Vertica Documentation – Partitioning tables

Admin: Defining partitions

Mon, 01 Jan 0001 00:00:00 +0000

You can specify partitioning for new and existing tables:

Define partitioning for a table with CREATE TABLE.
Specify partitioning for an existing table by modifying its definition with ALTER TABLE.
Create partition groups to consolidate partitions into logical subsets, minimizing the use of ROS storage.

Admin: Hierarchical partitioning

Mon, 01 Jan 0001 00:00:00 +0000

The meta-function CALENDAR_HIERARCHY_DAY leverages partition grouping. You specify this function as the partitioning GROUP BY expression. CALENDAR_HIERARCHY_DAY organizes a table's date partitions into a hierarchy of groups: the oldest date partitions are grouped by year, more recent partitions are grouped by month, and the most recent date partitions remain un-grouped. Grouping is dynamic: as recent data ages, the Tuple Mover merges their partitions into month groups, and eventually into year groups.

Managing timestamped data

Partition consolidation strategies are especially important for managing timestamped data, where the number of partitions can quickly escalate and risk ROS pushback. For example, the following statements create the store_orders table and load data into it. The CREATE TABLE statement includes a simple partition clause that specifies to partition data by date:

=> DROP TABLE IF EXISTS public.store_orders CASCADE;
=> CREATE TABLE public.store_orders
(
    order_no int,
    order_date timestamp NOT NULL,
    shipper varchar(20),
    ship_date date
)
UNSEGMENTED ALL NODES PARTITION BY order_date::DATE;
CREATE TABLE
=> COPY store_orders FROM '/home/dbadmin/export_store_orders_data.txt';
41834
(1 row)

As COPY loads the new table data into ROS storage, it executes this table's partition clause by dividing daily orders into separate partitions—in this case, 809 partitions, where each partition requires its own ROS container:

=> SELECT COUNT (DISTINCT ros_id) NumROS, node_name FROM PARTITIONS
    WHERE projection_name ilike '%store_orders_super%' GROUP BY node_name ORDER BY node_name;
 NumROS |    node_name
--------+------------------
    809 | v_vmart_node0001
    809 | v_vmart_node0002
    809 | v_vmart_node0003
(3 rows)

This is far above the recommended maximum of 50 partitions per projection. This number is also close to the default system limit of 1024 ROS containers per projection, risking ROS pushback in the near future.

You can approach this problem in several ways:

Consider consolidating table data into larger partitions—for example, partition by month instead of day. However, partitioning data at this level might limit effective use of partition management functions.
Regularly archive older partitions, and thereby minimize the number of accumulated partitions. However, this requires an extra layer of data management, and also inhibits access to historical data.

Alternatively, you can use CALENDAR_HIERARCHY_DAY to automatically merge partitions into a date-based hierarchy of partition groups. Each partition group is stored in its own set of ROS containers, apart from other groups. You specify this function in the table partition clause as follows:

PARTITION BY partition‑expression
  GROUP BY CALENDAR_HIERARCHY_DAY( partition‑expression [, active‑months[, active‑years] ] )

Important

Two requirements apply to using CALENDAR_HIERARCHY_DAY in a partition clause:

partition‑expression must be a DATE.
The partition expressions specified by the PARTITION BY clause and CALENDAR_HIERARCHY_DAY must be identical.

For example, given the previous table, you can repartition it as follows:


=> ALTER TABLE public.store_orders
      PARTITION BY order_date::DATE
      GROUP BY CALENDAR_HIERARCHY_DAY(order_date::DATE, 2, 2) REORGANIZE;

Grouping DATE data hierarchically

CALENDAR_HIERARCHY_DAY creates hierarchies of partition groups, and merges partitions into the appropriate groups. It does so by evaluating the partition expression of each table row with the following algorithm, to determine its partition group key:

GROUP BY (
CASE WHEN DATEDIFF('YEAR', partition-expression, NOW()::TIMESTAMPTZ(6)) >= active-years
       THEN DATE_TRUNC('YEAR', partition-expression::DATE)
     WHEN DATEDIFF('MONTH', partition-expression, NOW()::TIMESTAMPTZ(6)) >= active-months
       THEN DATE_TRUNC('MONTH', partition-expression::DATE)
     ELSE DATE_TRUNC('DAY', partition-expression::DATE) END);

In this example, the algorithm compares order_date in each store_orders row to the current date as follows:

Determines whether order_date is in an inactive year.

If order_date is in an inactive year, the row's partition group key resolves to that year. The row is merged into a ROS container for that year.
If order_date is an active year, CALENDAR_HIERARCHY_DAY evaluates order_date to determine whether it is in an inactive month.

If order_date is in an inactive month, the row's partition group key resolves to that month. The row is merged into a ROS container for that month.
If order_date is in an active month, the row's partition group key resolves to the order_date day. This row is merged into a ROS container for that day. Any rows where order_date is a future date is treated in the same way.

Important

The CALENDAR_HIERARCHY_DAY algorithm assumes that most table activity is focused on recent dates. Setting active‑years and active‑months to a low number ≥ 2 serves to isolate most merge activity to date-specific containers, and incurs minimal overhead. Vertica recommends that you use the default setting of 2 for active‑years and active‑months. For most users, these settings achieve an optimal balance between ROS storage and performance.

As a best practice, never set active‑years and active‑months to 0.

For example, if the current date is 2017‑09-26, CALENDAR_HIERARCHY_DAY resolves active‑years and active‑months to the following date spans:

active‑years: 2016-01-01 to 2017-12-31. Partitions in active years are grouped into monthly ROS containers or are merged into daily ROS containers. Partitions from earlier years are regarded as inactive and merged into yearly ROS containers.
active‑months: 2017-08-01 to 2017-09-30. Partitions in active months are merged into daily ROS containers.

Now, the total number of ROS containers is reduced to 40 per projection:

=> SELECT COUNT (DISTINCT ros_id) NumROS, node_name FROM PARTITIONS
    WHERE projection_name ilike '%store_orders_super%' GROUP BY node_name ORDER BY node_name;
 NumROS |    node_name
--------+------------------
     40 | v_vmart_node0001
     40 | v_vmart_node0002
     40 | v_vmart_node0003
(3 rows)

Note

Regardless of how the Tuple Mover groups and merges partitions, it always identifies one or more partitions or partition groups as active. For details, see Active and inactive partitions.

Dynamic regrouping

As shown earlier, CALENDAR_HIERARCHY_DAY references the current date when it creates partition group keys and merges partitions. As the calendar advances, the Tuple Mover reevaluates the partition group keys of tables that are partitioned with this function, and moves partitions as needed to different ROS containers.

Thus, given the previous example, on 2017-10-01 the Tuple Mover creates a monthly ROS container for August partitions. All partition keys between 2017-08-01 and 2017-08-31 are merged into the new ROS container 2017-08:

Likewise, on 2018-01-01, the Tuple Mover creates a ROS container for 2016 partitions. All partition keys between 2016-01-01 and 2016-12-31 that were previously grouped by month are merged into the new yearly ROS container:

Caution

After older partitions are grouped into months and years, any partition operation that acts on a subset of older partition groups is liable to split ROS containers into smaller ROS containers for each partition—for example, MOVE_PARTITIONS_TO_TABLE, where force‑split is set to true. These operations can lead to ROS pushback. If you anticipate frequent partition operations on hierarchically grouped partitions, consider modifying the partition expression so partitions are grouped no higher than months.

Customizing partition group hierarchies

Vertica provides a single function, CALENDAR_HIERARCHY_DAY, to facilitate hierarchical partitioning. Vertica stores the GROUP BY clause as a CASE statement that you can edit to suit your own requirements.

For example, Vertica stores the store_orders partition clause as follows:

=> ALTER TABLE public.store_orders
      PARTITION BY order_date::DATE
      GROUP BY CALENDAR_HIERARCHY_DAY(order_date::DATE, 2, 2);
=> select export_tables('','store_orders');
...
CREATE TABLE public.store_orders ( ... )

PARTITION BY ((store_orders.order_date)::date)
GROUP BY (
CASE WHEN ("datediff"('year', (store_orders.order_date)::date, ((now())::timestamptz(6))::date) >= 2)
       THEN (date_trunc('year', (store_orders.order_date)::date))::date
     WHEN ("datediff"('month', (store_orders.order_date)::date, ((now())::timestamptz(6))::date) >= 2)
       THEN (date_trunc('month', (store_orders.order_date)::date))::date
     ELSE (store_orders.order_date)::date END);

You can modify the CASE statement to customize the hierarchy of partition groups. For example, the following CASE statement creates a hierarchy of months, days, and hours:

=> ALTER TABLE store_orders
PARTITION BY (store_orders.order_date)
GROUP BY (
CASE WHEN DATEDIFF('MONTH', store_orders.order_date, NOW()::TIMESTAMPTZ(6)) >= 2
       THEN DATE_TRUNC('MONTH', store_orders.order_date::DATE)
     WHEN DATEDIFF('DAY', store_orders.order_date, NOW()::TIMESTAMPTZ(6)) >= 2
       THEN DATE_TRUNC('DAY', store_orders.order_date::DATE)
     ELSE DATE_TRUNC('hour', store_orders.order_date::DATE) END);

Admin: Partitioning and segmentation

Mon, 01 Jan 0001 00:00:00 +0000

In Vertica, partitioning and segmentation are separate concepts and achieve different goals to localize data:

Segmentation refers to organizing and distributing data across cluster nodes for fast data purges and query performance. Segmentation aims to distribute data evenly across multiple database nodes so all nodes participate in query execution. You specify segmentation with the CREATE PROJECTION statement's hash segmentation clause.
Partitioning specifies how to organize data within individual nodes for distributed computing. Node partitions let you easily identify data you wish to drop and help reclaim disk space. You specify partitioning with the CREATE TABLE statement's PARTITION BY clause.

For example: partitioning data by year makes sense for retaining and dropping annual data. However, segmenting the same data by year would be inefficient, because the node holding data for the current year would likely answer far more queries than the other nodes.

The following diagram illustrates the flow of segmentation and partitioning on a four-node database cluster:

Example table data
Data segmented by HASH(order_id)
Data segmented by hash across four nodes
Data partitioned by year on a single node

While partitioning occurs on all four nodes, the illustration shows partitioned data on one node for simplicity.

Admin: Managing partitions

Mon, 01 Jan 0001 00:00:00 +0000

You can manage partitions with the following operations:

Admin: Active and inactive partitions

Mon, 01 Jan 0001 00:00:00 +0000

The Tuple Mover assumes that all loads and updates to a partitioned table are targeted to one or more partitions that it identifies as active. In general, the partitions with the largest partition keys—typically, the most recently created partitions—are regarded as active. As the partition ages, its workload typically shrinks and becomes mostly read-only.

Setting active partition count

You can specify how many partitions are active for partitioned tables at two levels, in ascending order of precedence:

Configuration parameter ActivePartitionCount determines how many partitions are active for partitioned tables in the database. By default, ActivePartitionCount is set to 1. The Tuple Mover applies this setting to all tables that do not set their own active partition count.
Individual tables can supersede ActivePartitionCount by setting their own active partition count with CREATE TABLE and ALTER TABLE.

Partitioned tables in the same database can be subject to different distributions of update and load activity. When these differences are significant, it might make sense for some tables to set their own active partition counts.

For example, table store_orders is partitioned by month and gets its active partition count from configuration parameter ActivePartitionCount. If the parameter is set to 1, the Tuple Mover identifes the latest month—typically, the current one—as the table's active partition. If store_orders is subject to frequent activity on data for the current month and the one before it, you might want the table to supersede the configuration parameter, and set its active partition count to 2:

ALTER TABLE public.store_orders SET ACTIVEPARTITIONCOUNT 2;

Note

For tables partitioned by non-temporal attributes, set its active partition count to reflect the number of partitions that are subject to a high level of activity—for example, frequent loads or queries.

Identifying the active partition

The Tuple Mover typically identifies the active partition as the one most recently created. Vertica uses the following algorithm to determine which partitions are older than others:

If partition X was created before partition Y, partition X is older.
If partitions X and Y were created at the same time, but partition X was last updated before partition Y, partition X is older.
If partitions X and Y were created and last updated at the same time, the partition with the smaller key is older.

You can obtain the active partitions for a table by joining system tables PARTITIONS and STRATA and querying on its projections. For example, the following query gets the active partition for projection store_orders_super:

=> SELECT p.node_name, p.partition_key, p.ros_id, p.ros_size_bytes, p.ros_row_count, ROS_container_count
     FROM partitions p JOIN strata s ON p.partition_key = s.stratum_key AND p.node_name=s.node_name
     WHERE p.projection_name = 'store_orders_super' ORDER BY p.node_name, p.partition_key;
    node_name     | partition_key |      ros_id       | ros_size_bytes | ros_row_count | ROS_container_count
------------------+---------------+-------------------+----------------+---------------+---------------------
 v_vmart_node0001 | 2017-09-01    | 45035996279322851 |           6905 |           960 |                   1
 v_vmart_node0002 | 2017-09-01    | 49539595906590663 |           6905 |           960 |                   1
 v_vmart_node0003 | 2017-09-01    | 54043195533961159 |           6905 |           960 |                   1
(3 rows)

Active partition groups

If a table's partition clause includes a GROUP BY expression, Vertica applies the table's active partition count to its largest partition group key, and regards all the partitions in that group as active. If you group partitions with Vertica meta-function CALENDAR_HIERARCHY_DAY, the most recent date partitions are also grouped by day. Thus, the largest partition group key and largest partition key are identical. In effect, this means that only the most recent partitions are active.

For more information about partition grouping, see Partition grouping and Hierarchical partitioning.

Admin: Partition pruning

Mon, 01 Jan 0001 00:00:00 +0000

If a query predicate specifies a partitioning expression, the query optimizer evaluates the predicate against the ROS containers of the partitioned data. Each ROS container maintains the minimum and maximum values of its partition key data. The query optimizer uses this metadata to determine which ROS containers it needs to execute the query, and omits, or prunes, the remaining containers from the query plan. By minimizing the number of ROS containers that it must scan, the query optimizer enables faster execution of the query.

For example, a table might be partitioned by year as follows:

=> CREATE TABLE ... PARTITION BY EXTRACT(year FROM date);

Given this table definition, its projection data is partitioned into ROS containers according to year, one for each year—in this case, 2007, 2008, 2009.

The following query specifies the partition expression date:


=> SELECT ... WHERE date = '12-2-2009';

Given this query, the ROS containers that contain data for 2007 and 2008 fall outside the boundaries of the requested year (2009). The query optimizer prunes these containers from the query plan before the query executes:

Examples

Assume a table that is partitioned by time and will use queries that restrict data on time.

=> CREATE TABLE time ( tdate DATE NOT NULL, tnum INTEGER)
     PARTITION BY EXTRACT(year FROM tdate);
=> CREATE PROJECTION time_p (tdate, tnum) AS
=> SELECT * FROM time ORDER BY tdate, tnum UNSEGMENTED ALL NODES;

Note

Projection sort order has no effect on partition pruning.

=> INSERT INTO time VALUES ('03/15/04' , 1);
=> INSERT INTO time VALUES ('03/15/05' , 2);
=> INSERT INTO time VALUES ('03/15/06' , 3);
=> INSERT INTO time VALUES ('03/15/06' , 4);

The data inserted in the previous series of commands are loaded into three ROS containers, one per year, as that is how the data is partitioned:

=> SELECT * FROM time ORDER BY tnum;
   tdate    | tnum
------------+------
 2004-03-15 |    1  --ROS1 (min 03/01/04, max 03/15/04)
 2005-03-15 |    2  --ROS2 (min 03/15/05, max 03/15/05)
 2006-03-15 |    3  --ROS3 (min 03/15/06, max 03/15/06)
 2006-03-15 |    4  --ROS3 (min 03/15/06, max 03/15/06)
(4 rows)

Here's what happens when you query the time table:

In this query, Vertica can omit container ROS2 because it is only looking for year 2004:
```
=> SELECT COUNT(*) FROM time WHERE tdate = '05/07/2004';
```
In the next query, Vertica can omit two containers, ROS1 and ROS3:
```
=> SELECT COUNT(*) FROM time WHERE tdate = '10/07/2005';
```
The following query has an additional predicate on the tnum column for which no minimum/maximum values are maintained. In addition, the use of logical operator OR is not supported, so no ROS elimination occurs:
```
=> SELECT COUNT(*) FROM time WHERE tdate = '05/07/2004' OR tnum = 7;
```

Vertica Documentation – Partitioning tables

Admin: Defining partitions

Admin: Hierarchical partitioning

Managing timestamped data

Important

Grouping DATE data hierarchically

Important

Note

Dynamic regrouping

Caution

Customizing partition group hierarchies

Admin: Partitioning and segmentation

See also

Admin: Managing partitions

Admin: Active and inactive partitions

Setting active partition count

Note

Identifying the active partition

Active partition groups

Admin: Partition pruning

Examples

Note