see on GitHub

Change Vector

What is a Change Vector?

Change vectors is the RavenDB implementation of the Vector clock concept. They give us partial order over modifications of documents in a RavenDB cluster.

What are Change Vectors Constructed From?

A change vector is constructed from entries, one per database. . It looks like this:
[A:1-0tIXNUeUckSe73dUR6rjrA, B:7-kSXfVRAkKEmffZpyfkd+Zw]

This change vector is constructed from two entries,

A:1-0tIXNUeUckSe73dUR6rjrA and B:7-kSXfVRAkKEmffZpyfkd+Zw.

Each entry has the following structure Node tag:ETag-Database ID, so A:1-0tIXNUeUckSe73dUR6rjrA means that the document was modified on node A its local ETag is 1 and the database ID is 0tIXNUeUckSe73dUR6rjrA.

An ETag is a value that is scoped to a database instance on a particular node and is guranteed to always increase. It is used internally for many purposes and is the natural sort order for many operations (indexing, replication, ETL, etc). The database ID is used for cases where the node tag is not unique, which could happen when using external replication or restoring from backup.

How are Change Vectors used to Determine Order?

Given two change vectors X and Y we would say that X >= Y if foreach entry of X the value of the ETag is greater or equal to the corresponding entry in Y and Y has no entries that X doesn't have.
The same goes for <= we would say that X <= Y if foreach entry of X the value of the ETag is smaller or equal to the corresponding entry in Y and X has no entries that Y doesn't have.
We would say that X <> Y (no order or conflicted) if X has an entry with a higher ETag value than Y and Y has a different entry where its ETag value is greater than X.

Concurrency Control & Change Vectors

RavenDB defines some simple rules to determine how to handle concurrent operations on the same document across the cluster. It uses the document's change vector. This allows RavenDB to detect concurrency issues when writing to a document and conflicts when the document have been concurrently modified on different nodes during network partition.

Every document in RavenDB has a corresponding change vector. This change vector is updated by RavenDB every time the document is changed. This happens when on document creation and any modification such as PUT, PATCH, 'DELETE' or their bulk versions. A delete operation will also cause RavenDB to update the document change vector, however, at that point the change vector will belong to the document tombstone (since the document itself has already been deleted).

A change vector is present in the document's metadata and each time a document is updated, the server will update the change vector. This is mostly used internally inside RavenDB for many purposes (conflict detection, deciding what documents a particular subscription has already seen, what was sent to an ETL destination, etc) but can also be very useful for clients.

In particular, the change vector is guaranteed to change whenever the document changes and can be used as part of optimistic concurrency checks. A document modification can all specify an expected change vector for a document (with an empty change vector signifying that the document does not exists). In such a case, all operations in the transaction will be aborted and no changes will be applied to any of the documents modified in the transaction.

Change Vector Comparisons

Conceptually, comparing two change vectors means answering a question - which change vector refers to an earlier event.

The comparison is defined as follows:

  • Two change vectors are equal when and only when all etags equal between corresponding node and database IDs
  • Change vector A is larger than change vector B if and only if all etags are larger or equal between corresponding node and database IDs
  • Change vector A conflicts with change vector B if and only if at least one etags is larger, equal or has node etag (and the other don't) and at least one etags is smaller between corresponding node and database IDs

Note that the change vectors are unsortable for two reasons:

  • Change vectors are unsorted collections of node tags/etag tuples, they can be sorted in multiple ways
  • Conflicting change vectors cannot be compared

Example 1

Let us assume two change vectors, v1 = [A:8, B:10, C:34], v2 = [A:23, B:12, C:65]
When we compare v1 and v2, we will do three comparisons:

  • A --> 8 (v1) < 23 (v2)
  • B --> 10 (v1) < 12 (v2)
  • C --> 34 (v1) < 65 (v2)

Corresponding etags in v2 are greater than the ones in v1. This means that v1 < v2

Example 2

Let us assume two change vectors, v1 = [A:18, B:12, C:51], v2 = [A:23, B:12, C:65]
When we compare v1 and v2, we will do three comparisons:

  • A --> 18 (v1) < 23 (v2)
  • B --> 12 (v1) = 12 (v2)
  • C --> 51 (v1) < 65 (v2)

Corresponding etags in v2 are greater than the ones in v1. This means that v1 < v2

Example 3

Let us assume two change vectors, v1 = [A:18, B:12, C:65], v2 = [A:58, B:12, C:51]
When we compare v1 and v2, we will do three comparisons:

  • A --> 18 (v1) < 58 (v2)
  • B --> 12 (v1) = 12 (v2)
  • C --> 65 (v1) > 51 (v2)

Etag 'A' in v1 is smaller than in v2, and Etag 'C' is larger in v1 than in v2. This means that v1 conflicts with v2.

Concurrency Control at the Cluster

RavenDB implements a multi master strategy for handling database writes. This means that it will never reject a valid write to a document (under the assumption that if you tried to write data to the database, you're probably interested in keeping it). This behavior can lead to certain edge cases. In particular, under network partition scenario, it is possible for two clients to talk to two RavenDB nodes and to update a document with optimistic concurrency check.

The concurrency check is done at the local node level, to ensure that we can still process writes in the case of a network partition or partial failure scenario. That can mean that two writes to separate servers will both succeed, even if each write specified the same original change vector, because each server did the check independently. Under such a scenario, the generated change vectors for the document on each server will be different, and as soon as replication between these nodes will run the databases will detect this conflict and resolve it according to your conflict resolution strategy.

In practice, this kind of scenario is rare, since RavenDB attempts to direct all writes to the same node for each database under normal conditions to ensure that optimistic concurrency checks will always happen on the same machine.

Database Global Change Vector

The database global change vector is the same entity as the one used by the data it contains.
The value of the global change vector entries is determined by the maximum value of all the change vectors contained in its data.

E.g, a database containing the following documents:

  • Document A with change vector [A:1-0tIXNUeUckSe73dUR6rjrA, B:7-kSXfVRAkKEmffZpyfkd+Zw]
  • Document B with change vector [B:3-kSXfVRAkKEmffZpyfkd+Zw, C:13-ASFfVrAllEmzzZpyrtlrGq]

Will have the following global change vector:

  • [A:1-0tIXNUeUckSe73dUR6rjrA, B:7-kSXfVRAkKEmffZpyfkd+Zw, C:13-ASFfVrAllEmzzZpyrtlrGq]

The global change vector is used in the replication process to determine which data is already contained by the database. A document that all of his entries are lower or equal from/to the global change vector is considered contained.

Note that if data is considered contained by a database it doesn't mean it is present in the database, it may already be overwritten by more up-to-date data.