如何连接(合并)数据框(内部,外部,左侧,右侧)

给定两个数据帧:

df1 = data.frame(CustomerId = c(1:6), Product = c(rep("Toaster", 3), rep("Radio", 3)))
df2 = data.frame(CustomerId = c(2, 4, 6), State = c(rep("Alabama", 2), rep("Ohio", 1)))

df1
#  CustomerId Product
#           1 Toaster
#           2 Toaster
#           3 Toaster
#           4   Radio
#           5   Radio
#           6   Radio

df2
#  CustomerId   State
#           2 Alabama
#           4 Alabama
#           6    Ohio

如何进行数据库样式(即sql 样式)的联接 ?也就是说,我如何获得:

  • df1df2内部 df2
    仅返回右表中左表具有匹配键的行。
  • df1df2外部 df2
    返回两个表中的所有行,左边的连接记录在右边的表中具有匹配的键。
  • df1df2 左外部联接(或简称为左 df2
    返回左侧表中的所有行,以及右侧表中具有匹配键的所有行。
  • df1df2 右外部连接
    返回右侧表中的所有行,以及左侧表中具有匹配键的所有行。

额外信用:

如何执行 SQL 样式选择语句?

答案

通过使用merge功能及其可选参数:

内部联接 merge(df1, df2)将适用于这些示例,因为 R 通过通用变量名称自动联接框架,但是您很可能希望指定merge(df1, df2, by = "CustomerId")以确保您仅在您想要的字段上匹配。如果匹配的变量在不同的数据帧中具有不同的名称,则也可以使用by.xby.y参数。

外部 merge(x = df1, y = df2, by = "CustomerId", all = TRUE) merge(x = df1, y = df2, by = "CustomerId", all = TRUE)

左外部: merge(x = df1, y = df2, by = "CustomerId", all.x = TRUE)

右外部: merge(x = df1, y = df2, by = "CustomerId", all.y = TRUE)

交叉 merge(x = df1, y = df2, by = NULL) merge(x = df1, y = df2, by = NULL)

与内部联接一样,您可能希望将 “CustomerId” 显式传递给 R 作为匹配变量。我认为几乎总是最好明确声明要合并的标识符。如果输入 data.frames 发生意外更改,则更安全,以后更易于阅读。

您可以通过给多个列合并by一个载体,例如, by = c("CustomerId", "OrderId")

如果要合并的列名称不同,则可以指定,例如, by.x = "CustomerId_in_df1", by.y = "CustomerId_in_df2" ,其中CustomerId_in_df1是第一个数据框中的列名称,而CustomerId_in_df2是第二个数据框中的列名。 (如果您需要在多列上合并,这些也可以是向量。)

我建议您检查一下Gabor Grothendieck 的 sqldf 软件包 ,该软件包可让您用 SQL 表示这些操作。

library(sqldf)

## inner join
df3 <- sqldf("SELECT CustomerId, Product, State 
              FROM df1
              JOIN df2 USING(CustomerID)")

## left join (substitute 'right' for right join)
df4 <- sqldf("SELECT CustomerId, Product, State 
              FROM df1
              LEFT JOIN df2 USING(CustomerID)")

我发现 SQL 语法比 R 语法更简单,更自然(但这可能只反映了我的 RDBMS 偏见)。

有关联接的更多信息,请参见Gabor 的 sqldf GitHub

有一个内部联接的 data.table方法,这种方法非常节省时间和内存(对于某些较大的 data.frames 是必需的):

library(data.table)

dt1 <- data.table(df1, key = "CustomerId") 
dt2 <- data.table(df2, key = "CustomerId")

joined.dt1.dt.2 <- dt1[dt2]

merge也可以在 data.tables 上工作(因为它是通用的,并调用merge.data.table

merge(dt1, dt2)

在 stackoverflow 上记录的 data.table:
如何执行 data.table 合并操作
将外键上的 SQL 连接转换为 R data.table 语法
合并较大数据的有效替代方法
如何在 R 中使用 data.table 进行基本的左外部联接?

另一个选择是在plyr软件包中找到的join函数

library(plyr)

join(df1, df2,
     type = "inner")

#   CustomerId Product   State
# 1          2 Toaster Alabama
# 2          4   Radio Alabama
# 3          6   Radio    Ohio

type选项: innerleftrightfull

From ?join :与merge不同,无论使用哪种连接类型,[ join ] 都会保留 x 的顺序。

您也可以使用 Hadley Wickham 的超棒dplyr软件包进行联接。

library(dplyr)

#make sure that CustomerId cols are both type numeric
#they ARE not using the provided code in question and dplyr will complain
df1$CustomerId <- as.numeric(df1$CustomerId)
df2$CustomerId <- as.numeric(df2$CustomerId)

突变联接:使用 df2 中的匹配项将列添加到 df1 中

#inner
inner_join(df1, df2)

#left outer
left_join(df1, df2)

#right outer
right_join(df1, df2)

#alternate right outer
left_join(df2, df1)

#full join
full_join(df1, df2)

过滤联接:过滤出 df1 中的行,请勿修改列

semi_join(df1, df2) #keep only observations in df1 that match in df2.
anti_join(df1, df2) #drops all observations in df1 that match in df2.

R Wiki 上有一些很好的例子。我会在这里偷几个:

合并方法

由于您的键名相同,因此进行内部联接的简短方法是 merge():

merge(df1,df2)

可以使用 “all” 关键字创建完整的内部联接(来自两个表的所有记录):

merge(df1,df2, all=TRUE)

df1 和 df2 的左外部联接:

merge(df1,df2, all.x=TRUE)

df1 和 df2 的右外部连接:

merge(df1,df2, all.y=TRUE)

您可以向下翻转 “em”,“拍击” em 和 “摩擦” em 来获得您询问的其他两个外部联接:)

下标方法

使用下标方法在左侧带有 df1 的外部左连接为:

df1[,"State"]<-df2[df1[ ,"Product"], "State"]

外部联接的其他组合可以通过杂凑左外部联接下标示例来创建。 (是的,我知道这相当于说 “我将其留给读者练习……”)

2014 年的新功能:

尤其是如果您还对一般的数据操作(包括排序,过滤,子集,汇总等)感兴趣,那么您绝对应该看一下dplyr ,它具有多种旨在简化数据工作的功能框架和某些其他数据库类型。它甚至提供了相当复杂的 SQL 接口,甚至提供了将(大多数)SQL 代码直接转换为 R 的功能。

dplyr 软件包中与联接有关的四个功能是(引用):

  • inner_join(x, y, by = NULL, copy = FALSE, ...) :从 x 中返回所有行,其中 y 中有匹配值,并且返回 x 和 y 中的所有列
  • left_join(x, y, by = NULL, copy = FALSE, ...) :返回 x 的所有行以及 x 和 y 的所有列
  • semi_join(x, y, by = NULL, copy = FALSE, ...) :返回 x 中的所有行,其中 y 中有匹配的值,仅保留 x 中的列。
  • anti_join(x, y, by = NULL, copy = FALSE, ...) :从 x 返回所有行,其中 y 中没有匹配值,仅保留 x 中的列

这一切都在这里非常详细。

可以通过select(df,"column")完成select(df,"column") 。如果那还不足以满足您的 SQL 需求,那么您可以使用sql()函数,您可以按原样输入 SQL 代码,并且它将执行您指定的操作,就像您一直在 R 中编写代码一样(有关更多信息, ,请参阅dplyr / 数据库插图 )。例如,如果正确应用,则sql("SELECT * FROM hflights")将选择 “hflights” dplyr 表(“tbl”)中的所有列。

更新 data.table 方法以连接数据集。有关每种连接类型,请参见以下示例。有两种方法,一种是从[.data.table传递第二个 data.table 作为第一个参数传递给子集时,另一种方法是使用merge函数,该函数分派给 fast data.table 方法。

df1 = data.frame(CustomerId = c(1:6), Product = c(rep("Toaster", 3), rep("Radio", 3)))
df2 = data.frame(CustomerId = c(2L, 4L, 7L), State = c(rep("Alabama", 2), rep("Ohio", 1))) # one value changed to show full outer join

library(data.table)

dt1 = as.data.table(df1)
dt2 = as.data.table(df2)
setkey(dt1, CustomerId)
setkey(dt2, CustomerId)
# right outer join keyed data.tables
dt1[dt2]

setkey(dt1, NULL)
setkey(dt2, NULL)
# right outer join unkeyed data.tables - use `on` argument
dt1[dt2, on = "CustomerId"]

# left outer join - swap dt1 with dt2
dt2[dt1, on = "CustomerId"]

# inner join - use `nomatch` argument
dt1[dt2, nomatch=NULL, on = "CustomerId"]

# anti join - use `!` operator
dt1[!dt2, on = "CustomerId"]

# inner join - using merge method
merge(dt1, dt2, by = "CustomerId")

# full outer join
merge(dt1, dt2, by = "CustomerId", all = TRUE)

# see ?merge.data.table arguments for other cases

在基准测试以下,基于 R,sqldf,dplyr 和 data.table。
基准测试未加密 / 未索引的数据集。基准测试是对 50M-1 行数据集执行的,联接列上有 50M-2 个通用值,因此可以测试每种情况(内部,左,右,满),联接仍然不容易执行。这种连接类型很好地强调了连接算法。时间从sqldf:0.4.11dplyr:0.7.8 data.table:1.12.0data.table:1.12.0

# inner
Unit: seconds
   expr       min        lq      mean    median        uq       max neval
   base 111.66266 111.66266 111.66266 111.66266 111.66266 111.66266     1
  sqldf 624.88388 624.88388 624.88388 624.88388 624.88388 624.88388     1
  dplyr  51.91233  51.91233  51.91233  51.91233  51.91233  51.91233     1
     DT  10.40552  10.40552  10.40552  10.40552  10.40552  10.40552     1
# left
Unit: seconds
   expr        min         lq       mean     median         uq        max 
   base 142.782030 142.782030 142.782030 142.782030 142.782030 142.782030     
  sqldf 613.917109 613.917109 613.917109 613.917109 613.917109 613.917109     
  dplyr  49.711912  49.711912  49.711912  49.711912  49.711912  49.711912     
     DT   9.674348   9.674348   9.674348   9.674348   9.674348   9.674348       
# right
Unit: seconds
   expr        min         lq       mean     median         uq        max
   base 122.366301 122.366301 122.366301 122.366301 122.366301 122.366301     
  sqldf 611.119157 611.119157 611.119157 611.119157 611.119157 611.119157     
  dplyr  50.384841  50.384841  50.384841  50.384841  50.384841  50.384841     
     DT   9.899145   9.899145   9.899145   9.899145   9.899145   9.899145     
# full
Unit: seconds
  expr       min        lq      mean    median        uq       max neval
  base 141.79464 141.79464 141.79464 141.79464 141.79464 141.79464     1
 dplyr  94.66436  94.66436  94.66436  94.66436  94.66436  94.66436     1
    DT  21.62573  21.62573  21.62573  21.62573  21.62573  21.62573     1

请注意,您还可以使用data.table执行其他类型的联接:
- 更新连接 - 如果您想从另一个表到主表中查找值
- 聚集联接 - 如果要聚集正在联接的键,则不必实现所有联接结果
- 重叠连接 - 如果要按范围合并
- 滚动连接 - 如果您希望合并能够通过向前或向后滚动它们来匹配行之前 / 之后的值
- 平等加入 - 如果您的加入条件不相等

复制代码:

library(microbenchmark)
library(sqldf)
library(dplyr)
library(data.table)
sapply(c("sqldf","dplyr","data.table"), packageVersion, simplify=FALSE)

n = 5e7
set.seed(108)
df1 = data.frame(x=sample(n,n-1L), y1=rnorm(n-1L))
df2 = data.frame(x=sample(n,n-1L), y2=rnorm(n-1L))
dt1 = as.data.table(df1)
dt2 = as.data.table(df2)

mb = list()
# inner join
microbenchmark(times = 1L,
               base = merge(df1, df2, by = "x"),
               sqldf = sqldf("SELECT * FROM df1 INNER JOIN df2 ON df1.x = df2.x"),
               dplyr = inner_join(df1, df2, by = "x"),
               DT = dt1[dt2, nomatch=NULL, on = "x"]) -> mb$inner

# left outer join
microbenchmark(times = 1L,
               base = merge(df1, df2, by = "x", all.x = TRUE),
               sqldf = sqldf("SELECT * FROM df1 LEFT OUTER JOIN df2 ON df1.x = df2.x"),
               dplyr = left_join(df1, df2, by = c("x"="x")),
               DT = dt2[dt1, on = "x"]) -> mb$left

# right outer join
microbenchmark(times = 1L,
               base = merge(df1, df2, by = "x", all.y = TRUE),
               sqldf = sqldf("SELECT * FROM df2 LEFT OUTER JOIN df1 ON df2.x = df1.x"),
               dplyr = right_join(df1, df2, by = "x"),
               DT = dt1[dt2, on = "x"]) -> mb$right

# full outer join
microbenchmark(times = 1L,
               base = merge(df1, df2, by = "x", all = TRUE),
               dplyr = full_join(df1, df2, by = "x"),
               DT = merge(dt1, dt2, by = "x", all = TRUE)) -> mb$full

lapply(mb, print) -> nul

dplyr 自从 0.4 实现了所有这些连接,包括outer_join ,但是值得注意的是,在 0.4 之前的前几个版本中,它以前不提供outer_join ,因此存在很多非常糟糕的变通办法,用户代码在相当大的范围内徘徊。过了一会儿(您仍然可以在那个时期在 SO,Kaggle 的答案和 github 中找到这样的代码。因此,此答案仍然有用。)

与联接有关的发行要点

v0.5 (6/2016)

  • 处理 POSIXct 类型,时区,重复项,不同因子级别。更好的错误和警告。
  • 新的后缀参数可控制接收哪些后缀重复的变量名(#1296)

v0.4.0 (1/2015)

  • 实施右联接和外联接 (#96)
  • 突变联接,将新变量从另一个匹配的行添加到一个表。过滤联接,根据联接是否与另一表中的观测值匹配来过滤其中一个表中的观测值。

v0.3 (10/2014)

  • 现在可以通过每个表中的不同变量进行 left_join:df1%>%left_join(df2,c(“var1” =“var2”))

v0.2 (5/2014)

  • * _join()不再对列名重新排序(#324)

v0.1.3 (4/2014)

哈德利在该问题中的评论的变通办法:

  • right_join (x,y)在行方面与 left_join(y,x)相同,只是列的顺序不同。使用 select(new_column_order)轻松解决
  • external_join基本上是 union(left_join(x,y),right_join(x,y))- 即保留两个数据帧中的所有行。

在连接两个数据帧时,我分别惊奇地发现merge(..., all.x = TRUE, all.y = TRUE) ,每个数据帧各merge(..., all.x = TRUE, all.y = TRUE) 100 万行,一个有 2 列,而另一个有merge(..., all.x = TRUE, all.y = TRUE) dplyr::full_join() 。这是 dplyr v0.4

合并大约需要 17 秒,full_join 需要大约 65 秒。

不过有些吃,因为我通常默认使用 dplyr 进行操作。

对于左连接的基数为0..*:0..1或右连接的基数为0..1:0..*情况,可以从连接器中就地分配单边列( 0..1表)直接添加到连接对象( 0..*表)上,从而避免创建全新的数据表。这需要将参与者的键列匹配到参与者中,并相应地索引并排序参与者的行以进行分配。

如果键是单个列,则可以使用单个调用match()进行匹配。我将在此答案中介绍这种情况。

这是一个基于 OP 的示例,除了我向df2添加了额外的一行,其 ID 为 7,以测试连接器中不匹配键的情况。这实际上是df1df2

df1 <- data.frame(CustomerId=1:6,Product=c(rep('Toaster',3L),rep('Radio',3L)));
df2 <- data.frame(CustomerId=c(2L,4L,6L,7L),State=c(rep('Alabama',2L),'Ohio','Texas'));
df1[names(df2)[-1L]] <- df2[match(df1[,1L],df2[,1L]),-1L];
df1;
##   CustomerId Product   State
## 1          1 Toaster    <NA>
## 2          2 Toaster Alabama
## 3          3 Toaster    <NA>
## 4          4   Radio Alabama
## 5          5   Radio    <NA>
## 6          6   Radio    Ohio

在上面,我对键列是两个输入表的第一列的假设进行了硬编码。我认为,通常来说,这不是一个不合理的假设,因为如果您有一个带有键列的 data.frame,那么如果没有将它设置为 data.frame 的第一列,那将很奇怪。一开始。而且,您可以随时对列进行重新排序。这种假设的一个有益结果是,尽管我认为只是将一个假设替换为另一个假设,但是不必对键列的名称进行硬编码。简洁是整数索引以及速度的另一个优点。在下面的基准测试中,我将更改实现以使用字符串名称索引来匹配竞争的实现。

我认为,如果要对单个大表保留多个表,这是一个特别合适的解决方案。对于每个合并重复地重建整个表将是不必要且效率低下的。

另一方面,如果出于任何原因需要参加者通过此操作保持不变,则不能使用此解决方案,因为它直接修改了参加者。尽管在那种情况下,您可以简单地制作副本并在副本上执行就地分配。


附带说明一下,我简要研究了多列键的可能匹配解决方案。不幸的是,我找到的唯一匹配的解决方案是:

  • 无效的串联。例如match(interaction(df1$a,df1$b),interaction(df2$a,df2$b))或与paste()相同的想法。
  • 低效的笛卡尔连词,例如outer(df1$a,df2$a,`==`) & outer(df1$b,df2$b,`==`)
  • base R merge()和等效的基于包的合并函数,它们始终分配一个新表以返回合并的结果,因此不适合基于就地分配的解决方案。

例如,请参阅在不同数据帧上匹配多个列并获得其他列作为结果将两个列与另外两个列进行 匹配,在多个列上进行匹配 ,以及我最初提出就地解决方案Combine 的问题的重复R 中具有不同行数的两个数据帧


标杆管理

我决定做自己的基准测试,以了解就地分配方法与该问题中提供的其他解决方案的比较。

测试代码:

library(microbenchmark);
library(data.table);
library(sqldf);
library(plyr);
library(dplyr);

solSpecs <- list(
    merge=list(testFuncs=list(
        inner=function(df1,df2,key) merge(df1,df2,key),
        left =function(df1,df2,key) merge(df1,df2,key,all.x=T),
        right=function(df1,df2,key) merge(df1,df2,key,all.y=T),
        full =function(df1,df2,key) merge(df1,df2,key,all=T)
    )),
    data.table.unkeyed=list(argSpec='data.table.unkeyed',testFuncs=list(
        inner=function(dt1,dt2,key) dt1[dt2,on=key,nomatch=0L,allow.cartesian=T],
        left =function(dt1,dt2,key) dt2[dt1,on=key,allow.cartesian=T],
        right=function(dt1,dt2,key) dt1[dt2,on=key,allow.cartesian=T],
        full =function(dt1,dt2,key) merge(dt1,dt2,key,all=T,allow.cartesian=T) ## calls merge.data.table()
    )),
    data.table.keyed=list(argSpec='data.table.keyed',testFuncs=list(
        inner=function(dt1,dt2) dt1[dt2,nomatch=0L,allow.cartesian=T],
        left =function(dt1,dt2) dt2[dt1,allow.cartesian=T],
        right=function(dt1,dt2) dt1[dt2,allow.cartesian=T],
        full =function(dt1,dt2) merge(dt1,dt2,all=T,allow.cartesian=T) ## calls merge.data.table()
    )),
    sqldf.unindexed=list(testFuncs=list( ## note: must pass connection=NULL to avoid running against the live DB connection, which would result in collisions with the residual tables from the last query upload
        inner=function(df1,df2,key) sqldf(paste0('select * from df1 inner join df2 using(',paste(collapse=',',key),')'),connection=NULL),
        left =function(df1,df2,key) sqldf(paste0('select * from df1 left join df2 using(',paste(collapse=',',key),')'),connection=NULL),
        right=function(df1,df2,key) sqldf(paste0('select * from df2 left join df1 using(',paste(collapse=',',key),')'),connection=NULL) ## can't do right join proper, not yet supported; inverted left join is equivalent
        ##full =function(df1,df2,key) sqldf(paste0('select * from df1 full join df2 using(',paste(collapse=',',key),')'),connection=NULL) ## can't do full join proper, not yet supported; possible to hack it with a union of left joins, but too unreasonable to include in testing
    )),
    sqldf.indexed=list(testFuncs=list( ## important: requires an active DB connection with preindexed main.df1 and main.df2 ready to go; arguments are actually ignored
        inner=function(df1,df2,key) sqldf(paste0('select * from main.df1 inner join main.df2 using(',paste(collapse=',',key),')')),
        left =function(df1,df2,key) sqldf(paste0('select * from main.df1 left join main.df2 using(',paste(collapse=',',key),')')),
        right=function(df1,df2,key) sqldf(paste0('select * from main.df2 left join main.df1 using(',paste(collapse=',',key),')')) ## can't do right join proper, not yet supported; inverted left join is equivalent
        ##full =function(df1,df2,key) sqldf(paste0('select * from main.df1 full join main.df2 using(',paste(collapse=',',key),')')) ## can't do full join proper, not yet supported; possible to hack it with a union of left joins, but too unreasonable to include in testing
    )),
    plyr=list(testFuncs=list(
        inner=function(df1,df2,key) join(df1,df2,key,'inner'),
        left =function(df1,df2,key) join(df1,df2,key,'left'),
        right=function(df1,df2,key) join(df1,df2,key,'right'),
        full =function(df1,df2,key) join(df1,df2,key,'full')
    )),
    dplyr=list(testFuncs=list(
        inner=function(df1,df2,key) inner_join(df1,df2,key),
        left =function(df1,df2,key) left_join(df1,df2,key),
        right=function(df1,df2,key) right_join(df1,df2,key),
        full =function(df1,df2,key) full_join(df1,df2,key)
    )),
    in.place=list(testFuncs=list(
        left =function(df1,df2,key) { cns <- setdiff(names(df2),key); df1[cns] <- df2[match(df1[,key],df2[,key]),cns]; df1; },
        right=function(df1,df2,key) { cns <- setdiff(names(df1),key); df2[cns] <- df1[match(df2[,key],df1[,key]),cns]; df2; }
    ))
);

getSolTypes <- function() names(solSpecs);
getJoinTypes <- function() unique(unlist(lapply(solSpecs,function(x) names(x$testFuncs))));
getArgSpec <- function(argSpecs,key=NULL) if (is.null(key)) argSpecs$default else argSpecs[[key]];

initSqldf <- function() {
    sqldf(); ## creates sqlite connection on first run, cleans up and closes existing connection otherwise
    if (exists('sqldfInitFlag',envir=globalenv(),inherits=F) && sqldfInitFlag) { ## false only on first run
        sqldf(); ## creates a new connection
    } else {
        assign('sqldfInitFlag',T,envir=globalenv()); ## set to true for the one and only time
    }; ## end if
    invisible();
}; ## end initSqldf()

setUpBenchmarkCall <- function(argSpecs,joinType,solTypes=getSolTypes(),env=parent.frame()) {
    ## builds and returns a list of expressions suitable for passing to the list argument of microbenchmark(), and assigns variables to resolve symbol references in those expressions
    callExpressions <- list();
    nms <- character();
    for (solType in solTypes) {
        testFunc <- solSpecs[[solType]]$testFuncs[[joinType]];
        if (is.null(testFunc)) next; ## this join type is not defined for this solution type
        testFuncName <- paste0('tf.',solType);
        assign(testFuncName,testFunc,envir=env);
        argSpecKey <- solSpecs[[solType]]$argSpec;
        argSpec <- getArgSpec(argSpecs,argSpecKey);
        argList <- setNames(nm=names(argSpec$args),vector('list',length(argSpec$args)));
        for (i in seq_along(argSpec$args)) {
            argName <- paste0('tfa.',argSpecKey,i);
            assign(argName,argSpec$args[[i]],envir=env);
            argList[[i]] <- if (i%in%argSpec$copySpec) call('copy',as.symbol(argName)) else as.symbol(argName);
        }; ## end for
        callExpressions[[length(callExpressions)+1L]] <- do.call(call,c(list(testFuncName),argList),quote=T);
        nms[length(nms)+1L] <- solType;
    }; ## end for
    names(callExpressions) <- nms;
    callExpressions;
}; ## end setUpBenchmarkCall()

harmonize <- function(res) {
    res <- as.data.frame(res); ## coerce to data.frame
    for (ci in which(sapply(res,is.factor))) res[[ci]] <- as.character(res[[ci]]); ## coerce factor columns to character
    for (ci in which(sapply(res,is.logical))) res[[ci]] <- as.integer(res[[ci]]); ## coerce logical columns to integer (works around sqldf quirk of munging logicals to integers)
    ##for (ci in which(sapply(res,inherits,'POSIXct'))) res[[ci]] <- as.double(res[[ci]]); ## coerce POSIXct columns to double (works around sqldf quirk of losing POSIXct class) ----- POSIXct doesn't work at all in sqldf.indexed
    res <- res[order(names(res))]; ## order columns
    res <- res[do.call(order,res),]; ## order rows
    res;
}; ## end harmonize()

checkIdentical <- function(argSpecs,solTypes=getSolTypes()) {
    for (joinType in getJoinTypes()) {
        callExpressions <- setUpBenchmarkCall(argSpecs,joinType,solTypes);
        if (length(callExpressions)<2L) next;
        ex <- harmonize(eval(callExpressions[[1L]]));
        for (i in seq(2L,len=length(callExpressions)-1L)) {
            y <- harmonize(eval(callExpressions[[i]]));
            if (!isTRUE(all.equal(ex,y,check.attributes=F))) {
                ex <<- ex;
                y <<- y;
                solType <- names(callExpressions)[i];
                stop(paste0('non-identical: ',solType,' ',joinType,'.'));
            }; ## end if
        }; ## end for
    }; ## end for
    invisible();
}; ## end checkIdentical()

testJoinType <- function(argSpecs,joinType,solTypes=getSolTypes(),metric=NULL,times=100L) {
    callExpressions <- setUpBenchmarkCall(argSpecs,joinType,solTypes);
    bm <- microbenchmark(list=callExpressions,times=times);
    if (is.null(metric)) return(bm);
    bm <- summary(bm);
    res <- setNames(nm=names(callExpressions),bm[[metric]]);
    attr(res,'unit') <- attr(bm,'unit');
    res;
}; ## end testJoinType()

testAllJoinTypes <- function(argSpecs,solTypes=getSolTypes(),metric=NULL,times=100L) {
    joinTypes <- getJoinTypes();
    resList <- setNames(nm=joinTypes,lapply(joinTypes,function(joinType) testJoinType(argSpecs,joinType,solTypes,metric,times)));
    if (is.null(metric)) return(resList);
    units <- unname(unlist(lapply(resList,attr,'unit')));
    res <- do.call(data.frame,c(list(join=joinTypes),setNames(nm=solTypes,rep(list(rep(NA_real_,length(joinTypes))),length(solTypes))),list(unit=units,stringsAsFactors=F)));
    for (i in seq_along(resList)) res[i,match(names(resList[[i]]),names(res))] <- resList[[i]];
    res;
}; ## end testAllJoinTypes()

testGrid <- function(makeArgSpecsFunc,sizes,overlaps,solTypes=getSolTypes(),joinTypes=getJoinTypes(),metric='median',times=100L) {

    res <- expand.grid(size=sizes,overlap=overlaps,joinType=joinTypes,stringsAsFactors=F);
    res[solTypes] <- NA_real_;
    res$unit <- NA_character_;
    for (ri in seq_len(nrow(res))) {

        size <- res$size[ri];
        overlap <- res$overlap[ri];
        joinType <- res$joinType[ri];

        argSpecs <- makeArgSpecsFunc(size,overlap);

        checkIdentical(argSpecs,solTypes);

        cur <- testJoinType(argSpecs,joinType,solTypes,metric,times);
        res[ri,match(names(cur),names(res))] <- cur;
        res$unit[ri] <- attr(cur,'unit');

    }; ## end for

    res;

}; ## end testGrid()

这是我之前演示的基于 OP 的示例的基准:

## OP's example, supplemented with a non-matching row in df2
argSpecs <- list(
    default=list(copySpec=1:2,args=list(
        df1 <- data.frame(CustomerId=1:6,Product=c(rep('Toaster',3L),rep('Radio',3L))),
        df2 <- data.frame(CustomerId=c(2L,4L,6L,7L),State=c(rep('Alabama',2L),'Ohio','Texas')),
        'CustomerId'
    )),
    data.table.unkeyed=list(copySpec=1:2,args=list(
        as.data.table(df1),
        as.data.table(df2),
        'CustomerId'
    )),
    data.table.keyed=list(copySpec=1:2,args=list(
        setkey(as.data.table(df1),CustomerId),
        setkey(as.data.table(df2),CustomerId)
    ))
);
## prepare sqldf
initSqldf();
sqldf('create index df1_key on df1(CustomerId);'); ## upload and create an sqlite index on df1
sqldf('create index df2_key on df2(CustomerId);'); ## upload and create an sqlite index on df2

checkIdentical(argSpecs);

testAllJoinTypes(argSpecs,metric='median');
##    join    merge data.table.unkeyed data.table.keyed sqldf.unindexed sqldf.indexed      plyr    dplyr in.place         unit
## 1 inner  644.259           861.9345          923.516        9157.752      1580.390  959.2250 270.9190       NA microseconds
## 2  left  713.539           888.0205          910.045        8820.334      1529.714  968.4195 270.9185 224.3045 microseconds
## 3 right 1221.804           909.1900          923.944        8930.668      1533.135 1063.7860 269.8495 218.1035 microseconds
## 4  full 1302.203          3107.5380         3184.729              NA            NA 1593.6475 270.7055       NA microseconds

在这里,我以随机输入数据为基准,尝试在两个输入表之间使用不同的比例和不同的键重叠模式。此基准仍然仅限于单列整数键的情况。同样,为了确保就地解决方案适用于同一张表的左右联接,所有随机测试数据都使用0..1:0..1基数。这是通过在生成第二个 data.frame 的关键列时进行采样而不替换第一个 data.frame 的关键列来实现的。

makeArgSpecs.singleIntegerKey.optionalOneToOne <- function(size,overlap) {

    com <- as.integer(size*overlap);

    argSpecs <- list(
        default=list(copySpec=1:2,args=list(
            df1 <- data.frame(id=sample(size),y1=rnorm(size),y2=rnorm(size)),
            df2 <- data.frame(id=sample(c(if (com>0L) sample(df1$id,com) else integer(),seq(size+1L,len=size-com))),y3=rnorm(size),y4=rnorm(size)),
            'id'
        )),
        data.table.unkeyed=list(copySpec=1:2,args=list(
            as.data.table(df1),
            as.data.table(df2),
            'id'
        )),
        data.table.keyed=list(copySpec=1:2,args=list(
            setkey(as.data.table(df1),id),
            setkey(as.data.table(df2),id)
        ))
    );
    ## prepare sqldf
    initSqldf();
    sqldf('create index df1_key on df1(id);'); ## upload and create an sqlite index on df1
    sqldf('create index df2_key on df2(id);'); ## upload and create an sqlite index on df2

    argSpecs;

}; ## end makeArgSpecs.singleIntegerKey.optionalOneToOne()

## cross of various input sizes and key overlaps
sizes <- c(1e1L,1e3L,1e6L);
overlaps <- c(0.99,0.5,0.01);
system.time({ res <- testGrid(makeArgSpecs.singleIntegerKey.optionalOneToOne,sizes,overlaps); });
##     user   system  elapsed
## 22024.65 12308.63 34493.19

我编写了一些代码来创建上述结果的对数 - 对数图。我为每个重叠百分比生成了一个单独的图。有点混乱,但是我喜欢在同一图中表示所有解决方案类型和联接类型。

我使用样条曲线插值法显示了每个解决方案 / 连接类型组合的平滑曲线,并使用单独的 pch 符号绘制。连接类型由 pch 符号捕获,内部圆点,左侧和右侧尖括号的点分别代表左和右,菱形代表完整。解决方案类型由图例中所示的颜色捕获。

plotRes <- function(res,titleFunc,useFloor=F) {
    solTypes <- setdiff(names(res),c('size','overlap','joinType','unit')); ## derive from res
    normMult <- c(microseconds=1e-3,milliseconds=1); ## normalize to milliseconds
    joinTypes <- getJoinTypes();
    cols <- c(merge='purple',data.table.unkeyed='blue',data.table.keyed='#00DDDD',sqldf.unindexed='brown',sqldf.indexed='orange',plyr='red',dplyr='#00BB00',in.place='magenta');
    pchs <- list(inner=20L,left='<',right='>',full=23L);
    cexs <- c(inner=0.7,left=1,right=1,full=0.7);
    NP <- 60L;
    ord <- order(decreasing=T,colMeans(res[res$size==max(res$size),solTypes],na.rm=T));
    ymajors <- data.frame(y=c(1,1e3),label=c('1ms','1s'),stringsAsFactors=F);
    for (overlap in unique(res$overlap)) {
        x1 <- res[res$overlap==overlap,];
        x1[solTypes] <- x1[solTypes]*normMult[x1$unit]; x1$unit <- NULL;
        xlim <- c(1e1,max(x1$size));
        xticks <- 10^seq(log10(xlim[1L]),log10(xlim[2L]));
        ylim <- c(1e-1,10^((if (useFloor) floor else ceiling)(log10(max(x1[solTypes],na.rm=T))))); ## use floor() to zoom in a little more, only sqldf.unindexed will break above, but xpd=NA will keep it visible
        yticks <- 10^seq(log10(ylim[1L]),log10(ylim[2L]));
        yticks.minor <- rep(yticks[-length(yticks)],each=9L)*1:9;
        plot(NA,xlim=xlim,ylim=ylim,xaxs='i',yaxs='i',axes=F,xlab='size (rows)',ylab='time (ms)',log='xy');
        abline(v=xticks,col='lightgrey');
        abline(h=yticks.minor,col='lightgrey',lty=3L);
        abline(h=yticks,col='lightgrey');
        axis(1L,xticks,parse(text=sprintf('10^%d',as.integer(log10(xticks)))));
        axis(2L,yticks,parse(text=sprintf('10^%d',as.integer(log10(yticks)))),las=1L);
        axis(4L,ymajors$y,ymajors$label,las=1L,tick=F,cex.axis=0.7,hadj=0.5);
        for (joinType in rev(joinTypes)) { ## reverse to draw full first, since it's larger and would be more obtrusive if drawn last
            x2 <- x1[x1$joinType==joinType,];
            for (solType in solTypes) {
                if (any(!is.na(x2[[solType]]))) {
                    xy <- spline(x2$size,x2[[solType]],xout=10^(seq(log10(x2$size[1L]),log10(x2$size[nrow(x2)]),len=NP)));
                    points(xy$x,xy$y,pch=pchs[[joinType]],col=cols[solType],cex=cexs[joinType],xpd=NA);
                }; ## end if
            }; ## end for
        }; ## end for
        ## custom legend
        ## due to logarithmic skew, must do all distance calcs in inches, and convert to user coords afterward
        ## the bottom-left corner of the legend will be defined in normalized figure coords, although we can convert to inches immediately
        leg.cex <- 0.7;
        leg.x.in <- grconvertX(0.275,'nfc','in');
        leg.y.in <- grconvertY(0.6,'nfc','in');
        leg.x.user <- grconvertX(leg.x.in,'in');
        leg.y.user <- grconvertY(leg.y.in,'in');
        leg.outpad.w.in <- 0.1;
        leg.outpad.h.in <- 0.1;
        leg.midpad.w.in <- 0.1;
        leg.midpad.h.in <- 0.1;
        leg.sol.w.in <- max(strwidth(solTypes,'in',leg.cex));
        leg.sol.h.in <- max(strheight(solTypes,'in',leg.cex))*1.5; ## multiplication factor for greater line height
        leg.join.w.in <- max(strheight(joinTypes,'in',leg.cex))*1.5; ## ditto
        leg.join.h.in <- max(strwidth(joinTypes,'in',leg.cex));
        leg.main.w.in <- leg.join.w.in*length(joinTypes);
        leg.main.h.in <- leg.sol.h.in*length(solTypes);
        leg.x2.user <- grconvertX(leg.x.in+leg.outpad.w.in*2+leg.main.w.in+leg.midpad.w.in+leg.sol.w.in,'in');
        leg.y2.user <- grconvertY(leg.y.in+leg.outpad.h.in*2+leg.main.h.in+leg.midpad.h.in+leg.join.h.in,'in');
        leg.cols.x.user <- grconvertX(leg.x.in+leg.outpad.w.in+leg.join.w.in*(0.5+seq(0L,length(joinTypes)-1L)),'in');
        leg.lines.y.user <- grconvertY(leg.y.in+leg.outpad.h.in+leg.main.h.in-leg.sol.h.in*(0.5+seq(0L,length(solTypes)-1L)),'in');
        leg.sol.x.user <- grconvertX(leg.x.in+leg.outpad.w.in+leg.main.w.in+leg.midpad.w.in,'in');
        leg.join.y.user <- grconvertY(leg.y.in+leg.outpad.h.in+leg.main.h.in+leg.midpad.h.in,'in');
        rect(leg.x.user,leg.y.user,leg.x2.user,leg.y2.user,col='white');
        text(leg.sol.x.user,leg.lines.y.user,solTypes[ord],cex=leg.cex,pos=4L,offset=0);
        text(leg.cols.x.user,leg.join.y.user,joinTypes,cex=leg.cex,pos=4L,offset=0,srt=90); ## srt rotation applies *after* pos/offset positioning
        for (i in seq_along(joinTypes)) {
            joinType <- joinTypes[i];
            points(rep(leg.cols.x.user[i],length(solTypes)),ifelse(colSums(!is.na(x1[x1$joinType==joinType,solTypes[ord]]))==0L,NA,leg.lines.y.user),pch=pchs[[joinType]],col=cols[solTypes[ord]]);
        }; ## end for
        title(titleFunc(overlap));
        readline(sprintf('overlap %.02f',overlap));
    }; ## end for
}; ## end plotRes()

titleFunc <- function(overlap) sprintf('R merge solutions: single-column integer key, 0..1:0..1 cardinality, %d%% overlap',as.integer(overlap*100));
plotRes(res,titleFunc,T);

R合并基准单个列整数键,可选一对一99

R合并基准单个列整数键,可选一对一50

R合并基准单个列整数键,可选一对一


关于键列的数量和类型以及基数,这是第二个更大规模的基准测试。对于此基准测试,我使用三个关键列:一个字符,一个整数和一个逻辑,对基数没有限制(即0..*:0..* )。 (通常,由于浮点比较的复杂性,不建议使用双精度或复数值来定义键列,并且基本上没有人使用原始类型,键列的使用量少得多,因此我没有在键中包含这些类型另外,为了提供信息,我最初尝试通过包含 POSIXct 键列来使用四个键列,但是 POSIXct 类型由于某种原因在sqldf.indexed解决方案中不能很好地发挥作用,这可能是由于浮点比较异常,所以我将其删除。)

makeArgSpecs.assortedKey.optionalManyToMany <- function(size,overlap,uniquePct=75) {

    ## number of unique keys in df1
    u1Size <- as.integer(size*uniquePct/100);

    ## (roughly) divide u1Size into bases, so we can use expand.grid() to produce the required number of unique key values with repetitions within individual key columns
    ## use ceiling() to ensure we cover u1Size; will truncate afterward
    u1SizePerKeyColumn <- as.integer(ceiling(u1Size^(1/3)));

    ## generate the unique key values for df1
    keys1 <- expand.grid(stringsAsFactors=F,
        idCharacter=replicate(u1SizePerKeyColumn,paste(collapse='',sample(letters,sample(4:12,1L),T))),
        idInteger=sample(u1SizePerKeyColumn),
        idLogical=sample(c(F,T),u1SizePerKeyColumn,T)
        ##idPOSIXct=as.POSIXct('2016-01-01 00:00:00','UTC')+sample(u1SizePerKeyColumn)
    )[seq_len(u1Size),];

    ## rbind some repetitions of the unique keys; this will prepare one side of the many-to-many relationship
    ## also scramble the order afterward
    keys1 <- rbind(keys1,keys1[sample(nrow(keys1),size-u1Size,T),])[sample(size),];

    ## common and unilateral key counts
    com <- as.integer(size*overlap);
    uni <- size-com;

    ## generate some unilateral keys for df2 by synthesizing outside of the idInteger range of df1
    keys2 <- data.frame(stringsAsFactors=F,
        idCharacter=replicate(uni,paste(collapse='',sample(letters,sample(4:12,1L),T))),
        idInteger=u1SizePerKeyColumn+sample(uni),
        idLogical=sample(c(F,T),uni,T)
        ##idPOSIXct=as.POSIXct('2016-01-01 00:00:00','UTC')+u1SizePerKeyColumn+sample(uni)
    );

    ## rbind random keys from df1; this will complete the many-to-many relationship
    ## also scramble the order afterward
    keys2 <- rbind(keys2,keys1[sample(nrow(keys1),com,T),])[sample(size),];

    ##keyNames <- c('idCharacter','idInteger','idLogical','idPOSIXct');
    keyNames <- c('idCharacter','idInteger','idLogical');
    ## note: was going to use raw and complex type for two of the non-key columns, but data.table doesn't seem to fully support them
    argSpecs <- list(
        default=list(copySpec=1:2,args=list(
            df1 <- cbind(stringsAsFactors=F,keys1,y1=sample(c(F,T),size,T),y2=sample(size),y3=rnorm(size),y4=replicate(size,paste(collapse='',sample(letters,sample(4:12,1L),T)))),
            df2 <- cbind(stringsAsFactors=F,keys2,y5=sample(c(F,T),size,T),y6=sample(size),y7=rnorm(size),y8=replicate(size,paste(collapse='',sample(letters,sample(4:12,1L),T)))),
            keyNames
        )),
        data.table.unkeyed=list(copySpec=1:2,args=list(
            as.data.table(df1),
            as.data.table(df2),
            keyNames
        )),
        data.table.keyed=list(copySpec=1:2,args=list(
            setkeyv(as.data.table(df1),keyNames),
            setkeyv(as.data.table(df2),keyNames)
        ))
    );
    ## prepare sqldf
    initSqldf();
    sqldf(paste0('create index df1_key on df1(',paste(collapse=',',keyNames),');')); ## upload and create an sqlite index on df1
    sqldf(paste0('create index df2_key on df2(',paste(collapse=',',keyNames),');')); ## upload and create an sqlite index on df2

    argSpecs;

}; ## end makeArgSpecs.assortedKey.optionalManyToMany()

sizes <- c(1e1L,1e3L,1e5L); ## 1e5L instead of 1e6L to respect more heavy-duty inputs
overlaps <- c(0.99,0.5,0.01);
solTypes <- setdiff(getSolTypes(),'in.place');
system.time({ res <- testGrid(makeArgSpecs.assortedKey.optionalManyToMany,sizes,overlaps,solTypes); });
##     user   system  elapsed
## 38895.50   784.19 39745.53

使用上面给出的相同绘图代码生成的绘图:

titleFunc <- function(overlap) sprintf('R merge solutions: character/integer/logical key, 0..*:0..* cardinality, %d%% overlap',as.integer(overlap*100));
plotRes(res,titleFunc,F);

R合并基准各式各样的键,可选多对多99

R合并基准各式各样的键,可选多对多50

R合并基准各种键(可选,多对多1)