Features and Improvements in ArangoDB 3.2

The following list shows in detail which features have been added or improved in ArangoDB 3.2. ArangoDB 3.2 also contains several bugfixes that are not listed here.

Storage engines

ArangoDB 3.2 offers two storage engines:

  • the always-existing memory-mapped files storage engine
  • a new storage engine based on RocksDB

Memory-mapped files storage engine (MMFiles)

The former storage engine (named MMFiles engine henceforth) persists data in memory-mapped files.

Any data changes are done first in the engine’s write-ahead log (WAL). The WAL is replayed after a crash so the engine offers durability and crash-safety. Data from the WAL is eventually moved to collection-specific datafiles. The files are always written in an append-only fashion, so data in files is never overwritten. Obsolete data in files will eventually be purged by background compaction threads.

Most of this engine’s indexes are built in RAM. When a collection is loaded, this requires rebuilding the indexes in RAM from the data stored on disk. The MMFiles engine has collection-level locking.

This storage engine is a good choice when data (including the indexes) can fit in the server’s available RAM. If the size of data plus the in-memory indexes exceeds the size of the available RAM, then this engine may try to allocate more memory than available. This will either make the operating system swap out parts of the data (and cause disk I/O) or, when no swap space is configured, invoke the operating system’s out-of-memory process killer.

The locking strategy allows parallel reads and is often good enough in read-mostly workloads. Writes need exclusive locks on the collections, so they can block other operations in the same collection. The locking strategy also provides transactional consistency and isolation.

RocksDB storage engine

The RocksDB storage engine is new in ArangoDB 3.2. It is designed to store datasets that are bigger than the server’s available RAM. It persists all data (including the indexes) in a RocksDB instance.

That means any document read or write operations will be answered by RocksDB under the hood. RocksDB will serve the data from its own in-RAM caches or from disk. The RocksDB engine has a write-ahead log (WAL) and uses background threads for compaction. It supports data compression.

The RocksDB storage engine has document-level locking. Read operations do not block and are never blocked by other operations. Write operations only block writes on the same documents/index values. Because multiple writers can operate in parallel on the same collection, there is the possibility of write-write conflicts. If such write conflict is detected, one of the write operations is aborted with error 1200 (“conflict”). Client applications can then either abort the operation or retry, based on the required consistency semantics.

Storage engine selection

The storage engine to use in an ArangoDB cluster or a single-server instance must be selected initially. The default storage engine in ArangoDB 3.2 is the MMFiles engine if no storage engine is selected explicitly. This ensures all users upgrading from earlier versions can continue with the well-known MMFiles engine.

To select the storage-engine, there is the configuration option --server.storage-engine. It can be set to either mmfiles, rocksdb or auto. While the first two values will explicitly select a storage engine, the auto option will automatically choose the storage engine based on which storage engine was previously selected. If no engine was selected previously, auto will select the MMFiles engine. If an engine was previously selected, the selection will be written to a file ENGINE in the server’s database directory and will be read from there at any subsequent server starts.

Once the storage engine was selected, the selection cannot be changed by adjusting --server.storage-engine. In order to switch to another storage engine, it is required to re-start the server with another (empty) database directory. In order to use data created with the other storage engine, it is required to dump the data first with the old engine and restore it using the new storage engine. This can be achieved via invoking arangodump and arangorestore.

Unlike in MySQL, the storage engine selection in ArangoDB is for an entire cluster or an entire single-server instance. All databases and collections will use the same storage engine.

RocksDB storage engine: supported index types

The existing indexes in the RocksDB engine are all persistent. The following indexes are supported there:

  • primary: this type of index is automatically created. It indexes _id / _key

  • edge: this index is automatically created for edge collections. It indexes _from and _to

  • hash, skiplist, persistent: these are user-defined indexes, Despite their names, they are neither hash nor skiplist indexes. These index types map to the same RocksDB-based sorted index implementation. The same is true for the “persistent” index. The names “hash”, “skiplist” and “persistent” are only used for compatibility with the MMFiles engine where these indexes existed in previous and the current version of ArangoDB.

  • geo: user-defined index for proximity searches

  • fulltext: user-defined sorted reverted index on words occurring in documents


With SatelliteCollections, you can define collections to shard to a cluster and collections to replicate to each machine. The ArangoDB query optimizer knows where each shard is located and sends the requests to the DB-Servers involved, which then executes the query, locally. With this approach, network hops during join operations on sharded collections can be avoided and response times can be close to that of a single instance.

SatelliteCollections are available in the Enterprise Edition.

Memory management

  • make arangod start with less V8 JavaScript contexts

    This speeds up the server start and makes arangod use less memory at start. Whenever a V8 context is needed by a Foxx action or some other JavaScript operation and there is no usable V8 context, a new context will be created dynamically now.

    Up to --javascript.v8-contexts V8 contexts will be created, so this option will change its meaning. Previously as many V8 contexts as specified by this option were created at server start, and the number of V8 contexts did not change at runtime. Now up to this number of V8 contexts will be in use at the same time, but the actual number of V8 contexts is dynamic.

    The garbage collector thread will automatically delete unused V8 contexts after a while. The number of spare contexts will go down to as few as configured in the new option --javascript.v8-contexts-minimum. Actually that many V8 contexts are also created at server start.

    The first few requests in new V8 contexts may take longer than in contexts that have been there already. Performance may therefore suffer a bit for the initial requests sent to ArangoDB or when there are only few but performance- critical situations in which new V8 contexts need to be created. If this is a concern, it can easily be fixed by setting --javascipt.v8-contexts-minimum and --javascript.v8-contexts to a relatively high value, which will guarantee that many number of V8 contexts to be created at startup and kept around even when unused.

    Waiting for an unused V8 context will now also abort and write a log message in case no V8 context can be acquired/created after 60 seconds.

  • the number of pending operations in arangod can now be limited to a configurable number. If this number is exceeded, the server will now respond with HTTP 503 (service unavailable). The maximum size of pending operations is controlled via the startup option --server.maximal-queue-size. Setting it to 0 means “no limit”.

  • the in-memory document revisions cache was removed entirely because it did not provide the expected benefits. The 3.1 implementation shadowed document data in RAM, which increased the server’s RAM usage but did not speed up document lookups too much.

    This also obsoletes the startup options --database.revision-cache-chunk-size and --database.revision-cache-target-size.

    The MMFiles engine now does not use a document revisions cache but has in-memory indexes and maps documents to RAM automatically via mmap when documents are accessed. The RocksDB engine has its own mechanism for caching accessed documents.

Communication Layer

  • HTTP responses returned by arangod will now include the extra HTTP header x-content-type-options: nosniff to work around a cross-site scripting bug in MSIE

  • the default value for --ssl.protocol was changed from TLSv1 to TLSv1.2. When not explicitly set, arangod and all client tools will now use TLSv1.2.

  • the JSON data in all incoming HTTP requests in now validated for duplicate attribute names.

    Incoming JSON data with duplicate attribute names will now be rejected as invalid. Previous versions of ArangoDB only validated the uniqueness of attribute names inside incoming JSON for some API endpoints, but not consistently for all APIs.

  • Internal JavaScript REST actions will now hide their stack traces to the client unless in HTTP responses. Instead they will always log to the logfile.


  • updated V8 version to

  • change undocumented behavior in case of invalid revision ids in If-Match and If-None-Match headers from 400 (BAD) to 412 (PRECONDITION FAILED).

  • change default string truncation length from 80 characters to 256 characters for print/printShell functions in ArangoShell and arangod. This will emit longer prefixes of string values before truncating them with ..., which is helpful for debugging. This change is mostly useful when using the ArangoShell (arangosh).

  • the @arangodb module now provides a time function which returns the current time in seconds as a floating point value with microsecond precision.


  • There is now an official HTTP API for managing services, allowing services to be installed, modified, uninstalled and reconfigured without the administrative web interface.

  • It is now possible to upload a single JavaScript file instead of a zip archive if your service requires no configuration, additional files or setup. A minimal manifest will be generated automatically upon installation and the uploaded file will be used as the service’s main entry point.

Distributed Graph Processing

  • We added support for executing distributed graph algorithms aka Pregel.
  • Users can run arbitrary algorithms on an entire graph, including in cluster mode.
  • We implemented a number of algorithms for various well-known graph measures:
    • Connected Components
    • PageRank
    • Shortest Paths
    • Centrality Measures (Centrality and Betweenness)
    • Community Detection (via Label Propagation, Speakers-Listeners Label Propagation or DMID)
  • Users can contribute their own algorithms


Optimizer improvements

  • Geo indexes are now implicitly and automatically used when using appropriate SORT/FILTER statements in AQL, without the need to use the somewhat limited special-purpose geo AQL functions NEAR or WITHIN.

    Compared to using the special purpose AQL functions this approach has the advantage that it is more composable, and will also honor any LIMIT values used in the AQL query.

    The special purpose NEAR AQL function can now be substituted with the following AQL (provided there is a geo index present on the doc.latitude and doc.longitude attributes):

    FOR doc in geoSort
      SORT DISTANCE(doc.latitude, doc.longitude, 0, 0)
      LIMIT 5
      RETURN doc

    WITHIN can be substituted with the following AQL:

    FOR doc in geoFilter
      FILTER DISTANCE(doc.latitude, doc.longitude, 0, 0) < 2000
      RETURN doc

Miscellaneous improvements

  • added REGEX_REPLACE AQL function

    REGEX_REPLACE(text, search, replacement, caseInsensitive) → string

    Replace the pattern search with the string replacement in the string text, using regular expression matching.

    • text (string): the string to search in
    • search (string): a regular expression search pattern
    • replacement (string): the string to replace the search pattern with
    • returns string (string): the string text with the search regex pattern replaced with the replacement string wherever the pattern exists in text
  • added new startup option --query.fail-on-warning to make AQL queries abort instead of continuing with warnings.

    When set to true, this will make an AQL query throw an exception and abort in case a warning occurs. This option should be used in development to catch errors early. If set to false, warnings will not be propagated to exceptions and will be returned with the query results. The startup option can also be overriden on a per query-level.

  • the slow query list now contains the values of bind variables used in the slow queries. Bind variables are also provided for the currently running queries. This helps debugging slow or blocking queries that use dynamic collection names via bind parameters.

  • AQL breaking change in cluster: The SHORTEST_PATH statement using edge collection names instead of a graph names now requires to explicitly name the vertex collection names within the AQL query in the cluster. It can be done by adding WITH <name> at the beginning of the query.


    FOR v,e IN OUTBOUND SHORTEST_PATH @start TO @target edges [...]

    Now has to be:

    WITH vertices
    FOR v,e IN OUTBOUND SHORTEST_PATH @start TO @target edges [...]

    This change is due to avoid deadlock sitations in clustered case. An error stating the above is included.

Client tools

  • added data export tool, arangoexport.

    arangoexport can be used to export collections to json, jsonl or xml and export a graph or collections to xgmml.

  • added “jsonl” as input file type for arangoimp

  • added --translate option for arangoimp to translate attribute names from the input files to attriubte names expected by ArangoDB

    The --translate option can be specified multiple times (once per translation to be executed). The following example renames the “id” column from the input file to “_key”, and the “from” column to “_from”, and the “to” column to “_to”:

    arangoimp --type csv --file data.csv --translate "id=_key" --translate "from=_from" --translate "to=_to"

    --translate works for CSV and TSV inputs only.

  • added --threads option to arangoimp to specify the number of parallel import threads

  • changed default value for client tools option --server.max-packet-size from 128 MB to 256 MB. this allows transferring bigger result sets from the server without the client tools rejecting them as invalid.


  • added LDAP authentication (Enterprise Edition only)


  • added read only mode for users
  • collection level authorization rights

Read more in the overview.

Foxx and authorization

Miscellaneous Changes

  • arangod now validates several OS/environment settings on startup and warns if the settings are non-ideal. It additionally will print out ways to remedy the options.

    Most of the checks are executed on Linux systems only.

  • added “deduplicate” attribute for array indexes, which controls whether inserting duplicate index values from the same document into a unique array index will lead to an error or not:

    // with deduplicate = true, which is the default value:
    db.test.ensureIndex({ type: "hash", fields: ["tags[*]"], deduplicate: true });
    db.test.insert({ tags: ["a", "b"] });
    db.test.insert({ tags: ["c", "d", "c"] }); // will work, because deduplicate = true
    db.test.insert({ tags: ["a"] }); // will fail
    // with deduplicate = false
    db.test.ensureIndex({ type: "hash", fields: ["tags[*]"], deduplicate: false });
    db.test.insert({ tags: ["a", "b"] });
    db.test.insert({ tags: ["c", "d", "c"] }); // will not work, because deduplicate = false
    db.test.insert({ tags: ["a"] }); // will fail