Profiling and Hand-Optimizing AQL queries
To give you more insight into your query ArangoDB allows to execute your query with special instrumentation code enabled. This will then print a query plan with detailed execution statistics.
To use this in an interactive fashion on the shell you can use
db._profileQuery(..)
in arangosh. Alternatively, there is a button
Profile in the Query tab of the web interface.
The printed execution plan then contains three additional columns:
- Call: The number of times this query stage was executed
- Items: The number of temporary result rows (outputs) at this stage
- Filtered: The number of rows filtered away by this stage
- Runtime: The total time spent in this stage
Below the execution plan there are additional sections for the overall runtime statistics and the query profile.
Example: Simple AQL query
Assuming we got a collection named acollection
and insert 10000 documents
via for (let i=0; i < 10000;i++) db.acollection.insert({value:i})
.
Then a simple query filtering for value < 10
will return 10 results:
An AQL query is essentially executed in a pipeline that chains together different functional execution blocks. Each block gets the input rows from the parent above it, does some processing and then outputs a certain number of output rows.
Without any detailed insight into the query execution it is impossible to tell
how many results each pipeline-block had to work on and how long this took.
By executing the query with the query profiler (db._profileQuery()
or via
the Profile button in the web interface) you can check exactly how much work
each stage had to do.
Without any indexes this query should have to perform the following operations:
- Perform a full collection scan via a EnumerateCollectionNode and outputting
a row containing the document in
doc
. - Calculate the boolean expression
LET #1 = doc.value < 10
from all inputs via a CalculationNode - Filter out all input rows where
#1
is false via the FilterNode - Put the
doc
variable of the remaining rows into the result set via the ResultNode
The EnumerateCollectionNode processed and returned all 10k rows (documents), as did the CalculationNode. Because the AQL execution engine also uses an internal batch size of 1000 these blocks were also called 100 times each. The FilterNode as well as the ReturnNode however only ever returned 10 rows and only had to be called once, because the result size fits within a single batch.
Let us add a persistent index on value
to speed up the query:
db.acollection.ensureIndex({type:"persistent", fields:["value"]});
This results in replacing the collection scan and filter block with an
IndexNode
. The execution pipeline of the AQL query has become much shorter.
Also the number of rows processed by each pipeline block is only 10, because
we no longer need to look at all documents.
Example: AQL with Subquery
Let us consider a query containing a subquery:
The resulting query profile contains a SubqueryNode which has the runtime of all its children combined.
Actually, we cheated a little. The optimizer would have completely removed the
subquery if it had not been deactivated (rules:["-all"]
). The optimized
version would take longer in the “optimizing plan” stage, but should perform
better with a lot of results.
Example: AQL with Aggregation
Let us try a more advanced query, using a COLLECT statement. Assume we have a user collection with each document having a city, a username and an age attribute.
The following query gets us all age groups in buckets (0-9, 10-19, 20-29, …):
Without any indexes this query should have to perform the following operations:
- Perform a full collection scan via a EnumerateCollectionNode and outputting
a row containing the document in
doc
. - Compute the expression
LET #1 = FLOOR(u.age / 10) * 10
for all inputs via a CalculationNode - Perform the aggregations via the CollectNode
- Sort the resulting aggregated rows via a SortNode
- Build a result value via another CalculationNode
- Put the result variable into the result set via the ResultNode
Like within the example above, you can see that after the CalculationNode stage, from the originally 20 rows only a handful remained.
Typical AQL Performance Mistakes
With the new query profiler you should be able to spot typical performance mistakes that we see quite often:
- Not employing indexes to speed up queries with common filter expressions
- Not using shard keys in filter statements, when it is known (only a cluster problem)
- Using subqueries to calculate an intermediary result, but only using a few results
Bad example:
LET vertices = (
FOR v IN 1..2 ANY @startVertex GRAPH 'my_graph'
// <-- add a LIMIT 1 here
RETURN v
)
FOR doc IN collection
FILTER doc.value == vertices[0].value
RETURN doc
Adding a LIMIT 1
into the subquery should result in better performance,
because the traversal can be stopped after the first result instead of
computing all paths.
Another mistake is to start a graph traversal from the wrong side (if both ends are known).
Assume we have two vertex collections users and products as well as an
edge collection purchased. The graph model looks like this:
(users) <--[purchased]--> (products)
, i.e. every user is connected with an
edge in purchased to zero or more products.
If we want to know all users that have purchased the product playstation
as well as products of type
legwarmer we could use this query:
FOR prod IN products
FILTER prod.type == 'legwarmer'
FOR v,e,p IN 2..2 OUTBOUND prod purchased
FILTER v._key == 'playstation' // <-- last vertex of the path
RETURN p.vertices[1] // <-- the user
This query first finds all legwarmer products and then performs a traversal for each of them. But we could also inverse the traversal by starting of with the known playstation product. This way we only need a single traversal to achieve the same result:
FOR v,e,p IN 2..2 OUTBOUND 'product/playstation' purchased
FILTER v.type == 'legwarmer' // <-- last vertex of the path
RETURN p.vertices[1] // <-- the user