Version:

Table Of Contents

  1. Docs
  2. Features
  3. Concepts
  4. Window

Window

Kinetica supports the SQL concept of windows. It can do so through the /create/projection and /get/records/bycolumn endpoints. The window function is specified as a column name in the column_name parameter.

A window projection can be created from any table or view. If the source data set is replicated, the resulting window projection will also be replicated. The resulting window projection will be sharded in two cases:

  • If the source data set is sharded and the shard key is included in the projected column list; columns in a multi-column shard key must appear in the projected column list in the same relative order they appear in the shard key, though they do not need to be listed consecutively
  • If the projection call has a shard key specified

Function Details

A window in Kinetica is constructed like so:

<window function> OVER (<window specification>) AS <alias>

Function Type

A window function either be of aggregate or ranking type.

Aggregate

Function Description
AVG(expr) Calculates the average of the given expression expr over the specified window frame
COUNT(expr) Calculates the count of the given expression expr over the specified window frame
MAX(expr) Calculates the maximum value of the given expression expr over the specified window frame
MEAN(expr) Synonymous with AVG(). Calculates the average of the given expression expr over the specified window frame
MIN(expr) Calculates the minimum value of the given expression expr over the specified window frame
SUM(expr) Calculates the sum of the given expression expr over the specified window frame

Ranking

Function Description
CUME_DIST()

The relative position of the current row within the cumulative distribution of the selected partition, expressed as a percentage from 0 (exclusive) to 1 (inclusive). The formula for this calculation is as follows:

(count of <= records within partition) / (partition row count)

This function is shorthand for using the COUNT() function in separate partition statements to arrive at the same result:

DOUBLE
(
    COUNT(*) OVER
    (
        PARTITION BY <partition_column>
        ORDER BY <sort_column>
    )
) /
(
    COUNT(*) OVER
    (
        PARTITION BY <partition_column>
        ORDER BY <sort_column>
        ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING
    )
)
DENSE_RANK() Number of the current row within the selected partition except rows with identical values evaluate to different ranks. Starts at 1
FIRST_VALUE(<column>) The value found in the first row within a frame of the given expression
LAG(<column>[, <num>]) The value of the row before the given expression's value. Provide an additional comma-separated value to specify which row to select, e.g., LAG(vendor_id, 3) would list the value in the vendor_id column from three rows prior to the current row
LAST_VALUE(<column>) The value found in the last row within a frame of the given expression
LEAD(<column>[, <num>]) The value of the row after the given expression's value. Provide an additional comma-separated value to specify which row to select, e.g., LEAD(vendor_id, 3) would list the value in the vendor_id column from three rows after the current row
NTILE(<num of groups>) The group number of the row after partitioning the rows into num of groups groups. For example, NTILE(4) will partition data by quartiles and return the associated group number, 1 to 4.
PERCENT_RANK()

The rank of the current row within the selected partition, expressed as a percentage from 0 to 1, inclusive. The formula for this calculation is as follows:

(rank within partition - 1) / (partition row count - 1)

This function is shorthand for using the RANK() & COUNT() functions in separate partition statements to arrive at the same result:

DOUBLE
(
    RANK() OVER
    (
        PARTITION BY <partition_column>
        ORDER BY <sort_column>
    ) - 1
) /
(
    COUNT(*) OVER
    (
        PARTITION BY <partition_column>
        ORDER BY <sort_column>
        ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING
    ) - 1
)
RANK() Number of the current row within the selected partition. However, rows with identical values evaluate to the same rank. Starts at 1
ROW_NUMBER() Number of the current row within the selected partition. Starts at 1

Window Specification

A window specification is constructed like so:

PARTITION BY <column> [ORDER BY <ordering expression list> [<frame clause>]]

The window specification defines partition, frames and sort specifications. A few things to note:

  • PARTITION BY clauses can contain several comma-separated keys that define the partitions, e.g., PARTITION BY vendor_id, payment_type.
  • PARTITION BY null (or other constant value) will create a window spanning the entire data set, effectively windowing without a partition.
  • The ordering expression list, while optional, is useful in sorting the individual partitions. For RANGE frames, the ORDER BY can have a single column, which the frame will use for bounding. For ROWS frames, any number of columns can be included in the ORDER BY; e.g., ORDER BY vendor_id, payment_type. Sort order can be specified using ASC (ascending, the default) and DESC (descending) keywords. The default null ordering for an ascending sort is NULLS FIRST; the default for a descending sort is NULLS LAST.
  • All window functions except for FIRST_VALUE() & LAST_VALUE() require an ORDER BY clause.

Frame Clauses

A frame clause within a window specification is constructed like so:

<RANGE | ROWS> BETWEEN <frame start> AND <frame end>

Frames are defined by either ranges of values (RANGE) or number of rows (ROWS):

Range Frames

RANGE specifies that containment within a frame should be determined by the value in the column specified in the ORDER BY clause. This has several implications:

  • Peer rows (rows containing the same value in the ORDER BY column) are always included in a range frame; for any given row, an aggregate function applied to its frame will have the same result for all of its peer rows.
  • The bounds defined in the frame start & end are based on the column's values; e.g., BETWEEN 100 PRECEDING AND CURRENT ROW means a frame will contain records with ORDER BY column values greater than or equal to the current record's column value minus 100 and less than or equal to the current record's column value.
  • The ORDER BY clause can specify only one column, as this column will be used to bound the frame by its values.
Rows Frames

ROWS specifies that containment within a frame should be determined by the ordering of the rows, as specified in the ORDER BY clause. This also has several implications:

  • Peer rows are not considered when bounding a rows frame; only the relative ordering of the rows is considered.
  • The bounds defined in the frame start & end are based on the row's relative position within the ordered result set; e.g., BETWEEN 100 PRECEDING AND CURRENT ROW means a frame will contain 100 records before the current row as well as the current row.
  • The ORDER BY clause can contain multiple comma-separated columns.
Frame Bounds

A frame has a start and end specification:

  • Frame start keywords:
    • UNBOUNDED PRECEDING: the first row of the partition
    • <number> PRECEDING: either n rows before the current row (for rows-based frames), or n values less than the current row's value (for range-based frames)
    • CURRENT ROW: either the current row (for rows-based frames), or the current row and its peer rows (for range-based frames)
    • <number> FOLLOWING: either n rows after the current row (for rows-based frames), or n values greater than the current row's value (for range-based frames)
  • Frame end keywords:
    • UNBOUNDED FOLLOWING: the last row of the partition
    • <number> FOLLOWING: either n rows after the current row (for rows-based frames), or n values greater than the current row's value (for range-based frames)
    • CURRENT ROW: either the current row (for rows-based frames), or the current row and its peer rows (for range-based frames)
    • <number> PRECEDING: either n rows before the current row (for rows-based frames), or n values less than the current row's value (for range-based frames)

Important

The frame start must begin before the end of the frame end.

Frames that extend outside of the available data set will be narrowed to operate only over the available data. For example, a frame defined as 10 PRECEDING will, for the 5th record in the data set, be narrowed to consider just the first four records as the beginning of the frame in any calculations.

Important

If you use the <number> FOLLOWING or <number> PRECEDING keywords as both the frame start and end keywords, the first <number> used must be less than or equal to the second <number>

Example frame:

ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW

Examples

The following call to /get/records/bycolumn via Python would calculate a rolling sum of total amounts collected per vendor for the first two hours after midnight on January 1st, 2015, as well as the number of trips that occurred within 5 minutes of each trip:

records = h_db.get_records_by_column_and_decode(
    table_name = "nyctaxi",
    column_names = [
        "vendor_id",
        "STRING(pickup_datetime) AS pickup_datetime",
        "passenger_count",
        "total_amount",
        "SUM(total_amount) OVER" \
        "   (" \
        "      PARTITION BY vendor_id" \
        "      ORDER BY pickup_datetime" \
        "      ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW" \
        "   ) AS growing_sum",
        "COUNT(*) OVER" \
        "    (" \
        "        PARTITION BY vendor_id" \
        "        ORDER BY LONG(pickup_datetime)" \
        "        RANGE BETWEEN 300000 PRECEDING AND 300000 FOLLOWING" \
        "    ) AS trip_demand"
    ],
    offset = 0,
    limit = gpudb.GPUdb.END_OF_SET,
    options = {
        "expression": "pickup_datetime >= DATE('2015-01-01') AND pickup_datetime < DATETIME('2015-01-01 02:00:00')",
        "order_by": "vendor_id, pickup_datetime"
    }
)["records"]

The following call to /create/projection via Python would calculate a 5-before and 10-after moving average on the trip distance for 4-passenger trips, on a given day, per vendor:

h_db.create_projection(
    table_name = "nyctaxi",
    projection_name = "nyctaxi_local_avg_dist_by_vendor_4_passengers_20150101",
    column_names = [
        "vendor_id",
        "pickup_datetime",
        "trip_distance",
        "AVG(trip_distance) OVER" \
        "   (" \
        "      PARTITION BY vendor_id " \
        "      ORDER BY pickup_datetime " \
        "      ROWS BETWEEN 5 PRECEDING AND 10 FOLLOWING " \
        "   ) AS local_avg_dist"
    ],
    options = {
        "expression": "passenger_count = 4 AND " \
                      "pickup_datetime >= DATE('2015-01-01') AND " \
                      "pickup_datetime < DATE('2015-01-02')"
    }
)

The following call to /create/projection via Python would rank, by vendor, the total amounts collected from 3-passenger trips on a given day:

h_db.create_projection(
    table_name = "nyctaxi",
    projection_name = "nyctaxi_ranked_fares_by_vendor_3_passengers_20150111",
    column_names = [
        "vendor_id",
        "pickup_datetime",
        "dropoff_datetime",
        "total_amount",
        "RANK() OVER " \
        "   (PARTITION BY vendor_id ORDER BY total_amount) AS ranked_total",
        "PERCENT_RANK() OVER " \
        "   (PARTITION BY vendor_id ORDER BY total_amount) * 100 AS percent_ranked_total"
    ],
    options = {
        "expression": "passenger_count = 3 AND " \
                      "pickup_datetime >= DATE('2015-01-11') AND " \
                      "pickup_datetime < DATETIME('2015-01-12')"
    }
)

The following call to /create/projection via Python would compare each 5-passenger trip's tip amount to the lowest, average, & highest tip amounts for the vendor on a given day:

h_db.create_projection(
    table_name = "nyctaxi",
    projection_name = "nyctaxi_hi_lo_tip_compare_by_vendor_5_passengers_20150417",
    column_names = [
        "vendor_id",
        "pickup_datetime",
        "tip_amount",
        "tip_amount - " \
        "   FIRST_VALUE(tip_amount) OVER" \
        "      (PARTITION BY vendor_id ORDER BY tip_amount NULLS LAST) AS vs_lowest_tip",
        "tip_amount - " \
        "   AVG(tip_amount) OVER" \
        "      (" \
        "         PARTITION BY vendor_id" \
        "         ORDER BY tip_amount" \
        "         ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING" \
        "      ) AS vs_average_tip",
        "tip_amount - " \
        "   FIRST_VALUE(tip_amount) OVER" \
        "      (PARTITION BY vendor_id ORDER BY tip_amount DESC) AS vs_highest_tip"
    ],
    options = {
        "expression": "passenger_count = 5 AND " \
                      "pickup_datetime >= DATE('2015-04-17') AND " \
                      "pickup_datetime < DATE('2015-04-18') AND " \
                      "tip_amount > 0 AND " \
                      "trip_distance > 0"
    }
)

The following call to /create/projection via Python would calculate the quartiles for total amounts collected per vendor, the result of which could then be aggregated by calling /aggregate/groupby via Python and used to compare the overall average total collected vs. the average total within the interquartile range:

h_db.create_projection(
    table_name = "nyctaxi",
    projection_name = "nyctaxi_revenue_quartile_by_vendor",
    column_names = [
        "vendor_id",
        "total_amount",
        "NTILE(4) OVER (PARTITION BY vendor_id ORDER BY total_amount) AS quartile"
    ],
    options = {
    }
)

nyctaxi_rqbv = gpudb.GPUdbTable(name = 'nyctaxi_revenue_quartile_by_vendor', db = h_db)

records = nyctaxi_rqbv.aggregate_group_by(
    column_names = [
        "vendor_id",
        "AVG(total_amount) AS average_total_amount",
        "AVG(IF((quartile IN (2,3)), total_amount, null)) AS average_interq_range_total_amount"
    ],
    offset = 0,
    limit = -9999,
    options = {
        "sort_by": "key"
    }
)["data"]

Limitations

  • Empty window specifications are currently not supported; all window specifications must include a PARTITION BY clause. However, partitioning by a constant (e.g., PARTITION BY null) will result in a window spanning the entire data set.

    Note

    This will force all windowed data to a single node for processing and could consume a great deal of memory on that node.

  • A ranking type function does not accept frame clauses.

  • Any ranking function other than FIRST_VALUE() & LAST VALUE does not accept a frame clause.

  • Any ranking function other than FIRST_VALUE() & LAST VALUE requires an ORDER BY clause.

  • Window functions executed via /create/projection will have the corresponding projection limitations.

© Copyright 2020, Kinetica. Created using Sphinx.