Features and Improvements in ArangoDB 3.7
The following list shows in detail which features have been added or improved in ArangoDB 3.7. ArangoDB 3.7 also contains several bug fixes that are not listed here.
ArangoSearch
Wildcard search
ArangoSearch was extended to support the LIKE()
function and LIKE
operator
in AQL. This allows to check whether the given search pattern is contained in
specified attribute using wildcard matching (_
for any single character and
%
for any sequence of characters including none):
FOR doc IN viewName
SEARCH ANALYZER(LIKE(doc.text, "foo%b_r"), "text_en")
// or
SEARCH ANALYZER(doc.text LIKE "foo%b_r", "text_en")
// will match "foobar", "fooANYTHINGbor" etc.
RETURN doc.text
Covering Indexes
It is possible to directly store the values of document attributes in View
indexes now via a new View property storedValues
(not to be confused with
the existing storeValues
).
View indexes may fully cover SEARCH
queries for improved performance.
While late document materialization reduces the amount of fetched documents,
this new optimization can avoid to access the storage engine entirely.
{
"links": {
"articles": {
"fields": {
"categories": {}
}
}
},
"primarySort": [
{ "field": "publishedAt", "direction": "desc" }
],
"storedValues": [
{ "fields": [ "title", "categories" ] }
],
...
}
In above View definition, the document attribute categories is indexed for
searching, publishedAt is used as primary sort order and title as well as
categories are stored in the View using the new storedValues
property.
FOR doc IN articlesView
SEARCH doc.categories == "recipes"
SORT doc.publishedAt DESC
RETURN {
title: doc.title,
date: doc.publishedAt,
tags: doc.categories
}
The query searches for articles which contain a certain tag in the categories
array and returns title, date and tags. All three values are stored in the View
(publishedAt
via primarySort
and the two other via storedValues
), thus
no documents need to be fetched from the storage engine to answer the query.
This is shown in the execution plan as a comment to the EnumerateViewNode:
/* view query without materialization */
Execution plan:
Id NodeType Est. Comment
1 SingletonNode 1 * ROOT
2 EnumerateViewNode 1 - FOR doc IN articlesView SEARCH (doc.`categories` == "recipes") SORT doc.`publishedAt` DESC LET #1 = doc.`publishedAt` LET #7 = doc.`categories` LET #5 = doc.`title` /* view query without materialization */
5 CalculationNode 1 - LET #3 = { "title" : #5, "date" : #1, "tags" : #7 } /* simple expression */
6 ReturnNode 1 - RETURN #3
Indexes used:
none
Optimization rules applied:
Id RuleName
1 move-calculations-up
2 move-calculations-up-2
3 handle-arangosearch-views
See ArangoSearch Views.
Stemming support for more languages
The Snowball library was updated to the latest version 2, adding stemming support for the following languages:
- Arabic (
ar
) - Basque (
eu
) - Catalan (
ca
) - Danish (
da
) - Greek (
el
) - Hindi (
hi
) - Hungarian (
hu
) - Indonesian (
id
) - Irish (
ga
) - Lithuanian (
lt
) - Nepali (
ne
) - Romanian (
ro
) - Serbian (
sr
) - Tamil (
ta
) - Turkish (
tr
)
Create a custom Analyzer and set the locale
accordingly in the properties,
e.g. "el.utf-8"
for Greek. arangosh example:
var analyzers = require("@arangodb/analyzers");
analyzers.save("text_el", "text", {
locale: "el.utf-8",
stemming: true,
case: "lower",
accent: false,
stopwords: []
}, ["frequency", "norm", "position"]);
db._query(`RETURN TOKENS("αυτοκινητουσ πρωταγωνιστούσαν", "text_el")`)
// [ [ "αυτοκινητ", "πρωταγωνιστ" ] ]
Also see Analyzers: Supported Languages
Condition Optimization Option
The SEARCH
operation in AQL accepts a new option conditionOptimization
to
give users control over the search criteria optimization:
FOR doc IN myView
SEARCH doc.val > 10 AND doc.val > 5 /* more conditions */
OPTIONS { conditionOptimization: "none" }
RETURN doc
By default, all conditions get converted into disjunctive normal form (DNF).
Numerous optimizations can be applied, like removing redundant or overlapping
conditions (such as doc.val > 10
which is included by doc.val > 5
).
However, converting to DNF and optimizing the conditions can take quite some
time even for a low number of nested conditions which produce dozens of
conjunctions / disjunctions. It can be faster to just search the index without
optimizations.
See SEARCH operation.
Primary Sort Compression Option
There is a new option primarySortCompression
which can be set on View
creation to enable or disable the compression of the primary sort data:
{
"primarySort": [
{ "field": "date", "direction": "desc" },
{ "field": "title", "direction": "asc" }
],
"primarySortCompression": "none",
...
}
It defaults to LZ4 compression ("lz4"
), which was already used in ArangoDB
v3.5 and v3.6. Set it to "none"
on View creation to trade space for speed.
See ArangoSearch Views.
SatelliteGraphs
When doing joins involving graph traversals, shortest path or k-shortest paths computation in an ArangoDB cluster, data has to be exchanged between different servers. In particular graph traversals are usually executed on a Coordinator, because they need global information. This results in a lot of network traffic and potentially slow query execution.
SatelliteGraphs are the natural extension of the concept of SatelliteCollections to graphs. All of the usual benefits and caveats apply. SatelliteGraphs are synchronously replicated to all DB-Servers that are part of a cluster, which enables DB-Servers to execute graph traversals locally. This includes (k-)shortest path(s) computation and possibly joins with traversals and greatly improves performance for such queries.
SatelliteGraphs are only available in the Enterprise Edition.
Disjoint SmartGraphs
SmartGraphs have been extended with a new option isDisjoint
.
A Disjoint SmartGraph prohibits edges connecting different SmartGraph
components. If your graph doesn’t 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 in many cases the execution can be pushed down to a single DB-Server.
Disjoint SmartGraphs are only available in the Enterprise Edition.
AQL
Subquery optimizations
The execution process of AQL has been refactored internally. This especially pays off in subqueries. It will allow for more optimizations and better batching of requests.
The first stage of this refactoring has been part of 3.6 already where some
subqueries have gained a significant performance boost. 3.7 takes the next step
in this direction. AQL can now combine skipping and producing of outputs in a
single call, so all queries with a LIMIT offset or the fullCount option enabled
will benefit from this change straight away. This also holds true for subqueries,
hence the existing AQL optimizer rule splice-subqueries
is now able to
optimize all subqueries and is enabled by default.
The query planner can now also reuse internal registers that were allocated for storing temporary results inside subqueries, but not outside of subqueries.
Count optimizations
Subqueries can now use an optimized code path for counting documents if they are supposed to only return the number of matching documents. The optimization will be triggered for read-only subqueries that use a full collection scan or an index scan, without any additional filtering on document attributes (early pruning or document post-filtering) and without using LIMIT.
The optimization will help in the following situation (in case subCollection
is an edge collection):
FOR doc IN collection
LET count = COUNT(
FOR sub IN subCollection
FILTER sub._from == doc._id
RETURN sub
)
...
The restrictions are that the subquery result must only be used with the
COUNT
/LENGTH
AQL function and not for anything else. The subquery itself
must be read-only (no data-modification subquery), not use nested FOR loops,
no LIMIT clause and no FILTER condition or calculation that requires
accessing document data. Accessing index data is supported for filtering (as
in the above example that would use the edge index), but not for further
calculations.
In case a subquery does not match these criteria, it will not use the optimized code path for counting, but will execute normally.
If the optimization is triggered, it will show up in the query execution
plan under the rule name optimize-count
, and the subquery’s FOR loop will
be marked with a with count optimization
tag.
Traversal optimizations
Graph traversal performance is improved via some internal code refactoring:
- Traversal cursors are reused instead of recreated from scratch, if possible. This can save lots of calls to the memory management subsystem.
- Unnecessary checks have been removed from the cursors, by ensuring some invariants beforehand.
- Each vertex lookup needs to perform slightly less work.
The traversal speedups observed by these changes alone were around 8 to 10% for single-server traversals and traversals in OneShard setups. Cluster traversals will also benefit from these changes, but to a lesser extent. This is because the network roundtrips have a higher share of the total query execution times there.
Traversal performance can be further improved by not fetching the visited vertices from the storage engine in case the traversal query does not refer to them. For example, in the query:
FOR v, e, p IN 1..3 OUTBOUND 'collection/startVertex' edges
RETURN e
…the vertex variable (v
) is never accessed, making it unnecessary to fetch the
vertices from storage. If this optimization is applied, the traversal node will be
marked with /* vertex optimized away */
in the query’s execution plan output.
Unused edge and path variables (e
and p
) were already optimized away in
previous versions by the optimize-traversals
optimizer rule.
Traversal collection restrictions
AQL traversals now accept the options vertexCollections
and edgeCollections
to restrict the traversal to certain vertex or edge collections.
The use case for vertexCollections
is to not follow any edges that will point
to other than the specified vertex collections, e.g.
FOR v, e, p IN 1..3 OUTBOUND 'products/123' components
OPTIONS { vertexCollections: [ "bolts", "screws" ] }
RETURN v
The traversal’s start vertex is always considered valid, even if it not stored
in any of the collections listed in the vertexCollections
option.
The use case for edgeCollections
is to not take into consideration any edges
from edge collections other than the specified ones, e.g.
FOR v, e, p IN 1..3 OUTBOUND 'products/123' GRAPH 'components'
OPTIONS { edgeCollections: [ "productsToBolts", "productsToScrews" ] }
RETURN v
This is mostly useful in the context of named graphs, when the named graph
contains many edge collections. Not restricting the edge collections for the
traversal will make the traversal search for edges in all edge collections
of the graph, which can be expensive. In case it is known that only certain
edges from the named graph are needed, the edgeCollections
option can be a
handy performance optimization. It can replace less efficient post-filtering:
FOR v, e, p IN 1..3 OUTBOUND 'products/123' GRAPH 'components'
FILTER p.edges[* RETURN IS_SAME_COLLECTION("productsToBolts", CURRENT)
OR IS_SAME_COLLECTION("productsToScrews", CURRENT)] ALL == true
RETURN v
Also see AQL Traversal Options
Traversal parallelization (Enterprise Edition)
Nested traversals that run on a single server or a cluster DB-Server can now be executed in parallel.
Traversals have a new option parallelism
which can be used to specify the
level of parallelism:
FOR doc IN outerCollection
FOR v, e, p IN 1..3 OUTBOUND doc._id GRAPH 'components'
OPTIONS { parallelism: 4 }
...
Traversal parallelism is opt-in. If not specified, the parallelism
value
implicitly defaults to 1, which means no parallelism will be used. The maximum
value for parallelism
is capped to the number of available cores on the
target machine.
Due to the required synchronization for splitting up traversal inputs and merging results, using traversal parallelization may incur some overhead. So it is not a silver bullet for all use cases. However, parallelizing a traversal is normally useful when there are many inputs (start vertices) that the nested traversal can work on concurrently. This is often the case when a nested traversal is fed with several tens of thousands of start vertices, which can then be distributed randomly to worker threads for parallel execution.
Right now, traversal parallelization is limited to traversals in single server deployments and to cluster traversals that are running in a OneShard setup. Cluster traversals that run on a coordinator node and SmartGraph traversals are currently not parallelized.
AQL functions added
The following AQL functions have been added in ArangoDB 3.7:
- IN_RANGE()
(now available outside of
SEARCH
operations) - INTERLEAVE()
- IPV4_FROM_NUMBER()
- IPV4_TO_NUMBER()
- IS_IPV4()
- JACCARD()
- LEVENSHTEIN_MATCH()
- NGRAM_MATCH()
- NGRAM_POSITIONAL_SIMILARITY()
- NGRAM_SIMILARITY()
- PRODUCT()
- REPLACE_NTH()
Syntax enhancements
AQL now supports trailing commas in array and object definitions.
This is especially convenient for editing multi-line array/object definitions, since there doesn’t need to be a distinction between the last element and all others just for the comma. That means definitions such as the following are now supported:
[
1,
2,
3, // trailing comma
]
{
"a": 1,
"b": 2,
"c": 3, // trailing comma
}
Previous versions of ArangoDB did not support trailing commas in AQL queries and threw query parse errors when they were used.
AQL datetime parsing
The performance of parsing ISO 8601 date/time string values in AQL has improved significantly thanks to a specialized parser, replacing a regular expression.
Ternary operator
Improved the lazy evaluation capabilities of the ternary operator. If the second operand is left out, the expression of the condition is only evaluated once now, instead of once more for the true branch.
Other AQL improvements
“remove-unnecessary-calculations” optimizer rule
The AQL query optimizer now tries to not move potentially expensive AQL function
calls into loops in the remove-unnecessary-calculations
rule.
For example, in the query:
LET x = NOOPT(1..100)
LET values = SORTED(x)
FOR j IN 1..100
FILTER j IN values
RETURN j
… there is only one use of the values
variable. So the optimizer can remove
that variable and replace the filter condition with FILTER j IN SORTED(x)
.
However, that would move the potentially expensive function call into the
inner loop, which could be a pessimization.
The optimizer will not move the calculation of values into the loop anymore when
it merges calculations in the remove-unnecessary-calculations
optimizer rule.
“move-calculations-down” optimizer rule
The existing AQL optimizer rule move-calculations-down
is now able to also move
unrelated subqueries beyond SORT and LIMIT instructions, which can help avoid the
execution of subqueries for which the results are later discarded.
For example, in the query:
FOR doc IN collection1
LET sub1 = FIRST(FOR sub IN collection2 FILTER sub.ref == doc._key RETURN sub)
LET sub2 = FIRST(FOR sub IN collection3 FILTER sub.ref == doc._key RETURN sub)
SORT sub1
LIMIT 10
RETURN { doc, sub1, sub2 }
… the execution of the sub2
subquery can be delayed to after the SORT and LIMIT.
The query optimizer will automatically transform this query into the following:
FOR doc IN collection1
LET sub1 = FIRST(FOR sub IN collection2 FILTER sub.ref == doc._key RETURN sub)
SORT sub1
LIMIT 10
LET sub2 = FIRST(FOR sub IN collection3 FILTER sub.ref == doc._key RETURN sub)
RETURN { doc, sub1, sub2 }
Cluster
Incremental Plan Updates
In ArangoDB clusters, the Agency is the single source of truth for data definition (databases, collections, shards, indexes, views), the cluster configuration and the current cluster setup (e.g. shard distribution, shard leadership).
Coordinators and DB-Servers in the cluster maintain a local cache of the Agency’s
information, in order to access it in a performant way whenever they need any
information about the setup.
However, any change that was applied to the Plan
and Current
sections in the
Agency led to the server-local caches being invalidated, which triggered a full
reload of either Plan
or Current
by all Coordinators and DB-Servers.
The size of Plan
and Current
is proportional to the number of database objects,
so fully reloading the data from the Agency is an expensive operation for deployments
which have a high number of databases, collections, or shards.
In ArangoDB 3.7 the mechanism for filling the local caches on Coordinators and
DB-Servers with Agency data has changed fundamentally. Instead of invalidating the
entire cache and reloading the full Plan
or Current
section on every change,
each server is now using a permanent connection to the Agency and uses it to
poll for changes. Changes to the Agency data are sent over these connections as
soon as they are applied in the Agency, meaning that Coordinators and DB-Servers
can apply them immediately and incrementally. This removes the need for full
reloads. As a consequence, a significant reduction of overall network traffic between
Agents and other cluster nodes is expected, plus a significant reduction in CPU
usage on Agents for assembling and sending the Plan
or Current
parts.
Another positive side effect of this modification is that changes made to Agency
data should propagate faster in the cluster.
Parallel Move Shard
Shards can now move in parallel. The old locking mechanism was replaced by a
read-write lock and thus allows multiple jobs for the same destination server.
The actual transfer rates are still limited on DB-Server side but there is a
huge overall speedup. This also affects CleanOutServer
and
ResignLeadership
jobs.
General
Schema Validation for Documents
ArangoDB now supports validating documents on collection level using JSON Schema (draft-4).
In order to enforce a certain document structure in a collection we have
introduced the schema
collection property. It expects an object comprised
of a rule
(JSON Schema object), a level
and a message
that will be used
when validation fails. When documents are validated is controlled by the
validation level, which can be none
(off), new
(insert only), moderate
(on insert and modification, but existing documents can remain invalid)
or strict
(always).
See: Schema Validation
HTTP/2 support
The server now supports upgrading connections from HTTP 1.1 to HTTP 2. This should improve ArangoDBs compatibility with various L7 load balancers and modern cloud platforms like Kubernetes.
We also expect improved request throughput in cases where there are many concurrent requests.
Server Name Indication (Enterprise Edition)
Sometimes it is desirable to have the same server use different server keys
and certificates when it is contacted under different names. This is possible
with the Server Name Indication
(SNI) TLS extension. It is now supported by ArangoDB using a new startup option
--ssl.server-name-indication
.
JWT secret rotation (Enterprise Edition)
There are now new APIs and startup options for JWT secrets. The new option
--server.jwt-secret-folder
can be used to specify a path for more than one
JWT secret file.
Additionally the /_admin/server/jwt
API can be used to
reload the JWT secrets
of a local arangod process without having to restart it (hot-reload).
This may be used to roll out new JWT secrets throughout an ArangoDB cluster.
TLS key and certificate rotation
It is now possible to change the TLS keyfile (secret key as well as
public certificates) at run time. The API POST /_admin/server/tls
basically makes the arangod
server reload the keyfile from disk.
Furthermore, one can query the current TLS setup at runtime with the
GET /_admin/server/tls
API. The public certificates as well as a
SHA-256 hash of the private key is returned.
This allows rotation of TLS keys and certificates without a server restart.
Encryption at rest key rotation (Enterprise Edition)
It is possible to change the user supplied encryption key via the
HTTP API
by sending a POST request without payload to the new endpoint
/_admin/server/encryption
. The file supplied via --rocksdb.encryption-keyfile
will be reloaded and the internal encryption key will be re-encrypted with the
new user key. Note that this API is turned off by default. It can be enabled
via the --rocksdb.encryption-key-rotation
startup option.
Similarly the new option --rocksdb.encryption-keyfolder
can be used
to supply multiple user keys. By default, the first available user-supplied key
will be used as the internal encryption key. Alternatively, if the option
--rocksdb.encryption-gen-internal-key
is set to true
, a random internal
key will be generated and encrypted with each of the provided user keys.
Please be aware that the encryption at rest key rotation is an experimental feature, and its APIs and behavior are still subject to change.
Insert-Update and Insert-Ignore
ArangoDB 3.7 adds an insert-update operation that is similar to the already
existing insert-replace functionality. A new overwriteMode
flag has been
introduced to control the type of the overwrite operation in case of colliding
primary keys during the insert.
In the case of overwriteMode: "update"
, the parameters keepNull
and
mergeObjects
can be provided to control the update operation.
There is now also an insert-ignore operation that allows insert operations
to do nothing in case of a primary key conflict. This operation is an efficient
way of making sure a document with a specific primary key exists. If it does
not exist already, it will be created as specified. Should the document exist
already, nothing will happen and the insert will return without an error. No
write operations happens in this case, and only a single primary key lookup
needs to be performed in the storage engine.
This makes the insert-ignore operation the most efficient way
existing insert-replace functionality. A new overwriteMode
flag has been
introduced to control the type of the overwrite operation in case of colliding
primary keys during the insert.
The query options are available in AQL, the JS API and HTTP API.
Override detected total memory and CPU cores
arangod
detects the total amount of RAM present on the system and calculates
various default sizes based on this value. If you run it alongside other
services or in a container with a RAM limitation for its cgroup, then you
probably don’t want the server to detect and use all available memory.
An environment variable ARANGODB_OVERRIDE_DETECTED_TOTAL_MEMORY
can now be
set to restrict the amount of memory it will detect (also available in v3.6.3).
An environment variable ARANGODB_OVERRIDE_DETECTED_NUMBER_OF_CORES
can be
set to restrict the number of CPU cores that are visible to arangod.
See ArangoDB Server Environment Variables
RocksDB storage engine options exposed
Multiple additional RocksDB configuration options are now exposed to be configurable in arangod:
--rocksdb.cache-index-and-filter-blocks
to make the RocksDB block cache quota also include RocksDB memtable sizes--rocksdb.cache-index-and-filter-blocks-with-high-priority
to use cache index and filter blocks with high priority making index and filter blocks be less likely to be evicted than data blocks--rocksdb.pin-l0-filter-and-index-blocks-in-cache
to make filter and index blocks be pinned and only evicted from cache when the table reader is freed--rocksdb.pin-top-level-index-and-filter
to make the top-level index of partitioned filter and index blocks pinned and only be evicted from cache when the table reader is freed--rocksdb.target-file-size-base
: Per-file target file size for compaction (in bytes). the actual target file size for each level is--rocksdb.target-file-size-base
multiplied by--rocksdb.target-file-size-multiplier ^ (level - 1)
--rocksdb.target-file-size-multiplier
: Multiplier for--rocksdb.target-file-size
, a value of 1 means that files in different levels will have the same size (default)
Pregel
A new algorithm "wcc"
has been added to Pregel to find connected components.
There are now three algorithms to find connected components in a graph:
-
If your graph is effectively undirected (you have edges in both directions between vertices) then the simple connected components algorithm named
"connectedcomponents"
is suitable.It is a very simple and fast algorithm, but will only work correctly on undirected graphs. Your results on directed graphs may vary, depending on how connected your components are.
-
To find weakly connected components (WCC) you can now use the new algorithm named
"wcc"
. Weakly connected means that there exists a path from every vertex pair in that component.This algorithm will work on directed graphs but requires a greater amount of traffic between your DB-Servers.
-
to find strongly connected components (SCC) you can use the algorithm named
"scc"
. Strongly connected means every vertex is reachable from any other vertex in the same component.The algorithm is more complex than the WCC algorithm and requires more memory, because each vertex needs to store much more state.
Also see Pregel
Foxx
Foxx endpoints now provide the methods securityScheme
, securityScope
and
security
to allow defining Swagger security schemes.
Foxx routes now always have a Swagger operationId
. If the route is unnamed,
a distinct operationId will be generated based on the HTTP method and URL.
JavaScript
V8 and ICU library upgrades
The bundled V8 JavaScript engine was upgraded to version 7.9.317. The bundled Unicode character handling library ICU was upgraded to version 64.2.
The resource usage of V8 has improved a lot. Memory usage is down by 15%, spawning a new Isolate has become almost 10 times faster.
Here is the list of improvements that may matter to you as an ArangoDB user:
- JSON.parse improvements:
JSON parsing is roughly 60% faster compared to ArangoDB v3.6. Parsing JSON
is generally faster than parsing JavaScript because of the lower syntactic
complexity, but with the additional speedup of the JSON parser you should
consider to use
JSON.parse(string)
over JavaScript variable declarations for complex data:// Parsing a JSON string let structuredVar = JSON.parse('{"foo": "bar", …}'); // instead of using an object literal let structuredVar = {foo: "bar", …};
Also see Embedding JSON into JavaScript programs with JSON.parse.
- BigInt support in formatter:
Large integer numbers are now supported in number formatters:
const formatter = new Intl.NumberFormat('fr'); formatter.format(12345678901234567890n);
This no longer throws an
Cannot convert a BigInt value to a number
error. Note that ArangoDB does not support BigInt in general but only in JavaScript contexts. AQL, JSON etc. do not support BigInt. - Object.fromEntries support:
Performs the inverse operation of
Object.entries()
:const object = { x: 42, y: 50 }; const entries = Object.entries(object); // → [['x', 42], ['y', 50]] const result = Object.fromEntries(entries); // → { x: 42, y: 50 }
- Underscores for better readability of large numbers:
1_000_000_000_000 // → equals 1000000000000
- matchAll support for strings:
A convenient generator for a match object for each match:
const string = 'Favorite GitHub repos: tc39/ecma262 v8/v8.dev'; const regex = /\b(?<owner>[a-z0-9]+)\/(?<repo>[a-z0-9\.]+)\b/g; for (const match of string.matchAll(regex)) { console.log(`${match[0]} at ${match.index} with '${match.input}'`); console.log(`→ owner: ${match.groups.owner}`); console.log(`→ repo: ${match.groups.repo}`);
- ICU supports more languages and characters (Unicode 12.1), emoji handling was improved
Also see:
- V8 release blog posts (v7.2 to v7.9)
- V8 features (up to Chrome 79)
JavaScript APIs
The query
helper
was extended to support passing
query options:
require("@arangodb").query( { maxRuntime: 1 } )`RETURN SLEEP(2)`
Web UI
The interactive description of ArangoDB’s HTTP API (Swagger UI) shows the endpoint and model entries collapsed by default now for a better overview.
The bundled version of Swagger has been upgraded to 3.25.1. This change has also been backported to ArangoDB v3.6.4.
Metrics
The amount of exported metrics for monitoring has been extended and is now
available in a format compatible with Prometheus. You can now easily scrape
on /_admin/metrics
.
See Metrics HTTP API.
The following metrics have been added in ArangoDB 3.7:
Label | Description |
---|---|
arangodb_agency_append_hist |
Agency RAFT follower append histogram |
arangodb_agency_commit_hist |
Agency RAFT commit histogram |
arangodb_agency_compaction_hist |
Agency compaction histogram |
arangodb_agency_local_commit_index |
This agent’s commit index |
arangodb_agency_log_size_bytes |
Agency replicated log size (bytes) |
arangodb_agency_read_no_leader |
Agency read no leader |
arangodb_agency_read_ok |
Agency read ok |
arangodb_agency_supervision_accum_runtime_msec |
Accumulated Supervision Runtime |
arangodb_agency_supervision_accum_runtime_wait_for_replication_msec |
Accumulated Supervision wait for replication time |
arangodb_agency_supervision_failed_server_count |
Counter for FailedServer jobs |
arangodb_agency_supervision_runtime_msec |
Agency Supervision runtime histogram (ms) |
arangodb_agency_supervision_runtime_wait_for_replication_msec |
Agency Supervision wait for replication time (ms) |
arangodb_agency_term |
Agency’s term |
arangodb_agency_write_hist |
Agency write histogram (ms) |
arangodb_agency_write_no_leader |
Agency write no leader |
arangodb_agency_write_ok |
Agency write ok |
arangodb_agencycomm_request_time_msec |
Request time for Agency requests |
arangodb_aql_slow_query |
Number of AQL slow queries |
arangodb_aql_total_query_time_msec |
Total execution time of all queries |
arangodb_client_connection_statistics_io_time_bucket |
Request time needed to answer a request (ms) |
arangodb_client_connection_statistics_io_time_count |
Request time needed to answer a request (ms) |
arangodb_client_connection_statistics_io_time_sum |
Request time needed to answer a request (ms) |
arangodb_client_connection_statistics_queue_time_bucket |
Request time needed to answer a request (ms) |
arangodb_client_connection_statistics_queue_time_count |
Request time needed to answer a request (ms) |
arangodb_client_connection_statistics_queue_time_sum |
Request time needed to answer a request (ms) |
arangodb_client_connection_statistics_request_time_bucket |
Request time needed to answer a request (ms) |
arangodb_client_connection_statistics_request_time_count |
Request time needed to answer a request (ms) |
arangodb_client_connection_statistics_request_time_sum |
Request time needed to answer a request (ms) |
arangodb_client_connection_statistics_total_time_bucket |
Total time needed to answer a request (ms) |
arangodb_client_connection_statistics_total_time_count |
Total time needed to answer a request (ms) |
arangodb_client_connection_statistics_total_time_sum |
Total time needed to answer a request (ms) |
arangodb_dropped_followers_count |
Number of drop-follower events |
arangodb_heartbeat_failures |
Counting failed heartbeat transmissions |
arangodb_heartbeat_send_time_msec |
Time required to send heartbeat (ms) |
arangodb_http_request_statistics_async_requests |
Number of asynchronously executed HTTP requests |
arangodb_http_request_statistics_http_delete_requests |
Number of HTTP DELETE requests |
arangodb_http_request_statistics_http_get_requests |
Number of HTTP GET requests |
arangodb_http_request_statistics_http_head_requests |
Number of HTTP HEAD requests |
arangodb_http_request_statistics_http_options_requests |
Number of HTTP OPTIONS requests |
arangodb_http_request_statistics_http_patch_requests |
Number of HTTP PATCH requests |
arangodb_http_request_statistics_http_post_requests |
Number of HTTP POST requests |
arangodb_http_request_statistics_http_put_requests |
Number of HTTP PUT requests |
arangodb_http_request_statistics_other_http_requests |
Number of other HTTP requests |
arangodb_http_request_statistics_total_requests |
Total number of HTTP requests |
arangodb_load_current_accum_runtime_msec |
Accumulated Current loading time (ms) |
arangodb_load_current_runtime |
Current loading runtimes |
arangodb_load_plan_accum_runtime_msec |
Accumulated runtime of Plan loading (ms) |
arangodb_load_plan_runtime |
Plan loading runtimes |
arangodb_maintenance_action_accum_queue_time_msec |
Accumulated action queue time |
arangodb_maintenance_action_accum_runtime_msec |
Accumulated action runtime |
arangodb_maintenance_action_done_counter |
Counter of action that are done and have been removed from the registry |
arangodb_maintenance_action_duplicate_counter |
Counter of action that have been discarded because of a duplicate |
arangodb_maintenance_action_failure_counter |
Failure counter for the action |
arangodb_maintenance_action_queue_time_msec |
Time spend in the queue before execution |
arangodb_maintenance_action_registered_counter |
Counter of action that have been registered in the action registry |
arangodb_maintenance_action_runtime_msec |
Time spend execution the action |
arangodb_maintenance_agency_sync_accum_runtime_msec |
Accumulated runtime of agency sync phase |
arangodb_maintenance_agency_sync_runtime_msec |
Total time spend on agency sync |
arangodb_maintenance_phase1_accum_runtime_msec |
Accumulated runtime of phase one |
arangodb_maintenance_phase1_runtime_msec |
Maintenance Phase 1 runtime histogram |
arangodb_maintenance_phase2_accum_runtime_msec |
Accumulated runtime of phase two |
arangodb_maintenance_phase2_runtime_msec |
Maintenance Phase 2 runtime histogram |
arangodb_scheduler_awake_threads |
Number of awake worker threads |
arangodb_scheduler_num_worker_threads |
Number of worker threads |
arangodb_scheduler_queue_full_failures |
Number of times the scheduler queue was full and a task/request was rejected |
arangodb_scheduler_queue_length |
Server’s internal queue length |
arangodb_server_statistics_physical_memory |
Physical memory in bytes |
arangodb_server_statistics_server_uptime |
Number of seconds elapsed since server start |
arangodb_shards_leader_count |
Number of leader shards on this machine |
arangodb_shards_not_replicated |
Number of shards not replicated at all |
arangodb_shards_out_of_sync |
Number of leader shards not fully replicated |
arangodb_shards_total_count |
Number of shards on this machine |
arangodb_v8_context_alive |
Number of V8 contexts currently alive |
arangodb_v8_context_busy |
Number of V8 contexts currently busy |
arangodb_v8_context_created |
Number of V8 contexts created |
arangodb_v8_context_destroyed |
Number of V8 contexts destroyed |
arangodb_v8_context_dirty |
Number of V8 contexts currently dirty (waiting for garbage collection) |
arangodb_v8_context_enter_failures |
Number of times a V8 context could not be entered/acquired |
arangodb_v8_context_entered |
Number of times a V8 context was successfully entered |
arangodb_v8_context_exited |
Number of times a V8 context was successfully exited |
arangodb_v8_context_free |
Number of V8 contexts currently free |
arangodb_v8_context_max |
Maximum number of concurrent V8 contexts allowed |
arangodb_v8_context_min |
Minimum number of concurrent V8 contexts allowed |
Client tools
arangodump and arangorestore will now fail when using the --collection
option and none of the specified collections actually exist in the database (on dump)
or in the dump to restore (on restore). In case only some of the specified collections
exist, arangodump / arangorestore will issue warnings about the invalid collections,
but will continue to work for the valid collections.
These change were made to make end users more aware of that the executed commands for dumping or restoring data refer to non-existing collections and that backup or restore operations are potentially incomplete.
MMFiles storage engine
ArangoDB 3.7 does not contain the MMFiles storage engine anymore. In ArangoDB 3.7, the only available storage engine is the RocksDB storage engine, which is the default storage engine in ArangoDB since version 3.4. The MMFiles storage engine had been deprecated since the release of ArangoDB 3.6.
Any deployments that use the MMFiles storage engine will need to be migrated to the RocksDB storage engine using ArangoDB 3.6 (or earlier versions) in order to upgrade to ArangoDB 3.7.
All storage engine selection functionality has also been removed from the ArangoDB package installers. The RocksDB storage engine will be selected automatically for any new deployments created with ArangoDB 3.7.
This change simplifies the installation procedures and internal code paths.
Internal changes
Upgraded bundled RocksDB library version
The bundled version of the RocksDB library has been upgraded from 6.2 to 6.8.
Upgraded bundled OpenLDAP library version
The OpenLDAP version used for the LDAP integration in the ArangoDB Enterprise Edition has been upgraded to 2.4.50. This change has been backported to ArangoDB v3.6.5 as well.
Added libunwind library dependency
The Linux builds of ArangoDB now use the third-party library libunwind to get backtraces and to symbolize stack frames.
Building with libunwind can be turned off at compile time using the
-DUSE_LIBUNWIND
CMake variable.
Removed libcurl library dependency
The compile-time dependency on libcurl was removed. Cluster-internal communication is now performed using fuerte instead of libcurl.
Crash handler
The Linux builds of the arangod executable contain a built-in crash handler The crash handler is supposed to log basic crash information to the ArangoDB logfile in case the arangod process receives one of the signals SIGSEGV, SIGBUS, SIGILL, SIGFPE or SIGABRT. SIGKILL signals, which the operating system can send to a process in case of OOM (out of memory), are not interceptable and thus cannot be intercepted by the crash handler.
In case the crash handler receives one of the mentioned interceptable signals, it will write basic crash information to the logfile and a backtrace of the call site. The backtrace can be provided to the ArangoDB support for further inspection. Note that backtaces are only usable if debug symbols for ArangoDB have been installed as well.
After logging the crash information, the crash handler will execute the default action for the signal it has caught. If core dumps are enabled, the default action for these signals is to generate a core file. If core dumps are not enabled, the crash handler will simply terminate the program with a non-zero exit code.
The crash handler can be disabled at server start by setting the environment
variable ARANGODB_OVERRIDE_CRASH_HANDLER
to an empty string, 0
or off
.
Also see:
Supported compilers
Manually compiling ArangoDB from source will require a C++17-ready compiler.
Older versions of g++ that could be used to compile previous versions of ArangoDB, namely g++7, cannot be used anymore for compiling ArangoDB. g++9.2, g++9.3 and g++10 are known to work, and are the preferred compilers to build ArangoDB under Linux.
Under macOS, the official compiler is clang with a minimal target of macOS 10.14 (Mojave).
Under Windows, use the Visual C++ compiler of Visual Studio 2019 v16.5.0 or later. VS 2017 might still work, but is not officially supported any longer.
Documentation generation
The following features have been added for auto-generating documentation:
- the
--dump-options
command for arangod and the client tools now also emits an attributeos
which indicates on which operating system(s) the respective options are supported. - the
--dump-options
command for arangod now also emits an attributecomponent
which indicates for which node type(s) the respective options are supported (single
server,coordinator
,dbserver
,agent
).