并行 data.table 操作

Question

我的脚本中有一个 data.table object，我需要将单独的日期和时间列转换为 POSIXct object。 我正在使用fastPOSIXct()转换为POSIXct 。 我发现data.table操作存在一些瓶颈。 主要问题是在转换为POSIXct和POSIXct转换本身之前连接字符串。 我正在使用stri_c() paste0() 。 有没有办法并行化这个计算来加速它？ 正在读取的 csv 文件是一个大约 2gb 的大文件。

data.table结构

  index = match(file, csv_files)
  print(paste0("Starting File #", index, ". Running Active Future Filter."))
  csv1 = fread(file = file)
  # create row for date in Date format
  csv1[, date := as.Date(integer())]
  csv1[, trade_date := as.character(trade_date)]
  csv1[, date := as.Date(trade_date[1], "%Y%m%d"), by = trade_date]
  csv1[, active_exp := character()]
  csv1[, active_exp := get.active.future(date = date[1]), by = date]
  nrow(csv1)
  csv1 = csv1[csv1$active_exp == as.character(csv1$contract_delivery_date),]
  # Add POSIXct
  # keep only essential columns
  csv1 = csv1[,-c(3, 6, 9, 23, 11:21)]
  print(paste0("Running POSIX"))
  # establish date-time POSIX for tick
  csv1[, date_time_char := as.character()]
  # creates date time character type to be converted to POSIXct
  csv1[, date_time_char := stri_c(date[1], trade_time[1], sep = " "), by = list(date, trade_time)]
  
  # convert date time character type to POSIXct
  csv1[, date_time := fastPOSIXct(date_time_char[1]), by = date_time_char]
  # time shift for date when time is between 17:00 and 00:00
  print("Making time adjustment")
  csv1[,adj_date_time := time.shift(date_time[1]), by = date_time]
  # convert date
  csv1[, date := as.Date(adj_date_time, tz = "UTC")]
  # overwrite current loaded file

Answer 1

这应该会有所帮助（未经测试，因为该示例不可重现）：

index = match(file, csv_files)

csv1 = fread(file = file)
# create row for date in Date format
csv1[, trade_date := as.character(trade_date)]
csv1[,
     c("date", "active_exp") := {
       date1 = as.Date(.BY[[1L]], "%Y%m%d")
       active_exp = get.active.future(date1)
       .(date1, active_exp)},
     by = trade_date]

csv1 = csv1[active_exp == contract_delivery_date), -c(3, 6, 9, 23, 11:21)]

# creates date time character type to be converted to POSIXct
csv1[, 
     c("date_time", "adj_date_time", "date") = {
       date_time = fastPOSIXct(stri_c(.BY[[1L]]), .BY[[2L]], sep = " ")
       adj_date_time = time.shift(date_time)
       date = as.Date(adj_date_time, tz = "UTC")
     }
     , by = .(date, trade_time)
]

有多个地方对或多或少相同的数据进行了分组。 每次我们进行分组时，这意味着我们正在调用forder()并且必须对数据进行分组。 相反，我们可以尝试一次做所有事情。

为了使这更快，我建议只fread您需要的列。 相关，在导入文件后，我会让过滤器active_exp == contract_delivery_date 。

编辑您可以查看的一件事是使用IDate和ITime制作POSIXct列。 最慢的部分可能是stri_c / paste后跟分组或fasttime::fastPOSIXct 。 这是避免粘贴的方法，尽管它依赖于一些Rcpp来帮助解析时间戳列。

Rcpp::cppFunction("
IntegerVector to_time(std::vector< std::string > x) {
  //for format hh:mm:ss 
  int n = x.size();
  IntegerVector out(n);
  
  for (int i = 0; i < n; i++){;
    const std::string xi = x[i]; 
    const int hour = stoi(xi.substr(0, 2));
    const int minute = stoi(xi.substr(3, 2));
    const int second = stoi(xi.substr(6, 2));
    out[i] = hour * 60 * 60 + minute * 60 + second;
  }
  return(out);
}
")

csv1[,
     date_time := {
       date1 = as.IDate(as.character(trade_date), "%Y%m%d")
       time = as.ITime(to_time(trade_time))
       .(as.POSIXct(date1, time, tz = "UTC"))
     }]

当复制小数据集一百万次时，这需要 3 秒，而fasttime::fastPOSIXct()需要 20 秒。 将在下面发布基准测试。

您还应该考虑删除原始帖子的by =部分。 如果有很多重复的日期时间，那么是的，按组分组有时会更有效率。 但是，如果有很多独特的日期时间，您最好跳过分组步骤。

csv1 = csv1[rep(seq_len(.N), 1e6L)]

## for use in different use cases
date_col = as.Date(as.character(csv1$trade_date), "%Y%m%d")
IDate_col = as.IDate(date_col)
hour_col = as.ITime(to_time(csv1$trade_time))
bench::mark(
 as_char_to_date = as.Date(as.character(csv1$trade_date), "%Y%m%d")
 ,
 as.ITime(to_time(csv1$trade_time))
 ,
 as_date_to_idate = as.IDate(date_col)
 , as.POSIXct(IDate_col, hour_col, tz = "UTC")
 ,  use_fasttime = fasttime::fastPOSIXct(paste(date_col, csv1$trade_time), "UTC")
 , check = FALSE
)

## # A tibble: 5 x 13
##   expression                                       ## min   median `itr/sec`
##   <bch:expr>                                  <bch:tm> <bch:tm>     <dbl>
## 1 as_char_to_date                                3.03s    3.03s    0.330 
## 2 as.ITime(to_time(csv1$trade_time))             2.03s    2.03s    0.494 
## 3 as_date_to_idate                              13.2ms   13.5ms   23.1   
## 4 as.POSIXct(IDate_col, hour_col, tz = "UTC") 929.14ms 929.14ms    1.08  
## 5 use_fasttime                                  20.39s   20.39s    0.0490

Answer 2

编辑：正如@Cole 所建议的，我使用了这个 C++ function 和 IDate 和 ITime 来加快速度。 我没有对此进行基准测试，但可以说它的运行速度比以前的代码快 20 倍。 代码如下。

for(file in csv_files){

  index = match(file, csv_files)
  csv1 = fread(file = file, drop = c(3, 6, 9, 23, 11:21))
  
  options(warn = -1)
  # get date and time and classify as POSIXct
  csv1[,
       c("date_time", "date", "time") := {
         
         time = as.ITime(to_time(trade_time))
         date1 = as.IDate(as.character(trade_date), "%Y%m%d")
         if(hour(time) >= 17){
            POSIX = as.POSIXct(date1, time, tz = "UTC") - days(1)
            .(POSIX, as.Date(POSIX), time)
         }else{
            POSIX = as.POSIXct(date1, time, tz = "UTC")
            .(POSIX, as.Date(POSIX), time)
         }
       }]
  options(warn = 0)
  # covert trade date to character format
  csv1[, trade_date := as.character(trade_date)]
  # run active_exp
  csv1[,
       active_exp := get.active.future(as.Date(date[1])),
       by = date]
  # filter only active future
  csv1 = csv1[active_exp == as.character(contract_delivery_date)]
  # write file
  file_path = paste0("C:/Users/ocean/Documents/CME Data/Active Daily Data CSVs/", index, ".csv")
  fwrite(csv1, file = file_path)
  
  print(paste0("Progress = ", round(index/length(csv_files), 3)*100, "%"))
}

OLD：我使用了@Cole 提供的代码，但仍然运行了很长时间。 我稍微编辑了代码以修复一些语法，如下所示。 文件 = csv_files[1] 索引 = 匹配（文件，csv_files）

csv1 = fread(file = file, drop = c(3, 6, 9, 23, 11:21))
# create row for date in Date format
csv1[, trade_date := as.character(trade_date)]
csv1[,
     c("date", "active_exp") := {
       date1 = as.Date(.BY[[1L]], "%Y%m%d")
       active_exp = get.active.future(date1)
       .(date1, active_exp)},
     by = trade_date]

csv1 = csv1[active_exp == contract_delivery_date]

# creates date time character type to be converted to POSIXct
csv1[, 
     c("date_time", "adj_date_time", "date") := {
       date_time = fastPOSIXct(stri_c(.BY[[1L]], .BY[[2L]], sep = " "))
       adj_date_time = time.shift(date_time)
       date = as.Date(adj_date_time, tz = "UTC")
     }
     , by = .(date, trade_time)
]

为了尝试提供一个可重现的示例，数据的dput()如下所示。

    structure(list(trade_date = c(20200115L, 20200115L, 20200115L, 
20200115L, 20200115L, 20200115L), trade_time = c("17:00:00", 
"17:00:00", "17:00:00", "17:00:00", "17:00:00", "17:00:00"), 
    trade_sequence_number = c(9028350L, 9028357L, 9028366L, 9028394L, 
    9028397L, 9028400L), session_indicator = c("E", "E", "E", 
    "E", "E", "E"), ticker_symbol = c("ES", "ES", "ES", "ES", 
    "ES", "ES"), future_option_index_indicator = c("F", "F", 
    "F", "F", "F", "F"), contract_delivery_date = c(2003L, 2003L, 
    2003L, 2003L, 2003L, 2003L), trade_quantity = c(176L, 0L, 
    3L, 2L, 4L, 10L), strike_price = c(0L, 0L, 0L, 0L, 0L, 0L
    ), trade_price = c(3287.75, 3287.75, 3288, 3288, 3288, 3288
    ), ask_bid_type = c(NA, NA, NA, NA, NA, NA), indicative_quote_type = c(NA, 
    NA, NA, NA, NA, NA), market_quote = c(NA, NA, NA, NA, NA, 
    NA), close_open_type = c("", "O", "", "", "", ""), valid_open_exception = c(NA, 
    NA, NA, NA, NA, NA), post_close = c(NA, NA, NA, NA, NA, NA
    ), cancel_code_type = c(NA, NA, NA, NA, NA, NA), insert_code_type = c(NA, 
    NA, NA, NA, NA, NA), fast_late_indicator = c(NA, NA, NA, 
    NA, NA, NA), cabinet_indicator = c(NA, NA, NA, NA, NA, NA
    ), book_indicator = c(NA, NA, NA, NA, NA, NA), entry_date = c(20200114L, 
    20200114L, 20200114L, 20200114L, 20200114L, 20200114L), exchange_code = c("XCME", 
    "XCME", "XCME", "XCME", "XCME", "XCME")), row.names = c(NA, 
-6L), class = c("data.table", "data.frame"), .internal.selfref = <pointer: 0x000001ef853c1ef0>)

鉴于文件的大小（38 个文件为 2-4gb），我只需要学习 C++ 还是让它在一夜之间运行？