Features and Improvements in ArangoDB 3.9

The following list shows in detail which features have been added or improved in ArangoDB 3.9. ArangoDB 3.9 also contains several bug fixes that are not listed here.

Hybrid (Disjoint) SmartGraphs (Enterprise Edition)

SmartGraphs have been extended with a new option to create Hybrid SmartGraphs. Hybrid SmartGraphs are capable of using SatelliteCollections within their graph definition and therefore can make use of all the benefits of SatelliteCollections.

Edge definitions can now be created between SmartCollections and SatelliteCollections. As SatelliteCollections are globally replicated to each participating DB-Server, regular graph traversals, weighted traversals, shortest path, and k shortest paths queries can partially be executed locally on each DB-Server. This means that query execution can be fully local whenever actual data from the SatelliteCollections is being processed. This can improve data locality and reduce the number of network hops between cluster nodes.

In case you do have collections that are needed in almost every traversal but are small enough to be copied over to every participating DB-Server, Hybrid SmartGraphs are the perfect fit, as this will increase the amount of local query execution.

Hybrid SmartGraphs can also be disjoint. A Disjoint SmartGraph prohibits edges connecting different SmartGraph components. The same rule applies to Hybrid Disjoint SmartGraphs. If your graph does not need edges between vertices with different SmartGraph attribute values, then you should enable this option. This topology restriction allows the query optimizer to improve traversal execution times, because the execution can be pushed down to a single DB-Server in many cases.

Hybrid SmartGraphs are only available in the Enterprise Edition.


Segmentation and Collation Analyzers

The new segmentation Analyzer type allows you to tokenize text in a language-agnostic manner as per Unicode Standard Annex #29, making it suitable for mixed language strings. It can optionally preserve all non-whitespace or all characters instead of keeping alphanumeric characters only, as well as apply case conversion.

The collation Analyzer converts the input into a set of language-specific tokens. This makes comparisons follow the rules of the respective language, most notable in range queries against Views.



Analyzers in Web Interface

A new menu item ANALYZERS has been added to the side navigation bar of the Web UI. Through this page, you can view existing Analyzers as well create new Analyzers. The UI is full-featured and lets you feed in all parameters and options that you could otherwise input through the HTTP or JavaScript API.

It also lets you copy configuration from an existing Analyzer, allowing for a much quicker workflow when your new Analyzer is very similar to an existing one.

It offers two edit/view modes - a form mode where a standard web form is used to capture user input, and a JSON mode where experienced users can directly write the raw Analyzer configuration in JSON format.

Additional Fields in Views Creation Form

ArangoSearch Views have 6 immutable fields (apart from name and type) that can only be set once at the time of creation. The web interface now includes these fields in the View creation form, so that you can set them when creating Views through the UI:

  • primarySort
  • primarySortCompression
  • storedValues
  • writebufferIdle
  • writebufferActive
  • writebufferSizeMax

Web interface session handling

The previously inactive startup parameter --server.session-timeout was revived and now controls the timeout for web interface sessions (and other sessions that are based on JWTs created by the /_open/auth API).

For security reasons, the default timeout value for web interface sessions has been reduced to one hour, after which a session is ended automatically. Web interface sessions that are active (i.e. that have any user activity) are automatically extended until the user ends the session explicitly or if there is a period of one hour without any user activity.

The timeout value for web interface sessions can be adjusted via the --server.session-timeout startup parameter (in seconds).

Configurable root redirect

Added two options to arangod to allow HTTP redirection customization for root (/) call of the HTTP API:

  • --http.permanently-redirect-root: if true (default), use a permanent redirection (use HTTP 301 code), if false fall back to temporary redirection (use HTTP 302 code).

  • --http.redirect-root-to: redirect of root URL to a specified path. Redirects to /_admin/aardvark/index.html if not set (default).

These options are useful to override the built-in web interface with some user-defined action.


Prune Variable

Added an option to store PRUNE expression as a variable. Now, a PRUNE condition can be stored in a variable and be used later in the query without having to repeat the PRUNE condition:

FOR v, e, p IN 10 OUTBOUND @start GRAPH "myGraph"
  PRUNE pruneCondition = v.isRelevant == true
  FILTER pruneCondition

The condition v.isRelevant == true is stored in the variable pruneCondition, and later used as a condition for FILTER.

See Pruning.

Upsert with Index Hint

Added support for the indexHint and forceIndexHint options to the UPSERT operation. It will be used as a hint for the document lookup that is performed as part of the UPSERT operation, and can help in cases such as UPSERT not picking the best index automatically.

UPSERT { a: 1234 }
  INSERT { a: 1234, name: "AB"}
  UPDATE {name: "ABC"} IN myCollection
  OPTIONS { indexHint: "index_name", forceIndexHint: true }

See UPSERT Options

Decay Functions

Added three decay functions to AQL:

Decay functions calculate a score with a function that decays depending on the distance of a numeric value from a user given origin.

DECAY_GAUSS(41, 40, 5, 5, 0.5) // 1
DECAY_LINEAR(5, 0, 10, 0, 0.2) // 0.6
DECAY_EXP(2, 0, 10, 0, 0.2)    // 0.7247796636776955

Vector Functions

Added three vector functions to AQL for calculating the cosine similarity, Manhattan distance, and Euclidean distance:

COSINE_SIMILARITY([0,1], [1,0]) // 0
L1_DISTANCE([-1,-1], [2,2]) // 6
L2_DISTANCE([1,1], [5,2]) // 4.1231056256176606

Traversal filtering optimizations

A post-filter on the vertex and/or edge result of a traversal will now be applied during the traversal to avoid generating the full output for AQL. This will have a positive impact on performance when filtering on the vertex/edge but still returning the path.

Previously all paths were produced even for non-matching vertices/edges. The new optimization now will check on the vertex/edge filter condition first and only produce the remaining paths.

For example, the query

FOR v, e, p IN 10 OUTBOUND @start GRAPH "myGraph"
  FILTER v.isRelevant == true

can now be optimized, and the traversal statement will only produce paths for which the last vertex satisfied isRelevant == true.

Traversal partial path buildup

There is now a performance optimization for traversals in which the path is returned, but only a specific sub-attribute of the path is used later (e.g. vertices, edges, or weight sub-attribute).

For example, the query

FOR v, e, p IN 1..3 OUTBOUND @start GRAPH "myGraph"
  RETURN p.vertices

only requires the buildup of the vertices sub-attribute of the path result, but not the buildup of the edges sub-attribute.

This optimization should have a positive impact on performance for larger traversal result sets.

Warnings on invalid OPTIONS

Invalid use of OPTIONS in AQL queries will now raise a warning when the query is parsed. This is useful to detect misspelled attribute names in OPTIONS, e.g.

INSERT ... INTO collection
  OPTIONS { overwrightMode: 'ignore' } /* should have been 'overwriteMode' */

It is also useful to detect the usage of valid OPTIONS attribute names that are used at a wrong position in the query, e.g.

FOR doc IN collection
  FILTER doc.value == 1234
  INSERT doc INTO other
    OPTIONS { indexHint: 'myIndex' } /* should have been used above for FOR */

In case options are used incorrectly, a warning with code 1575 will be raised during query parsing or optimization. By default, warnings are reported but do not lead to the query being aborted. This can be toggled by the startup option --query.fail-on-warnings or the per-query runtime option failOnWarnings.

Memory usage tracking

The AQL operations K_SHORTEST_PATHS and SHORTEST_PATH are now included in the memory usage tracking performed by AQL, so that memory acquired by these operations will be accounted for and checked against the configured memory limit (options --query.memory-limit and --query.memory-limit-global).

Execution of complex queries

Very large queries (in terms of query execution plan complexity) are now split into multiple segments that are executed using separate stacks. This avoids potential stack overflow. The number of execution nodes after that such stack splitting is performed can be configured via the startup option --query.max-nodes-per-callstack. The default value is 200 for macOS, and 250 for the other supported platforms. The value can be adjusted per query via the maxNodesPerCallstack query option.

Query complexity limits

AQL now has some hard-coded query complexity limits, to prevent large programmatically generated queries from causing trouble (too deep recursion, enormous memory usage, long query optimization and distribution passes etc.).

The following limits have been introduced:

  • a recursion limit for AQL query expressions. An expression can now be up to 500 levels deep. An example expression is 1 + 2 + 3 + 4, which is 3 levels deep 1 + (2 + (3 + 4)). The expression recursion is limited to 500 levels.
  • a limit for the number of execution nodes in the initial query execution plan. The number of execution nodes is limited to 4,000.

RocksDB block cache control

The new query option fillBlockCache can be used to control the population of the RocksDB block cache with data read by the query. The default value for this per-query option is true, which means that any data read by the query will be inserted into the RocksDB block cache if not already present in there. This mimics the previous behavior and is a sensible default.

Setting the option to false allows to not store any data read by the query in the RocksDB block cache. This is useful for queries that read a lot of (cold) data which would lead to the eviction of the hot data from the block cache.

Multi-dimensional Indexes (experimental)

ArangoDB 3.9 features a new index type zkd. It can be created like other indexes on collections. In contrast to the persistent index type (same for hash and skiplist, which today are just aliases for persistent), it lifts the following restriction.

A persistent index can only be used with query filters where a conjunction of equalities on a prefix of indexed fields covers the filter. For example, given a collection with a persistent index on the fields ["a", "b"]. Then the following filters can be satisfied by the index:

  • FILTER doc.a == @a
  • FILTER doc.a == @a && doc.b == @b
  • FILTER doc.a == @a && @bl <= doc.b && doc.b <= @bu

While the following filters cannot, or only partially, be satisfied by a persistent index:

  • FILTER doc.b == @b
  • FILTER @bl <= doc.b && doc.b <= @bu
  • FILTER @al <= doc.a && doc.a <= @au && @bl <= doc.b && doc.b <= @bu

A zkd index can be used to satisfy them all. An example where this is useful are documents with an assigned time interval, where a query should find all documents that contain a given time point, or overlap with some time interval.

There are also drawbacks in comparison with persistent indexes. For one, the zkd index is not sorted. Secondly, it has a significantly higher overhead, and the emerging performance is much more dependent on the distribution of the dataset, making it less predictable.

Multi-dimensional Indexes are an experimental feature.

Server options

Extended naming convention for databases

There is a new startup option --database.extended-names-databases to allow database names to contain most UTF-8 characters. This feature is experimental in ArangoDB 3.9, but will become the norm in a future version.

Running the server with the option enabled provides support for database names that are not comprised within the ASCII table, such as Japanese or Arabic letters, emojis, letters with accentuation. Also, many ASCII characters that were formerly banned in the traditional naming convention are now accepted.

Example database names that can be used with the new naming convention: "España", "😀", "犬", "كلب", "@abc123", "København", "München", "Россия", "abc? <> 123!"

The ArangoDB client tools arangobench, arangodump, arangoexport, arangoimport, arangorestore, and arangosh ship with full support for the extended database naming convention.

Note that the default value for --database.extended-names-databases is false for compatibility with existing client drivers and applications that only support ASCII names according to the traditional database naming convention used in previous ArangoDB versions. Enabling the feature may lead to incompatibilities up to the ArangoDB instance becoming inaccessible for such drivers and client applications.

Please be aware that dumps containing extended database names cannot be restored into older versions that only support the traditional naming convention. In a cluster setup, it is required to use the same database naming convention for all Coordinators and DB-Servers of the cluster. Otherwise the startup will be refused. In DC2DC setups it is also required to use the same database naming convention for both datacenters to avoid incompatibilities.

Also see Database Naming Conventions.


The server now has two flags for retaining or escaping control and Unicode characters in the log. The flag --log.escape is now deprecated and, instead, the new flags --log.escape-control-chars and --log.escape-unicode-chars should be used.

  • --log.escape-control-chars:

    This flag applies to the control characters, that have hex codes below \x20, and also the character DEL with hex code \x7f.

    When the flag value is set to false, control characters will be retained when they have a visible representation, and replaced with a space character in case they do not have a visible representation. For example, the control character \n is visible, so a \n will be displayed in the log. Contrary, the control character BEL is not visible, so a space will be displayed instead.

    When the flag value is set to true, the hex code for the character is displayed, for example, the BEL character will be displayed as its hex code, \x07.

    The default value for this flag is true to ensure compatibility with previous versions.

  • --log.escape-unicode-chars:

    If its value is set to false, Unicode characters will be retained and written to the log as-is. For example, will be logged as . If the flag value is set to true, any Unicode characters are escaped, and the hex codes for all Unicode characters are logged instead. For example, would be logged as its hex code, \u72AC.

    The default value for this flag is set to false for compatibility with previous versions.

Also see Logging.

Version information

The arangod server now provides a command --version-json to print version information in JSON format. This output can be used by tools that need to programmatically inspect an arangod executable.

A pseudo log topic "all" was added. Setting the log level for the “all” log topic will adjust the log level for all existing log topics. For example, --log.level all=debug will set all log topics to log level “debug”.

Overload control

Starting with version 3.9.0, ArangoDB returns an x-arango-queue-time-seconds HTTP header with all responses. This header contains the most recent request queueing/dequeuing time (in seconds) as tracked by the server’s scheduler. This value can be used by client applications and drivers to detect server overload and react on it.

The arangod startup option --http.return-queue-time-header can be set to false to suppress these headers in responses sent by arangod.

In a cluster, the value returned in the x-arango-queue-time-seconds header is the most recent queueing/dequeuing request time of the Coordinator the request was sent to, except if the request is forwarded by the Coordinator to another Coordinator. In that case, the value will indicate the current queueing/dequeuing time of the forwarded-to Coordinator.

In addition, client applications and drivers can optionally augment the requests they send to arangod with the header x-arango-queue-time-seconds. If set, the value of the header should contain the maximum server-side queuing time (in seconds) that the client application is willing to accept. If the header is set in an incoming request, arangod will compare the current dequeuing time from its scheduler with the maximum queue time value contained in the request header. If the current queueing time exceeds the value set in the header, arangod will reject the request and return HTTP 412 (precondition failed) with the error code 21004 (queue time violated). In a cluster, the x-arango-queue-time-seconds request header will be checked on the receiving Coordinator, before any request forwarding.

Support info API

A new HTTP REST API endpoint GET /_admin/support-info was added for retrieving deployment information for support purposes. The endpoint returns data about the ArangoDB version used, the host (operating system, server ID, CPU and storage capacity, current utilization, a few metrics) and the other servers in the deployment (in case of active failover or cluster deployments).

As this API may reveal sensitive data about the deployment, it can only be accessed from inside the _system database. In addition, there is a policy control startup option --server.support-info-api that controls if and to whom the API is made available. This option can have the following values:

  • disabled: support info API is disabled.
  • jwt: support info API can only be accessed via superuser JWT.
  • hardened (default): if --server.harden is set, the support info API can only be accessed via superuser JWT. Otherwise it can be accessed by admin users only.
  • public: everyone with access to the _system database can access the support info API.

Miscellaneous changes

Collection statuses

The previously existing collection statuses “new born”, “loading”, “unloading” and “unloaded” were removed, as they weren’t actively used in arangod.

These statuses were last relevant with the MMFiles storage engine, when it was important to differentiate which collections were present in main memory and which weren’t. With the RocksDB storage engine, all that is automatically handled anyway, and the mentioned statuses are not important anymore.

The “Load” and “Unload” buttons for collections have also been removed from the web interface. This change also obsoletes the load() and unload() calls for collections as well as their HTTP API equivalents. The APIs will remain in place for now for downwards-compatibility but have been changed to no-ops. They will eventually be removed in a future version of ArangoDB.

Cluster-internal timeouts

The internal timeouts for inactive cluster transactions on DB servers was increased from 3 to 5 minutes.

Previously transactions on DB servers could expire quickly, which led to spurious “query ID not found” or “transaction ID not found” errors on DB servers for multi-server queries/transactions with unbalanced access patterns for the different participating DB servers.

Transaction timeouts on Coordinators remain unchanged, so any queries/transactions that are abandoned will be aborted there, which will also be propagated to DB-Servers.

Client tools

Increased default number of threads

The default value for the --threads startup parameter was changed from 2 to the maximum of 2 and the number of available CPU cores for the following client tools:

  • arangodump
  • arangoimport
  • arangorestore

This change can help to improve performance of imports, dumps or restore processes on machines with multiple cores in case the --threads parameter was not previously used. As a trade-off, the change may lead to an increased load on servers, so any scripted imports, dumps or restore processes that want to keep the server load under control should set the number of client threads explicitly when invoking any of the above client tools.


arangoimport received a new startup option --merge-attributes that allows you to create additional attributes in CSV/TSV imports based on other attribute values and hard-coded string literals/separators.

The following example would add a new attribute named fullName that consists of the values of the firstName and lastName columns, separated by a colon character :, as well as as an additional attribute nameAndId that builds on the new fullName attribute and concatenates it with a hyphen - and the value of the id column:

arangoimport \
  --merge-attributes fullName=[firstName]:[lastName] \
  --merge-attributes nameAndId=[fullName]-[id] \

Also see Merging Attributes.

arangoimport also provides a new --datatype startup option, in order to fix the datatypes for certain attributes in CSV/TSV imports. For example, in the the following CSV input file, it is unclear if the numeric values should be imported as numbers or as stringified numbers for the individual attributes:


To determine the datatypes for the individual columns, arangoimport can be invoked with the --datatype startup option, once for each attribute:

--datatype key=string
--datatype price=number
--datatype weight=number
--datatype fk=string

This will turn the numeric-looking values in the key attribute into strings but treat the attributes price and weight as numbers. Finally, the values in attribute fk will be treated as strings again.

See Overriding data types per attribute.


arangobench now prints a short description of the test case started, so it is easier to figure out what operations are carried out by a test case. Several test cases in arangobench have been deprecated because they do not target real world use cases but were rather writing for some internal testing. The deprecated test cases will be removed in a future version to clear up the list of test cases.

arangobench now supports multiple Coordinators. The flag --server.endpoint can be specified multiple times, as in the example below:

arangobench \
  --server.endpoint tcp://[::1]::8529 \
  --server.endpoint tcp://[::1]::8530 \
  --server.endpoint tcp://[::1]::8531 \

This does not compromise the use of the other client tools, that preserve the behavior of having one Coordinator and one endpoint.

Also see arangobench Options


arangodump now supports multiple Coordinators. The flag --server.endpoint can be used multiple times, as in the example below:

arangodump \
  --server.endpoint tcp://[::1]::8529 \
  --server.endpoint tcp://[::1]::8530 \
  --server.endpoint tcp://[::1]::8531 \

This does not compromise the use of the other client tools that preserve the behavior of having one Coordinator and one endpoint.

Also see arangodump examples


arangorestore now supports multiple Coordinators. The flag --server.endpoint can be used multiple times, as in the example below:

arangorestore \
  --server.endpoint tcp://[::1]::8529 \
  --server.endpoint tcp://[::1]::8530 \
  --server.endpoint tcp://[::1]::8531 \

This does not compromise the use of the other client tools that preserve the behavior of having one Coordinator and one endpoint.

Also see arangorestore examples


The arangovpack utility supports more input and output formats (JSON and VelocyPack, plain or hex-encoded). The former options --json and --pretty have been removed and have been replaced with separate options for specifying the input and output types:

  • --input-type (json, json-hex, vpack, vpack-hex)
  • --output-type (json, json-pretty, vpack, vpack-hex)

The former option --print-non-json has been replaced with the new option --fail-on-non-json which makes arangovpack fail when trying to emit non-JSON types to JSON output.

Internal changes

The compiler version used to build the ArangoDB Linux executables has been upgraded from g++ 9.3.0 to g++ 10.2.1. g++ 10 is also the expected version of g++ when compiling ArangoDB from source.

The bundled version of the Snappy compression library was upgraded from version 1.1.8 to version 1.1.9.

The minimum architecture requirements have been raised from the Westmere architecture to the Sandy Bridge architecture. 256-bit AVX instructions are now expected to be present on all targets that run ArangoDB 3.9 executables. If a target does not support AVX instructions, it may fail with SIGILL at runtime.