The AQL query optimizer

AQL queries are sent through an optimizer before execution. The task of the optimizer is to create an initial execution plan for the query, look for optimization opportunities and apply them. As a result, the optimizer might produce multiple execution plans for a single query. It will then calculate the costs for all plans and pick the plan with the lowest total cost. This resulting plan is considered to be the optimal plan, which is then executed.

The optimizer is designed to only perform optimizations if they are safe, in the meaning that an optimization should not modify the result of a query. A notable exception to this is that the optimizer is allowed to change the order of results for queries that do not explicitly specify how results should be sorted.

Execution plans

The explain command can be used to query the optimal executed plan or even all plans the optimizer has generated. Additionally, explain can reveal some more information about the optimizer's view of the query.

Inspecting plans using the explain helper

The explain method of ArangoStatement as shown in the next chapters creates very verbose output. You can work on the output programmatically, or use this handsome tool that we created to generate a more human readable representation.

You may use it like this: (we disable syntax highlighting here)

arangosh> db._create("test"); 
arangosh> for (i = 0; i < 100; ++i) { db.test.save({ value: i }); }
arangosh> db.test.ensureIndex({ type: "skiplist", fields: [ "value" ] });
arangosh> var explain = require("@arangodb/aql/explainer").explain;
arangosh> explain("FOR i IN test FILTER i.value > 97 SORT i.value RETURN i.value", {colors:false});
show execution results

Execution plans in detail

Let's have a look at the raw json output of the same execution plan using the explain method of ArangoStatement:

arangosh> stmt = db._createStatement("FOR i IN test FILTER i.value > 97 SORT i.value RETURN i.value");
arangosh> stmt.explain();
show execution results

As you can see, the result details are very verbose so we will not show them in full in the next sections. Instead, let's take a closer look at the results step by step.

Execution nodes

In general, an execution plan can be considered to be a pipeline of processing steps. Each processing step is carried out by a so-called execution node

The nodes attribute of the explain result contains these execution nodes in the execution plan. The output is still very verbose, so here's a shorted form of it:

arangosh> stmt.explain().plan.nodes.map(function (node) { return node.type; });
show execution results

Note that the list of nodes might slightly change in future versions of ArangoDB if new execution node types get added or the optimizer create somewhat more optimized plans).

When a plan is executed, the query execution engine will start with the node at the bottom of the list (i.e. the ReturnNode).

The ReturnNode's purpose is to return data to the caller. It does not produce data itself, so it will ask the node above itself, this is the CalculationNode in our example. CalculationNodes are responsible for evaluating arbitrary expressions. In our example query, the CalculationNode will evaluate the value of i.value, which is needed by the ReturnNode. The calculation will be applied for all data the CalculationNode gets from the node above it, in our example the FilterNode.

FilterNodes will only let certain documents pass. Normally, filters are based on the evaluation of an expression. The filters expression result (i.value > 97) is calculated in the CalculationNode above the FilterNode.

Finally, all of this needs to be done for documents of collection test. This is where the IndexNode enters the game. It will use an index (thus its name) to find certain documents in the collection and ship it down the pipeline in the order required by SORT i.value. The IndexNode itself has a SingletonNode as its input. The sole purpose of a SingletonNode node is to provide a single empty document as input for other processing steps. It is always the end of the pipeline.

Here's a summary:

  • SingletonNode: produces an empty document as input for other processing steps.
  • IndexNode: iterates over the index on attribute value in collection test in the order required by SORT i.value.
  • CalculationNode: evaluates the result of the calculation i.value > 97 to true or false
  • FilterNode: only lets documents pass where above calculation returned true
  • CalculationNode: calculates return value i.value
  • ReturnNode: returns data to the caller

Optimizer rules

Note that in the example, the optimizer has optimized the SORT statement away. It can do it safely because there is a sorted skiplist index on i.value, which it has picked in the IndexNode. As the index values are iterated over in sorted order anyway, the extra SortNode would have been redundant and was removed.

Additionally, the optimizer has done more work to generate an execution plan that avoids as much expensive operations as possible. Here is the list of optimizer rules that were applied to the plan:

arangosh> stmt.explain().plan.rules;
show execution results

Here is the meaning of these rules in context of this query:

  • move-calculations-up: moves a CalculationNode as far up in the processing pipeline as possible
  • move-filters-up: moves a FilterNode as far up in the processing pipeline as possible
  • remove-redundant-calculations: replaces references to variables with references to other variables that contain the exact same result. In the example query, i.value is calculated multiple times, but each calculation inside a loop iteration would produce the same value. Therefore, the expression result is shared by several nodes.
  • remove-unnecessary-calculations: removes CalculationNodes whose result values are not used in the query. In the example this happens due to the remove-redundant-calculations rule having made some calculations unnecessary.
  • use-index: use an index to iterate over a collection instead of performing a full collection scan. In the example case this makes sense, as the index can be used for filtering and sorting.
  • use-index-for-sort: removes a SORT operation if it is already satisfied by traversing over a sorted index

Note that some rules may appear multiple times in the list, with number suffixes. This is due to the same rule being applied multiple times, at different positions in the optimizer pipeline.

Collections used in a query

The list of collections used in a plan (and query) is contained in the collections attribute of a plan:

arangosh> stmt.explain().plan.collections
show execution results

The name attribute contains the name of the collection, and type is the access type, which can be either read or write.

Variables used in a query

The optimizer will also return a list of variables used in a plan (and query). This list will contain auxiliary variables created by the optimizer itself. This list can be ignored by end users in most cases.

Cost of a query

For each plan the optimizer generates, it will calculate the total cost. The plan with the lowest total cost is considered to be the optimal plan. Costs are estimates only, as the actual execution costs are unknown to the optimizer. Costs are calculated based on heuristics that are hard-coded into execution nodes. Cost values do not have any unit.

Retrieving all execution plans

To retrieve not just the optimal plan but a list of all plans the optimizer has generated, set the option allPlans to true:

This will return a list of all plans in the plans attribute instead of in the plan attribute:

arangosh> stmt.explain({ allPlans: true });
show execution results

Retrieving the plan as it was generated by the parser / lexer

To retrieve the plan which closely matches your query, you may turn off most optimization rules (i.e. cluster rules cannot be disabled if you're running the explain on a cluster coordinator) set the option rules to -all:

This will return an unoptimized plan in the plan:

arangosh> stmt.explain({ optimizer: { rules: [ "-all" ] } });
show execution results

Note that some optimizations are already done at parse time (i.e. evaluate simple constant calculation as 1 + 1)

Turning specific optimizer rules off

Optimizer rules can also be turned on or off individually, using the rules attribute. This can be used to enable or disable one or multiple rules. Rules that shall be enabled need to be prefixed with a +, rules to be disabled should be prefixed with a -. The pseudo-rule all matches all rules.

Rules specified in rules are evaluated from left to right, so the following works to turn on just the one specific rule:

arangosh> stmt.explain({ optimizer: { rules: [ "-all", "+use-index-range" ] } });
show execution results

By default, all rules are turned on. To turn off just a few specific rules, use something like this:

arangosh> stmt.explain({ optimizer: { rules: [ "-use-index-range", "-use-index-for-sort" ] } });
show execution results

The maximum number of plans created by the optimizer can also be limited using the maxNumberOfPlans attribute:

arangosh> stmt.explain({ maxNumberOfPlans: 1 });
show execution results

Optimizer statistics

The optimizer will return statistics as a part of an explain result.

The following attributes will be returned in the stats attribute of an explain result:

  • plansCreated: total number of plans created by the optimizer
  • rulesExecuted: number of rules executed (note: an executed rule does not indicate a plan was actually modified by a rule)
  • rulesSkipped: number of rules skipped by the optimizer

Warnings

For some queries, the optimizer may produce warnings. These will be returned in the warnings attribute of the explain result:

arangosh> var stmt = db._createStatement("FOR i IN 1..10 RETURN 1 / 0")
arangosh> stmt.explain().warnings;
show execution results

There is an upper bound on the number of warning a query may produce. If that bound is reached, no further warnings will be returned.

Things to consider for optimizing queries

While the optimizer can fix some things in queries, its not allowed to take some assumptions, that you, the user, knowing what queries are intended to do can take. It may pull calculations to the front of the execution, but it may not cross certain borders.

So in certain cases you may want to move calculations in your query, so they're cheaper. Even more expensive is if you have calculacions that are executed in javascript:

arangosh> db._explain('FOR x IN 1..10 LET then=DATE_NOW() FOR y IN 1..10 LET now=DATE_NOW() LET nowstr=CONCAT(now, x, y, then) RETURN nowstr', {}, {colors: false})
arangosh> db._explain('LET now=DATE_NOW() FOR x IN 1..10 FOR y IN 1..10 LET nowstr=CONCAT(now, x, y, now) RETURN nowstr', {}, {colors: false})
show execution results

You can see, that the optimizer found 1..10 is specified twice, but can be done first one time.

While you may see time passing by during the execution of the query and its calls to DATE_NOW() this may not be the desired thing in first place. The queries V8 Expressions will however also use significant resources, since its executed 10 x 10 times => 100 times. Now if we don't care for the time ticking by during the query execution, we may fetch the time once at the startup of the query, which will then only give us one V8 expression at the very start of the query.

Next to bringing better performance, this also obeys the DRY principle.

Optimization in a cluster

When you're running AQL in the cluster, the parsing of the query is done on the coordinator. The coordinator then chops the query into snipets, which are to remain on the coordinator, and others that are to be distributed over the network to the shards. The cutting sites are interconnected via Scatter-, Gather- and RemoteNodes.

These nodes mark the network borders of the snippets. The optimizer strives to reduce the amount of data transfered via these network interfaces by pushing FILTERs out to the shards, as it is vital to the query performance to reduce that data amount to transfer over the network links.

Snippets marked with DBS are executed on the shards, COOR ones are excuted on the coordinator.

As usual, the optimizer can only take certain assumptions for granted when doing so, i.e. user-defined functions have to be executed on the coordinator. If in doubt, you should modify your query to reduce the number interconnections between your snippets.

When optimizing your query you may want to look at simpler parts of it first.

List of execution nodes

The following execution node types will appear in the output of explain:

  • SingletonNode: the purpose of a SingletonNode is to produce an empty document that is used as input for other processing steps. Each execution plan will contain exactly one SingletonNode as its top node.
  • EnumerateCollectionNode: enumeration over documents of a collection (given in its collection attribute) without using an index.
  • IndexNode: enumeration over one or many indexes (given in its indexes attribute) of a collection. The index ranges are specified in the condition attribute of the node.
  • EnumerateListNode: enumeration over a list of (non-collection) values.
  • FilterNode: only lets values pass that satisfy a filter condition. Will appear once per FILTER statement.
  • LimitNode: limits the number of results passed to other processing steps. Will appear once per LIMIT statement.
  • CalculationNode: evaluates an expression. The expression result may be used by other nodes, e.g. FilterNode, EnumerateListNode, SortNode etc.
  • SubqueryNode: executes a subquery.
  • SortNode: performs a sort of its input values.
  • AggregateNode: aggregates its input and produces new output variables. This will appear once per COLLECT statement.
  • ReturnNode: returns data to the caller. Will appear in each read-only query at least once. Subqueries will also contain ReturnNodes.
  • InsertNode: inserts documents into a collection (given in its collection attribute). Will appear exactly once in a query that contains an INSERT statement.
  • RemoveNode: removes documents from a collection (given in its collection attribute). Will appear exactly once in a query that contains a REMOVE statement.
  • ReplaceNode: replaces documents in a collection (given in its collection attribute). Will appear exactly once in a query that contains a REPLACE statement.
  • UpdateNode: updates documents in a collection (given in its collection attribute). Will appear exactly once in a query that contains an UPDATE statement.
  • UpsertNode: upserts documents in a collection (given in its collection attribute). Will appear exactly once in a query that contains an UPSERT statement.
  • NoResultsNode: will be inserted if FILTER statements turn out to be never satisfiable. The NoResultsNode will pass an empty result set into the processing pipeline.

For queries in the cluster, the following nodes may appear in execution plans:

  • ScatterNode: used on a coordinator to fan-out data to one or multiple shards.
  • GatherNode: used on a coordinator to aggregate results from one or many shards into a combined stream of results.
  • DistributeNode: used on a coordinator to fan-out data to one or multiple shards, taking into account a collection's shard key.
  • RemoteNode: a RemoteNode will perform communication with another ArangoDB instances in the cluster. For example, the cluster coordinator will need to communicate with other servers to fetch the actual data from the shards. It will do so via RemoteNodes. The data servers themselves might again pull further data from the coordinator, and thus might also employ RemoteNodes. So, all of the above cluster relevant nodes will be accompanied by a RemoteNode.

List of optimizer rules

The following optimizer rules may appear in the rules attribute of a plan:

  • move-calculations-up: will appear if a CalculationNode was moved up in a plan. The intention of this rule is to move calculations up in the processing pipeline as far as possible (ideally out of enumerations) so they are not executed in loops if not required. It is also quite common that this rule enables further optimizations to kick in.
  • move-filters-up: will appear if a FilterNode was moved up in a plan. The intention of this rule is to move filters up in the processing pipeline as far as possible (ideally out of inner loops) so they filter results as early as possible.
  • sort-in-values: will appear when the values used as right-hand side of an IN operator will be pre-sorted using an extra function call. Pre-sorting the comparison array allows using a binary search in-list lookup with a logarithmic complexity instead of the default linear complexity in-list lookup.
  • remove-unnecessary-filters: will appear if a FilterNode was removed or replaced. FilterNodes whose filter condition will always evaluate to true will be removed from the plan, whereas FilterNode that will never let any results pass will be replaced with a NoResultsNode.
  • remove-redundant-calculations: will appear if redundant calculations (expressions with the exact same result) were found in the query. The optimizer rule will then replace references to the redundant expressions with a single reference, allowing other optimizer rules to remove the then-unneeded CalculationNodes.
  • remove-unnecessary-calculations: will appear if CalculationNodes were removed from the query. The rule will removed all calculations whose result is not referenced in the query (note that this may be a consequence of applying other optimizations).
  • remove-redundant-sorts: will appear if multiple SORT statements can be merged into fewer sorts.
  • interchange-adjacent-enumerations: will appear if a query contains multiple FOR statements whose order were permuted. Permutation of FOR statements is performed because it may enable further optimizations by other rules.
  • remove-sort-rand: will appear when a SORT RAND() expression is removed by moving the random iteration into an EnumerateCollectionNode.
  • remove-collect-variables: will appear if an INTO clause was removed from a COLLECT statement because the result of INTO is not used. May also appear if a result of a COLLECT statement's AGGREGATE variables is not used.
  • propagate-constant-attributes: will appear when a constant value was inserted into a filter condition, replacing a dynamic attribute value.
  • replace-or-with-in: will appear if multiple OR-combined equality conditions on the same variable or attribute were replaced with an IN condition.
  • remove-redundant-or: will appear if multiple OR conditions for the same variable or attribute were combined into a single condition.
  • use-indexes: will appear when an index is used to iterate over a collection. As a consequence, an EnumerateCollectionNode was replaced with an IndexNode in the plan.
  • remove-filters-covered-by-index: will appear if a FilterNode was removed or replaced because the filter condition is already covered by an IndexNode.
  • use-index-for-sort: will appear if an index can be used to avoid a SORT operation. If the rule was applied, a SortNode was removed from the plan.
  • move-calculations-down: will appear if a CalculationNode was moved down in a plan. The intention of this rule is to move calculations down in the processing pipeline as far as possible (below FILTER, LIMIT and SUBQUERY nodes) so they are executed as late as possible and not before their results are required.
  • patch-update-statements: will appear if an UpdateNode was patched to not buffer its input completely, but to process it in smaller batches. The rule will fire for an UPDATE query that is fed by a full collection scan, and that does not use any other indexes and subqueries.
  • optimize-traversals: will appear if either the edge or path output variable in an AQL traversal was optimized away, or if a FILTER condition from the query was moved in the TraversalNode for early pruning of results.
  • inline-subqueries: will appear when a subquery was pulled out in its surrounding scope, e.g. FOR x IN (FOR y IN collection FILTER y.value >= 5 RETURN y.test) RETURN x.a would become FOR tmp IN collection FILTER tmp.value >= 5 LET x = tmp.test RETURN x.a

The following optimizer rules may appear in the rules attribute of cluster plans:

  • distribute-in-cluster: will appear when query parts get distributed in a cluster. This is not an optimization rule, and it cannot be turned off.
  • scatter-in-cluster: will appear when scatter, gather, and remote nodes are inserted into a distributed query. This is not an optimization rule, and it cannot be turned off.
  • distribute-filtercalc-to-cluster: will appear when filters are moved up in a distributed execution plan. Filters are moved as far up in the plan as possible to make result sets as small as possible as early as possible.
  • distribute-sort-to-cluster: will appear if sorts are moved up in a distributed query. Sorts are moved as far up in the plan as possible to make result sets as small as possible as early as possible.
  • remove-unnecessary-remote-scatter: will appear if a RemoteNode is followed by a ScatterNode, and the ScatterNode is only followed by calculations or the SingletonNode. In this case, there is no need to distribute the calculation, and it will be handled centrally.
  • undistribute-remove-after-enum-coll: will appear if a RemoveNode can be pushed into the same query part that enumerates over the documents of a collection. This saves inter-cluster roundtrips between the EnumerateCollectionNode and the RemoveNode.

Note that some rules may appear multiple times in the list, with number suffixes. This is due to the same rule being applied multiple times, at different positions in the optimizer pipeline.