Features and Improvements in ArangoDB 3.5

The following list shows in detail which features have been added or improved in ArangoDB 3.5. ArangoDB 3.5 also contains several bug fixes that are not listed here.


Configurable Analyzers

Analyzers can split string values into smaller parts and perform additional processing such as word stemming and case conversion. In ArangoDB 3.4 there is a fixed set of text analyzers for 12 different languages, which tokenize strings into case-insensitive word stems using language-dependent rules based on the chosen locale, without discarding any stop-words (common words which carry little meaning such as “the”). An additional no-operation analyzer identity is available to keep the input unaltered in its entirety.

In 3.5, analyzers can be customized as well as used independent of ArangoSearch Views in AQL. It is possible to tokenize strings without word stemming, remove user-defined stop-words, split by a delimiting character only, perform case conversion and/or removal of diacritic characters against the full input without tokenization and more.

See Analyzers for all available options.

Sorted Index

The index behind an ArangoSearch View can have a primary sort order. A direction can be specified upon View creation for each uniquely named attribute (ascending or descending), to enable an optimization for AQL queries which iterate over a view and sort by one of the attributes. If the index direction matches the requested SORT direction, then the data can be read in order directly from the index without actual sort operation.

View definition example:

  "links": {
    "coll1": {
      "fields": {
        "text": {
    "coll2": {
      "fields": {
        "text": {
  "primarySort": [
      "field": "text",
      "direction": "desc"

AQL query example:

FOR doc IN viewName
  SORT doc.text DESC
  RETURN doc

Execution plan without a sorted index being used:

Execution plan:
 Id   NodeType            Est.   Comment
  1   SingletonNode          1   * ROOT
  2   EnumerateViewNode      1     - FOR doc IN viewName   /* view query */
  3   CalculationNode        1       - LET #1 = doc.`val`   /* attribute expression */
  4   SortNode               1       - SORT #1 DESC   /* sorting strategy: standard */
  5   ReturnNode             1       - RETURN doc

Execution plan with the primary sort order of the index being utilized:

Execution plan:
 Id   NodeType            Est.   Comment
  1   SingletonNode          1   * ROOT
  2   EnumerateViewNode      1     - FOR doc IN viewName SORT doc.`val` DESC   /* view query */
  5   ReturnNode             1       - RETURN doc

Note that the primarySort option is immutable: it can not be changed after View creation. It is therefore not possible to configure it through the Web UI. The View needs to be created via the HTTP or JavaScript API (arangosh) to set it.

See Primary Sort Order of ArangoSearch Views.

AQL Integration

Some small new features give more control over ArangoSearch from AQL.

  • The scoring functions BM25() and TFIDF() are not limited to the SORT operation anymore, but can also be returned as part of the query result:

    FOR doc IN viewName
      SEARCH ...
      LET score = BM25(doc)
      SORT score DESC
      RETURN { doc, score }
  • The score can be manipulated to influence the ranking based on attribute values and using numeric AQL functions:

    FOR movie IN imdbView
      SEARCH PHRASE(movie.title, "Star Wars", "text_en")
      SORT BM25(movie) * LOG(movie.runtime + 1) DESC
      RETURN movie
  • The SEARCH operation accepts an options object to restrict the search to certain collections indexed by a View:

    FOR doc IN viewName
      SEARCH ... OPTIONS { collections: ["coll1", "coll2"] }
      RETURN doc
  • A new function IN_RANGE() was added for matching values within defined boundaries, to enable easy range searching primarily for numbers and strings.



Pruning in Traversals

With PRUNE you can stop walking down certain paths early in a graph traversal to improve its efficiency. This is different to FILTER, which would perform a post-filtering after the actual traversal was carried out already in most cases. Using PRUNE, the traverser will not follow any more edges on the current path if the pruning condition is met, but will emit the traversal variables for whatever stopped it.

See: Graph Traversal Pruning

SORT-LIMIT optimization

A new SORT-LIMIT optimization has been added. This optimization will be pulled off by the query optimizer if there is a SORT statement followed by a LIMIT node, and the overall number of documents to return is relatively small in relation to the total number of documents to be sorted. In this case, the optimizer will use a size-constrained heap for keeping only the required number of results in memory, which can drastically reduce memory usage and, for some queries, also execution time for the sorting.

If the optimization is applied, it will show as “sort-limit” rule in the query execution plan.

Also see:

Index hints in AQL

Users may now take advantage of the indexHint inline query option to override the internal optimizer decision regarding which index to use to serve content from a given collection. The index hint works with the named indices feature, making it easy to specify which index to use.


Sorted primary index (RocksDB engine)

The query optimizer can now make use of the sortedness of primary indexes if the RocksDB engine is used. This means the primary index can be utilized for queries that sort by either the _key or _id attributes of a collection and also for range queries on these attributes.

In the list of documents for a collection in the web interface, the documents will now always be sorted in lexicographical order of their _key values. An exception for keys representing quasi-numerical values has been removed when doing the sorting in the web interface. Removing this exception can also speed up the display of the list of documents.

This change potentially affects the order in which documents are displayed in the list of documents overview in the web interface. A document with a key value “10” will now be displayed before a document with a key value of “9”. In previous versions of ArangoDB this was exactly opposite.

Edge index query optimization (RocksDB engine)

An AQL query that uses the edge index only and returns the opposite side of the edge can now be executed in a more optimized way, e.g.

FOR edge IN edgeCollection FILTER edge._from == "v/1" RETURN edge._to

is fully covered by the RocksDB edge index.

For MMFiles this rule does not apply.

AQL syntax improvements

AQL now allows the usage of floating point values without leading zeros, e.g. .1234. Previous versions of ArangoDB required a leading zero in front of the decimal separator, i.e 0.1234.

Also see: AQL Numeric Literals

k Shortest Paths queries

AQL now allows to perform k Shortest Paths queries, that is, query a number of paths of increasing length from a start vertex to a target vertex.

See: AQL k Shortest Paths


The SmartJoins feature available in the ArangoDB Enterprise Edition allows running joins between two sharded collections with performance close to that of a local join operation.

The prerequisite for this is that the two collections have an identical sharding setup, established via the distributeShardsLike attribute of one of the collections.

Quick example setup for two collections with identical sharding:

> db._create("products", { numberOfShards: 3, shardKeys: ["_key"] });
> db._create("orders", { distributeShardsLike: "products", shardKeys: ["productId"] });
> db.orders.ensureIndex({ type: "hash", fields: ["productId"] });

Now an AQL query that joins the two collections via their shard keys will benefit from the SmartJoins optimization, e.g.

FOR p IN products 
  FOR o IN orders 
    FILTER p._key == o.productId 
    RETURN o

In this query’s execution plan, the extra hop via the coordinator can be saved that is normally there for generic joins. Thanks to the SmartJoins optimization, the query’s execution is as simple as:

Execution plan:
 Id   NodeType                  Site  Est.   Comment
  1   SingletonNode             DBS      1   * ROOT
  3   EnumerateCollectionNode   DBS      9     - FOR o IN orders   /* full collection scan, 3 shard(s) */
  7   IndexNode                 DBS      0       - FOR p IN products   /* primary index scan, scan only, 3 shard(s) */
 10   RemoteNode                COOR     0         - REMOTE
 11   GatherNode                COOR     0         - GATHER
  6   ReturnNode                COOR     0         - RETURN o

Without the SmartJoins optimization, there will be an extra hop via the coordinator for shipping the data from each shard of the one collection to each shard of the other collection, which will be a lot more expensive:

Execution plan:
 Id   NodeType        Site  Est.   Comment
  1   SingletonNode   DBS      1   * ROOT
 16   IndexNode       DBS      3     - FOR p IN products   /* primary index scan, index only, projections: `_key`, 3 shard(s) */
 14   RemoteNode      COOR     3       - REMOTE
 15   GatherNode      COOR     3       - GATHER
  8   ScatterNode     COOR     3       - SCATTER
  9   RemoteNode      DBS      3       - REMOTE
  7   IndexNode       DBS      3       - FOR o IN orders   /* hash index scan, 3 shard(s) */
 10   RemoteNode      COOR     3         - REMOTE
 11   GatherNode      COOR     3         - GATHER
  6   ReturnNode      COOR     3         - RETURN o

In the end, SmartJoins can optimize away a lot of the inter-node network requests normally required for performing a join between sharded collections. The performance advantage of SmartJoins compared to regular joins will grow with the number of shards of the underlying collections.

In general, for two collections with n shards each, the minimal number of network requests for the general join (no SmartJoins optimization) will be n * (n + 2). The number of network requests increases quadratically with the number of shards.

SmartJoins can get away with a minimal number of n requests here, which scales linearly with the number of shards.

SmartJoins will also be especially advantageous for queries that have to ship a lot of data around for performing the join, but that will filter out most of the data after the join. In this case SmartJoins should greatly outperform the general join, as they will eliminate most of the inter-node data shipping overhead.

Also see:

Background Index Creation

Creating new indexes is by default done under an exclusive collection lock. This means that the collection (or the respective shards) are not available for write operations as long as the index is created. This “foreground” index creation can be undesirable, if you have to perform it on a live system without a dedicated maintenance window.

Starting with ArangoDB 3.5, indexes can also be created in “background”, not using an exclusive lock during the entire index creation. The collection remains basically available, so that other CRUD operations can run on the collection while the index is being created. This can be achieved by setting the inBackground attribute when creating an index.

To create an index in the background in arangosh just specify inBackground: true, like in the following example:

db.collection.ensureIndex({ type: "hash", fields: [ "value" ], inBackground: true });

Indexes that are still in the build process will not be visible via the ArangoDB APIs. Nevertheless it is not possible to create the same index twice via the ensureIndex API while an index is still being created. AQL queries also will not use these indexes until the index reports back as fully created. Note that the initial ensureIndex call or HTTP request will still block until the index is completely ready. Existing single-threaded client programs can thus safely set the inBackground option to true and continue to work as before.

Should you be building an index in the background you cannot rename or drop the collection. These operations will block until the index creation is finished. This is equally the case with foreground indexing.

After an interrupted index build (i.e. due to a server crash) the partially built index will the removed. In the ArangoDB cluster the index might then be automatically recreated on affected shards.

Background index creation might be slower than the “foreground” index creation and require more RAM. Under a write heavy load (specifically many remove, update or replace operations), the background index creation needs to keep a list of removed documents in RAM. This might become unsustainable if this list grows to tens of millions of entries.

Building an index is always a write-heavy operation, so it is always a good idea to build indexes during times with less load.

Please note that background index creation is useful only in combination with the RocksDB storage engine. With the MMFiles storage engine, creating an index will always block any other operations on the collection.

Also see: Creating Indexes in Background

TTL (time-to-live) Indexes

The new TTL indexes feature provided by ArangoDB can be used for automatically removing expired documents from a collection.

TTL indexes support eventual removal of documents which are past a configured expiration timepoint. The expiration timepoints can be based upon the documents’ original insertion or last-updated timepoints, with adding a period during which to retain the documents. Alternatively, expiration timepoints can be specified as absolute values per document. It is also possible to exclude documents from automatic expiration and removal.

Please also note that TTL indexes are designed exactly for the purpose of removing expired documents from collections. It is not recommended to rely on TTL indexes for user-land AQL queries. This is because TTL indexes internally may store a transformed, always numerical version of the index attribute value even if it was originally passed in as a datestring. As a result TTL indexes will likely not be used for filtering and sort operations in user-land AQL queries.

Also see: TTL Indexes


All collections now support a minimum replication factor (minReplicationFactor) property. This is default set to 1, which is identical to previous behavior. If in a failover scenario a shard of a collection has less than minReplicationFactor many in sync followers it will go into “read-only” mode and will reject writes until enough followers are in sync again.

In more detail:

  • Having minReplicationFactor == 1 means as soon as a “master-copy” is available of the data writes are allowed.
  • Having minReplicationFactor > 1 requires additional in sync copies on follower servers to allow writes.

The feature is used to reduce the diverging of data in case of server failures and to help new followers to catch up.

Also see: Collection Database Methods

HTTP API extensions

Extended index API

The HTTP API for creating indexes at POST /_api/index has been extended two-fold:

  • to create a TTL (time-to-live) index, it is now possible to specify a value of ttl in the type attribute. When creating a TTL index, the attribute expireAfter is also required. That attribute contains the expiration time (in seconds), which is based on the documents’ index attribute value.

  • to create an index in background, the attribute inBackground can be set to true.

API for querying the responsible shard

The HTTP API for collections has got an additional route for retrieving the responsible shard for a document at PUT /_api/collection/<name>/responsibleShard.

When calling this route, the request body is supposed to contain the document for which the responsible shard should be determined. The response will contain an attribute shardId containing the ID of the shard that is responsible for that document.

A method collection.getResponsibleShard(document) was added to the JS API as well.

It does not matter if the document actually exists or not, as the shard responsibility is determined from the document’s attribute values only.

Please note that this API is only meaningful and available on a cluster coordinator.


Foxx API for running tests

The HTTP API for running Foxx service tests now supports a filter attribute, which can be used to limit which test cases should be executed.

Stream Transaction API

There is a new HTTP API for transactions. This API allows clients to add operations to a transaction in a streaming fashion. A transaction can consist of a series of supported transactional operations, followed by a commit or abort command. This allows clients to construct transactions in a more natural way than with JavaScript-based transactions.

Note that this requires client applications to abort transactions which are no longer necessary. Otherwise resources and locks acquired by the transactions will be in use until the server decides to garbage-collect them.

In order to keep resource usage low, a maximum lifetime and transaction size for stream transactions is enforced on the coordinator to ensure that transactions cannot block the cluster from operating properly:

  • Maximum idle timeout of 10 seconds between operations
  • Maximum transaction size of 128 MB per DBServer

These limits are also enforced for stream transactions on single servers.

Enforcing the limits is useful to free up resources used by abandoned transactions, for example from transactions that are abandoned by client applications due to programming errors or that were left over because client connections were interrupted. Also see Known Issues

See: Stream Transaction HTTP API

Minimal replication Factor

Within the properties of a collection we can now define a minReplicationFactor. This affects all routes that can create or modify the properties of a collection, including the graph API _api/gharial. All places where a replicationFactor can be modified, can now modify the minReplicationFactor as well.

Web interface

When using the RocksDB engine, the selection of index types “hash” and “skiplist” has been removed from the web interface when creating new indexes.

The index types “hash”, “skiplist” and “persistent” are just aliases of each other when using the RocksDB engine, so there is no need to offer them all. In the web interface there remains the index of type “persistent”, which is feature-wise identical with “hash” and “skiplist” indexes for the RocksDB engine. Existing “hash” and “skiplist” indexes will remain fully functional.


V8 updated

The bundled version of the V8 JavaScript engine has been upgraded from 5.7.492.77 to 7.1.302.28.

Among other things, the new version of V8 provides a native JavaScript BigInt type which can be used to store arbitrary-precision integers. However, to store such BigInt objects in ArangoDB, they need to be explicitly converted to either strings or simple JavaScript numbers. Converting BigInts to strings for storage is preferred because converting a BigInt to a simple number may lead to precision loss.

// will fail with "bad parameter" error:
value = BigInt("123456789012345678901234567890");
db.collection.insert({ value });

// will succeed:
db.collection.insert({ value: String(value) });

// will succeed, but lead to precision loss:
db.collection.insert({ value: Number(value) });

The new V8 version also changes the default timezone of date strings to be conditional on whether a time part is included:

> new Date("2019-04-01");
Mon Apr 01 2019 02:00:00 GMT+0200 (Central European Summer Time)

> new Date("2019-04-01T00:00:00");
Mon Apr 01 2019 00:00:00 GMT+0200 (Central European Summer Time)

If the timezone is explicitly set in the date string, then the specified timezone will always be honored:

> new Date("2019-04-01Z");
Mon Apr 01 2019 02:00:00 GMT+0200 (Central European Summer Time)
> new Date("2019-04-01T00:00:00Z");
Mon Apr 01 2019 02:00:00 GMT+0200 (Central European Summer Time)

JavaScript Dependencies

More than a dozen JavaScript dependencies were updated in 3.5 (changelog).

The most significant one is the update of joi from 9.2.0 to 14.3.1. See the respective release notes to see if there are breaking changes for you.

Note that you can bundle your own version of joi if you need to rely on version-dependent features.

JavaScript Security Options

ArangoDB 3.5 provides several new options for restricting the functionality of JavaScript application code running in the server, with the intent to make a setup more secure.

There now exist startup options for restricting which environment variables and values of which configuration options JavaScript code is allowed to read. These options can be set to prevent leaking of confidential information from the environment or the setup into the JavaScript application code. Additionally there are options to restrict outbound HTTP connections from JavaScript applications to certain endpoints and to restrict filesystem access from JavaScript applications to certain directories only.

Finally there are startup options to turn off the REST APIs for managing Foxx services, which can be used to prevent installation and uninstallation of Foxx applications on a server. A separate option is provided to turn off access and connections to the central Foxx app store via the web interface.

A complete overview of the security options can be found in Security Options.


Request credentials are now exposed via the auth property:

const tokens = context.collection("tokens");
router.get("/authorized", (req, res) => {
  if (!req.auth || !req.auth.bearer || !tokens.exists(req.auth.bearer)) {
    res.throw(403, "Not authenticated");
  // ...

API improvements

Collections now provide the documentId method to derive document ids from keys.


const collection = context.collection("users");
const documentKey = "my-document-key";
const documentId = `${collection.name()}/${documentKey}`;


const collection = context.collection("users");
const documentKey = "my-document-key";
const documentId = collection.documentId(documentKey);

Client tools

Dump and restore all databases

arangodump got an option --all-databases to make it dump all available databases instead of just a single database specified via the option --server.database.

When set to true, this makes arangodump dump all available databases the current user has access to. The option --all-databases cannot be used in combination with the option --server.database.

When --all-databases is used, arangodump will create a subdirectory with the data of each dumped database. Databases will be dumped one after the after. However, inside each database, the collections of the database can be dumped in parallel using multiple threads. When dumping all databases, the consistency guarantees of arangodump are the same as when dumping multiple single database individually, so the dump does not provide cross-database consistency of the data.

arangorestore got an option --all-databases to make it restore all databases from inside the subdirectories of the specified dump directory, instead of just the single database specified via the option --server.database.

Using the option for arangorestore only makes sense for dumps created with arangodump and the --all-databases option. As for arangodump, arangorestore cannot be invoked with the both options --all-databases and --server.database at the same time. Additionally, the option --force-same-database cannot be used together with --all-databases.

If the to-be-restored databases do not exist on the target server, then restoring data into them will fail unless the option --create-database is also specified for arangorestore. Please note that in this case a database user must be used that has access to the _system database, in order to create the databases on restore.

Also see:

Warning if connected to DBServer

Under normal circumstances there should be no need to connect to a database server in a cluster with one of the client tools, and it is likely that any user operations carried out there with one of the client tools may cause trouble.

The client tools arangosh, arangodump and arangorestore will now emit a warning when connecting with them to a database server node in a cluster.

Startup option changes

The value type of the hidden startup option --rocksdb.recycle-log-file-num has been changed from numeric to boolean in ArangoDB 3.5, as the option is also a boolean option in the underlying RocksDB library.

Client configurations that use this configuration variable should adjust their configuration and set this variable to a boolean value instead of to a numeric value.


Improved overview of available program options

The --help-all command-line option for all ArangoDB executables will now also show all hidden program options.

Previously hidden program options were only returned when invoking arangod or a client tool with the cryptic --help-. option. Now --help-all simply returns them as well.

Fewer system collections

The system collections _frontend, _modules and _routing are not created anymore for new databases by default.

_modules and _routing are only needed for legacy functionality. Existing _routing collections will not be touched as they may contain user-defined entries, and will continue to work.

Existing _modules collections will also remain functional.

The _frontend collection may still be required for actions triggered by the web interface, but it will automatically be created lazily if needed.

Named indices

Indices now have an additional name field, which allows for more useful identifiers. System indices, like the primary and edge indices, have default names (primary and edge, respectively). If no name value is specified on index creation, one will be auto-generated (e.g. idx_13820395). The index name cannot be changed after index creation. No two indices on the same collection may share the same name, but two indices on different collections may.

ID values in log messages

By default, ArangoDB and its client tools now show a 5 digit unique ID value in any of their log messages, e.g.

2019-03-25T21:23:19Z [8144] INFO [cf3f4] ArangoDB (version 3.5.0 enterprise [linux]) is ready for business. Have fun!.

In this message, the cf3f4 is the message’s unique ID value. ArangoDB users can use this ID to build custom monitoring or alerting based on specific log ID values. Existing log ID values are supposed to stay constant in future releases of arangod.

Additionally the unique log ID values can be used by the ArangoDB support to find out which component of the product exactly generated a log message. The IDs also make disambiguation of identical log messages easier.

The presence of these ID values in log messages may confuse custom log message filtering or routing mechanisms that parse log messages and that rely on the old log message format.

This can be fixed adjusting any existing log message parsers and making them aware of the ID values. The ID values are always 5 byte strings, consisting of the characters [0-9a-f]. ID values are placed directly behind the log level (e.g. INFO).

Alternatively, the log IDs can be suppressed in all log messages by setting the startup option --log.ids false when starting arangod or any of the client tools.


We have moved from C++11 to C++14, which allows us to use some of the simplifications, features and guarantees that this standard has in stock. To compile ArangoDB from source, a compiler that supports C++14 is now required.

The bundled JEMalloc memory allocator used in ArangoDB release packages has been upgraded from version 5.0.1 to version 5.2.0.

The bundled version of the RocksDB library has been upgraded from 5.16 to 6.2.

The unit test framework has been changed from catch to googletest. This change also renames a CMake configuration variable from USE_CATCH_TESTS to USE_GOOGLE_TESTS.