { version: 4, deleted: false },
    { version: 5, deleted: true }
  ]
}

With each of these approaches you can store an up-to-date and flattened version in one collection and the history data in a separate collection. This should improve query times if you're only interested in the latest version of a document. But when you need both the latest version and historical data, you'll need to perform two queries, rather than one. So the choice of using a single collection vs. two separate collections should depend on how often your application needs the historical versions.

Most of this answer is just a brain dump of my thoughts, I haven't actually tried any of this yet. Looking back on it, the first option is probably the easiest and best solution, unless the overhead of duplicate data is very significant for your application. The second option is quite complex and probably isn't worth the effort. The third option is basically an optimization of option two and should be easier to implement, but probably isn't worth the implementation effort unless you really can't go with option one.

Looking forward to feedback on this, and other people's solutions to the problem :)

回答2:

We have partially implemented this on our site and we use the 'Store Revisions in a separate document" (and separate database). We wrote a custom function to return the diffs and we store that. Not so hard and can allow for automated recovery.

回答3:

Why not a variation on Store changes within the document ?

Instead of storing versions against each key pair, the current key pairs in the document always represents the most recent state and a 'log' of changes is stored within a history array. Only those keys which have changed since creation will have an entry in the log.

{
  _id: "4c6b9456f61f000000007ba6"
  title: "Bar",
  body: "Is this thing on?",
  tags: [ "test", "trivial" ],
  comments: [
    { key: 1, author: "joe", body: "Something cool" },
    { key: 2, author: "xxx", body: "Spam", deleted: true },
    { key: 3, author: "jim", body: "Not bad at all" }
  ],
  history: [
    { 
      who: "joe",
      when: 20160101,
      what: { title: "Foo", body: "What should I write?" }
    },
    { 
      who: "jim",
      when: 20160105,
      what: { tags: ["test", "test2"], comments: { key: 3, body: "Not baaad at all" }
    }
  ]
}

回答4:

One can have a current NoSQL database and a historical NoSQL database. There will be a an nightly ETL ran everyday. This ETL will record every value with a timestamp, so instead of values it will always be tuples (versioned fields). It will only record a new value if there was a change made on the current value, saving space in the process. For example, this historical NoSQL database json file can look like so:

{
  _id: "4c6b9456f61f000000007ba6"
  title: [
    { date: 20160101, value: "Hello world" },
    { date: 20160202, value: "Foo" }
  ],
  body: [
    { date: 20160101, value: "Is this thing on?" },
    { date: 20160102, value: "What should I write?" },
    { date: 20160202, value: "This is the new body" }
  ],
  tags: [
    { date: 20160101, value: [ "test", "trivial" ] },
    { date: 20160102, value: [ "foo", "test" ] }
  ],
  comments: [
    {
      author: "joe", // Unversioned field
      body: [
        { date: 20160301, value: "Something cool" }
      ]
    },
    {
      author: "xxx",
      body: [
        { date: 20160101, value: "Spam" },
        { date: 20160102, deleted: true }
      ]
    },
    {
      author: "jim",
      body: [
        { date: 20160101, value: "Not bad" },
        { date: 20160102, value: "Not bad at all" }
      ]
    }
  ]
}