Use subquery joins to combine queries

Much of the data stored within New Relic relates to other data: Transaction and TransactionError, PageView and PageAction, Log and infrastructure events, and more. You can perform analysis and calculate correlations between these events using subquery joins.

How to write a subquery join

A subquery is a query that is nested inside another query. With subquery joins, you can combine the result of a subquery with the result of its outer query based on a key, allowing for analysis and enrichment across datasets.

A subquery join requires three components: two datasets and a primary key to link the two together.

FROM Event [INNER|LEFT] JOIN (subquery) ON [key =] key SELECT ...

Subquery joins contain simple rules to the syntax:

• The JOIN clause must always follow immediately after the FROM clause.
• You can prefix the JOIN with the join type. INNER or LEFT is optional, and defaults to INNER when omitted.
• Parentheses containing a subquery must immediately follow the JOIN clause.
• The ON clause must immediately follow the subquery and has two forms (more details below).

You can have multiple JOIN clauses in one query, too. For example, this query uses two JOINs within subqueries:

FROM JavaScriptError
  JOIN (
    FROM PageAction
      JOIN (
        FROM PageView SELECT count(*) FACET session AS pageViewSession, city
        LIMIT MAX
      ) ON session = pageViewSession
    SELECT count(*) FACET city, currentUrl, session AS pageActionSession
  ) ON session = pageActionSession
SELECT count(*) FACET city, currentUrl, session, errorClass

The image below contains two datasets: the average CPU percentage of infrastructure containers (ProcessSample) and the average duration of app transactions by container.

Often, data from different sources is correlated. In this case, you can determine if a container's higher CPU usage is causing slower transactions by using the following subquery join:

FROM Transaction
  JOIN (FROM ProcessSample SELECT average(cpuPercent) AS cpu
  FACET containerId LIMIT MAX) ON containerId
SELECT average(duration)/latest(cpu) FACET containerId, containerName

With this query, you can see the containers that have a higher average transaction duration given their CPU usage, and investigate outliers to see if there is a bug to fix or optimizations to be made.

Subquery join limitations

Subquery joins have the following limitations:

• The joined subquery will continue to have a default LIMIT of 10, with a maximum LIMIT of 5,000. Note that the outer query's LIMIT does not affect the inner query.
• The use of TIMESERIES in the joined subquery is not supported. If your outer query uses TIMESERIES, keep in mind that the joined subquery will provide a single result for the full query timespan.
• The use of COMPARE WITH in the joined subquery is not supported. If your outer query uses COMPARE WITH, keep in mind that the joined subquery will provide a single result based on the queries base timespan, and will not provide a separate value for the outer queries compare with timespan.
• Like all subqueries, joined subqueries cannot be used in alert conditions.
• While SELECT * is supported in the parent query, it is not supported in the joined subquery.
• The cardinality of the join is limited to 1:100, meaning a single join key cannot map to more than one hundred rows in the subquery result.
• The ON clause only supports equality conditions.
• The JOIN key cannot be a complex attribute, like a metric value.
• We do no coercion of attribute types in the JOIN condition. The left side of the JOIN's ON condition needs to be the same type as the right side of the ON condition.
• Metric wildcards are not supported in the JOIN's ON condition
• The subquery cannot be a metric row-wise query.
• The right-hand side of the JOIN's ON condition must be an identifier projected by the query. It cannot use a function or mathematical operation.
• The joined subquery cannot project a uniques() result.

Subquery join examples

Here are some example subquery joins:

Add missing data

In this example, APM data is in two places: PageView provides information on a page visited by an end user, while PageAction provides the actions taken on the page. These two events both have session IDs that identify the session of an end user, but some data is contained in one event and not the other.

Here, city is an attribute in PageView but not PageAction, whereas currentUrl is an attribute in PageAction but not PageView.

With subquery joins, you can enrich the PageAction dataset with the missing data. By combining the data using the session ID, you can see not only the URL with the most clicks, but also where in the world these actions came from:

FROM PageAction
  JOIN (FROM PageView SELECT count(*) FACET session, city
  LIMIT MAX) ON session
SELECT count(*) FACET city, currentUrl

Copy

Tip: Include LIMIT MAX in the inner subquery to maximize the number of results joined to the outer query. The current max result limit is 5,000 rows.

Calculations across datasets

In this example, two log datasets will be parsed to match one another, then calculations will be run to find the log error percentage.

Logs can often have data buried within the log message. Using enhanced string parsing functions like aparse() you can extract key values. In this case, the primary key application ID is within the log message of both Log and Log_Error:

In order to find the error percentage by application, a LEFT JOIN is required. This is because not all of the applications have errors, and an INNER JOIN would exclude these apps.

WITH numeric(aparse(message, '%applicationId: * %')) AS application_Id
FROM Log
  LEFT JOIN (FROM Log_Error SELECT count(*) AS errCnt
     FACET numeric(aparse(message, '%appId: * %')) AS app_Id
     SINCE 1 day ago LIMIT MAX) ON application_Id=app_Id
SELECT (latest(errorCount) OR 0)/count(*) AS errorPercentage
FACET app_name, application_Id SINCE 1 day ago

Copy

Even though there was data stored in two different log partitions and the application ID was hidden in the log messages, subquery joins allowed parsing out the needed data and calculating the error percentage. Using latest() was also helpful to make sure the aggregations of the inner query were calculated correctly with the outer query.