CouchDB: Document Conflicts

Three years ago, if someone asked me what couchDB was, I’d probably tell them it was some fancy piece of furniture. But that is three years ago, and since then, I have started a career, got married, learnt a few things, and most importantly realized that couchDB has nothing to do with furniture.

CouchDB is in fact, a NOSQL database that plays an integral role in my current company’s tech stack. It has also been the source of many headaches over the years; And having messed around with it for the better part of three years, I have picked up a couple of things about couchDB that might help you avoid a few headaches of your own.

But first, lets first understand three key aspects of couchDB: The document and the revision tree, the changes feed, and the conflict resolution model.

The Document and the Revision Tree

Data in couchDB is represented as individual documents. A document is simply a JSON structure with a set of key value pairs, and a unique key for identification. Any update performed on a document will be tracked via a revision number.

A bare-bones couchDB document may look like this:

    "_id": "my_document",
    "_rev": "3-a",
    "name": "Priyath Gregory",
    "age": 28

CouchDB tracks the revision history of a document via a revision tree. Each document will have its own revision tree, with each node corresponding to a previous revision of the document.

For example, my_document (updated thrice) may have a revision tree similar to this:

1-a -> 2-a -> 3-a

The changes Feed

CouchDB tracks updates to its database via the changes feed. The changes feed is a list of all the changes made to the documents in the database, in the order they were made. If you have access to a couchDB instance, you can use the _changes resource to obtain the changes feed of the corresponding database.

A typical request and response may look like this:

curl -X GET

Each entry or sequence in the results array represents a document update of the database. It is important to note that the changes feed only lists the sequence corresponding to the latest revision of a document. For example, if we perform another update on my_document the response would be something like this:


Both sequences 1 and 2 are removed from the result set as they correspond to stale revisions of our updated document.

Replication and Conflict Resolution

Without a doubt, the standout feature of couchDB is in its ability to synchronize two copies of the same database. This is done through an in-built replication protocol that depends on the changes feed, as well as the revision-based update model used by couchDB to track document updates.

For example, consider two synchronized databases, A and B. Any document update on A will be synced to B and vice-versa. This is fairly straightforward, but what happens if the same document is updated in both A and B at the same time? The replication mechanism will ensure that both revisions of the document are copied to databases A and B; but clearly, both revisions cannot exist together.

my_document revision tree:

                  4-a (conflicting update performed in database A)
1-a -> 2-a -> 3-a 
                  4-b (conflicting update performed in database B)

Enter couchDB’s conflict resolution model.

Conflicting revisions result in conflicting documents. And it is up-to couchDB’s conflict resolution model to fix this. The fix is fairly simple. The conflict is resolved by arbitrarily selecting a winning revision from the two conflicting revisions, while marking the other as conflicted. CouchDB does not try to merge these conflicting revisions.

A conflict resolved document’s revision tree may look like this:

                  4-a (winning revision)
1-a -> 2-a -> 3-a 
                  4-b (conflicted revision)

Now for all intents and purposes, the winning revision will remain as the active version of the document across both databases A and B.. but what many users do not realize is that the conflicted revision will also “live” within the databases in a dormant state.

And herein lies the root cause for many issues that we have faced with regards to couchDB, the existence (or rather, our ignorance to the existence of!) conflicting revisions.

Conflicting Revisions and Document Deletion

Deletes in couchDB are soft deletes. What this means is a deleted document in couchDB will still remain in the database, just with the a flag (_deleted) set to true. Any attempt at retrieval of a deleted document will result in a response similar to this:



There’s also a more extreme version of deletion supported by couchDB called the purge.

Purging a document will remove all traces of it from the database. Purging is generally not recommended since couchDB replication (for that specific document), completely falls apart once a purge is performed.

Things get interesting when we delete (or purge) a document that has conflicting revisions.

Consider the deletion of our document my_document, with the latest revision 4-a and a conflicting revision 4-b. It is logical to assume that post-deletion, the document will be.. well, deleted from the database.

But couchDB is weird, and it does weird things.

What actually happens is, once we delete the document by specifying its latest revision (4-a in this case), the revision will get deleted as expected.. BUT the dormant conflicting leaf revision 4-b will immediately revert as the active revision of the document. And for all intents and purposes, my_document will still remain active in the database with 4-b as its latest revision.


                  4-a (winning revision)
1-a -> 2-a -> 3-a 
                  4-b (conflicted revision)


                  4-a (deleted revision)
1-a -> 2-a -> 3-a 
                  4-b (winning revision)

The implications of this if not handled properly can be severe. Deleted items in your web application may seemingly re-appear with a completely random state from the past.

Handling Document Conflicts

  • First and foremost, try to avoid document conflicts like the plague. If you are running a single instance of couchDB, this is a no-issue. CouchDB will reject a conflicting document update with a 409 response. But if you have multiple database copies in sync through replication, and if each instance can receive its own document updates, understand that you will always be susceptible to document conflicts.
  • Make sure to specify all leaf revisions of the revision tree during document deletion. This can be accomplished by using GET /db/documentId?conflicts=true to obtain a response with the conflicted leaf revisions of the document, and then specifying a bulk deletion of the retrieved revisions using the POST /{db}/_bulk_docs endpoint.
  • There can also be cases where a conflicting revision may actually contain important data. After all, the winning revision is an arbitrary selection. If this is the case, you can:
    • retrieve all conflicting revisions of the document
    • Retrieve the document content corresponding to each conflicting revision.
    • Perform application-specific merging
    • Use the _bulk_docs endpoint to update the first revision of the document (with the merged results) while deleting the others.
  • You can also use an application-specific custom flag to mark deleted documents without performing an actual delete. The downside of this approach is you will need application logic as well as a mechanism in place to handle your “deleted” documents in a sensible manner.

If you are interested in a more detailed look at what we have discussed here, I suggest you check out the official couchDB documentation on Replication and the Conflict Model.

One Response to“CouchDB: Document Conflicts”

  1. Lahiru Thennakoon
    February 1, 2020 at 2:49 am #

    Great post.Really well explained.
    Thank you for sharing .

Comments are closed.