ArangoDB v3.13 is under development and not released yet. This documentation is not final and potentially incomplete.

Multi-dimensional indexes

Multi-dimensional indexes allow you to index two- or higher-dimensional data such as time ranges, for efficient intersection of multiple range queries

A multi-dimensional index maps multi-dimensional data in the form of multiple numeric attributes to one dimension while mostly preserving locality so that similar values in all of the dimensions remain close to each other in the mapping to a single dimension. Queries that filter by multiple value ranges at once can be better accelerated with such an index compared to a persistent index.

You can choose between two subtypes of multi-dimensional indexes:

  • An mdi index with a fields property that describes which document attributes to use as dimensions
  • An mdi-prefixed index with a fields property as well as a prefixFields property to specify one or more mandatory document attributes to narrow down the search space using equality checks, see Prefix fields

Both subtypes require that the attributes described by fields have numeric values. You can optionally omit documents from the index that have any of the fields or prefixFields attributes not set or set to null by declaring the index as sparse with sparse: true.

Multi-dimensional indexes can be created with a uniqueness constraint with unique: true. This only allows a single document with a given combination of attribute values, using all of the fields attributes and (for mdi-prefixed indexes) prefixFields. Documents omitted because of sparse: true are exempt.

You can store additional attributes in multi-dimensional indexes with the storedValues property. They can be used for projections (unlike the fields attributes) so that indexes can cover more queries without having to access the full documents.

Non-unique mdi indexes have a fixed selectivity estimate of 1. For mdi indexes with unique: true as well as for mdi-prefixed indexes, you can control whether index selectivity estimates are maintained for the index. It is enabled by default and you can disable it with estimates: false. Not maintaining index selectivity estimates can have a slightly positive impact on write performance but the query optimizer is not able to determine the usefulness of different competing indexes in AQL queries when there are multiple candidate indexes to choose from.

The mdi index type was previously called zkd.

Querying documents within a 3D box

Assume we have documents in a collection points of the form

{"x": 12.9, "y": -284.0, "z": 0.02}

and we want to query all documents that are contained within a box defined by [x0, x1] * [y0, y1] * [z0, z1].

To do so one creates a multi-dimensional index on the attributes x, y and z, e.g. in arangosh:

db.points.ensureIndex({
  type: "mdi",
  fields: ["x", "y", "z"],
  fieldValueTypes: "double"
});

Unlike with other indexes, the order of the fields does not matter.

fieldValueTypes is required and the only allowed value is "double" to use a double-precision (64-bit) floating-point format internally.

Now we can use the index in a query:

FOR p IN points
  FILTER x0 <= p.x && p.x <= x1
  FILTER y0 <= p.y && p.y <= y1
  FILTER z0 <= p.z && p.z <= z1
  RETURN p

Possible range queries

Having an index on a set of fields does not require you to specify a full range for every field. For each field you can decide if you want to bound it from both sides, from one side only (i.e. only an upper or lower bound) or not bound it at all.

Furthermore you can use any comparison operator. The index supports <= and >= naturally, == will be translated to the bound [c, c]. Strict comparison is translated to their non-strict counterparts and a post-filter is inserted.

FOR p IN points
  FILTER 2 <= p.x && p.x < 9
  FILTER p.y >= 80
  FILTER p.z == 4
  RETURN p

The null value is less than all other values in AQL. Therefore, range queries without a lower bound need to include null but sparse indexes do not include null values. For example, FILTER p.x < 9 sets an upper bound, but unless you also set a lower bound like AND p.x >= 2, the lower bound is basically p.x >= null.

You can explicitly exclude null in range queries so that sparse indexes can be utilized: FILTER p.x < 9 AND p.x != null

Example Use Case

If you build a calendar using ArangoDB you could create a collection for each user that contains the appointments. The documents would roughly look as follows:

{
  "from": 345365,
  "to": 678934,
  "what": "Dentist",
}

from/to are the timestamps when an appointment starts/ends. Having an multi-dimensional index on the fields ["from", "to"] allows you to query for all appointments within a given time range efficiently.

Finding all appointments within a time range

Given a time range [f, t] we want to find all appointments [from, to] that are completely contained in [f, t]. Those appointments clearly satisfy the condition

f <= from and to <= t

Thus our query would be:

FOR app IN appointments
  FILTER f <= app.from
  FILTER app.to <= t
  RETURN app

Finding all appointments that intersect a time range

Given a time range [f, t] we want to find all appointments [from, to] that intersect [f, t]. Two intervals [f, t] and [from, to] intersect if and only if

f <= to and from <= t

Thus our query would be:

FOR app IN appointments
  FILTER f <= app.to
  FILTER app.from <= t
  RETURN app

Prefix fields

Multi-dimensional indexes can accelerate range queries well but they are inefficient for queries that check for equality of values. For use cases where you have a combination of equality and range conditions in queries, you can use the mdi-prefixed subtype instead of mdi. It has all the features of the mdi subtype but additionally lets you define one or more document attributes you want to use for equality checks. This combination allows to efficiently narrow down the search space to a subset of multi-dimensional index data before performing the range checking.

The attributes for equality checking are specified via the prefixFields property of mdi-prefix indexes. These attributes can have non-numeric values, unlike the attributes you use as fields.

db.<collection>.ensureIndex({
  type: "mdi-prefixed",
  prefixFields: ["v", "w"]
  fields: ["x", "y"],
  fieldValueTypes: "double"
});

You need to specify all of the prefixFields attributes in your queries to utilize the index.

FOR p IN points
  FILTER p.v == "type"
  FILTER p.w == "group"
  FILTER 2 <= p.x && p.x < 9
  FILTER p.y >= 80
  RETURN p

You can create mdi-prefixed indexes on edge collections with the _from or _to edge attribute as the first prefix field. Graph traversals with range filters can then utilize such indexes. See Vertex-centric indexes for details.

Storing additional values in indexes

Introduced in: v3.12.0

Multi-dimensional indexes allow you to store additional attributes in the index that can be used to satisfy projections of the document. They cannot be used for index lookups or for sorting, but for projections only. They allow multi-dimensional indexes to fully cover more queries and avoid extra document lookups. This can have a great positive effect on index scan performance if the number of scanned index entries is large.

You can set the storedValues option and specify the additional attributes as an array of attribute paths when creating a new mdi or mdi-prefixed index, similar to the fields option:

db.<collection>.ensureIndex({
  type: "mdi",
  fields: ["x", "y"],
  fieldValueTypes: "double",
  storedValues: ["y", "z"]
});

This indexes the x and y attributes so that the index can be used for range queries by these attributes. Using these document attributes like for returning them from the query is not covered by the index, however, unless you add the attributes to storedValues in addition to fields. The reason is that the index doesn’t store the original values of the attributes.

You can have the same attributes in storedValues and fields as the attributes in fields cannot be used for projections, but you can also store additional attributes that are not listed in fields. The above example stores the y and z attribute values in the index using storedValues. The index can thus supply the values for projections without having to look up the full document.

Attributes in storedValues cannot overlap with the attributes specified in prefixFields. There is no reason to store them in the index because you need to specify them in queries in order to use mdi-prefixed indexes.

In unique indexes, only the index attributes in fields and (for mdi-prefixed indexes) prefixFields are checked for uniqueness. The index attributes in storedValues are not checked for their uniqueness.

You cannot create multiple multi-dimensional indexes with the same sparse, unique, fields and (for mdi-prefixed indexes) prefixFields attributes but different storedValues settings. That means the value of storedValues is not considered by index creation calls when checking if an index is already present or needs to be created.

Non-existing attributes are stored as null values.

The maximum number of attributes that you can use in storedValues is 32.

Lookahead Index Hint

Introduced in: v3.10.0

Using the lookahead index hint can increase the performance for certain use cases. Specifying a lookahead value greater than zero makes the index fetch more documents that are no longer in the search box, before seeking to the next lookup position. Because the seek operation is computationally expensive, probing more documents before seeking may reduce the number of seeks, if matching documents are found. Please keep in mind that it might also affect performance negatively if documents are fetched unnecessarily.

You can specify the lookahead value using the OPTIONS keyword:

FOR app IN appointments OPTIONS { lookahead: 32 }
    FILTER @to <= app.to
    FILTER app.from <= @from
    RETURN app

Limitations

  • Using array expansions for attributes is not possible (e.g. array[*].attr)
  • You can only index numeric values that are representable as IEEE-754 double.
  • A high number of dimensions (more than 5) can impact the performance considerably.
  • The performance can vary depending on the dataset. Densely packed points can lead to a high number of seeks. This behavior is typical for indexing using space filling curves.