Features and Improvements in ArangoDB 3.0

Redesigned cluster core architecture, graph traversals using the AQL query language, persistent indexes, overhauled Foxx microservices and web interface

The following list shows in detail which features have been added or improved in ArangoDB 3.0. ArangoDB 3.0 also contains several bugfixes that are not listed here.

Internal data format changes

ArangoDB now uses VelocyPack  for storing documents, query results and temporarily computed values. Using a single data format removed the need for some data conversions in the core that slowed operations down previously.

The VelocyPack format is also quite compact, and reduces storage space requirements for “small” values such as boolean, integers, short strings. This can speed up several operations inside AQL queries.

VelocyPack document entries stored on disk are also self-contained, in the sense that each stored document will contain all of its data type and attribute name descriptions. While this may require a bit more space for storing the documents, it removes the overhead of fetching attribute names and document layout from shared structures as in previous versions of ArangoDB. It also simplifies the code paths for storing and reading documents.

AQL improvements

Syntax improvements

LIKE string-comparison operator

AQL now provides a LIKE operator and can be used to compare strings like this, for example inside filter conditions:

value LIKE search

This change makes LIKE an AQL keyword. Using LIKE as an attribute or collection name in AQL thus requires quoting the name from now on.

The LIKE operator is currently implemented by calling the already existing AQL function LIKE, which also remains operational in 3.0. Use the LIKE function in case you want to search case-insensitive (optional parameter), as the LIKE operator always compares case-sensitive.

AQL array comparison operators

All AQL comparison operators now also exist in an array variant. In the array variant, the operator is preceded with one of the keywords ALL, ANY or NONE. Using one of these keywords changes the operator behavior to execute the comparison operation for all, any, or none of its left hand argument values. It is therefore expected that the left hand argument of an array operator is an array.


[ 1, 2, 3 ] ALL IN [ 2, 3, 4 ]   // false
[ 1, 2, 3 ] ALL IN [ 1, 2, 3 ]   // true
[ 1, 2, 3 ] NONE IN [ 3 ]        // false
[ 1, 2, 3 ] NONE IN [ 23, 42 ]   // true
[ 1, 2, 3 ] ANY IN [ 4, 5, 6 ]   // false
[ 1, 2, 3 ] ANY IN [ 1, 42 ]     // true
[ 1, 2, 3 ] ANY == 2             // true
[ 1, 2, 3 ] ANY == 4             // false
[ 1, 2, 3 ] ANY > 0              // true
[ 1, 2, 3 ] ANY <= 1             // true
[ 1, 2, 3 ] NONE < 99            // false
[ 1, 2, 3 ] NONE > 10            // true
[ 1, 2, 3 ] ALL > 2              // false
[ 1, 2, 3 ] ALL > 0              // true
[ 1, 2, 3 ] ALL >= 3             // false
["foo", "bar"] ALL != "moo"      // true
["foo", "bar"] NONE == "bar"     // false
["foo", "bar"] ANY == "foo"      // true

Regular expression string-comparison operators

AQL now supports the operators =~ and !~ for testing strings against regular expressions. =~ tests if a string value matches a regular expression, and !~ tests if a string value does not match a regular expression.

The two operators expect their left-hand operands to be strings, and their right-hand operands to be strings containing valid regular expressions as specified below.

The regular expressions may consist of literal characters and the following characters and sequences:

  • . – the dot matches any single character except line terminators. To include line terminators, use [\s\S] instead to simulate . with DOTALL flag.
  • \d – matches a single digit, equivalent to [0-9]
  • \s – matches a single whitespace character
  • \S – matches a single non-whitespace character
  • \t – matches a tab character
  • \r – matches a carriage return
  • \n – matches a line-feed character
  • [xyz] – set of characters. matches any of the enclosed characters (i.e. x, y or z in this case
  • [^xyz] – negated set of characters. matches any other character than the enclosed ones (i.e. anything but x, y or z in this case)
  • [x-z] – range of characters. Matches any of the characters in the specified range, e.g. [0-9A-F] to match any character in 0123456789ABCDEF
  • [^x-z] – negated range of characters. Matches any other character than the ones specified in the range
  • (xyz) – defines and matches a pattern group
  • (x|y) – matches either x or y
  • ^ – matches the beginning of the string (e.g. ^xyz)
  • $ – matches the end of the string (e.g. xyz$)

Note that the characters ., *, ?, [, ], (, ), {, }, ^, and $ have a special meaning in regular expressions and may need to be escaped using a backslash (\\). A literal backslash should also be escaped using another backslash, i.e. \\\\.

Characters and sequences may optionally be repeated using the following quantifiers:

  • x* – matches zero or more occurrences of x
  • x+ – matches one or more occurrences of x
  • x? – matches one or zero occurrences of x
  • x{y} – matches exactly y occurrences of x
  • x{y,z} – matches between y and z occurrences of x
  • x{y,} – matches at least y occurrences of x

Enclosing identifiers in forward ticks

AQL identifiers can now optionally be enclosed in forward ticks in addition to using backward ticks. This allows convenient writing of AQL queries in JavaScript template strings (which are delimited with backticks themselves), e.g.

var q = `FOR doc IN ´collection´ RETURN doc.´name´`;

Functions added

The following AQL functions have been added in 3.0:

  • REGEX_TEST(value, regex): tests whether the string value matches the regular expression specified in regex. Returns true if it matches, and false otherwise.

    The syntax for regular expressions is the same as for the regular expression operators =~ and !~.

  • HASH(value): Calculates a hash value for value. value is not required to be a string, but can have any data type. The calculated hash value will take the data type of value into account, so for example the number 1 and the string “1” will have different hash values. For arrays the hash values will be created if the arrays contain exactly the same values (including value types) in the same order. For objects the same hash values will be created if the objects have exactly the same attribute names and values (including value types). The order in which attributes appear inside objects is not important for hashing. The hash value returned by this function is a number. The hash algorithm is not guaranteed to remain the same in future versions of ArangoDB. The hash values should therefore be used only for temporary calculations, e.g. to compare if two documents are the same, or for grouping values in queries.

  • TYPENAME(value): Returns the data type name of value. The data type name can be either null, bool, number, string, array or object.

  • LOG(value): Returns the natural logarithm of value. The base is Euler’s constant (2.71828…).

  • LOG2(value): Returns the base 2 logarithm of value.

  • LOG10(value): Returns the base 10 logarithm of value.

  • EXP(value): Returns Euler’s constant (2.71828…) raised to the power of value.

  • EXP2(value): Returns 2 raised to the power of value.

  • SIN(value): Returns the sine of value.

  • COS(value): Returns the cosine of value.

  • TAN(value): Returns the tangent of value.

  • ASIN(value): Returns the arcsine of value.

  • ACOS(value): Returns the arccosine of value.

  • ATAN(value): Returns the arctangent of value.

  • ATAN2(y, x): Returns the arctangent of the quotient of y and x.

  • RADIANS(value): Returns the angle converted from degrees to radians.

  • DEGREES(value): Returns the angle converted from radians to degrees.

Optimizer improvements

“inline-subqueries” rule

The AQL optimizer rule “inline-subqueries” has been added. This rule can pull out certain subqueries that are used as an operand to a FOR loop one level higher, eliminating the subquery completely. This reduces complexity of the query’s execution plan and will likely enable further optimizations. For example, the query

FOR i IN (
    FOR j IN [1,2,3]
      RETURN j

will be transformed by the rule to:

FOR i IN [1,2,3]

The query

FOR name IN (
  FOR doc IN _users
    FILTER doc.status == 1
    RETURN doc.name
  RETURN name

will be transformed into

FOR tmp IN _users
  FILTER tmp.status == 1
  RETURN tmp.name

The rule will only fire when the subquery is used as an operand to a FOR loop, and if the subquery does not contain a COLLECT with an INTO variable.

“remove-unnecessary-calculations” rule

The AQL optimizer rule “remove-unnecessary-calculations” now fires in more cases than in previous versions. This rule removes calculations from execution plans, and by having less calculations done, a query may execute faster or requires less memory.

The rule will now remove calculations that are used exactly once in other expressions (e.g. LET a = doc RETURN a.value) and calculations, or calculations that are just references to other variables (e.g. LET a = b).

“optimize-traversals” rule

The AQL optimizer rule “merge-traversal-filter” was renamed to “optimize-traversals”. The rule will remove unused edge and path result variables from the traversal in case they are specified in the FOR section of the traversal, but not referenced later in the query. This saves constructing edges and paths results that are not used later.

AQL now uses VelocyPack internally for storing intermediate values. For many value types it can now get away without extra memory allocations and less internal conversions. Values can be passed into internal AQL functions without copying them. This can lead to reduced query execution times for queries that use C++-based AQL functions.

“replace-or-with-in” and “use-index-for-sort” rules

These rules now fire in some additional cases, which allows simplifying index lookup conditions and removing SortNodes from execution plans.

Cluster state management

The cluster’s internal state information is now also managed by ArangoDB instances. Earlier versions relied on third party software being installed for the storing the cluster state. The state is managed by dedicated ArangoDB instances, which can be started in a special agency mode. These instances can operate in a distributed fashion. They will automatically elect one of them to become their leader, being responsible for storing the state changes sent from servers in the cluster. The other instances will automatically follow the leader and will transparently stand in should it become unavailable. The Agency instances are also self-organizing: they will continuously probe each other and re-elect leaders. The communication between the Agency instances use the consensus-based RAFT protocol.

The operations for storing and retrieving cluster state information are now much less expensive from an ArangoDB cluster node perspective, which in turn allows for faster cluster operations that need to fetch or update the overall cluster state.

_from and _to attributes of edges are updatable and usable in indexes

In ArangoDB prior to 3.0 the attributes _from and _to of edges were treated specially when loading or storing edges. That special handling led to these attributes being not as flexible as regular document attributes. For example, the _from and _to attribute values of an existing edge could not be updated once the edge was created. Now this is possible via the single-document APIs and via AQL.

Additionally, the _from and _to attributes could not be indexed in user-defined indexes, e.g. to make each combination of _from and _to unique. Finally, as _from and _to referenced the linked collections by collection id and not by collection name, their meaning became unclear once a referenced collection was dropped. The collection id stored in edges then became unusable, and when accessing such edge the collection name part of it was always translated to _undefined.

In ArangoDB 3.0, the _from and _to values of edges are saved as regular strings. This allows using _from and _to in user-defined indexes. Additionally, this allows to update the _from and _to values of existing edges. Furthermore, collections referenced by _from and _to values may be dropped and re-created later. Any _from and _to values of edges pointing to such dropped collection are unaffected by the drop operation now.

Unified APIs for CRUD operations

The CRUD APIs for documents and edge have been unified. Edges can now be inserted and modified via the same APIs as documents. _from and _to attribute values can be passed as regular document attributes now:

db.myedges.insert({ _from: "myvertices/some", _to: "myvertices/other", ... });

Passing _from and _to separately as it was required in earlier versions is not necessary anymore but will still work:

db.myedges.insert("myvertices/some", "myvertices/other", { ... });

The CRUD operations now also support batch variants that works on arrays of documents/edges, e.g.

  { _from: "myvertices/some", _to: "myvertices/other", ... },
  { _from: "myvertices/who", _to: "myvertices/friend", ... },
  { _from: "myvertices/one", _to: "myvertices/two", ... },

The batch variants are also available in ArangoDB’s HTTP API. They can be used to more efficiently carry out operations with multiple documents than their single-document equivalents, which required one HTTP request per operation. With the batch operations, the HTTP request/response overhead can be amortized across multiple operations.

Persistent indexes

ArangoDB 3.0 provides an experimental persistent index feature. Persistent indexes store the index values on disk instead of in-memory only. This means the indexes do not need to be rebuilt in-memory when a collection is loaded or reloaded, which should improve collection loading times.

The persistent indexes in ArangoDB are based on the RocksDB engine. To create a persistent index for a collection, create an index of type “rocksdb” as follows:

db.mycollection.ensureIndex({ type: "rocksdb", fields: [ "fieldname" ]});

The persistent indexes are sorted, so they allow equality lookups and range queries. Note that the feature is still highly experimental and has some known deficiencies. It will be finalized until the release of the 3.0 stable version.

Upgraded V8 version

The V8 engine that is used inside ArangoDB to execute JavaScript code has been upgraded from version 4.3.61 to The new version makes several more ES6 features available by default, including

  • arrow functions
  • computed property names
  • rest parameters
  • array destructuring
  • numeric and object literals

Web Admin Interface

The ArangoDB 3.0 web interface is significantly improved. It now comes with a more responsive design, making it easier to use on different devices. Navigation and menus have been simplified, and related items have been regrouped to stay closer together and allow tighter workflows.

The AQL query editor is now much easier to use. Multiple queries can be started and tracked in parallel, while results of earlier queries are still preserved. Queries still running can be canceled directly from the editor. The AQL query editor now allows the usage of bind parameters too, and provides a helper for finding collection names, AQL function names and keywords quickly.

The web interface now keeps track of whether the server is offline and of which server-side operations have been started and are still running. It now remains usable while such longer-running operations are ongoing. It also keeps more state about user’s choices (e.g. windows sizes, whether the tree or the code view was last used in the document editor).

Cluster statistics are now integrated into the web interface as well. Additionally, a menu item “Help us” has been added to easily provide the ArangoDB team feedback about the product.

The frontend may now be mounted behind a reverse proxy on a different path. For this to work the proxy should send a X-Script-Name header containing the path.

A backend configuration for haproxy might look like this:

reqadd X-Script-Name:\ /arangodb

The frontend will recognize the subpath and produce appropriate links. ArangoDB will only accept paths from trusted frontend proxies. Trusted proxies may be added on startup:

--frontend.proxy-request-check true --frontend.trusted-proxy

--frontend.trusted-proxy may be any address or netmask.

To disable the check and blindly accept any x-script-name set --frontend.proxy-request-check to false.

Foxx improvements

The Foxx framework has been completely rewritten for 3.0 with a new, simpler and more familiar API. The most notable changes are:

  • Legacy mode for 2.8 services

    Stuck with old code? You can continue using your 2.8-compatible Foxx services with 3.0 by adding "engines": {"arangodb": "^2.8.0"} (or similar version ranges that exclude 3.0 and up) to the service manifest.

  • No more global variables and magical comments

    The applicationContext is now module.context. Instead of magical comments just use the summary and description methods to document your routes.

  • Repository and Model have been removed

    Instead of repositories just use ArangoDB collections directly. For validation simply use the joi schemas (but wrapped in joi.object()) that previously lived inside the model. Collections and queries return plain JavaScript objects.

  • Controllers have been replaced with nestable routers

    Create routers with require('@arangodb/foxx/router')(), attach them to your service with module.context.use(router). Because routers are no longer mounted automagically, you can export and import them like any other object. Use router.use('/path', subRouter) to nest routers as deeply as you want.

  • Routes can be named and reversed

    No more memorizing URLs: add a name to your route like router.get('/hello/:name', function () {...}, 'hello') and redirect to the full URL with res.redirect(req.resolve('hello', {name: 'world'})).

  • Simpler express-like middleware

    If you already know express, this should be familiar. Here’s a request logger in three lines of code:

    router.use(function (req, res, next) {
      var start = Date.now();
      try {next();}
      finally {console.log(`${req.method} ${req.url} ${res.statusCode} ${Date.now() - start}ms`);}
  • Sessions and auth without dependencies

    To make it easier to get started, the functionality previously provided by the simple-auth, oauth2, sessions-local and sessions-jwt services have been moved into Foxx as the @arangodb/foxx/auth, @arangodb/foxx/oauth2 and @arangodb/foxx/sessions modules.


ArangoDB’s logging is now grouped into topics. The log verbosity and output files can be adjusted per log topic. For example

--log.level startup=trace --log.level queries=trace --log.level info

will log messages concerning startup at trace level, AQL queries at trace level and everything else at info level. --log.level can be specified multiple times at startup, for as many topics as needed.

Some relevant log topics available in 3.0 are:

  • collector: information about the WAL collector’s state
  • compactor: information about the collection datafile compactor
  • datafiles: datafile-related operations
  • mmap: information about memory-mapping operations (including msync)
  • queries: executed AQL queries, slow queries
  • replication: replication-related info
  • requests: HTTP requests
  • startup: information about server startup and shutdown
  • threads: information about threads

This also allows directing log output to different files based on topics. For example, to log all AQL queries to a file “queries.log” one can use the options:

--log.level queries=trace --log.output queries=file:///path/to/queries.log

To additionally log HTTP request to a file named “requests.log” add the options:

--log.level requests=info --log.output requests=file:///path/to/requests.log

Build system

ArangoDB now uses the cross-platform build system CMake for all its builds. Previous versions used two different build systems, making development and contributions harder than necessary. Now the build system is unified, and all targets (Linux, Windows, macOS) are built from the same set of build instructions.


The documentation has been enhanced and re-organized to be more intuitive.

A new introduction for beginners should bring you up to speed with ArangoDB in less than an hour. Additional topics have been introduced and will be extended with upcoming releases.

The topics AQL and HTTP API are now separated from the manual for better searchability and less confusion. A version switcher makes it easier to jump to the version of the docs you are interested in.