delete rows by reference

NAMESPACE

@@ -11,7 +11,7 @@ export(setindex, setindexv, indices)
 export(as.data.table,is.data.table,test.data.table)
 export(last,first,like,"%like%","%ilike%","%flike%","%plike%",between,"%between%",inrange,"%inrange%", "%notin%")
 export(timetaken)
-export(truelength, setalloccol, alloc.col, ":=", let)
+export(truelength, setalloccol, setallocrow, alloc.col, ":=", let)
 export(setattr, setnames, setcolorder, set, setDT, setDF)
 export(setorder, setorderv)
 export(setNumericRounding, getNumericRounding)
@@ -28,7 +28,7 @@ export(tstrsplit)
 export(frank)
 export(frankv)
 export(address)
-export(.SD,.N,.I,.GRP,.NGRP,.BY,.EACHI, measure, measurev, patterns)
+export(.SD,.N,.I,.GRP,.NGRP,.BY,.EACHI,.ROW, measure, measurev, patterns)
 # TODO(#6197): Export these.
 # export(., J)
 export(rleid)

R/data.table.R

@@ -11,7 +11,7 @@ methods::setPackageName("data.table",.global)
 #   (1) add to man/special-symbols.Rd
 #   (2) export() in NAMESPACE
 #   (3) add to vignettes/datatable-importing.Rmd#globals section
-.SD = .N = .I = .GRP = .NGRP = .BY = .EACHI = NULL
+.SD = .N = .I = .GRP = .NGRP = .BY = .EACHI = .ROW = NULL
 # These are exported to prevent NOTEs from R CMD check, and checkUsage via compiler.
 # But also exporting them makes it clear (to users and other packages) that data.table uses these as symbols.
 # And NULL makes it clear (to the R's mask check on loading) that they're variables not functions.
@@ -1559,6 +1559,19 @@ replace_dot_alias = function(e) {
           names(jsub)=""
           jsub[[1L]]=as.name("list")
         }
+
+        # Check for .ROW := NULL pattern (delete rows by reference)
+        if ((is.character(lhs) && length(lhs)==1L && lhs==".ROW") ||
+            (is.name(lhs) && identical(lhs, quote(.ROW)))) {
+          if (!is.null(jsub) && !identical(jsub, quote(NULL)))
+            stopf(".ROW can only be used with := NULL to delete rows")
+          if (is.null(irows))
+            stopf(".ROW := NULL requires i= condition to specify rows to delete")
+          if (!missingby)
+            stopf(".ROW := NULL does not support 'by' or 'keyby'. To delete rows using by grouping, first compute the row indices (e.g. rows = DT[, .I[cond], by=grp]$V1) and then delete them DT[rows, .ROW := NULL].")
+          .Call(CdeleteRows, x, irows)
+          return(suppPrint(x))
+        }
         av = all.vars(jsub,TRUE)
         if (!is.atomic(lhs)) stopf("LHS of := must be a symbol, or an atomic vector (column names or positions).")
         if (is.character(lhs)) {
@@ -2850,6 +2863,10 @@ setalloccol = alloc.col = function(DT, n=getOption("datatable.alloccol"), verbos
   ans
 }
 
+setallocrow = function(DT, n=-1L) {
+  invisible(.Call(Callocrowwrapper, DT, as.integer(n)))
+}
+
 selfrefok = function(DT,verbose=getOption("datatable.verbose")) {
   .Call(Cselfrefokwrapper,DT,verbose)
 }

inst/tests/tests.Rraw

@@ -21520,3 +21520,95 @@ test(2365.1, melt(df_melt, id.vars=1:2), melt(dt_melt, id.vars=1:2))
 df_dcast = data.frame(a = c("x", "y"), b = 1:2, v = 3:4)
 dt_dcast = data.table(a = c("x", "y"), b = 1:2, v = 3:4)
 test(2365.2, dcast(df_dcast, a ~ b, value.var = "v"), dcast(dt_dcast, a ~ b, value.var = "v"))
+
+# delete rows by reference #635
+# atomic types and list columns
+dt = data.table(
+  int = 1:5,
+  real = c(1.1, 2.2, 3.3, 4.4, 5.5),
+  char = letters[1:5],
+  lgl = c(TRUE, FALSE, TRUE, FALSE, TRUE),
+  cplx = as.complex(1:5),
+  raw_col = as.raw(1:5),
+  list_col = list(1L, 1:2, 1:3, 1:4, 1:5)
+)
+test(2366.01, copy(dt)[1L, .ROW := NULL], dt[-1])
+test(2366.02, copy(dt)[1, .ROW := NULL], dt[-1])
+test(2366.03, copy(dt)[c(TRUE, FALSE, FALSE, TRUE, FALSE), .ROW := NULL], dt[-c(1,4)])
+test(2366.04, copy(dt)[int==1L, .ROW := NULL], dt[-1])
+test(2366.05, copy(dt)[int<2L, .ROW := NULL], dt[-1])
+test(2366.06, copy(dt)[-1, .ROW := NULL], dt[1])
+# zero row or empty data.tables
+dt = data.table()
+test(2366.07, dt[logical(0), .ROW := NULL], dt)
+dt = data.table(a=integer(0), b=character(0))
+test(2366.08, dt[logical(0), .ROW := NULL], dt)
+# multirow
+dt = data.table(a=1:5, b=letters[1:5])
+test(2366.09, copy(dt)[c(1L, 3L), .ROW := NULL], dt[c(2,4,5)])
+test(2366.11, copy(dt)[1:2, .ROW := NULL], dt[3:5])
+test(2366.12, copy(dt)[1:5, .ROW := NULL], dt[0])
+# NA handling and edges case
+dt = data.table(a=1:5, b=letters[1:5])
+test(2366.13, copy(dt)[c(1L, NA_integer_, 3L), .ROW := NULL], dt[c(2,4,5)])
+test(2366.14, copy(dt)[c(NA_integer_, NA_integer_), .ROW := NULL], dt)
+test(2366.15, copy(dt)[c(TRUE, NA, FALSE, NA, TRUE), .ROW := NULL], dt[c(2,3,4)])
+test(2366.16, copy(dt)[integer(0), .ROW := NULL], dt)
+test(2366.17, copy(dt)[logical(0), .ROW := NULL], dt)
+test(2366.18, copy(dt)[c(FALSE, FALSE, FALSE, FALSE, FALSE), .ROW := NULL], dt)
+test(2366.19, copy(dt)[a > 100, .ROW := NULL], dt)  # no matches
+# Duplicate indices
+dt = data.table(a=1:5, b=letters[1:5])
+test(2366.20, copy(dt)[c(1L, 1L), .ROW := NULL], dt[-1])
+test(2366.21, copy(dt)[c(1L, 1L, 2L, 2L), .ROW := NULL], dt[3:5])
+test(2366.22, copy(dt)[c(3L, 1L, 3L, 1L), .ROW := NULL], dt[c(2,4,5)])
+# integer64
+if (test_bit64) {
+  dt = data.table(a=1:5, b=as.integer64(11:15))
+  test(2366.23, copy(dt)[c(1L, 3L), .ROW := NULL], dt[-c(1L,3L)])
+  test(2366.24, copy(dt)[1:5, .ROW := NULL], data.table(a=integer(0), b=integer64(0)))
+}
+# Date/IDate/ITime columns
+dt = data.table(a=1:5, d=as.Date("2024-01-01") + 0:4, t=as.ITime(paste0(10:14, ":00:00")), dt=as.POSIXct("2024-01-01 12:00:00") + 3600*0:4)
+test(2366.25, copy(dt)[c(1L, 3L), .ROW := NULL], dt[c(2,4,5)])
+test(2366.26, copy(dt)[c(2L, 4L), .ROW := NULL]$d, as.Date("2024-01-01") + c(0,2,4))
+# Factor columns
+dt = data.table(a=1:5, f=factor(letters[1:5], levels=letters[1:10]))
+test(2366.27, copy(dt)[c(1L, 3L), .ROW := NULL], dt[-c(1L,3L)])
+test(2366.28, levels(copy(dt)[c(1L, 3L), .ROW := NULL]$f), letters[1:10])
+dt = data.table(a=1:5, of=ordered(letters[1:5], levels=letters[5:1]))
+test(2366.29, copy(dt)[c(2L, 4L), .ROW := NULL], dt[-c(2L,4L)])
+test(2366.30, is.ordered(copy(dt)[c(2, 4L), .ROW := NULL]$of))
+# Keys - should be cleared after deletion
+dt = data.table(a=5:1, b=letters[1:5], key="a")
+test(2366.31, key(copy(dt)[1L, .ROW := NULL]), NULL)
+test(2366.32, haskey(copy(dt)[1L, .ROW := NULL]), FALSE)
+# Indices - should be cleared after deletion
+dt = data.table(a=1:5, b=letters[1:5], c=5:1)
+setindex(dt, b)
+test(2366.33, indices(copy(dt)[1L, .ROW := NULL]), NULL)
+# row names
+dt = data.table(a=1:5, b=letters[1:5])
+test(2366.34, attr(copy(dt)[c(1L, 3L), .ROW := NULL], "row.names"), 1:3)
+# selfref check
+test(2366.35, selfrefok(copy(dt)[1L, .ROW := NULL]), 1L)
+# errors
+dt = data.table(a=1:4, g=1:2)
+test(2366.36, dt[1L, .ROW := 1L], error=".ROW can only be used with := NULL")
+test(2366.37, dt[1L, .ROW := "delete"], error=".ROW can only be used with := NULL")
+test(2366.38, dt[1L, .ROW := FALSE], error=".ROW can only be used with := NULL")
+test(2366.39, dt[, .ROW := NULL], error=".ROW := NULL requires i= condition")
+test(2366.40, dt[1L, .ROW := NULL, by=g], error=".ROW := NULL does not support 'by' or 'keyby'")
+# large table
+dt = data.table(a=1:20000, b=rep(letters, length.out=20000))
+idx = seq(1L, 20000L, by=2L)
+test(2366.41, copy(dt)[idx, .ROW := NULL], dt[-idx])
+# Chaining and complexer i expressions
+dt = data.table(a=1:10, b=letters[1:10])
+test(2366.42, copy(dt)[a>2, .ROW := NULL][b=="a"], data.table(a=1L, b="a"))
+test(2366.43, copy(dt)[a %% 2 == 0, .ROW := NULL], dt[a %% 2 != 0])
+test(2366.44, copy(dt)[!(a < 5 & b != "d"), .ROW := NULL], dt[1:3])
+# make columns resizable
+dt = data.table(a=1:3)
+test(2366.91, truelength(dt$a), 0L)
+test(2366.92, {setallocrow(dt); truelength(dt$a)}, 3L)

man/assign.Rd

@@ -56,12 +56,16 @@ set(x, i = NULL, j, value)
     DT[i, colC := mean(colB), by = colA]          # update (or add) column called "colC" by reference by group. A major feature of `:=`.
     DT[,`:=`(new1 = sum(colB), new2 = sum(colC))] # Functional form
     DT[, let(new1 = sum(colB), new2 = sum(colC))] # New alias for functional form.
+    DT[i, .ROW := NULL]                           # delete rows by reference.
 }
 
 The \code{\link{.Last.updated}} variable contains the number of rows updated by the most recent \code{:=} or \code{set} calls, which may be useful, for example, in production settings for testing assumptions about the number of rows affected by a statement; see \code{\link{.Last.updated}} for details.
 
 Note that for efficiency no check is performed for duplicate assignments, i.e. if multiple values are passed for assignment to the same index, assignment to this index will occur repeatedly and sequentially; for a given use case, consider whether it makes sense to create your own test for duplicates, e.g. in production code.
 
+Note that \code{.ROW := NULL} is a special case used to delete rows by reference. Unlike column assignment, this requires an \code{i} expression to specify which rows to delete, and does not support \code{by} or \code{keyby}.
+To delete rows using a per-group condition, first compute the indices (\code{rows = DT[, .I[cond], by=grp]$V1}) then delete \code{DT[rows, .ROW := NULL]}. See \code{\link{.ROW}} or \code{\link{special-symbols}} for details.
+
 All of the following result in a friendly error (by design) :
 
 \preformatted{
@@ -158,6 +162,13 @@ set(DT, j = c("b", "d"), value = list(200L, 300L))
 ## Set values for multiple columns with multiple specified rows.
 set(DT, c(1L, 3L), c("b", "d"), value = list(500L, 800L))
 
+# Delete rows by reference
+DT = data.table(a=1:10, b=letters[1:10])
+DT[c(2,4,6), .ROW := NULL]      # delete rows 2, 4, and 6
+DT
+DT[a>5, .ROW := NULL]           # delete rows where a>5
+DT
+
 \dontrun{
 # Speed example:
 

man/special-symbols.Rd

@@ -9,10 +9,12 @@
 \alias{.EACHI}
 \alias{.NGRP}
 \alias{.NATURAL}
+\alias{.ROW}
 \title{ Special symbols }
 \description{
     \code{.SD}, \code{.BY}, \code{.N}, \code{.I}, \code{.GRP}, and \code{.NGRP} are \emph{read-only} symbols for use in \code{j}. \code{.N} can be used in \code{i} as well. \code{.I} can be used in \code{by} as well. See the vignettes, Details and Examples here and in \code{\link{data.table}}.
     \code{.EACHI} is a symbol passed to \code{by}; i.e. \code{by=.EACHI}, \code{.NATURAL} is a symbol passed to \code{on}; i.e. \code{on=.NATURAL}
+    \code{.ROW} is a symbol used with \code{:= NULL} to delete rows by reference. See also \code{\link{assign}}.
 }
 \details{
     The bindings of these variables are locked and attempting to assign to them will generate an error. If you wish to manipulate \code{.SD} before returning it, take a \code{\link{copy}(.SD)} first (see FAQ 4.5). Using \code{:=} in the \code{j} of \code{.SD} is reserved for future use as a (tortuously) flexible way to update \code{DT} by reference by group (even when groups are not contiguous in an ad hoc by).
@@ -32,6 +34,8 @@
 
     \code{.NATURAL} is defined as \code{NULL} but its value is not used. Its usage is \code{on=.NATURAL} (alternative of \code{X[on=Y]}) which joins two tables on their common column names, performing a natural join; see \code{\link{data.table}}'s \code{on} argument for more details.
 
+    \code{.ROW} is a symbol that can only be used with \code{:= NULL} to delete rows by reference. When you use \code{DT[i, .ROW := NULL]}, the rows matching the \code{i} expression are removed from \code{DT} in-place. This is an efficient way to delete rows without copying the entire data.table. The \code{i} argument is required and \code{by}/\code{keyby} are not supported. After deletion, any keys and indices on \code{DT} are cleared. See \code{\link{:=}} for more on reference semantics.
+
     Note that \code{.N} in \code{i} is computed up-front, while that in \code{j} applies \emph{after filtering in \code{i}}. That means that even absent grouping, \code{.N} in \code{i} can be different from \code{.N} in \code{j}. See Examples.
 
     Note also that you should consider these symbols read-only and of limited scope -- internal data.table code might manipulate them in unexpected ways, and as such their bindings are locked. There are subtle ways to wind up with the wrong object, especially when attempting to copy their values outside a grouping context. See examples; when in doubt, \code{copy()} is your friend.
@@ -72,5 +76,12 @@ DT[, .(min(.SD[,-1])), by=.I]
 # Do not expect this to correctly append the value of .BY in each group; copy(.BY) will work.
 by_tracker = list()
 DT[, { append(by_tracker, .BY); sum(v) }, by=x]
+
+# .ROW to delete rows by reference
+DT = data.table(a=1:5, b=letters[1:5])
+DT[c(2,4), .ROW := NULL]
+DT
+DT[a>2, .ROW := NULL]
+DT
 }
 \keyword{ data }

man/truelength.Rd

@@ -2,6 +2,7 @@
 \alias{truelength}
 \alias{setalloccol}
 \alias{alloc.col}
+\alias{setallocrow}
 \title{ Over-allocation access }
 \description{
     These functions are experimental and somewhat advanced. By \emph{experimental} we mean their names might change and perhaps the syntax, argument names and types. So if you write a lot of code using them, you have been warned! They should work and be stable, though, so please report problems with them. \code{alloc.col} is just an alias to \code{setalloccol}. We recommend to use \code{setalloccol} (though \code{alloc.col} will continue to be supported) because the \code{set*} prefix in \code{setalloccol} makes it clear that its input argument is modified in-place.
@@ -14,11 +15,14 @@ setalloccol(DT,
 alloc.col(DT,
     n = getOption("datatable.alloccol"),        # default: 1024L
     verbose = getOption("datatable.verbose"))   # default: FALSE
+setallocrow(DT, n = -1L)
 }
 \arguments{
 \item{x}{ Any type of vector, including \code{data.table} which is a \code{list} vector of column pointers. }
 \item{DT}{ A \code{data.table}. }
-\item{n}{ The number of spare column pointer slots to ensure are available. If \code{DT} is a 1,000 column \code{data.table} with 24 spare slots remaining, \code{n=1024L} means grow the 24 spare slots to be 1024. \code{truelength(DT)} will then be 2024 in this example. }
+\item{n}{ For \code{setalloccol} and \code{alloc.col}: the number of spare column pointer slots to ensure are available. If \code{DT} is a 1,000 column \code{data.table} with 24 spare slots remaining, \code{n=1024L} means grow the 24 spare slots to be 1024. \code{truelength(DT)} will then be 2024 in this example.
+
+  For \code{setallocrow}: the total number of rows to allocate. If \code{n >= 0}, allocates capacity for exactly \code{n} rows in total. If \code{n == -1} (default), shrinks columns to exact current size to free excess memory. }
 \item{verbose}{ Output status and information. }
 }
 \details{
@@ -34,6 +38,12 @@ alloc.col(DT,
     (perhaps in your .Rprofile); e.g., \code{options(datatable.alloccol=10000L)}.
 
     Please note: over-allocation of the column pointer vector is not for efficiency \emph{per se}; it is so that \code{:=} can add columns by reference without a shallow copy.
+
+    \code{setallocrow} is a utility function that prepares columns for fast row operations (delete or insert) by reference and manages row capacity. (Note that 'insert' by reference is not yet implemented)
+    Before deleting or inserting rows by reference, columns must be resizable.
+    \code{setallocrow} ensures all columns are in the appropriate state by converting ALTREP columns to materialized form and reallocating
+    columns to have the target capacity. When \code{n >= 0}, columns are allocated to hold exactly \code{n} rows in total, providing capacity for future row additions up to that count.
+    When \code{n == -1} (default), columns are shrunk to exact size to free unused memory. This operation modifies \code{DT} by reference.
 }
 \value{
     \code{truelength(x)} returns the length of the vector allocated in memory. \code{length(x)} of those items are in use. Currently, it is just the list vector of column
@@ -43,6 +53,8 @@ alloc.col(DT,
 
     \code{setalloccol} \emph{reallocates} \code{DT} by reference. This may be useful for efficiency if you know you are about to going to add a lot of columns in a loop.
     It also returns the new \code{DT}, for convenience in compound queries.
+
+    \code{setallocrow} modifies \code{DT} by reference to ensure all columns are resizable.
 }
 \seealso{ \code{\link{copy}} }
 \examples{

src/assign.c

@@ -592,7 +592,7 @@ SEXP assign(SEXP dt, SEXP rows, SEXP cols, SEXP newcolnames, SEXP values)
           Rprintf(_("RHS for item %d has been duplicated because MAYBE_REFERENCED==%d MAYBE_SHARED==%d ALTREP==%d, but then is being plonked. length(values)==%d; length(cols)==%d\n"),
                   i+1, MAYBE_REFERENCED(thisvalue), MAYBE_SHARED(thisvalue), ALTREP(thisvalue), length(values), length(cols));
         }
-        thisvalue = copyAsPlain(thisvalue);   // PROTECT not needed as assigned as element to protected list below.
+        thisvalue = copyAsPlain(thisvalue, -1);   // PROTECT not needed as assigned as element to protected list below.
       } else {
         if (verbose) Rprintf(_("Direct plonk of unnamed RHS, no copy. MAYBE_REFERENCED==%d, MAYBE_SHARED==%d\n"), MAYBE_REFERENCED(thisvalue), MAYBE_SHARED(thisvalue));  // e.g. DT[,a:=as.character(a)] as tested by 754.5
       }

src/coalesce.c

@@ -52,7 +52,7 @@ SEXP coalesce(SEXP x, SEXP inplaceArg, SEXP nan_is_na_arg) {
       error(_("Item %d is length %d but the first item is length %d. Only singletons are recycled."), i+2, length(item), nrow);
   }
   if (!inplace) {
-    first = PROTECT(copyAsPlain(first)); nprotect++;
+    first = PROTECT(copyAsPlain(first, -1)); nprotect++;
     if (verbose) Rprintf(_("coalesce copied first item (inplace=FALSE)\n"));
   }
   const void **valP = (const void **)R_alloc(nval, sizeof(*valP));

src/data.table.h

@@ -330,8 +330,12 @@ bool allNA(SEXP x, bool errorForBadType);
 SEXP colnamesInt(SEXP x, SEXP cols, SEXP check_dups, SEXP skip_absent);
 bool INHERITS(SEXP x, SEXP char_);
 void copyVectorElements(SEXP dst, SEXP src, R_xlen_t n, bool deep_copy, const char *caller);
-SEXP copyAsPlain(SEXP x);
+SEXP copyAsPlain(SEXP x, R_xlen_t overalloc);
+SEXP allocrow(SEXP dt, R_xlen_t n);
 void copySharedColumns(SEXP x);
+
+// deleterows.c
+SEXP deleteRows(SEXP dt, SEXP rows_to_delete);
 SEXP lock(SEXP x);
 SEXP unlock(SEXP x);
 bool islocked(SEXP x);
@@ -406,6 +410,7 @@ SEXP assign(SEXP, SEXP, SEXP, SEXP, SEXP);
 SEXP copy(SEXP);
 SEXP setdt_nrows(SEXP);
 SEXP alloccolwrapper(SEXP, SEXP, SEXP);
+SEXP allocrowwrapper(SEXP, SEXP);
 SEXP selfrefokwrapper(SEXP, SEXP);
 SEXP truelength(SEXP);
 SEXP setcharvec(SEXP, SEXP, SEXP);