faster bmerge numeric and roll

.ci/atime/tests.R

@@ -424,5 +424,6 @@ test.list <- atime::atime_test_list(
     seconds.limit = 1,
     expr = data.table:::`[.data.table`(dt, , base::max(V1, na.rm = TRUE), by = id)),
 
+
   tests=extra.test.list)
 # nolint end: undesirable_operator_linter.

NEWS.md

@@ -12,6 +12,8 @@
 
 3. options `"datatable.old.matrix.autoname"` is now `FALSE` by default, meaning `names(data.table(x=1, cbind(1)))` is now `c("x", "V2")`. Toggle the option to retain the old behavior for now; future releases will work to remove this possibility. See the release notes for 1.18.0, item 1 under `NOTE OF INTENDED FUTURE POTENTIAL BREAKING CHANGES`.
 
+4. Joins on numeric (`double`, including `POSIXct`) columns are no longer affected by `setNumericRounding()`, and always use exact matching as with the default `setNumericRounding(0L)`. If you relied on `setNumericRounding(1L)` or `setNumericRounding(2L)` to match nearly-equal numeric values in joins, round those columns explicitly before joining. `setNumericRounding()` continues to affect grouping and ordering. Numeric joins, especially rolling joins, are faster as a result. Thanks @jangorecki for the PR.
+
 ### NEW FEATURES
 
 1. `nafill()`, `setnafill()` extended to work on logical, factor and character vectors (part of [#3992](https://github.com/Rdatatable/data.table/issues/3992)). Includes support for `Date`, `IDate`, `POSIXct` and character vectors. Thanks @jangorecki for the request and @jangorecki, @MichaelChirico and @ben-schwen for the PRs.

inst/tests/tests.Rraw

@@ -2741,8 +2741,14 @@ test(932, DT[J(c(0,2,6,8)), roll=+Inf, rollends=TRUE, v], INT(1,1,6,7))
 test(933, DT[J(c(0,2,6,8)), roll=-Inf, rollends=TRUE, v], INT(1,2,7,7))
 test(934, DT[J(c(0,2,6,8)), roll=+Inf, v], INT(NA,1,6,7))
 test(935, DT[J(c(0,2,6,8)), roll=-Inf, v], INT(1,2,7,NA))
-test(936, DT[J(c(-10,-1,2,12,13)), roll=5, rollends=TRUE, v], INT(NA,1,1,7,NA))
-test(937, DT[J(c(-10,-1,2,12,13)), roll=-5, rollends=TRUE, v], INT(NA,1,2,7,NA))
+test(936.1, DT[J(c(-10,-1,2,12,13)), roll=5, rollends=TRUE, v], INT(NA,1,1,7,NA))
+test(936.2, DT[J(c(-10,-1,-4,2,12,9,13)), roll=5, rollends=TRUE, v], INT(NA,1,1,1,7,7,NA))  # include -4 (exactly row1-roll) and out-of-order too
+test(936.3, DT[J(c(-10,-1,2,12,13)), roll=5, v], INT(NA,NA,1,7,NA))  # default rollends when roll>0 is c(FALSE,TRUE)
+test(936.4, DT[J(c(-10,-1,-4,2,12,13,9)), roll=5, v], INT(NA,NA,NA,1,7,NA,7))
+test(937.1, DT[J(c(-10,-1,2,12,13)), roll=-5, rollends=TRUE, v], INT(NA,1,2,7,NA))
+test(937.2, DT[J(c(-10,-1,-4,2,12,9,13)), roll=-5, rollends=TRUE, v], INT(NA,1,1,2,7,7,NA))
+test(937.3, DT[J(c(-10,-1,2,12,13)), roll=-5, v], INT(NA,1,2,NA,NA))
+test(937.4, DT[J(c(-10,-1,-4,2,12,9,13)), roll=-5, v], INT(NA,1,1,2,NA,NA,NA))
 test(938, DT[J(c(-10,2,6,7,8)), roll="nearest", v], INT(1,1,7,7,7))
 test(939, DT[J(c(-10,2,6,7,8)), roll="nearest", rollends=c(TRUE,FALSE), v], INT(1,1,7,7,NA))
 test(940, DT[J(c(-10,2,6,7,8)), roll="nearest", rollends=c(FALSE,TRUE), v], INT(NA,1,7,7,7))
@@ -4349,7 +4355,7 @@ test(1196.1, DT[,.N,by=val]$N, INT(1,1,1,1))
 test(1196.2, DT[.(x),.N], 1L)
 old_rounding = setNumericRounding(2L) # default is 0
 test(1197.1, DT[,.N,by=val]$N, INT(1,1,2))
-test(1197.2, DT[.(x),.N], 2L)
+test(1197.2, DT[.(x),.N], 1L)  # bmerge ignores setNumericRounding.
 setNumericRounding(old_rounding)
 
 DT = data.table(id=INT(1,2,1), val1=3:1, val2=3:1, val3=list(2:3,4:6,7:10))   # 5380

man/setNumericRounding.Rd

@@ -3,7 +3,7 @@
 \alias{getNumericRounding}
 \title{ Change or turn off numeric rounding }
 \description{
-Change rounding to 0, 1 or 2 bytes when joining, grouping or ordering numeric
+Change rounding to 0, 1 or 2 bytes when grouping or ordering numeric
 (i.e. double, POSIXct) columns.
 }
 \usage{
@@ -15,9 +15,9 @@ getNumericRounding()
 }
 \details{
 Computers cannot represent some floating point numbers (such as 0.6)
-precisely, using base 2. This leads to unexpected behaviour when joining or
-grouping columns of type 'numeric'; i.e. 'double', see example below. In
-cases where this is undesirable, data.table allows rounding such data up to
+precisely, using base 2. This can lead to unexpected behaviour when grouping
+columns of type 'numeric'; i.e. 'double', see example below. In cases where this
+is undesirable, data.table allows rounding such data up to
 approximately 11 significant figures which is plenty of digits for many cases.
 This is achieved by rounding the last 2 bytes off the significand. Other possible
 values are 1 byte rounding, or no rounding (full precision, default).
@@ -50,7 +50,8 @@ DT[.(0.6)]   # no match, can be confusing since 0.6 is clearly there in DT
              # happens due to floating point representation limitations
 
 setNumericRounding(2)   # round off last 2 bytes
-DT[.(0.6)]   # works
+DT[.(0.6)]   # numeric joins always use full precision, hence, still no match
+DT[,.N,by=a] # grouping uses rounding
 
 # using type 'numeric' for integers > 2^31 (typically ids)
 DT = data.table(id = c(1234567890123, 1234567890124, 1234567890125), val=1:3)