In xiaodaigh/disk.frame: Larger-than-RAM Disk-Based Data Manipulation Framework

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  eval=TRUE,
  include=TRUE
)

Custom one-stage group-by

Group-by is one the most useful and commonly used operations in data manipulation. The most common to perform group with dplyr syntax would be something similar to the below

dataf %>% 
  group_by(grpA) %>% 
  summarize(mean(X))

However, the above nice and concise group-by syntax creates a challenge in {disk.frame}, because {disk.frame} organises the data in chunks and these chunks are often manipulated independently to each other which makes it difficult to compute the mean of any column if the group-by column is stored in different chunks. In this chapter, I will explain how {disk.frame} achieve the one-stage group-by syntax that you see, and how you can define custom one-stage group-by functions for use with {disk.frame}.

At a glance

{disk.frame} allows the user to create custom one-stage group-by functions. To make a function fn one stage, one needs to define two functions:

1. fn_df.chunk_agg.disk.frame which applies itself to each chunk
2. fn_df.collected_agg.disk.frame which accepts a list of returns from fn_df.chunk_agg.disk.frame and finalize the computation

For example, to make mean a one-stage group-by function, {disk.frame} has defined mean_df.chunk_agg.disk.frame and mean_df.collected_agg.disk.frame, which we will illustrate with examples below.

But first, we shall explain some theory behind {disk.frame} to help you better understand "why does {disk.frame} do it like that?".

How does {disk.frame} work

One may ask, how come only a few functions are supported for one-stage group-by? And why are some functions like median only produce estimates instead of producing the exact figure? To answer these questions, we need to have an understanding of how {disk.frame} works.

A disk.frame is organized as chunks stored on disk. Each chunk is a file stored in fst format. The {future} package is used to apply the same function to each chunk, each of these operations are carried out in a separate R session. These R sessions cannot communicate with each other during the execution of the operations.

Once the operations have been performed on all chunks, the results will be bought back to the session from which the operations were called. This is the only point of inter-process communication.

To summarize, the two phases of a df %>% some_fn %>% collect operation are

1. The some_fn is applied to each chunk, and the result is assumed to be a data.frame
2. collect then row-binds (rbind/bind_rows/rbindlist) the results together to form a data.frame in the main session

How group-by works

Except for passing the result back to the main session, communication between worker sessions are not allowed. This limits how group-by operations can be performed, hence why group-by was done in two stages prior to {disk.frame} v0.3.0. However, R's meta-programming abilities allow us to rewrite code to perform the two-stage group-bys using one-stage group-by code. For example, consider:

df %>% 
  group_by(grp1) %>% 
  summarize(sum(x)) %>% 
  collect

we can use meta-programming to transform that to

df %>% 
  chunk_group_by(grp1) %>% 
  chunk_summarize(__tmp1__= sum(x)) %>% 
  collect() %>% 
  group_by(grp1) %>% 
  summarize(x = sum(__tmp1__))

Basically, we are "compiling" one-stage group-by code to two-stage group-by code, and then executing it.

For mean, it's trickier, as one needs to keep track on the numerator and the denominator separately in computing mean(x) = sum(x)/length(x).

Therefore, {disk.frame} compiles

df %>% 
  group_by(grp1) %>% 
  summarize(meanx = mean(x)) %>% 
  collect

to

df %>% 
  chunk_group_by(grp1) %>% 
  chunk_summarize(__tmp1__ = list(mean_df.chunk_agg.disk.frame(x))) %>% 
  collect %>% 
  group_by(grp1) %>% 
  chunk_summarize(meanx = mean_df.chunk_agg.disk.frame(__tmp1__))

where mean_df.chunk_agg.disk.frame defines what needs to be done to each chunk, as you can see, the return value is a vector where the elements are named sumx and lengthx. Also note because the return value is not a scalar, we need to write it in a list (line 3).

Here is an example implementation of mean.chunk_agg.disk.frame

mean_df.chunk_agg.disk.frame <- function(x, na.rm = FALSE, ...) {
  sumx = sum(x, na.rm = na.rm)
  lengthx = length(x) - ifelse(na.rm, sum(is.na(x)), 0)
  c(sumx = sumx, lengthx = lengthx)
}

The mean_df.collected_agg.disk.frame receives a list of outputs from mean_df.chunk_agg.disk.frame. Recall that mean.chunk_agg.disk.frame returns a vector for each chunk, so the input to mean_df.collected_agg.disk.frame is a list of vectors

mean_df.collected_agg.disk.frame <- function(listx) {
  sum(sapply(listx, function(x) x["sumx"]))/sum(sapply(listx, function(x) x["lengthx"]))
}

How to define custom one-stage group-by functions

Now that we have seen two examples, namely sum and mean, we are ready to summarize how group-by functions are implemented.

Given the below

df %>% 
  group_by(grp1) %>% 
  summarize(namex = fn(x)) %>% 
  collect

{disk.frame} compiles it to

df %>% 
  chunk_group_by(grp1) %>% 
  chunk_summarize(__tmp1__ = list(fn_df.chunk_agg.disk.frame(x))) %>% 
  collect %>% 
  group_by(grp1) %>% 
  chunk_summarize(namex = fn_df.collected_agg.disk.frame(__tmp1__))

Based on the above information, to make fn a one-stage group-by function, the user has to

1. Define fn_df.chunk_agg.disk.frame which is a function to be applied to each chunk
2. Define fn_df.collected_agg.disk.frame which is a function to be applied to a list containing the returns from fn.chunk_agg.disk.frame applied to each chunk

Example of implementing sum:

1. Define sum_df.chunk_agg.disk.frame

sum_df.chunk_agg.disk.frame <- function(x, na.rm = FALSE) {
  sum(x, na.rm=na.rm)
}

1. Define sum_df.collected_agg.disk.frame, which needs to accept a list of sum(x, na.rm), but sum(x, na.rm) is just a numeric, so

sum_df.collected_agg.disk.frame <- function(list_sum) {
  sum(unlist(list_sum))
}

Example of implementing n_distinct:

The dplyr::n_distinct function counts the number of distint values from a vector x

1. Define n_distinct_df.chunk_agg.disk.frame, to return a list of unique values. Because the same value can appear in multiple chunks, to ensure that we don't double count, we simply return all the unique values from each chunk which is then de-duplicated in the next phase

n_distinct_df.chunk_agg.disk.frame <- function(x, na.rm = FALSE) {
  if(na.rm) {
    setdiff(unique(x), NA)
  } else {
    unique(x)
  }
}

1. Define n_distinct_df.collected_agg.disk.frame, which de-duplicates the unique values

n_distinct_df.collected_agg.disk.frame <- function(list_of_chunkwise_uniques) {
  dplyr::n_distinct(unlist(list_of_chunkwise_uniques))
}

Limitations

We have seen that {disk.frame} performs operations in two phases

1. apply the same function to each chunk
2. row-bind the results

and there are no communication between the sessions that applies the functions at chunk level.

Hence, it is generally difficult to compute rank based summarizations like median exactly. Hence most rank based calculations are estimates only. This is also true of distributed data system like Spark whose median function is also estimates only.

Another limitation for now is that summarization that is more complex then f(x) is not supported. E.g. sum(x) + 1, sum(x + mean(x)), sum(x) + mean(x), and fn(sum(x)) are not yet supported as arguments in the summarize function.

xiaodaigh/disk.frame documentation built on Feb. 3, 2023, 10:04 p.m.