truncation_strategy#
Defines strategies for performing truncation in private joins.
Classes#
Strategies for performing truncation in private joins. |
- class TruncationStrategy#
Strategies for performing truncation in private joins.
These are used to determine the sensitivity of a private join between two tables having
AddMaxRowsas a protected change. The formula for the sensitivity of the table resulting from a private join is:\(sensitivity=(T_{left}*S_{right}*M_{left}) + (T_{right}*S_{left}*M_{right})\)
where:
\(T_{left}\) and \(T_{right}\) are the truncation thresholds for the left and right truncation strategies, respectively. This value is 1 for
DropNonUnique.\(S_{left}\) and \(S_{right}\) are the stability of the left and right truncation strategies, respectively. This value is 2 for
DropExcessand 1 forDropNonUnique.\(M_{left}\) and \(M_{right}\) are the
max_rowsparameters of theAddMaxRowsprotected changes of the left and right tables, respectively.
- class DropExcess#
Bases:
TruncationStrategy.TypeDrop rows with matching join keys above a threshold.
This truncation strategy drops rows such that no more than
max_rowsrows have the same join key. Which rows are kept is deterministic and does not depend on the order in which they appear in the private data. For example, using theDropExcess(1)strategy while joining on columns A and B in the below table:A
B
Val
a
b
1
a
c
2
a
b
3
b
a
4
causes it to be treated as one of the below tables:
A
B
Val
a
b
1
a
c
2
b
a
4
A
B
Val
a
b
3
a
c
2
b
a
4
This is generally the preferred truncation strategy, even when the
DropNonUniquestrategy could also be used, because it results in fewer dropped rows.- max_rows :int#
Maximum number of rows to keep.
- class DropNonUnique#
Bases:
TruncationStrategy.TypeDrop all rows with non-unique join keys.
This truncation strategy drops all rows which share join keys with another row in the dataset. It is similar to the
DropExcess(1)strategy, but doesn’t keep any of the rows with duplicate join keys. For example, using theDropNonUniquestrategy while joining on columns A and B in the below table:A
B
Val
a
b
1
a
c
2
a
b
3
b
a
4
causes it to be treated as:
A
B
Val
a
c
2
b
a
4
This truncation strategy results in less noise than
DropExcess(1). However, it also drops more rows in datasets where many rows have non-unique join keys. In most cases, DropExcess is the preferred strategy.