Features and Improvements in ArangoDB 3.8
The following list shows in detail which features have been added or improved in ArangoDB 3.8. ArangoDB 3.8 also contains several bug fixes that are not listed here.
AQL
AQL window operations
The WINDOW
keyword can be used for aggregations over related rows, usually
preceding and / or following rows.
The WINDOW
operation performs a COLLECT AGGREGATE
-like operation on a set
of query rows. However, whereas a COLLECT
operation groups multiple query
rows into a single result group, a WINDOW
operation produces a result for
each query row:
- The row for which function evaluation occurs is called the current row.
- The query rows related to the current row over which function evaluation occurs, comprise the window frame for the current row.
Window frames are determined with respect to the current row:
- By defining a window frame to be all rows from the query start to the current row, you can compute running totals for each row.
- By defining a frame as extending N rows on either side of the current row, you can compute rolling averages.
See WINDOW
operation.
Weighted Traversals
The graph traversal option bfs
is now deprecated and superseded by the new
option order
. It supports a new traversal type "weighted"
, which enumerate
paths by increasing weights.
The cost of an edge can be read from an attribute which can be specified with
the weightAttribute
option.
FOR x, v, p IN 0..10 OUTBOUND "places/York" GRAPH "kShortestPathsGraph"
OPTIONS {
order: "weighted",
weightAttribute: "travelTime",
uniqueVertices: "path"
}
FILTER p.edges[*].travelTime ALL < 3
LET totalTime = LAST(p.weights)
FILTER totalTime < 6
SORT totalTime DESC
RETURN {
path: p.vertices[*]._key,
weight: LAST(p.weights),
weights: p.edges[*].travelTime
}
path |
weight |
weights |
---|---|---|
["York","London","Birmingham","Carlisle"] |
5.3 |
[1.8,2.5,1] |
["York","London","Birmingham"] |
4.3 |
[1.8,2.5] |
["York","London","Brussels"] |
4.3 |
[1.8,2.5] |
["York","London"] |
1.8 |
[1.8] |
["York"] |
0 |
[] |
Weighted traversals do not support negative weights. If a document attribute (as specified by weightAttribute
) with a negative value is encountered during traversal, or if defaultWeight
is set to a negative number, then the query is aborted with an error.
The preferred way to start a breadth-first search from now on is with
order: "bfs"
. The default remains depth-first search if no order
is
specified, but can also be explicitly requested with order: "dfs"
.
Also see AQL graph traversals
k Paths
A new graph traversal method K_PATHS
was added to AQL. It will enumerate all
paths between a source and a target vertex that match the given path length.
For example, the query:
FOR path IN 2..4 OUTBOUND K_PATHS "v/source" TO "v/target" GRAPH "g"
RETURN path
… will yield all paths in the format:
{
"vertices": ["v/source", ... , "v/target"],
"edges": ["v/source" -> "v/1", ... , "v/n" -> "v/target"]
}
… that have length of exactly 2 or 3 or 4, start at v/source
and end at
v/target
. No order is guaranteed for those paths in the result set.
For more details see AQL k Paths
AQL bit functions
ArangoDB 3.8 adds the following bit handling functions to AQL:
BIT_AND()
: and-combine two or more numbersBIT_OR()
: or-combine two or more numbersBIT_XOR()
: xor-combine two or more numbersBIT_NEGATE()
: bitwise negationBIT_TEST()
: test if bit is set at positionBIT_POPCOUNT()
: number of bits setBIT_SHIFT_LEFT()
: bitwise shift-leftBIT_SHIFT_RIGHT()
: bitwise shift-rightBIT_CONSTRUCT()
: construct a number with bits set at given positionsBIT_DECONSTRUCT()
: deconstruct a number into an array with the positions of its set bitsBIT_TO_STRING()
: create a bitstring representation from a numeric valueBIT_FROM_STRING()
: parse a bitstring representation into a number
Also see Bit functions.
BIT_AND()
, BIT_OR()
and BIT_XOR()
are also available as aggregate
functions for usage inside COLLECT AGGREGATE
.
All above bit operations support unsigned integer values with up to 32 bits.
Using values outside the supported range will make any of these bit functions
return null
and register a warning.
This functionality has been backported to v3.7.7 as well.
AQL binary and hexadecimal integer literals
ArangoDB 3.8 allows using binary (base 2) and hexadecimal (base 16) integer literals in AQL. These literals can be used where regular (base 10) integer literals can be used.
- The prefix for binary integer literals is
0b
, e.g.0b10101110
. - The prefix for hexadecimal integer literals is
0x
, e.g.0xabcdef02
.
Binary and hexadecimal integer literals can only be used for unsigned integers.
The maximum supported value is 232 - 1, i.e.
0b11111111111111111111111111111111
(binary) or 0xffffffff
(hexadecimal).
This functionality has been backported to v3.7.7 as well.
Projections on sub-attributes
AQL now also support projections on sub-attributes (e.g. a.b.c
).
In previous versions of ArangoDB, projections were only supported on top-level attributes. For example, in the query:
FOR doc IN collection
RETURN doc.a.b
… the projection that was used was just a
. Now the projection will be a.b
,
which can help reduce the amount of data to be extracted from documents, when
only some sub-attributes are accessed.
In addition, indexes can now be used to extract the data of sub-attributes
for projections. If for the above example query an index on a.b
exists,
it will be used now. Previously, no index could be used for this projection.
Projections now can also be fed by any attribute in a combined index. For example, in the query:
FOR doc IN collection
RETURN doc.b
… the projection can be satisfied by a single-attribute index on attribute b
,
but now also by a combined index on attributes a
and b
(or b
and a
).
AQL optimizer improvements
The “move-calculations-up” optimizer rule was improved so that it can move calculations out of subqueries into the outer query, so that they will be executed less often.
In queries or subqueries that return only constant values and/or that assign constant values to variables, these constant values are now stored only once per query and not once input row. This can slightly improve memory usage and execution time of such queries.
Explaining a query now also shows the query optimizer rules with the highest execution times in the explain output.
AQL performance improvements
The performance of AQL standard
sort operations has been improved in ArangoDB
3.8. This is true for sorts carried out explicitly by using the SORT
keyword
and for sorts that are implicitly executed due to using a sorting COLLECT
operation. Sort performance is especially better for sorting numeric values.
The improvements are limited to SortNodes with the standard
sorting strategy.
SortNodes using the constrained heap
strategy may not see a speedup.
There are also performance improvements for COLLECT
operations that only
count values or that aggregate values using AGGREGATE
. The exact mileage
can vary, but is substantial for some queries.
AQL usability options
Requiring WITH
statements
The new startup option --query.require-with
will make AQL queries in single
server mode also require WITH
clauses in AQL queries where a cluster
installation would require them.
The option is set to false by default, but can be turned on in single servers
to remove this behavior difference between single servers and clusters, making
a later transition from single server to cluster easier.
Allowing the usage of collection names in AQL expressions
The new startup option --query.allow-collections-in-expressions
controls
whether using collection names in arbitrary places in AQL expressions is
allowed, although using collection names like this is very likely unintended.
For example, consider the query
FOR doc IN collection RETURN collection
Here, the collection name is collection, and its usage in the FOR
loop is
intended and valid. However, collection is also used in the RETURN
statement, which is legal but potentially unintended. It should likely be
RETURN doc
or RETURN doc.someAttribute
instead. Otherwise, the entire
collection will be materialized and returned as many times as there are
documents in the collection. This can take a long time and even lead to
out-of-memory crashes in the worst case.
Setting the option --query.allow-collections-in-expression
to false will
prohibit such unintentional usage of collection names in queries, and instead
make the query fail with error 1568 (“collection used as expression operand”).
The default value of the option is true in 3.8, meaning that potentially unintended usage of collection names in queries is still allowed. The default value for the option will change to false in 3.9. The option will also be deprecated in 3.9 and removed in future versions. From then on, unintended usage of collection names will always be disallowed.
Also see ArangoDB Server Query Options
ArangoSearch
Pipeline Analyzer
Added new Analyzer type "pipeline"
for chaining effects of multiple Analyzers
into one. It allows you to combine text normalization for a case insensitive
search with n-gram tokenization, or to split text at multiple delimiting
characters followed by stemming.
See ArangoSearch Pipeline Analyzer
AQL Analyzer
Added new Analyzer type "aql"
capable of running an AQL query (with some
restrictions) to perform data manipulation/filtering.
Geo-spatial queries
Added two Geo Analyzers "geojson"
and "geopoint"
as well as the
following ArangoSearch Geo functions
which enable geo-spatial queries backed by View indexes:
GEO_CONTAINS()
GEO_DISTANCE()
GEO_IN_RANGE()
GEO_INTERSECTS()
Stop words Analyzer
Introduced in: v3.8.1
Added new Analyzer "stopwords"
capable of removing specified tokens from the
input. It can be used standalone or be combined with other Analyzers via a
pipeline Analyzer to add stop word functionality to them. Previously, only the
text Analyzer type provided stop word support.
See ArangoSearch stopwords
Analyzer
Approximate count
Added a new option countApproximate
for SEARCH
queries to control how the
total count of rows is calculated if the fullCount
option is enabled for a
query or when a COLLECT WITH COUNT
clause is executed:
"exact"
(default): rows are actually enumerated for a precise count."cost"
: a cost based approximation is used. Does not enumerate rows and returns an approximate result with O(1) complexity. Gives a precise result if theSEARCH
condition is empty or if it contains a single term query only (e.g.SEARCH doc.field == "value"
), the usual eventual consistency of Views aside.
Also see: AQL SEARCH
Operation
This feature was also backported to v3.7.6.
ArangoSearch thread control
Added new command line options for fine-grained control over ArangoSearch’s maintenance threads, now allowing to set the minimum and maximum number of threads for committing and consolidation separately:
--arangosearch.commit-threads
--arangosearch.commit-threads-idle
--arangosearch.consolidation-threads
--arangosearch.consolidation-threads-idle
They supersede the options --arangosearch.threads
and
--arangosearch.threads-limit
. See
ArangoDB Server ArangoSearch Options.
This feature was also backported to v3.7.5.
Web interface
Dashboard
The cluster nodes overview in the web interface will now also display all Agent instances. Agent failures are now visible there, too. The overview also shows which agent is currently the leader.
Shard distribution overview
The web interface now provides a shard distribution overview for the entire cluster. The overview includes general details about cluster-wide distribution as well as per-server figures for the number of leader and follower shards.
The shard distribution overview is only available in the _system
database.
Cluster maintenance mode
Inside the _system
database of a cluster, the web interface now displays the
cluster supervision maintenance status. This can be used to check if the cluster
is currently in maintenance mode. For users with sufficient privileges, it is
also possible to toggle the maintenance mode from there.
Collection figures
The web interface now displays the approximate size of the data in a collection for both indexes and documents, based on the estimates provided by RocksDB.
These are estimates which are intended to be calculated quickly, but are not perfectly accurate. The estimates can still be useful to get an idea of how “big” a collection approximately is. The sizing information is provided in the Info tab of each collection’s detail view.
For collections in a cluster, the web interface now displays the number of documents in each shard (data distribution) plus the leader and follower DB-Servers for each shard.
Server logs in cluster
The web interface can now display the most recent server log entries for
Coordinators and DB-Servers in a cluster. Logs are made available in the
_system
database via the Nodes menu item. Up to 2048 log entries will be
kept on each instance.
The privileges for accessing server logs in the web interface are identical
to the privileges required for accessing logs via the GET /_admin/log
HTTP
REST API. If security is a concern, in-memory logs buffering can be turned
off entirely using the startup option --log.in-memory false
, plus the log
API can be turned off or restricted via the --log.api-enabled false
or
--log.api-enabled jwt
startup options.
Server metrics
The statistics view in the web interface now provide some key metrics for DB-Servers in case the metrics API is enabled. Different statistics may be visible depending on the operating system.
The web interface can now display the servers’ current metrics (as exposed
via the /_admin/metrics/v2
API) for all Coordinators and DB-Servers in a cluster.
The current metrics are provided in a tabular format output and can be downloaded
from the UI for further analysis. This is not meant to be a 100% replacement for
Grafana, but rather as a quick self-service alternative to check the servers’
statuses.
Shard synchronization status
The shard synchronization overview in the web interface now provides a better overview of what the shard synchronization is currently doing, and what its progress is.
For shards that are currently not in sync it will display whether the followers are currently syncing or waiting for their turn to come (because the amount of parallelism for syncing multiple shards can be restricted). The progress values displayed for shard synchronization should also be more helpful for shards with more than one follower and in situations where one follower is in sync and the other is not (yet).
Memory usage
Agency memory usage
The in-memory object sizes for Agency data have been reduced in ArangoDB 3.8, which should reduce the memory usage of Agent instances for clusters with a larger amount of databases/collections/shards. On-disk sizes or sizes of Agency dumps retrieved via APIs should not change, however.
The change also helps Coordinators and DB-Servers, which since v3.7.4 also maintain an in-memory cache of Agency data so that they can reduce the number of requests to the Agency.
The default RocksDB settings for Agency instances have been adjusted so that the Agency memory usage consumed by RocksDB is limited to a 1 GB RocksDB block cache and to 512 MB for the total write buffer size. Previously, Agency memory usage could grow a lot higher for systems with a lot of memory if the startup parameters were not set explicitly.
Default per-query memory limit
A default per-query memory limit has been introduced for queries, to prevent rogue AQL queries from consuming the too much memory of an arangod instance.
The per-query limit is introduced via changing the default value of the option
--query.memory-limit
from previously 0
(meaning no limit) to a dynamically
calculated value. The per-query memory limit defaults are now (depending on the
amount of available RAM):
Available memory: 0 (0MiB) Limit: 0 unlimited, %mem: n/a
Available memory: 134217728 (128MiB) Limit: 33554432 (32MiB), %mem: 25.0
Available memory: 268435456 (256MiB) Limit: 67108864 (64MiB), %mem: 25.0
Available memory: 536870912 (512MiB) Limit: 201326592 (192MiB), %mem: 37.5
Available memory: 805306368 (768MiB) Limit: 402653184 (384MiB), %mem: 50.0
Available memory: 1073741824 (1024MiB) Limit: 603979776 (576MiB), %mem: 56.2
Available memory: 2147483648 (2048MiB) Limit: 1288490189 (1228MiB), %mem: 60.0
Available memory: 4294967296 (4096MiB) Limit: 2576980377 (2457MiB), %mem: 60.0
Available memory: 8589934592 (8192MiB) Limit: 5153960755 (4915MiB), %mem: 60.0
Available memory: 17179869184 (16384MiB) Limit: 10307921511 (9830MiB), %mem: 60.0
Available memory: 25769803776 (24576MiB) Limit: 15461882265 (14745MiB), %mem: 60.0
Available memory: 34359738368 (32768MiB) Limit: 20615843021 (19660MiB), %mem: 60.0
Available memory: 42949672960 (40960MiB) Limit: 25769803776 (24576MiB), %mem: 60.0
Available memory: 68719476736 (65536MiB) Limit: 41231686041 (39321MiB), %mem: 60.0
Available memory: 103079215104 (98304MiB) Limit: 61847529063 (58982MiB), %mem: 60.0
Available memory: 137438953472 (131072MiB) Limit: 82463372083 (78643MiB), %mem: 60.0
Available memory: 274877906944 (262144MiB) Limit: 164926744167 (157286MiB), %mem: 60.0
Available memory: 549755813888 (524288MiB) Limit: 329853488333 (314572MiB), %mem: 60.0
As before, a per-query memory limit value of 0
means no limitation.
The limit values are per AQL query, so they may still be too high in case
queries run in parallel. The defaults are intentionally high in order to not
stop any valid, previously working queries from succeeding.
Using a per-query memory limit by default is a downwards-incompatible change in
ArangoDB 3.8 and may make queries fail if they use a lot of memory. If this
happens, it may be useful to increase the value of --query.memory-limit
or
even set it to 0
(meaning no limitation).
There is a metric arangodb_aql_local_query_memory_limit_reached
that can be
used to check how many times queries reached the per-query memory limit.
There is now also a startup option --query.memory-limit-override
which can be
used to control whether individual AQL queries can increase their memory limit
via the memoryLimit
query option. This is the default, so a query that
increases its memory limit is allowed to use more memory than set via the
--query.memory-limit
startup option value. If the option is set to false
,
individual queries can only lower their maximum allowed memory usage but not
increase it.
Global AQL query memory limit
The new startup option --query.global-memory-limit
can be used to set a limit
on the combined estimated memory usage of all AQL queries (in bytes). If this
option has a value of 0
, then no global memory limit is in place. This is
also the default value and the same behavior as in previous versions of ArangoDB.
Setting the option to a value greater than zero will mean that the total memory usage of all AQL queries will be limited approximately to the configured value. The limit is enforced by each server in a cluster independently, i.e. it can be set separately for Coordinators, DB-Servers etc. The memory usage of a query that runs on multiple servers in parallel is not summed up, but tracked separately on each server.
If a memory allocation in a query would lead to the violation of the configured global memory limit, then the query is aborted with error code 32 (“resource limit exceeded”).
The global memory limit is approximate, in the same fashion as the per-query
memory limit exposed by the option --query.memory-limit
is. Some operations,
namely calls to AQL functions and their intermediate results, are currently not
properly tracked.
The global query memory limit in option --query.global-memory-limit
has a
default value that depends on the amount of available RAM:
Available memory: 0 (0MiB) Limit: 0 unlimited, %mem: n/a
Available memory: 134217728 (128MiB) Limit: 33554432 (32MiB), %mem: 25.0
Available memory: 268435456 (256MiB) Limit: 67108864 (64MiB), %mem: 25.0
Available memory: 536870912 (512MiB) Limit: 255013683 (243MiB), %mem: 47.5
Available memory: 805306368 (768MiB) Limit: 510027366 (486MiB), %mem: 63.3
Available memory: 1073741824 (1024MiB) Limit: 765041049 (729MiB), %mem: 71.2
Available memory: 2147483648 (2048MiB) Limit: 1785095782 (1702MiB), %mem: 83.1
Available memory: 4294967296 (4096MiB) Limit: 3825205248 (3648MiB), %mem: 89.0
Available memory: 8589934592 (8192MiB) Limit: 7752415969 (7393MiB), %mem: 90.2
Available memory: 17179869184 (16384MiB) Limit: 15504831938 (14786MiB), %mem: 90.2
Available memory: 25769803776 (24576MiB) Limit: 23257247908 (22179MiB), %mem: 90.2
Available memory: 34359738368 (32768MiB) Limit: 31009663877 (29573MiB), %mem: 90.2
Available memory: 42949672960 (40960MiB) Limit: 38762079846 (36966MiB), %mem: 90.2
Available memory: 68719476736 (65536MiB) Limit: 62019327755 (59146MiB), %mem: 90.2
Available memory: 103079215104 (98304MiB) Limit: 93028991631 (88719MiB), %mem: 90.2
Available memory: 137438953472 (131072MiB) Limit: 124038655509 (118292MiB), %mem: 90.2
Available memory: 274877906944 (262144MiB) Limit: 248077311017 (236584MiB), %mem: 90.2
Available memory: 549755813888 (524288MiB) Limit: 496154622034 (473169MiB), %mem: 90.2
Using a global memory limit for all queries by default is a
downwards-incompatible change in ArangoDB 3.8 and may make queries fail if they
use a lot of memory. If this happens, it may be useful to increase the value of
--query.global-memory-limit
or even set it to 0
(meaning no limitation).
There is a metric arangodb_aql_global_query_memory_limit_reached
that can be
used to check how many times queries reached the global memory limit.
If both --query.global-memory-limit
and --query.memory-limit
are set, the
former must be set at least as high as the latter.
Shard synchronization
Improvements for initial synchronization
The initial replication of collections/shards data is now faster by not wrapping
each document in a separate {"type":2300,"data":...}
envelope. In addition, the
follower side of the replication will request initial shard data from leaders in
VelocyPack format if the leader is running at least version 3.8.
Stripping the envelopes and using VelocyPack for data transfer allows for smaller data sizes when exchanging the documents and for faster processing, and thus can lead to time savings in document packing and unpacking as well as a reduction in the number of required roundtrips.
The shard synchronization protocol was also improved by only transferring the required parts of the inventory from leader to follower. Previously, for each shard the entire inventory was exchanged, which included all shards of the respective database with all their details. This change helps to reduce memory usage and speed up initial synchronization for databases with lots of collections or shards.
In addition, 3 cluster-internal requests are now saved per shard in the initial shard synchronization protocol by reusing already existing information in the different steps of the replication process. All these changes can speed up the getting-in-sync of followers after a server restart, or when provisioning new replicas.
Replication protocol based on Merkle trees
For collections created with ArangoDB 3.8, a new internal data format is used that allows for a very fast synchronization of differences between the leader and a follower that is trying to reconnect.
The new format used in 3.8 is based on Merkle trees, making it more efficient to pin-point the data differences between the leader and a follower that is trying to reconnect.
The algorithmic complexity of the new protocol is determined by the amount of differences between the leader and follower shard data, meaning that if there are no or very few differences, the getting-in-sync protocol will run very fast. In previous versions of ArangoDB, the complexity of the protocol was determined by the number of documents in the shard, and the protocol required a scan over all documents in the shard on both the leader and the follower to find the differences.
The new protocol is used automatically for all collections/shards created with ArangoDB 3.8. Collections/shards created with earlier versions will use the old protocol, which is still fully supported.
New deployments created with ArangoDB 3.8 will automatically benefit from the new protocol, and existing deployments will benefit from the new protocol for any collections that are created with 3.8 onwards. Existing collections created with previous versions of ArangoDB will only benefit from the new protocol if the collections are dumped and recreated/restored using arangodump and arangorestore.
Index selectivity estimates
Compressed estimates format
When index selectivity estimates are updated and written to disk, they are now written in a compressed format. This can greatly reduce the amount of data written to disk for each index estimate update. The compressed format is used automatically in ArangoDB 3.8 for all selectivity estimate writes.
Less impact of selectivity estimate updates for system collections
Previous versions of ArangoDB could suffer from an “idle writes” problem, in which an otherwise idle arangod instance would still write a lot of data to disk over time. These writes happened because the server statistics feature periodically stored the current statistics in some system collections, so that they can be retrieved later and also be inspected from the web interface at any point.
With ArangoDB 3.8 these background writes to the statistics collections will still happen, but their impact has been greatly reduced: if the statistics collections are created with ArangoDB 3.8 (this will happen when creating a new deployment based on 3.8), there will be no updates to the index selectivity estimates of the statistics collections at all. This will save the majority of the write payload size. For deployments created with earlier versions of ArangoDB, the index selectivity estimates for the statistics collections will still be updated periodically, but they are written in the compressed index selectivity estimates format (see above).
Optional selectivity estimates for new indexes
For any user-defined index of type “persistent”, it is now also possible to
disable index selectivity estimates for the index, by setting the estimates
flag to false
when creating the index, e.g.
db.myCollection.ensureIndex({ type: "persistent", fields: ["value"], estimates: false });
By default index selectivity estimates are maintained for all newly created indexes. Turning them off can have a slightly positive performance impact for write operations. The downside of turning off index selectivity estimates will be that the query optimizer will not be able to determine the usefulness of different competing indexes in AQL queries when there are multiple candidate indexes to choose from.
Encryption at Rest
The Encryption at Rest feature in the ArangoDB 3.8 Enterprise Edition will now automatically use hardware acceleration for encryption and decryption if available.
The AES-NI instruction set (Advanced Encryption Standard New Instructions) will be used if available on the target platform. This instruction set is available on major Intel and AMD processors for around a decade.
The benefits of using the hardware-accelerated version of AES are better performance than for a software-only implementation, plus resistance to side-channel attacks.
All other things equal, deployments that use Encryption at Rest should see a reduction of CPU usage by using the hardware-accelerated encryption.
HTTP security options
ArangoDB 3.8 provides a new startup option --cluster.api-jwt-policy
that
allows additional checking for valid JWTs in all requests to sub-routes of
the /_admin/cluster
REST API endpoint.
This is a security option to restrict access to these cluster APIs to
operator tools and privileged users.
The possible values for the startup option are:
jwt-all
: requires a valid JWT for all accesses to/_admin/cluster
and its sub-routes. If this configuration is used, the CLUSTER and NODES sections of the web interface will be disabled, as they are relying on the ability to read data from several cluster APIs.jwt-write
: requires a valid JWT for write accesses (all HTTP methods except HTTP GET) to/_admin/cluster
. This setting can be used to allow privileged users to read data from the cluster APIs, but not to do any modifications. All existing permissions checks for the cluster API routes are still in effect with this setting, meaning that read operations without a valid JWT may still require dedicated other permissions (as in v3.7).jwt-compat
: no additional access checks are in place for the cluster APIs. However, all existing permissions checks for the cluster API routes are still in effect with this setting, meaning that all operations may still require dedicated other permissions (as in v3.7).
The default value for the option is jwt-compat
, which means this option will
not cause any extra JWT checks compared to v3.7.
JavaScript security options
The following startup options have been added to optionally limit certain areas of JavaScript code execution:
-
--javascript.tasks
: the default value for this option istrue
, meaning JavaScript tasks are available as before. However, with this option they can be turned off by admins to limit the amount of JavaScript user code that is executed. -
--javascript.transactions
: the default value for this option istrue
, meaning JavaScript Transactions are available as before. However, with this option they can be turned off by admins to limit the amount of JavaScript user code that is executed.
Metrics
3.8 features a new metrics API under /_admin/metrics/v2
.
This became necessary, since the old metrics output was not following
all Prometheus conventions for metrics. For example, the naming
convention says that the name of a counters must end in _total
.
Furthermore, the histogram bucket counts must be reported cumulated.
Fixing all these is a breaking change, therefore we continue to serve
the old metrics output (with old names and uncumulated histograms)
under /_admin/metrics
and deprecate this API in 3.8. It will be
removed in future versions.
The new API under /_admin/metrics/v2
should be used from now on and
we publish new dashboards for Grafana for it. We have defined multiple
“personas” and build individual dashboards which each include a certain
subset of the metrics tailored for the particular persona. So for
example, a database admin would only see metrics which are relevant
for the database administration work. Of course, there is also a
dashboard with all metrics, neatly sorted into categories. In 3.8,
we have over 200 metrics and nearly 300 graphs in the complete
dashboard.
The complete list of metrics together with documentation can be found in the Metrics HTTP API documentation.
The list of renamed metrics can be found under API Changes in 3.8.
For the description of a seamless upgrade path see Incompatible changes in 3.8.
Traffic accounting metrics
Introduced in: v3.8.9
The following metrics for traffic accounting were added:
Label | Description |
---|---|
arangodb_client_user_connection_statistics_bytes_received |
Bytes received for requests, only user traffic. |
arangodb_client_user_connection_statistics_bytes_sent |
Bytes sent for responses, only user traffic. |
arangodb_http1_connections_total |
Total number of HTTP/1.1 connections accepted. |
Logging
New options for logging
The following logging-related options have been added:
-
added option
--log.use-json-format
to switch log output to JSON format. Each log message then produces a separate line with JSON-encoded log data, which can be consumed by applications.The attributes produced for each log message JSON object are:
Key Value time
date/time of log message, in format specified by --log.time-format
prefix
only emitted if --log.prefix
is setpid
process id, only emitted if --log.process
is settid
thread id, only emitted if --log.thread
is setthread
thread name, only emitted if --log.thread-name
is setrole
server role (1 character), only emitted if --log.role
is setlevel
log level (e.g. "WARN"
,"INFO"
)file
source file name of log message, only emitted if --log.line-number
is setline
source file line of log message, only emitted if --log.line-number
is setfunction
source file function name, only emitted if --log.line-number
is settopic
log topic name id
log id (5 digit hexadecimal string), only emitted if --log.ids
is sethostname
hostname if --log.hostname
is setmessage
the actual log message payload -
added option
--log.process
to toggle the logging of the process id (pid) in log messages. Logging the process ID is useless when running arangod in Docker containers, as the pid will always be 1. So one may as well turn it off in these contexts with the new option. -
added option
--log.hostname
to optionally log the current host’s name at the beginning of each log message (or inside thehostname
attribute for JSON-based logging). Setting--log.hostname
to a value ofauto
will automatically determine the hostname and use that for logging. -
added option
--log.in-memory
to toggle storing log messages in memory, from which they can be consumed via the/_admin/log
HTTP API and by the Web UI. By default, this option is turned on, so log messages are consumable via the API and UI. Turning this option off will disable that functionality, save a tiny bit of memory for the in-memory log buffers and prevent potential log information leakage via these means. -
added option
--log.in-memory-level
to control which log messages are preserved in memory (in case --log.in-memory is set to true). The default value isinfo
, meaning all log messages of typesinfo
,warning
,error
andfatal
will be stored by an instance in memory. By setting this option towarning
, only warning log messages will be preserved in memory, and by setting the option toerror
only error messages will be kept. This option is useful because the number of in-memory log messages is limited to the latest 2048 messages, and these slots are by default shared between informational, warning and error messages. -
added option
--log.max-entry-length
to control the maximum line length for individual log messages that are written into normal logfiles by arangod (note: this does not include audit log messages). Any log messages longer than the specified value will be truncated and the suffix ‘…’ will be added to them. The purpose of this parameter is to shorten long log messages in case there is not a lot of space for logfiles, and to keep rogue log messages from overusing resources. The default value is 128 MB, which is very high and should effectively mean downwards-compatibility with previous arangod versions, which did not restrict the maximum size of log messages. -
added option
--audit.max-entry-length
to control the maximum line length for individual audit log messages that are written into audit logs by arangod. Any audit log messages longer than the specified value will be truncated and the suffix ‘…’ will be added to them. The default value is 128 MB, which is very high and should effectively mean downwards-compatibility with previous arangod versions, which did not restrict the maximum size of log messages. -
added option
--audit.queue
to control audit logging queuing behavior (Enterprise Edition only):The option controls whether audit log messages are submitted to a queue and written to disk in batches or if they should be written to disk directly without being queued. Queueing audit log entries may be beneficial for latency, but can lead to unqueued messages being lost in case of a power loss or crash. Setting this option to
false
mimics the behavior from 3.7 and before, where audit log messages were not queued but written in a blocking fashion. -
any occurrence of
$PID
inside a log output value (e.g.--log.output
or--audit.output
) will be replaced at runtime with the actual process id. This enables logging to process-specific files.Please note that the dollar sign in
$PID
may need extra escaping when specified from inside shells such as Bash.
Other logging improvements
-
The maximum size of log messages buffered in memory was increased from 256 bytes per log message to 512 bytes per log message. This should prevent most in-memory log messages returned by the
/_admin/log
HTTP API from being truncated unnecessarily. -
Audit logging and slow query logging for AQL queries now also include the query’s result code (success or error code in case the query ran into an error). This can be used to find queries which ran into errors (audit logging) or long-running queries which ran into errors (normal logging).
-
Audit logging now also honors the configured logging date/time output format for the regular logger. Previously the audit logging always logged date/time value in the server’s local time, and used the format
YYYY-MM-DDTHH:MM:SS
.From 3.8 onwards, the audit logger will use the format specified via the
--log.time-format
option, which defaults toutc-datestring
. The means the audit logging will by default log all dates/times in UTC time. To restore the pre-3.8 behavior, please set the option tolocal-datestring
, which will make the audit logger (and all other server log messages) use the server’s local time.
Timezone conversion
Added IANA timezone database tzdata.
The following AQL functions have been added for converting datetimes in UTC to any timezone in the world including historical daylight saving times and vice versa. An optional detail flag returns the timezone information including effect range, abbreviation, offset to UTC and whether daylight saving time is active:
-
RETURN DATE_UTCTOLOCAL("2020-10-15T01:00:00.999Z", "America/New_York") // [ "2020-10-14T21:00:00.999" ]
RETURN DATE_UTCTOLOCAL("2020-10-15T01:00:00.999Z", "America/New_York", true) /* { "local": "2020-10-14T21:00:00.999", "tzdb": "2020f", "zoneInfo": { "name": "EDT", "begin": "2020-03-08T07:00:00.000Z", "end": "2020-11-01T06:00:00.000Z", "save": true, "offset": -14400 } } */
-
RETURN DATE_LOCALTOUTC("2020-10-14T21:00:00.999", "America/New_York") // [ "2020-10-15T01:00:00.999Z" ]
RETURN DATE_LOCALTOUTC("2020-10-14T21:00:00.999", "America/New_York") /* { "utc": "2020-10-15T01:00:00.999Z", "tzdb": "2020f", "zoneInfo": { "name": "EDT", "begin": "2020-03-08T07:00:00.000Z", "end": "2020-11-01T06:00:00.000Z", "save": true, "offset": -14400 } } */
Also some functions have been added to acquire the system timezone ArangoDB is running on and to list all valid IANA timezone names including canonical, aliases and deprecated ones.
-
RETURN DATE_TIMEZONE() // [ "Etc/UTC" ]
-
RETURN DATE_TIMEZONES() // [ "Africa/Abidjan", ..., "Europe/Berlin", ..., "Zulu" ]
Client tools
arangodump concurrency / shard-parallelism
Since v3.4.0, arangodump can use multiple threads for dumping database data in parallel. arangodump versions prior to v3.8.0 distribute dump jobs for individual collections to concurrent worker threads, which is optimal for dumping many collections of approximately the same size, but does not help for dumping few large collections or few large collections with many shards.
Starting with v3.8.0, arangodump can also dispatch dump jobs for individual shards of each collection, allowing higher parallelism if there are many shards to dump but only few collections.
Also see arangodump Threads.
arangodump output format
Since its inception, arangodump wrapped each dumped document into an extra JSON envelope, such as follows:
{"type":2300,"key":"test","data":{"_key":"test","_rev":..., ...}}
In case a dump taken with v3.8.0 or higher is known to never be used in older
ArangoDB versions, the JSON envelopes can be turned off with the new startup
option --envelope false
to reduce the dump size and use a bit less memory
and bandwidth:
{"_key":"test","_rev":..., ...}
Also see arangodump Dump Output Format.
Using the new non-enveloped dump format also allows arangorestore to parallelize restore operations for individual collections. This is not possible with the old, enveloped format.
arangorestore parallelization for single collections
arangorestore can now parallelize restore operations even for single collections, which can lead to increased restore performance. This requires that a dump in the new non-enveloped dump format is used, and that there are enough arangorestore threads to employ.
The dump format can be configured by specifying the --envelope false
option
when invoking arangodump, and the number of restore threads can be adjusted by
setting arangorestore’s --threads
option.
arangodump dumping of individual shards
arangodump can now optionally dump individual shards only, by specifying the
--shard
option one or multiple times. This option can be used to split the
dump of a large collection with multiple shards into multiple separate dump
processes, which could be run against different Coordinators etc.
arangodump and arangorestore with JWT secret
arangodump and arangorestore can now also be invoked by providing the cluster’s
JWT secret instead of the username/password combination. Both tools now provide
the options --server.jwt-secret-keyfile
(to read the JWT secret from a file)
and --server.ask-jwt-secret
(to enter it manually).
arangobench with custom queries
In addition to executing the predefined benchmarks, the arangobench client tool
now offers a new test case named custom-query
for running arbitrary AQL
queries against an ArangoDB installation.
To run a custom AQL query, the query needs to be specified in either the
--custom-query
option or the --custom-query-file
option. In the former case
the query string can be passed on the command-line, in the latter case the
query string will be read from a file.
Continuing arangorestore operations
arangorestore now provides a --continue
option. Setting it will make
arangorestore keep track of the restore progress, so if the restore process
gets aborted it can later be continued from the point it left off.
Controlling the number of documents per batch for arangoexport
arangoexport now has a --documents-per-batch
option that can be used to limit
the number of documents to be returned in each batch from the server. This is
useful if a query is run on overly large documents, which would lead to the
response sizes getting out of hand with the default number of documents per
batch (1000).
Controlling the maximum query runtime of arangoexport
arangoexport now has a --query-max-runtime
option to limit the runtime of
queries it executes.
arangorestore option to enable revision trees
Introduced in: v3.8.7, v3.9.2
A new --enable-revision-trees
option has been added to arangorestore, which
adds the syncByRevision
and usesRevisionsAsDocumentIds
attributes to the
collection structure if they are missing. As a consequence, these collections
created by arangorestore are able to use revision trees and a faster
getting-in-sync procedure after a restart.
The option defaults to true
, meaning that the attributes are added if they are
missing. If you set the option to false
, the attributes are not added to the
collection structure. If the attributes are already present in the dump data, they
are not modified by arangorestore, irrespective of the setting of this option.
Miscellaneous
Cluster support for two APIs
-
Added cluster support for the JavaScript API method
collection.checksum()
and the REST HTTP API endpointGET /_api/collection/{collection-name}/checksum
, which calculate CRC checksums for collections. -
Added cluster support for the JavaScript API method
db._engineStats()
and the REST HTTP API endpointGET /_api/engine/stats
, which provide runtime information about the storage engine state.
I/O heartbeat
Introduced in: v3.8.7, v3.9.2
An I/O heartbeat has been added which checks that the underlying volume is writable with reasonable performance. The test is done every 15 seconds and can be switched off.
Use the accompanying new metrics to check for test failures:
Label | Description |
---|---|
arangodb_ioheartbeat_delays_total |
Total number of delayed I/O heartbeats. |
arangodb_ioheartbeat_duration |
Histogram of execution times in microseconds. |
arangodb_ioheartbeat_failures_total |
Total number of failures. |
These metrics are only populated if the new --database.io-heartbeat
startup
option is set to true
(which is the default).
Internal changes
Library version upgrades
The bundled version of the Snappy compression/decompression library has been upgraded to 1.1.8.
The bundled version of libunwind has been upgraded to 1.5.
For ArangoDB 3.8, the bundled version of rclone is 1.51.0.
Spliced subqueries
The AQL optimizer rule “splice-subqueries” is now mandatory, in the sense that
it cannot be disabled anymore. As a side effect of this change, there will no
query execution plans created by 3.8 that contain execution nodes of type
SubqueryNode
. SubqueryNode
s will only be used during query planning and
optimization, but at the end of the query optimization phase will all have
been replaced with nodes of types SubqueryStartNode
and SubqueryEndNode
.
The code to execute non-spliced subqueries remains in place so that 3.8 can still execute queries planned on a 3.7 instance with the “splice-subqueries” optimizer rule intentionally turned off. The code for executing non-spliced subqueries can be removed in 3.9.
Query register usage
There is an AQL query execution plan register usage optimization that may positively affect some AQL queries that use a lot of variables that are only needed in certain parts of the query. The positive effect will come from saving registers, which directly translates to saving columns in AqlItemBlocks.
Previously, the number of registers that were planned for each depth level of the query never decreased when going from one level to the next. Even though unused registers were recycled since 3.7, this did not lead to unused registers being completely dismantled.
Now there is an extra step at the end of the register planning that keeps track of the actually used registers on each depth, and that will shrink the number of registers for the depth to the id of the maximum register. This is done for each depth separately. Unneeded registers on the right hand side of the maximum used register are now discarded. Unused registers on the left-hand side of the maximum used register id are not discarded, because we still need to guarantee that registers from depths above stay in the same slot when starting a new depth.
Better protection against overwhelm
The cluster now protects itself better against being overwhelmed by too many concurrent requests.
This is mostly achieved by limiting the total amount of requests from
the low priority queue which are ongoing concurrently. There is a new option
--server.ongoing-low-priority-multiplier
(default is 4), which
essentially says that only 4 times as many requests may be ongoing
concurrently as there are worker threads. The default is chosen such
that it is sensible for most workloads, but in special situations it
can help to adjust the value.
See ArangoDB Server Server Options for details and hints for configuration.
There have been further improvements, in particular to ensure that certain APIs to diagnose the situation in the cluster still work, even when a lot of normal requests are piling up. For example, the cluster health API will still be available in such a case.
Furthermore, followers will now be dropped much later and only if they are actually failed, which leads to a lot fewer shard re-synchronizations in case of very high load.
Overall, these measures should all be below the surface and not be visible to the user at all (apart from preventing problems under high load).