Skip to main content

LQL Overview

LQL is an SQL-like query language for specifying the selection, filtering, and manipulation of data. Queries let you interactively request information from specified curated datasources. Queries have a defined structure for authoring detections.

Find non-compliant resources or suspicious activity by executing LQL queries against data ingested from cloud providers, Kubernetes, CloudTrail activity logs, and the Lacework agent. Additional datasources will become available in the future.

Then you can associate queries with policies, which contain rich reporting metadata.

note

Currently, Lacework has introduced LQL for the configuration of policies and queries for some datasources only.

Example Query

The following comes from Lacework's implementation of a query:

Example
 {
source {
LW_CFG_AWS_EC2_EBS_ENCRYPTION_BY_DEFAULT
}
filter {
RESOURCE_CONFIG:EbsEncryptionByDefault = 'false'
}
return distinct {
ACCOUNT_ALIAS,
ACCOUNT_ID as RESOURCE_KEY,
RESOURCE_REGION,
RESOURCE_TYPE,
SERVICE,
'EbsEncryptionNotEnabledByDefault' as COMPLIANCE_FAILURE_REASON
}
}

The query is delimited by { } and contains three sections:

  • source is the data you are querying (the datasource(s) where the query looks for data). The example's source is the LW_CFG_AWS_EC2_EBS_ENCRYPTION_BY_DEFAULT datasource. LQL queries can specify other datasources and connect multiple datasources, so you can configure custom policies to target the necessary datasource(s).

  • filter specifies the data you are looking for (i.e. the records of interest). The example's filter specifies the records available in LW_CFG_AWS_EC2_EBS_ENCRYPTION_BY_DEFAULT for those whose resource config for EbsEncryptionByDefault = 'false'.

  • return is the information you need to return (i.e. the fields the query exposes). The example's return lists several fields. Because there may be unwanted duplicates among result records when Lacework composes them from just a small number of columns, the query includes the distinct modifier. This behaves like an SQL SELECT DISTINCT. Each returned column in this case is a field that is present in the query, but results can be composed by manipulating strings, dates, JSON, and numbers as well.

The resulting data is depicted as a table. The table's columns are named based on the columns selected. If desired, you could alias them to other names as well.

The previous example query filters and returns results from a single datasource.

Data Types

LQL processes and produces typed data. The data types currently supported or handled are:

  • String, identified in LQL by String
  • Timestamp, identified in LQL by Timestamp
  • Number, identified in LQL by Number
  • JSON, identified in LQL by Json
  • Boolean
  • Null

These identifiers are case-insensitive within LQL. For example, you can write STRING or string.

Though Boolean is a data type processed by LQL, it cannot be returned from a datasource.

In processing, data may be coerced into another type. Coercion means that the system performs the conversion on your behalf, without you needing to specify a conversion. A null identifier can be coerced to be used as another type. In general, a value from JSON can be automatically coerced into a value of one of the other types if it is the correct kind of JSON value. (A JSON number can be automatically coerced into a number, a JSON string value into a string, etc.)

Syntax and Semantics

General Layout

LQL is syntax-oriented. Whitespace has no meaning other than to separate symbols. This means you can use an outline format, as depicted in the first example, or put the entire query on a single line.

Keywords

Keywords in LQL are case-insensitive. For example, FILTER and filter mean the same thing.

Reserved Keywords

The following keywords are reserved. Attempting to use them as names in your own queries will cause an error. If these words appear in JSON paths within queries, you need to surround them in double quotes.

  • EXPR
  • JOIN
  • LIMIT
  • OUTER
  • PARAMINFO
  • PROPERTIES
  • SELECT
  • SQL
  • TYPE
  • VARIANT
  • WHERE

Identifiers

Identifiers in LQL start with a letter, and can contain letters, digits, and the underscore character. Two examples are LW_CFG_AWS_EC2_EBS_ENCRYPTION_BY_DEFAULT and ACCOUNT_ALIAS.

note

Identifiers in LQL are case-sensitive. ACCOUNT_ALIAS and account_alias are two different identifiers, and a column that is named with one of them cannot be accessed by the other.

Comments

-- and // start a single line comment. Text between -- or // and the end of the line is ignored.

/* starts a multiline comment and */ ends a multiline comment. The text between /* and */ is ignored.

Strings

Strings in LQL are delimited by the ' character. Between the ' delimiters, any characters can appear. If you need to encode a ' character within a literal string, you must double it. For example, 'iam.amazonaws.com' encodes the literal string iam.amazonaws.com, and the literal string 'That''s amazing' encodes That's amazing.

Numbers

Literal numbers in LQL are either integral or decimal.

Boolean Values

The symbols true and false represent their Boolean logic counterparts.

Null

The symbol null represents the absence of a known value. You can return it explicitly from various functions such as CASE and COALESCE. As with SQL, you will test for something being null with is nullL rather than = null.

Null occurs both in columns of data and in JSON objects. It is important to note that null for functions and data columns are treated differently than a JSON null (as in {"ERROR_CODE": null}). Therefore, null and "JSON null" are different things. There is no symbol in LQL for a JSON null. For more information, go to the documentation on the is null and is JSON null comparisons and the key_exists function.

Operators and Functions

LQL uses conventional SQL notation for arithmetic, comparisons, and logical operations.

Arithmetic

  • The following operators have the ordinary meanings and precedence for arithmetic: + - * /
  • The % operator represents the modulus operation.
  • The - operator also can be used for unary negation. For example, -x returns the negative for whatever value x has.

LQL follows SQL semantics for arithmetic; if either input to an arithmetic operation is null, the result is null.

Grouping Parentheses

Parentheses (( and )) are used to clearly group expressions and subexpressions, and to establish order of operations. For example, (x + y) * z adds x and y together and then multiplies their sum by z. x + y * z will multiply y and z together, then add their product to x.

Comparison

You can compare numbers, strings, and timestamps.

Comparison Operators

LQL follows SQL semantics for comparisons; if either input to a comparison operation is null, the result is null, not true or false.

The following comparison operators have their ordinary meanings: < <= > >= = <>. The <> operator means "not-equals".

IN, NOT IN

To test if something is equal to some value out of a collection of values, use the IN (...) operator, as in the following form:

IN ( list of literal values ), NOT IN ( list of literal values )

For example, EVENT_SOURCE in ('A1', 'B2') results in true if EVENT_SOURCE has either the string value of A1 or B2. All values in the list of literals must be of the same type, and neither null nor of JSON type.

To test if something is not in a list of possibilities, use NOT IN. For example, STATUS NOT IN (401, 403).

BETWEEN

The BETWEEN operator is shorthand for ( something >= low_bound AND something <= high_bound ).

Use this comparison in the following form:

something BETWEEN low_bound AND high_bound

BETWEEN works with numbers, timestamps, and strings. All arguments must be coercible to numbers or coercible to timestamps.

LIKE

Similar to SQL LIKE, this operator tests if the string on the left matches the pattern on the right in a case-sensitive manner, in the following form:

something LIKE pattern

Within the pattern, the % character is a wildcard meaning "any characters". The _ character is a wildcard that means "any single character". The entire left string is compared with the pattern, so if you want to match a substring and it could be anywhere in the left string, start and end the pattern with %. For example, EVENT_SOURCE LIKE '%aws.com' matches any string that ends with aws.com.

ILIKE

Similar to SQL ILIKE, this operator tests if the string on the left matches the pattern on the right in a case-insensitive manner, in the following form:

something ILIKE pattern

The wildcard semantics for LIKE apply to ILIKE as well.

RLIKE

Similar to SQL RLIKE, this operator tests if the string on the left matches the pattern on the right. The pattern is written using POSIX extended regular expressions, in the following form:

something RLIKE pattern

The entire left string is compared with the pattern, so if you want to match a substring and it could be anywhere in the left string, start and end the pattern with .*. For example, EVENT_SOURCE RLIKE '.*\.amazonaws\.com' matches any string that ends with .amazonaws.com.

LIKE ANY

Checks in a case-sensitive manner whether a string on the left matches any pattern specified on the right, in the following form:

something LIKE ANY (pattern_0, pattern_1, ..., pattern_n)

This is logically equivalent to something LIKE pattern_0 OR something LIKE pattern_1 ... OR something LIKE pattern_n.

ILIKE ANY

Checks in a case-insensitive manner whether a string on the left matches any pattern specified on the right, in the following form:

something ILIKE ANY (pattern_0, pattern_1, ..., pattern_n)

This is logically equivalent to something ILIKE pattern_0 OR something ILIKE pattern_1 ... OR something ILIKE pattern_n.

RLIKE ANY

Checks whether a pattern on the left matches any pattern specified on the right. The patterns may be POSIX-extended regular expressions.

Use this comparison operator in the following form:

something RLIKE ANY (pattern_0, pattern_1, ..., pattern_n)

This is logically equivalent to something RLIKE pattern_0 OR something RLIKE pattern_1 ... OR something RLIKE pattern_n.

Comparing with standard null

The IS NULL and IS NOT NULL comparisons test whether a value from a non-JSON column or function is or is not null. Similar to SQL, you do not compare anything as being equal to null. For example, ERROR_CODE is null tests whether ERROR_CODE has a null.

Use this comparison operator in the following form:

something IS NULL, something IS NOT NULL

Comparing with JSON null

The IS JSON NULL and IS NOT JSON NULL comparisons test whether a JSON value is "JSON null". For example, RESOURCE_CONFIG:SecondaryServer is json null tests whether the JSON object in the RESOURCE_CONFIG column has a null set for its "SecondaryServer" field, that is, matching { "SecondaryServer": null }. If the key "SecondaryServer" is not present in the JSON object, is json null results in null itself, not false.

Use this comparison operator in the following form:

something IS JSON NULL, something IS NOT JSON NULL

Logical

The following operations have their standard logical meanings: AND OR NOT

AND has higher precedence than OR; A AND B OR C returns true if A and B are both true, or if C is true. NOT has higher precedence than AND; NOT A AND B returns true if A is false and B is true.

If either input to a logical operation is null, the result is null.

Conditional

CASE WHEN THEN ELSE END

  • CASE WHEN test THEN value ...
  • CASE value WHEN value THEN value ...

The CASE construct follows SQL syntax. All THEN clauses, and the ELSE clause if provided, must return the same datatype or a datatype that can be coerced to a common data type.

Type Conversion

something :: as_type

The :: operator converts a data value into a specific data type. For example, COUNT_OBJECTS::String converts the numeric count into a string, and '2021-06-03T00:47:33Z'::Timestamp creates a timestamp value for a specific time.

Referring to Data

All datasources that are sources for LQL, and all datasources produced by LQL, have named columns (referred to as fields).

Every datasource is a namespace for its columns. If you write a query that refers to the LW_CFG_AWS_EC2_EBS_ENCRYPTION_BY_DEFAULT datasource, you can refer to its ACCOUNTID field as either ACCOUNT_ID or LW_CFG_AWS_EC2_EBS_ENCRYPTION_BY_DEFAULT.ACCOUNT_ID. If the source datasource is aliased, you can refer to a data column _qualified by its alias (for example, MyAlias.ACCOUNT_ID). For more information, go to Source Clause Details.

Referring to JSON Fields

LQL provides simple JSON field access directly in the syntax. JSON field access works only on datasource fields of type JSON.

  • field : json_path

The : operator references elements within a JSON object. The json_path is a .-separated sequence of JSON keys. For example, given an EVENT field of type JSON: EVENT:requestParameters refers to the requestParameters field of the JSON object held in EVENT, and EVENT:requestParameters.ipPermissions refers to the ipPermissions field of that object.

Because field names in JSON may not be legal identifiers in LQL, they can be enclosed in double-quotes (for example, "). The above could be written as EVENT:"requestParameters"."ipPermissions" as well. An example of using these for illegal JSON identifiers would be EVENT:"RESOURCE_TAG"."aws:cloudformation:stack-name", as the string aws:cloudformation:stack-name is not a legal LQL identifier.

Because JSON has a free-form structure, any element within a field's JSON object structure is itself of type JSON. The requestParameters is a nested object in the example above. To force LQL to treat the result as a particular type, use the :: type-cast operator. For example, EVENT:eval_guid::String retrieves the eval_guid field and casts it to a string.

Working with a JSON array in an expression is not currently supported. You can expand JSON arrays to rows for filtering and projecting purposes in the source clause. For more information, see Source Clause Details.

Data Type Casting Support

FROM_TYPETO_TYPENotes
BooleanJSONSupported
BooleanNumberUnsupported
BooleanStringUnsupported
Use a CASE expression to convert, such as CASE when IS_ARRAY(VOLUME_MAP) then 'True' else 'False' end as ARRY_BOOL.
BooleanTimestampUnsupported
JSONBooleanSupported
JSONNumberSupported
JSONStringSupported
JSONTimestampSupported
NumberBooleanSupported
NumberJSONSupported
NumberStringSupported
NumberTimestampSupported
The provided numbers are interpreted as epoch seconds.
StringBooleanSupported
StringJSONSupported
StringNumberSupported
StringTimestampSupported
The string is converted based on its contents: A string in a standard datetime format such as ISO 8601 is converted according to standard rules. A string containing a string of digits is interpreted as a stringified epoch number - if it contains 13 digits, it is interpreted as epoch milliseconds and if it contains 10 or fewer digits it is interpreted as epoch seconds.
TimestampBooleanUnsupported
TimestampJSONSupported
TimestampNumberUnsupported
TimestampStringSupported
Stringified as a textual date format (not as a stringified epoch milliseconds/seconds number). The specific format may be subject to change over time.

source Clause Details

Description

The source clause specifies the source data for further manipulating. It specifies a table-like set of tuples.

A datasource that is referenced by source, such as LW_CFG_AWS_EC2_EBS_ENCRYPTION_BY_DEFAULT, can be aliased to another name. For example,

source { LW_CFG_AWS_EC2_EBS_ENCRYPTION_BY_DEFAULT }

refers to it without aliasing, you can access its columns by an unqualified name such as ACCOUNT_ID or a qualified name such as LW_CFG_AWS_EC2_EBS_ENCRYPTION_BY_DEFAULT.ACCOUNT_ID. However, if it is aliased, as in the following:

source { LW_CFG_AWS_EC2_EBS_ENCRYPTION_BY_DEFAULT a }

a qualified field reference uses a.ACCOUNT_ID instead.

Joining Multiple Datasources

You can connect the data of two or more datasources in an LQL query using the WITH ... ON construct. For detailed information, go to Connecting Multiple Datasources in Queries.

Expanding JSON Arrays into Rows

When a JSON field contains an array, it can be expanded into records for the source. For example:

Example
 {
source {
LW_CFG_AWS_EC2_SECURITY_GROUPS a,
array_to_rows(a.RESOURCE_CONFIG:IpPermissions) as (ip_permissions),
array_to_rows(a.RESOURCE_CONFIG:IpPermissionsEgress) as (ip_permissions_egress)
}
filter {
RESOURCE_CONFIG:GroupName = 'default'
and (ip_permissions <> '[]'
or ip_permissions_egress <> '[]')
}
return distinct {
ACCOUNT_ALIAS,
ACCOUNT_ID,
ARN as RESOURCE_KEY,
RESOURCE_REGION,
RESOURCE_TYPE,
SERVICE,
'DefaultSecurityGroupAllowsTraffic' as COMPLIANCE_FAILURE_REASON
}
}

In the above example, the array referenced by a.RESOURCE_CONFIG:IpPermissions is expanded into rows, and the values of that array are exposed in a column named ip_permissions. The array referenced by a.RESOURCE_CONFIG:IpPermissionsEgress is expanded into rows, and the values of that array are exposed in a column named ip_permissions_egress. Both the ip_permissions and ip_permissions_egress columns are available in the source for filtering and projecting.

In general, the array_to_rows and array_to_rows_non_empty functions operate on a JSON array derived from a source named earlier in the source clause. Expanding does two things:

  1. It expands the set of rows returned by the source. For each row of the datasource where the JSON array comes from, its array of n elements causes expansion to n rows. (An array of three elements in a row results in three rows, an array of four elements results in four rows, etc.)
  2. It creates a logical column in the source section. The colAlias is the name of this column (this is why colAlias is required).

The colAlias name is in the namespace of all columns produced from the source clause, so if its name matches the name of any other column, it must be given a dtName name. The dtName must be unique within all of the source names/aliases.

The logical column that results from flattening is an JSON entity (object, string, number, boolean, null). It can be used like any other JSON column: returned, used as a source for path expressions, and so on.

For example, consider the following input records:

Table A

IDOBJECT
1{"servicePerms" : [{"service": "web", "access": "root"}, {"service", "smtp", "access" : "mail"}]}
2{"servicePerms" : [{"service": "cron", "access": "root"}]}

The following source clause:

source { A, array_to_rows(A:OBJECT.servicePerms) as (SERVICE_PERMS) }

results in A being expanded to the following set of records:

IDOBJECTSERVICE_PERMS
1{"servicePerms" : [{"service": "web", "access": "root"}, {"service", "smtp"}, {"access" : "mail"}]}{"service": "web", "access": "root"}
1{"servicePerms" : [{"service": "web", "access": "root"}, {"service", "smtp"}, {"access" : "mail"}]}{"service", "smtp", "access" : "mail"}
2{"servicePerms" : [{"service": "cron", "access": "root"}]}{"service": "cron", "access": "root"}

array_to_rows vs array_to_rows_non_empty

Expanding works on JSON arrays derived from a datasource. The expression that produces the array may produce NULL (normal or JSON null) instead.

  • array_to_rows_non_empty: When the expression produces NULL, the row is eliminated from the output. (This is similar to an inner join lacking a matching tuple.)
  • array_to_rows: When the expression produces NULL, a single row is eliminated from the output, with a null value for the value of colALias. (This is similar to a SQL outer join in the presence of a missing match).

When the query includes one or more ARRAY_TO_ROWS source functions, the query might result in many records for each source record. Go to Limitations On Queries in Policy Evaluation for restrictions that apply to such queries when used in LQL policies.

Chained Expansion

An array that is extracted from the result of expansion can be itself expanded. You can use this to expand nested relationships.

Branched Expansion

Expansion can be applied multiple times to a named source. The effect is to produce, per source record, a cartesian product of records from each expansion of that table. For example, if one expansion expands one record to three, and another expansion of that same source record expands that one record to four, the result of applying both of those expansions produces twelve tuples.

filter Clause Details

Description

The filter clause specifies a boolean expression that filters the source records. You can use any valid expression here, as long as the result type of it is boolean.

return Clause Details

Description

The return clause specifies the column(s) to return from the query.

Each column in a result has a name, so that it can be identified later. All column names defined or resulting from a return clause must be unique.

The name for a column is inferred if the column is a simple reference to a column from the source, such as EVENT or CloudTrailRawEvents.EVENT_ID. References that are the result of expressions, such as EVENT:eventTime or coalesce(ERROR_CODE, 0), must be aliased, as with EVENT:eventTime AS eventTime and coalesce(ERROR_CODE, 0) as errorCode. You can alias any column reference to another name.

A field with a JPath reference is named by default with the field name, not with the name of the last step in the JPath. For example, in the following, the name of the column returned is SomeJsonField, not objKey:

{
source {
MySource
}
return {
MySource.SomeJsonField:objKey
}
}

Returning Distinct Records With return distinct

Similar to an SQL query, you may need to reduce a set of records that come from the source to only the unique or distinct ones. For example, you used array_to_rows to create multiple records, and then filtered those records based on some criterion, but now you want to return only the unexpanded set of records that matched the filter. Adding the distinct keyword causes the resulting records to be deduplicated.

Returning All Columns of a Datasource With .*

Similar to an SQL query, all columns of a named datasource can be returned by using the .* wildcard. For example, the following returns all columns from LW_CFG_AWS_EC2_SECURITY_GROUPS:

return {
LW_CFG_AWS_EC2_SECURITY_GROUPS.*
}

Connecting Multiple Datasources in Queries

You can connect the data of two or more datasources in an LQL query using the WITH ... ON construct. For example, the following query uses data of the LW_HE_FILES datasource with its related LW_HE_MACHINES data:

Sample_LQL {
source {
LW_HE_FILES files with LW_HE_MACHINES machines
}
filter {
ENDS_WITH (FILE_NAME, 'exe') AND OS like '%Linux%'
}
return {
PATH,
HOST_NAME
}
}

In the example query, the FILE_NAME and PATH attributes are from LW_HE_FILES, and the OS and HOST_NAME attributes are from LW_HE_MACHINES.

The behavior of WITH is to select records from the source on the left side with any matching records from source(s) on the right. All values for the right side's source's fields will be NULL if there is no match in the right side. (If you are familiar with SQL, this is equivalent to a "left outer join" between them.)

Lacework knows which datasources connect to each other and how they connect. The connections available from each datasource to other datasources are provided with the datasource through the metadata APIs.

Named and Default Paths

Every defined connection from one datasource to another has a name. If a connection is considered to be the default connection, it has the name (default). If two or more connections from one datasource to another exist, each connection has its own name.

To use the name, use the ON keyword. The following example uses the (default) name in the source clause. The connection's name is in single-quotes just like a literal string.

source {
LW_HE_FILES with LW_HE_MACHINES on '(default)'
}

Connecting Three or More Datasources

Your query may be accessing the data of one source with the data of another source, and can also connect the data of that other source with one that is connected to it. Or, your query can access the data of one source with that of two related datasources. You can chain and branch the connections as described in this section.

Chaining Connections

For example, the following source clause connects LW_HA_FILE_CHANGES records with their LW_HE_FILES records, and LW_HE_FILES records with their LW_HE_MACHINES records. This creates a chain of LW_HA_FILE_CHANGES → LW_HE_FILES → LW_HE_MACHINES.

source {
LW_HA_FILE_CHANGES with LW_HE_FILES with LW_HE_MACHINES
}

You can use parentheses to group the connections for clarity. For example:

    LW_HA_FILE_CHANGES with (LW_HE_FILES with LW_HE_MACHINES)

Note that they chain from right to left. The following is not correct. The reason is that the left side of WITH must be a single datasource.

    // Will be rejected
(LW_HA_FILE_CHANGES with LW_HE_FILES) with LW_HE_MACHINES

When you chain multiple steps, using grouping parentheses can help keep the connections clear. For example, both of the following are correct and mean the same thing, but one is easier to follow:

    LW_HA_FILE_CHANGES
with LW_HE_FILES with LW_HE_MACHINES on '(default)'
on '(default)'

LW_HA_FILE_CHANGES
with (LW_HE_FILES with LW_HE_MACHINES on '(default)')
on '(default)'

Forked Connection Paths

The previous examples showed retrieving from three datasources using a chain of connections. Because LW_HA_FILE_CHANGES connects directly to LW_HE_MACHINES as well as LW_HE_FILES, you could use a branching connection as well:

source {
LW_HA_FILE_CHANGES with (
LW_HE_FILES,
LW_HE_MACHINES
)
}

In branching connections, the query brings back all combinations of the target datasources' records per record of the left-side source. You can picture the branches as LW_HA_FILE_CHANGES → (combinations of LW_HE_FILES and LW_HE_MACHINES).

Consider a query for machines data with their related files and process data that filters for Linux machines running processes whose executables are links ending in .exe:

query {
source {
LW_HE_MACHINES machines with (
LW_HE_FILES files,
LW_HE_PROCESSES processes
)
}
filter {
machines.OS like ('%Linux%')
AND files.IS_LINK <> 0
AND ENDS_WITH(processes.EXE_PATH, '.exe')
}
return distinct {
machines.MID,
machines.HOSTNAME,
machines.TAGS
}
}

This filters across all combinations of files for a machine and processes for that same machine, and returns just the machines that meet the condition.

Branching connections also use named connections. You could write the above source clause as:

    source {
LW_HE_MACHINES machines with (
LW_HE_FILES files on '(default)',
LW_HE_PROCESSES processes on '(default)'
)
}

Connections and Data Cardinality

The metadata associated with each connection includes whether the relationship produces at most one record, or many records. In the metadata for each datasource's connections to other sources, these are designated by the following:

CodeMeaning
ONERelationship leads to at most one record
MANYRelationship leads to many records

Limitations On Queries in Policy Evaluation

The first (or root) datasource among your connected datasources corresponds to the type of data the policy is for. If the source specification is LW_HE_MACHINES with LW_HE_PROCESSES, the root datasource is LW_HE_MACHINES, and policies using this query are policies about machines.

LQL policies assume that each result represents a single violating resource. If the datasource connections create many results for each root source's records, misleading results may occur. Therefore, the following validations take place on queries being used in policies:

  • If the query uses only connections with a cardinality of ONE, and does not use any ARRAY_TO_ROWS source functions, any data from the datasources can be returned in the result.
  • If the query uses any connections with a cardinality of MANY, and/or if the query uses any ARRAY_TO_ROWS source functions, the query must obey the following:
    • The query must return distinct results, using return distinct.
    • The query can return result columns derived only from the root datasource and/or datasources whose connection to the root is ONE. The query cannot return results derived from the datasources connected to the root by MANY, nor can it return results derived from ARRAY_TO_ROWS. There are no limitations on data used in the filter clause, regardless.