Joins

Kinetica supports the SQL concept of joining data sets, via join views, connecting related records between two or more tables.

A join view name must adhere to the standard naming criteria. Each join view exists within a schema and follows the standard name resolution rules for tables.

The supported join types are:

  • INNER - matching rows between two tables
  • LEFT - matching rows between two tables, and rows in the left-hand table with no matching rows in the right-hand table
  • RIGHT - matching rows between two tables, and rows in the right-hand table with no matching rows in the left-hand table
  • FULL OUTER matching rows between two tables, and rows in both tables with no matching rows in the other
  • Cross - all rows in one table matched against all rows in the other; this is the join that is executed when no join conditions are specified--there is no keyword for specifying this type

Join Execution Types

Two types of joins are available, depending on need:

Native Joins

A native join isolates the join operation and creates a native join view, whose purpose is to connect the tables involved in the join. The native join consists of three operations:

  1. Joining of records between tables
  2. Filtering of joined records
  3. Selection of columns to include in the resulting join view

The /create/jointable endpoint is used to request a native join.

Once the join view has been created, adding or deleting data from the joined view base tables does not change the data queried from the join view, though updates will be reflected. The join view may be used as a source table in most native API calls.

Joins can be performed on any number and combination of tables and views.

These limitations and others are discussed in further detail in the Limitations & Cautions section.

Any column expressions used in a join will be evaluated on querying the join view. More detail about the impact of using column expressions can be found under Memory Implications.

SQL Joins

A SQL join is performed automatically when the database receives any SQL request containing a JOIN clause. It will generate a native join view, which will either be used as the base table for subsequent operations contained in the original query or as the source of records in the final result set. In either case, the join view is transient and will be removed, by default, after the query that caused it to be generated is complete.

SQL joins can be executed through either the /execute/sql endpoint or ODBC/JDBC.

Join Distribution Types

There are two types of join distribution:

  1. Node-Local Joins - each node executes its portion of the overall join operation; faster, but join distribution criteria must be met
  2. Distributed Joins - data is redistributed across the nodes so that node-local joins can be performed; slower and requiring more memory, but with no distribution restrictions

Node-Local Joins

Tables being joined together must either be replicated, or be sharded on the columns being used to join the tables to avoid extensive interprocessor communication. While a join may include any number of replicated tables and sharded tables, all sharded tables must be joined by equating their respective shard key columns.

For instance, given a table A and a table B, both sharded on their respective id columns, the only node-local join possible is between A.id and B.id. No other columns from either table can be used to connect the two tables.

When a join view is created, it will contain the results of all join and filter clauses specified. It is recommended that all tables specified in the join view creation be joined within the specified expression. Any replicated tables not joined in the specified expression will be merged into the join view as cross-products. The /filter endpoint can be used subsequently to connect the unmerged tables.

Distributed Joins

Distributed joins allow the connection of data sets, regardless of their distribution scheme. These joins are less performant due to interprocessor communication overhead and require more memory & disk space to process. This scheme makes the following possible:

  • Inner join between sharded tables on columns other than their shard keys
  • Left outer join from a replicated table to a sharded table
  • Full outer join between two tables that are not both replicated
  • Non-equi join (A.id > A.id) between two tables that are not both replicated

Important

Distributed joins are only available via the /execute/sql endpoint or SQL using ODBC/JDBC. The /create/jointable endpoint only processes local joins.

Distributed joins will reshard or replicate data, depending on four criteria:

  1. Whether the join condition is an equality-based or inequality-based
  2. Whether the left-hand table is:
    • Replicated
    • Being joined on its shard key
    • Neither replicated nor being joined on its shard key
  3. Whether the right-hand table is:
    • Replicated
    • Being joined on its shard key
    • Neither replicated nor being joined on its shard key
  4. The type of join:
    • INNER
    • LEFT (RIGHT)
    • FULL

Equality-Based Join Rules

The following chart shows the resulting distribution action for equality-based joins, given the distribution schemes of the left-hand & right-hand tables. In parentheses are the join types that will result in a distribution operation.

Note

Cells marked with (All) indicate that a distribution operation will occur for all join types, INNER, LEFT, RIGHT, & FULL. Also, cells marked LEFT can be used to infer the distribution for RIGHT joins, by simply reversing the left & right tables in the query.

Join Criteria CompositionRight Table
Shard KeyNon-Shard KeyReplicated
Left TableShard Key<Node-local join>Reshard Right (All)Reshard Right (FULL)
Non-Shard KeyReshard Left (All)Reshard Both (All)Replicate Left (All)
ReplicatedReshard Left (LEFT,FULL)Replicate Right (LEFT,FULL)<Node-local join>

Some example scenarios:

  • If a sharded table A is joined on its shard key x to a replicated table B on its column y, and the join type is FULL OUTER, the right-side table B will be sharded on its column y to match the sharding of the left-side table.
  • If a sharded table A is joined on columns w & x, which are not its shard key, to a sharded table B on its shard key columns y & z, and the join is of any type, the left-side table A will be resharded on its columns w & x to match the sharding of the right-side table.
  • If a replicated table is joined to a replicated table, and the join is of any type, it will be a node-local join and not require redistribution of any data.

Inequality-Based Join Rules

The following chart shows the resulting distribution action for inequality-based joins, given the distribution schemes of the left-hand & right-hand tables. In parentheses are the join types that will result in a distribution operation.

Note

Cells marked LEFT can be used to infer the distribution for RIGHT joins, by simply reversing the left & right tables in the query.

Join Criteria CompositionRight Table
Shard KeyNon-Shard KeyReplicated
Left TableShard Key

Replicate Right (INNER,LEFT)

Replicate Both (FULL)

Replicate Right (INNER,LEFT)

Replicate Both (FULL)

Replicate Left (FULL)
Non-Shard Key

Replicate Right (INNER,LEFT)

Replicate Both (FULL)

Replicate Right (INNER,LEFT)

Replicate Both (FULL)

Replicate Left (FULL)
ReplicatedReplicate Right (LEFT,FULL)Replicate Right (LEFT,FULL)<Node-local join>

Some example scenarios:

  • If a sharded table A is joined on its shard key x to a replicated table B on its column y, and the join type is FULL OUTER, the left-side table A will be replicated.
  • If a sharded table A is joined on columns w & x, which are not its shard key, to a sharded table B on its shard key columns y & z, redistribution of data will occur based on join type:
    • For INNER & LEFT joins, the right-side table B will be replicated
    • For FULL OUTER joins, both tables will be replicated
  • If a replicated table is joined to a replicated table, and the join is of any type, it will be a node-local join and not require redistribution of any data.

Creating a Join View

To create a join view, the /create/jointable endpoint requires four parameters:

  1. the name of the join view to create
  2. the list of member tables to be joined with optional aliases (TableA as A, TableB as B, ..., TableZ)
  3. the list of columns and/or column expressions selected from the joined tables with optional aliases (TableA.id, B.id as B_id, ..., TableY.total + TableZ.total as YZ_total)
  4. the SQL-style expression by which the tables can be joined and, optionally, filtered (TableA.b_id = B.id and ... A.z_id = TableZ.id)

Given a table clause of TableA as A, the a_id column of that table can be accessed via any of the following means:

  • unique name, assuming the name is unique across the tables being joined (a_id)
  • name with table prefix (TableA.a_id)
  • name with table alias (A.a_id)

Note

All sharded tables must be associated through the given expression.

Examples

Inner Join

In Python, given tables customer, order and lineitem, a join view can be created via:

1
2
3
4
5
6
h_db.create_join_table(
    join_table_name = 'example.customer_order_item',
    table_names = ['example.customer as c','example.orders as o','example.lineitem as l'],
    column_names = ['c_name','o_orderstatus','l_partkey','l_extendedprice','l_quantity','l_orderkey'],
    expressions = ['c_custkey = o_custkey','o_orderkey = l_orderkey']
)

The view produced is an equi-join of the member tables, containing the specified columns from all three tables, accessible by their simple column names, given no name collisions among the columns selected. The result would match what would be produced by the SQL:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
SELECT
    c_name,
    o_orderstatus,
    l_partkey,
    l_extendedprice,
    l_quantity
FROM
    example.customer c,
    example.orders o,
    example.lineitem l
WHERE
    c_custkey = o_custkey AND
    o_orderkey = l_orderkey

or:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
SELECT
    c_name,
    o_orderstatus,
    l_partkey,
    l_extendedprice,
    l_quantity
FROM
    example.customer c
INNER JOIN
    example.orders o ON c_custkey = o_custkey
INNER JOIN
    example.lineitem l ON o_orderkey = l_orderkey

Outer Join

To perform a left outer join of orders with or without line items, given tables order & lineitem, a join view can be created via:

1
2
3
4
5
6
h_db.create_join_table(
    join_table_name = 'example.order_lineitem',
    table_names = ['example.orders as o','example.lineitem as l'],
    column_names = ['o_orderkey','l_partkey','l_extendedprice','l_quantity'],
    expressions = ['left join o, l on (o_orderkey = l_orderkey)']
)

The view produced would match what would be produced by the SQL:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
SELECT
    o_orderkey,
    l_partkey,
    l_extendedprice,
    l_quantity
FROM
    example.orders o
LEFT JOIN
    example.lineitem l
ON
    o_orderkey = l_orderkey

Operating on a Join View

Since join views are read-only, no inserts, updates, or deletes can be issued against them. Additionally, the following data processing endpoints cannot be applied to a join view.

Examples

The examples in this section will be in Python. For easier use of the API, a handle to the join view created in the Creating a Join View section can be established:

1
join_view = gpudb.GPUdbTable(name = 'example.customer_order_item', db = h_db)

A Python example filter on the join view created in the Creating a Join View section for customer orders containing multiples of the same product is:

1
2
3
4
join_view.filter(
    view_name = 'example.customer_order_item_multiple',
    expression = 'l_quantity > 1'
)

When executed against a join view, the filter endpoint produces a view that is also a join view, having the same endpoint restrictions mentioned above. A chain of such filters can be used to create more and more restrictive views of the original data set. The filter endpoint can also be used to merge any tables not merged during the creation of the original join view or during subsequent filtering of the view.

To call the group-by endpoint, in Python:

1
2
3
4
5
6
7
8
9
records = join_view.aggregate_group_by(
    column_names = [
        'c_name',
        'SUM(l_quantity) as total_items',
        'SUM(l_quantity * l_extendedprice) as total_spent'
    ],
    offset = 0,
    options = {'sort_by':'key'}
)['data']

To call the statistics endpoint, in Python:

1
2
3
4
revenue = join_view.aggregate_statistics(
    column_name = 'l_quantity * l_extendedprice',
    stats = 'sum'
)['stats']['sum']

To call the endpoint to retrieve data from specific join view columns, in Python:

1
2
3
4
5
records = join_view.get_records_by_column(
    column_names = ['c_name','o_orderstatus','l_partkey','l_extendedprice','l_quantity'],
    offset = 0,
    limit = gpudb.GPUdb.END_OF_SET
)

SQL Conversion Example

This example shows how the TPC-H benchmark's Query 5 SQL statement can be implemented in the Kinetica Python interface. The SQL for Query 5 is:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
SELECT
    n_name,
    SUM(l_extendedprice * (1 - l_discount)) AS revenue
FROM
    example.customer,
    example.orders,
    example.lineitem,
    example.supplier,
    example.nation,
    example.region
WHERE
    c_custkey = o_custkey AND
    l_orderkey = o_orderkey AND
    l_suppkey = s_suppkey AND
    c_nationkey = s_nationkey AND
    s_nationkey = n_nationkey AND
    n_regionkey = r_regionkey AND
    r_name = 'ASIA' AND
    o_orderdate >= '1994-01-01' AND
    o_orderdate < DATE('1994-01-01') + INTERVAL '1' YEAR
GROUP BY
    n_name
ORDER BY
    revenue DESC

To implement the above statement, a join view needs to be created with all of the tables listed in the FROM clause, and an expression, based on the WHERE clause, which binds the tables together and applies the necessary filters:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
h_db.create_join_table(
    join_table_name = 'example.tpch_query_5_join_filter',
    table_names = [
        'example.customer',
        'example.orders',
        'example.lineitem',
        'example.supplier',
        'example.nation',
        'example.region'
    ],
    column_names = ['n_name','l_extendedprice * (1 - l_discount) as effective_price'],
    expressions = [
        'c_custkey = o_custkey',
        'l_orderkey = o_orderkey',
        'l_suppkey = s_suppkey',
        'c_nationkey = s_nationkey',
        's_nationkey = n_nationkey',
        'n_regionkey = r_regionkey',
        'r_name = \'ASIA\'',
        'o_orderdate >= \'1994-01-01\'',
        "o_orderdate < TIMESTAMPADD(YEAR, 1, DATE('1994-01-01'))"
    ]
)

Neither the table names nor the column names of the join view need to be aliased, as all column names are unique across all tables in this example. However, the derived column expression still requires an alias. Note also that columns used in the expressions list do not need to appear in the column_names list.

This join view could also have been done in two steps, a join and a filter:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
h_db.create_join_table(
    join_table_name = 'example.tpch_query_5_join',
    table_names = [
        'example.customer',
        'example.orders',
        'example.lineitem',
        'example.supplier',
        'example.nation',
        'example.region'
    ],
    column_names = [
        'n_name',
        'l_extendedprice * (1 - l_discount) as effective_price',
        'r_name',
        'o_orderdate',
        'l_orderkey'
    ],
    expressions = [
        'c_custkey = o_custkey',
        'l_orderkey = o_orderkey',
        'l_suppkey = s_suppkey',
        'c_nationkey = s_nationkey',
        's_nationkey = n_nationkey',
        'n_regionkey = r_regionkey'
    ]
)

The join view can then have the WHERE clause filters applied in a separate step:

1
2
3
4
5
6
join_view.filter(
     view_name = 'example.tpch_query_5_join_filter',
     expression = 'r_name = \'ASIA\' and ' \
                  "o_orderdate >= '1994-01-01' and " \
                  "o_orderdate < TIMESTAMPADD(YEAR, 1, DATE('1994-01-01'))"
)

Note that the columns used in the filter's expression must appear in the join's column_names list, so that they are available for subsequent filter calls to use.

Regardless of whether the join/filter was done in one step or two, the results need to be grouped and summed to match the TPC-H query. The following call is given the directive to sort_by the value, which will sort on the aggregation column:

1
2
3
4
5
6
7
8
records = join_view.aggregate_group_by(
    column_names = [
        'n_name',
        'SUM(effective_price) as revenue'
    ],
    offset = 0,
    options = {"sort_by":"value", "sort_order":"descending"}
)['data']

The results can be displayed by calling:

1
2
for record in zip(records["n_name"], records["revenue"]):
    print "{:<13s} {:7.2f}".format(record[0], record[1])

Performance Optimization

In order to improve the performance of a join operation, the columns used to make the relation should be indexed.

When a join is performed along a foreign key relation, it is indexed automatically as a function of the foreign key itself--no additional setup is required.

When a join is performed along a relation where no foreign key exists, the indexing must be done as a separate configuration step. Both columns, on either side of the join relation, should have a column index applied. An error may result if a join is attempted across a relation where only one side of the relation is indexed.

When a composite equi-join is performed, only one index will be used. For example: you have two tables, a and b, that each have the same columns, x and y. You also have an attribute index on each column. If you use the expression a.x == b.x and a.y == b.y, only the first expression, a.x == b.x, will use the index.

Limitations & Cautions

  • Joins cannot be used to render geospatial tracks; read more about geospatial objects here.
  • While a node-local join may include any number of replicated tables and sharded tables, all sharded tables must be joined by equating their respective shard key columns.
  • A sharded table can be left (outer) joined to a replicated table via node-local join, but a replicated table can only be left (outer) joined to a sharded table via distributed join.
  • When a data set is cleared (dropped) all the join views created from the data set are also automatically cleared.
  • Records cannot be added to a join view, nor can records be updated or deleted.
  • A join view is transient, by default, and will expire after the default TTL setting.
  • A join view is not persisted, and will not survive a database restart.

Memory Implications

Joins have a variable memory footprint depending on the type of join. Kinetica will attempt to create the most efficient join as possible. The amount of memory a join requires depends on the type of join and the source data sets being joined together.

  • A primary key join occupies the least amount of memory in Kinetica. This type is created from a join between two data sets with a primary key - foreign key relationship, allowing for an index on the lookup between the two data sets.
  • A bitmask join occupies the second least amount of memory in Kinetica. This type results from a join between two data sets with no filtering of either table.
  • An equijoin typically occupies the third least amount of memory in Kinetica. This type results from a join between two data sets containing an equality operator, returning only the records with equivalent values for the specified columns.
  • A predicate join can be the most expensive or one of the cheapest for your memory; it entirely depends on how sparse the join is. A predicate join comprises a list of indices, or matched pairs of values, from the source data sets. This type results from a join between two data sets that are filtered prior to joining them.

Any column expression used in the column list for a join will be evaluated when the join is queried, requiring no persistent memory to hold the calculated values. While the memory footprint is small, the performance impact could be large, as the expression will need to be re-evaluated each time the join is queried.

Operations on Underlying Tables

Join views are affected by data modifications on underlying tables in different ways, based on the type of data modification.

  • inserts - if a record is added to an underlying table, it will not appear in queries of the join view, whether or not the applied filter would include it
  • updates - if a record that already contributes to a join view record set is modified in its underlying table, it will appear with the modified values when the join view is re-queried; if a record that does not contribute to a join view record set is modified in its underlying table in a way that would make it part of the join view record set, it will not appear when the join view is re-queried
  • deletes - if a record that already contributes to a join view record set is deleted from its underlying table, it will still appear when the join view is re-queried

Overall, data manipulation within the tables supporting a join view will not cause the number of records returned by the join view to change, though updates to records already part of the join view record set will be reflected in the data values retrieved from the join view.