Skip to main content

Datasources

Datasources in Apache Druid are things that you can query. The most common kind of datasource is a table datasource, and in many contexts the word "datasource" implicitly refers to table datasources. This is especially true during data ingestion, where ingestion is always creating or writing into a table datasource. But at query time, there are many other types of datasources available.

The word "datasource" is generally spelled dataSource (with a capital S) when it appears in API requests and responses.

Datasource type

table

SELECT column1, column2 FROM "druid"."dataSourceName"

The table datasource is the most common type. This is the kind of datasource you get when you perform data ingestion. They are split up into segments, distributed around the cluster, and queried in parallel.

In Druid SQL, table datasources reside in the druid schema. This is the default schema, so table datasources can be referenced as either druid.dataSourceName or simply dataSourceName.

In native queries, table datasources can be referenced using their names as strings (as in the example above), or by using JSON objects of the form:

"dataSource": {
"type": "table",
"name": "dataSourceName"
}

To see a list of all table datasources, use the SQL query SELECT * FROM INFORMATION_SCHEMA.TABLES WHERE TABLE_SCHEMA = 'druid'.

lookup

SELECT k, v FROM lookup.countries

Lookup datasources correspond to Druid's key-value lookup objects. In Druid SQL, they reside in the lookup schema. They are preloaded in memory on all servers, so they can be accessed rapidly. They can be joined onto regular tables using the join operator.

Lookup datasources are key-value oriented and always have exactly two columns: k (the key) and v (the value), and both are always strings.

To see a list of all lookup datasources, use the SQL query SELECT * FROM INFORMATION_SCHEMA.TABLES WHERE TABLE_SCHEMA = 'lookup'.

info

Performance tip: Lookups can be joined with a base table either using an explicit join, or by using the SQL LOOKUP function. However, the join operator must evaluate the condition on each row, whereas the LOOKUP function can defer evaluation until after an aggregation phase. This means that the LOOKUP function is usually faster than joining to a lookup datasource.

Refer to the Query execution page for more details on how queries are executed when you use table datasources.

union

SELECT column1, column2
FROM (
SELECT column1, column2 FROM table1
UNION ALL
SELECT column1, column2 FROM table2
UNION ALL
SELECT column1, column2 FROM table3
)

Unions allow you to treat two or more tables as a single datasource. In SQL, this is done with the UNION ALL operator applied directly to tables, called a "table-level union". In native queries, this is done with a "union" datasource.

With SQL table-level unions the same columns must be selected from each table in the same order, and those columns must either have the same types, or types that can be implicitly cast to each other (such as different numeric types). For this reason, it is more robust to write your queries to select specific columns.

With the native union datasource, the tables being unioned do not need to have identical schemas. If they do not fully match up, then columns that exist in one table but not another will be treated as if they contained all null values in the tables where they do not exist.

In either case, features like expressions, column aliasing, JOIN, GROUP BY, ORDER BY, and so on cannot be used with table unions.

Refer to the Query execution page for more details on how queries are executed when you use union datasources.

Dynamic table append

SELECT column1, column2, column3
FROM TABLE(APPEND('table1','table2','table3'))

Perform dynamic table appends in SQL using TABLE(APPEND(...)). This simplifies SQL syntax to match columns by name from multiple tables. The native query syntax remains the same as for native union datasources. Suppose you have three tables:

  • table1 has column1
  • table2 has column2
  • table3 has column1, column2, column3

You can create a union view of all the tables by using table-level union:

SELECT * from (
SELECT column1,NULL AS column2,NULL AS column3 FROM table1
UNION ALL
SELECT NULL AS column1,column2,NULL AS column3 FROM table2
UNION ALL
SELECT column1,column2,column3 FROM table3
) t

However depending on the size of the table's schema it might be quite complicated to do that; TABLE(APPEND('table1','table2','table3')) represents the same in a more compact form.

info

Only tables defined in the catalog are supported in TABLE(APPEND()) - due to that; common table expressions result in table not found errors for queries like:

WITH cte_table AS (SELECT * from TABLE1) SELECT * FROM TABLE(APPEND('cte_table'))

inline

SELECT * from (VALUES ('United States', 'San Francisco'),
('Canada', 'Calgary')
) t (country, city)

Inline datasources allow you to query a small amount of data that is embedded in the query itself. They are useful when you want to write a query on a small amount of data without loading it first. They are also useful as inputs into a join. Druid also uses them internally to handle subqueries that need to be inlined on the Broker. See the query datasource documentation for more details.

There are two fields in an inline datasource: an array of columnNames and an array of rows. Each row is an array that must be exactly as long as the list of columnNames. The first element in each row corresponds to the first column in columnNames, and so on.

Inline datasources are not available in Druid SQL.

Refer to the Query execution page for more details on how queries are executed when you use inline datasources.

query

-- Uses a subquery to count hits per page, then takes the average.
SELECT
AVG(cnt) AS average_hits_per_page
FROM
(SELECT page, COUNT(*) AS hits FROM site_traffic GROUP BY page)

Query datasources allow you to issue subqueries. In native queries, they can appear anywhere that accepts a dataSource (except underneath a union). In SQL, they can appear in the following places, always surrounded by parentheses:

  • The FROM clause: FROM (<subquery>).
  • As inputs to a JOIN: <table-or-subquery-1> t1 INNER JOIN <table-or-subquery-2> t2 ON t1.<col1> = t2.<col2>.
  • In the WHERE clause: WHERE <column> { IN | NOT IN } (<subquery>). These are translated to joins by the SQL planner.
info

Performance tip: In most cases, subquery results are fully buffered in memory on the Broker and then further processing occurs on the Broker itself. This means that subqueries with large result sets can cause performance bottlenecks or run into memory usage limits on the Broker. See the Query execution page for more details on how subqueries are executed and what limits will apply.

join

-- Joins "sales" with "countries" (using "store" as the join key) to get sales by country.
SELECT
store_to_country.v AS country,
SUM(sales.revenue) AS country_revenue
FROM
sales
INNER JOIN lookup.store_to_country ON sales.store = store_to_country.k
GROUP BY
countries.v

Join datasources allow you to do a SQL-style join of two datasources. Stacking joins on top of each other allows you to join arbitrarily many datasources.

In Druid {DRUIDVERSION}, joins in native queries are implemented with a broadcast hash-join algorithm. This means that all datasources other than the leftmost "base" datasource must fit in memory. In native queries, the join condition must be an equality. In SQL, any join condition is accepted, but only equalities of a certain form (see Joins in SQL) execute efficiently as part of a native join. For other kinds of conditions, planner will try to re-arrange condition such that some of the sub-conditions are evaluated as a filter on top of join and other sub-conditions are left out in the join condition. In worst case scenario, SQL will execute the join condition as a cross join (cartesian product) plus a filter.

This feature is intended mainly to allow joining regular Druid tables with lookup, inline, and query datasources. Refer to the Query execution page for more details on how queries are executed when you use join datasources.

Joins in SQL

SQL joins take the form:

<o1> [ INNER | LEFT [OUTER] ] JOIN <o2> ON <condition>

Any condition is accepted, but only certain kinds of conditions execute efficiently as a native join. The condition must be a single clause like the following, or an AND of clauses involving at least one of the following:

  • Equality between fields of the same type on each side, like t1 JOIN t2 ON t1.x = t2.x.
  • Equality between a function call on one side, and a field on the other side, like t1 JOIN t2 ON LOWER(t1.x) = t2.x.
  • The equality operator may be = (which does not match nulls) or IS NOT DISTINCT FROM (which does match nulls).

In other cases, Druid will either insert a subquery below the join, or will use a cross join (cartesian product) followed by a filter. Joins executed in these ways may run into resource or performance constraints. To determine if your query is using one of these execution paths, run EXPLAIN PLAN FOR <query> and look for the following:

  • query type datasources under the left or right key of your join datasource.
  • join type datasource with condition set to "1" (cartesian product) followed by a filter that encodes the condition you provided.

In these cases, you may be able to improve the performance of your query by rewriting it.

For more information about how Druid translates SQL to native queries, refer to the Druid SQL documentation.

Joins in native queries

Native join datasources have the following properties. All are required.

FieldDescription
leftLeft-hand datasource. Must be of type table, join, lookup, query, or inline. Placing another join as the left datasource allows you to join arbitrarily many datasources.
rightRight-hand datasource. Must be of type lookup, query, or inline. Note that this is more rigid than what Druid SQL requires.
rightPrefixString prefix that will be applied to all columns from the right-hand datasource, to prevent them from colliding with columns from the left-hand datasource. Can be any string, so long as it is nonempty and is not be a prefix of the string __time. Any columns from the left-hand side that start with your rightPrefix will be shadowed. It is up to you to provide a prefix that will not shadow any important columns from the left side.
conditionExpression that must be an equality where one side is an expression of the left-hand side, and the other side is a simple column reference to the right-hand side. Note that this is more rigid than what Druid SQL requires: here, the right-hand reference must be a simple column reference; in SQL it can be an expression.
joinTypeINNER or LEFT.

Join performance

Joins are a feature that can significantly affect performance of your queries. Some performance tips and notes:

  1. Joins are especially useful with lookup datasources, but in most cases, the LOOKUP function performs better than a join. Consider using the LOOKUP function if it is appropriate for your use case.
  2. When using joins in Druid SQL, keep in mind that it can generate subqueries that you did not explicitly include in your queries. Refer to the Druid SQL documentation for more details about when this happens and how to detect it.
  3. One common reason for implicit subquery generation is if the types of the two halves of an equality do not match. For example, since lookup keys are always strings, the condition druid.d JOIN lookup.l ON d.field = l.field will perform best if d.field is a string.
  4. As of Druid {DRUIDVERSION}, the join operator must evaluate the condition for each row. In the future, we expect to implement both early and deferred condition evaluation, which we expect to improve performance considerably for common use cases.
  5. Currently, Druid does not support pushing down predicates (condition and filter) past a Join (i.e. into Join's children). Druid only supports pushing predicates into the join if they originated from above the join. Hence, the location of predicates and filters in your Druid SQL is very important. Also, as a result of this, comma joins should be avoided.

Future work for joins

Joins are an area of active development in Druid. The following features are missing today but may appear in future versions:

  • Reordering of join operations to get the most performant plan.
  • Preloaded dimension tables that are wider than lookups (i.e. supporting more than a single key and single value).
  • RIGHT OUTER and FULL OUTER joins in the native query engine. Currently, they are partially implemented. Queries run but results are not always correct.
  • Performance-related optimizations as mentioned in the previous section.
  • Join conditions on a column containing a multi-value dimension.

unnest

Use the unnest datasource to unnest a column with multiple values in an array. For example, you have a source column that looks like this:

Nested
[a, b]
[c, d]
[e, [f,g]]

When you use the unnest datasource, the unnested column looks like this:

Unnested
a
b
c
d
e
[f, g]

When unnesting data, keep the following in mind:

  • The total number of rows will grow to accommodate the new rows that the unnested data occupy.
  • You can unnest the values in more than one column in a single unnest datasource, but this can lead to a very large number of new rows depending on your dataset.

The unnest datasource uses the following syntax:

  "dataSource": {
"type": "unnest",
"base": {
"type": "table",
"name": "nested_data"
},
"virtualColumn": {
"type": "expression",
"name": "output_column",
"expression": "\"column_reference\""
},
"unnestFilter": "optional_filter"
}
  • dataSource.type: Set this to unnest.
  • dataSource.base: Defines the datasource you want to unnest.
    • dataSource.base.type: The type of datasource you want to unnest, such as a table.
  • dataSource.virtualColumn: Virtual column that references the nested values. The output name of this column is reused as the name of the column that contains unnested values. You can replace the source column with the unnested column by specifying the source column's name or a new column by specifying a different name. Outputting it to a new column can help you verify that you get the results that you expect but isn't required.
  • unnestFilter: A filter only on the output column. You can omit this or set it to null if there are no filters.

To learn more about how to use the unnest datasource, see the unnest tutorial.