Distributed in-database vectorized operations using user defined table functions
The subject technology receives first data. The subject technology transforms first data to a wide format, the wide format comprising a second table. The subject technology splits a set of rows into a set of shards of rows. The subject technology, for each shard from the set of shards, sends a particular user defined table function (UDTF), including a particular shard of rows, to a different execution node to perform a set of operations for determining a rolling correlation over a window size. The subject technology provides a set of output values of each particular UDTF corresponding to each shard from the set of shards in a second UDTF. The subject technology sends the second UDTF to a particular execution node to perform an aggregate operation. The subject technology receives a value of the rolling correlation from the particular execution node based on the aggregate operation.
1. A system comprising:
at least one hardware processor; and
a memory storing instructions that cause the at least one hardware processor to perform operations comprising:
receiving first data comprising a first table, the first table comprising a first set of columns;
transforming the first data to a wide format, the wide format comprising a second table, the second table comprising a second set of columns, each column representing a particular asset, the second set of columns corresponding to a set of rows, and each row of the set of rows corresponding to a particular date;
splitting the set of rows into a set of shards of rows;
determining, based on a first number of the set of shards, a first number of execution nodes for processing the set of shards;
for each shard from the set of shards, sending a particular user defined table function (UDTF), including a particular shard of rows, to a different execution node to perform a set of operations for determining a rolling correlation over a window size, a particular vector processor of the different execution node being utilized to perform the set of operations for determining the rolling correlation over the window size;
providing a set of output values of each particular UDTF corresponding to each shard from the set of shards in a second UDTF;
sending the second UDTF to a particular execution node to perform an aggregate operation using the set of output values of each particular UDTF; and
receiving a value of the rolling correlation from the particular execution node based on the aggregate operation.
2. The system of claim 1 , wherein the first set of columns comprises a first column corresponding to a date, a second column corresponding to a financial asset, and a third column corresponding to a return.
3. The system of claim 2 , wherein each shard from the set of shards includes a same number of rows.
4. The system of claim 1 , wherein the window size comprises a period of time, the period of time comprising a particular number of days.
5. The system of claim 1 , wherein each shard from the set of shards of rows is not stored entirely in memory.
6. The system of claim 1 , wherein sending the particular UDTF further comprises:
for each shard from the set of shards, sending a shard identifier to the different execution node.
7. The system of claim 1 , wherein the operations further comprise:
performing a join operation based on an individual financial asset table to a correlation table; and
determining an aggregate value at risk using matrix multiplication.
8. The system of claim 1 , wherein the operations further comprise:
determining a set of overlap rows, the set of overlap rows comprising a particular number of rows between two different shards; and
sending the set of overlap rows to another different node to compute rolling correlation for those rows.
9. The system of claim 1 , wherein each execution node comprises a different server instance or is included in a different virtual warehouse.
10. A method comprising:
receiving first data comprising a first table, the first table comprising a first set of columns;
transforming the first data to a wide format, the wide format comprising a second table, the second table comprising a second set of columns, each column representing a particular asset, the second set of columns corresponding to a set of rows, and each row of the set of rows corresponding to a particular date;
splitting the set of rows into a set of shards of rows;
determining, based on a first number of the set of shards, a first number of execution nodes for processing the set of shards;
for each shard from the set of shards, sending a particular user defined table function (UDTF), including a particular shard of rows, to a different execution node to perform a set of operations for determining a rolling correlation over a window size, a particular vector processor of the different execution node being utilized to perform the set of operations for determining the rolling correlation over the window size;
providing a set of output values of each particular UDTF corresponding to each shard from the set of shards in a second UDTF;
sending the second UDTF to a particular execution node to perform an aggregate operation using the set of output values of each particular UDTF; and
receiving a value of the rolling correlation from the particular execution node based on the aggregate operation.
11. The method of claim 10 , wherein the first set of columns comprises a first column corresponding to a date, a second column corresponding to a financial asset, and a third column corresponding to a return.
12. The method of claim 11 , wherein each shard from the set of shards includes a same number of rows.
13. The method of claim 10 , wherein the window size comprises a period of time, the period of time comprising a particular number of days.
14. The method of claim 10 , wherein each shard from the set of shards of rows is not stored entirely in memory.
15. The method of claim 10 , wherein sending the particular UDTF further comprises:
for each shard from the set of shards, sending a shard identifier to the different execution node.
16. The method of claim 10 , further comprising:
performing a join operation based on an individual financial asset table to a correlation table; and
determining an aggregate value at risk using matrix multiplication.
17. The method of claim 10 , further comprising:
determining a set of overlap rows, the set of overlap rows comprising a particular number of rows between two different shards; and
sending the set of overlap rows to another different node to compute the rolling correlation for the set of overlap rows.
18. The method of claim 10 , wherein each execution node comprises a different server instance or is included in a different virtual warehouse.
19. A non-transitory computer-storage medium comprising instructions that, when executed by one or more processors of a machine, configure the machine to perform operations comprising:
receiving first data comprising a first table, the first table comprising a first set of columns;
transforming the first data to a wide format, the wide format comprising a second table, the second table comprising a second set of columns, each column representing a particular asset, the second set of columns corresponding to a set of rows, and each row of the set of rows corresponding to a particular date;
splitting the set of rows into a set of shards of rows;
determining, based on a first number of the set of shards, a first number of execution nodes for processing the set of shards;
for each shard from the set of shards, sending a particular user defined table function (UDTF), including a particular shard of rows, to a different execution node to perform a set of operations for determining a rolling correlation over a window size, a particular vector processor of the different execution node being utilized to perform the set of operations for determining the rolling correlation over the window size;
providing a set of output values of each particular UDTF corresponding to each shard from the set of shards in a second UDTF;
sending the second UDTF to a particular execution node to perform an aggregate operation using the set of output values of each particular UDTF; and
receiving a value of the rolling correlation from the particular execution node based on the aggregate operation.
20. The non-transitory computer-storage medium of claim 19 , wherein the first set of columns comprises a first column corresponding to a date, a second column corresponding to a financial asset, and a third column corresponding to a return.
21. The non-transitory computer-storage medium of claim 20 , wherein each shard from the set of shards includes a same number of rows.
22. The non-transitory computer-storage medium of claim 19 , wherein the window size comprises a period of time, the period of time comprising a particular number of days.
23. The non-transitory computer-storage medium of claim 19 , wherein each shard from the set of shards of rows is not stored entirely in memory.
24. The non-transitory computer-storage medium of claim 19 , wherein sending the particular UDTF further comprises:
for each shard from the set of shards, sending a shard identifier to the different execution node.
25. The non-transitory computer-storage medium of claim 19 , wherein the operations further comprise:
performing a join operation based on an individual financial asset table to a correlation table; and
determining an aggregate value at risk using matrix multiplication.
26. The non-transitory computer-storage medium of claim 19 , wherein the operations further comprise:
determining a set of overlap rows, the set of overlap rows comprising a particular number of rows between two different shards; and
sending the set of overlap rows to another different node to compute the rolling correlation for the set of overlap rows.
27. The non-transitory computer-storage medium of claim 19 , wherein each execution node comprises a different server instance or is included in a different virtual warehouse.