Fluid Data

John Mount, Win-Vector LLC

November 11, 2017

Introduction

The cdata R package provides a powerful extension of the “fluid data” (or “coordinatized data”) concept (please see here for some notes ) that goes way beyond the concepts of pivot/un-pivot.

The fluid data concept is:

  1. Data cells have coordinates, and the dependence of these coordinates on a given data representation (a table or map) is an inessential detail to be abstracted out.
  2. There may not be one “preferred” shape (or frame of reference) for data: you have to anticipate changing data shape many times to adapt to the tasks and tools (data relativity).

cdata supplies two general operators for fluid data work at database scale (and Spark big data scale):

  1. cdata::rowrecs_to_blocks(): operators centered around SQL cross-join semantics. un-pivot, tidyr::gather(), and cdata::unpivot_to_blocks() are special cases of this general operator.
  2. blocks_to_rowrecs(): an operator centered around SQL group by semantics. pivot, tidyr::spread(), cdata::pivot_to_rowrecs(), transpose, and one-hot-encoding are special cases of this general operator.

Because these operators are powerful, they are fairly general, and at first hard to mentally model (especially if you insist on think of them in only in terms of more a specialized operator such as pivot, instead of more general relational concepts such as “cross join” and “group by”). These operators are thin wrappers populating and enforcing a few invariants over a large SQL statement. That does not mean that these operators are trivial, they are thin because SQL is powerful and we have a good abstraction.

Due to the very detailed and explicit controls used in these operators- they are very comprehensible once studied. We will follow-up later with additional training material to make quicker comprehension available to more readers. This document is limiting itself to being a mere concise statement of and demonstration of the operators.

Data coordinate notation theory

We are going to introduce a explicit, dense, and powerful data coordinate notation.

Consider the following table that we call a “control table”:

suppressPackageStartupMessages(library("cdata"))
library("wrapr")
library("rqdatatable")
## Loading required package: rquery
library("rquery")
options(width = 160) 
tng <- wrapr::mk_tmp_name_source('fdexample')
controlTable <- wrapr::build_frame(
   "group", "col1", "col2" |
   "aa"   , "c1"  , "c2"   |
   "bb"   , "c3"  , "c4"   )
knitr::kable(controlTable)
group col1 col2
aa c1 c2
bb c3 c4

Control tables partially specify a change of data shape or change of data cell coordinates.

The specification is interpreted as follows:

The region controlTable[ , 2:ncol(controlTable)] specifies partial coordinates of data cells in another table. In our example these partial coordinates are “c1”, “c2”, “c3”, and “c4” treated as column names. For example if our data is:

dat1 <- wrapr::build_frame(
   "ID" , "c1"        , "c2"        , "c3"        , "c4"         |
   "id1", "val_id1_c1", "val_id1_c2", "val_id1_c3", "val_id1_c4" |
   "id2", "val_id2_c1", "val_id2_c2", "val_id2_c3", "val_id2_c4" |
   "id3", "val_id3_c1", "val_id3_c2", "val_id3_c3", "val_id3_c4" )
knitr::kable(dat1)
ID c1 c2 c3 c4
id1 val_id1_c1 val_id1_c2 val_id1_c3 val_id1_c4
id2 val_id2_c1 val_id2_c2 val_id2_c3 val_id2_c4
id3 val_id3_c1 val_id3_c2 val_id3_c3 val_id3_c4

Then each data cell in dat1 (excluding the key-columns, in this case “ID”) is named by the row-id (stored in the ID column) plus the column-name (“c1”, “c2”, “c3”, and “c4”). Knowing ID plus the column name unique identifies the data-caring cell in table dat1.

However, there is an alternate cell naming available from the controlTable notation. Each name in the region controlTable[ , 2:ncol(controlTable)] is itself uniquely named by the group entry and column name of the control table itself. This means we have the following correspondence from the partial names “c1”, “c2”, “c3”, and “c4” to a new set of partial names:

namePairings <- expand.grid(seq_len(nrow(controlTable)), 
                     2:ncol(controlTable))
colnames(namePairings) <- c("controlI", "controlJ")
namePairings$coords_style1 <- 
  vapply(seq_len(nrow(namePairings)),
         function(ii) {
           as.character(paste("column:",
                              controlTable[namePairings$controlI[[ii]], 
                                           namePairings$controlJ[[ii]]]))
         },
         character(1))
namePairings$coords_style2 <- 
  vapply(seq_len(nrow(namePairings)),
         function(ii) {
           paste("group:",
                 controlTable$group[[namePairings$controlI[[ii]]]],
                 ", column:",
                 colnames(controlTable)[[namePairings$controlJ[[ii]]]])
         },
         character(1))
as.matrix(namePairings[ , c("coords_style1", "coords_style2")])
##      coords_style1 coords_style2             
## [1,] "column: c1"  "group: aa , column: col1"
## [2,] "column: c3"  "group: bb , column: col1"
## [3,] "column: c2"  "group: aa , column: col2"
## [4,] "column: c4"  "group: bb , column: col2"

The idea is the control table is a very succinct description of the pairing of the namePairings$coords_style1 cell partial coordinates and the namePairings$coords_style2 cell partial coordinates. As we have said the namePairings$coords_style1 cell partial coordinates become full cell coordinates for the data cells in dat1 when combined with the dat1 ID column. The namePairings$coords_style2 are part of a natural naming for the data cells in the following table:

my_db <- DBI::dbConnect(RSQLite::SQLite(), ":memory:")
dat1db <- DBI::dbWriteTable(my_db, 'dat1db', dat1, 
                            overwrite = TRUE,
                            temporary = TRUE)
dat2 <- cdata::rowrecs_to_blocks_q(wideTable = 'dat1db',
                                   controlTable = controlTable,
                                   my_db = my_db,
                                   columnsToCopy = "ID",
                                   tempNameGenerator = tng)
DBI::dbReadTable(my_db, dat2) %.>%
  orderby(., c("ID", "group")) %.>%
  knitr::kable(.)
ID group col1 col2
id1 aa val_id1_c1 val_id1_c2
id1 bb val_id1_c3 val_id1_c4
id2 aa val_id2_c1 val_id2_c2
id2 bb val_id2_c3 val_id2_c4
id3 aa val_id3_c1 val_id3_c2
id3 bb val_id3_c3 val_id3_c4

For dat2 the composite row-key (ID, group) plus the column name (one of col1 or col2) gives us the positions of the data carrying cells.

So essentially the two readings of controlTable are a succinct representation of the explicit pairing of data cell coordinates shown in the namePairings table.

The Operators

In terms of the above notation/theory our two operators rowrecs_to_blocks() and blocks_to_rowrecs() are (in principle) easy to describe:

The above is certainly succinct, but carries a lot of information and allows for a lot of different possible applications. Many important applications are derived from how these two operators interact with row-operations and column-operations.

We give simple examples of each of the operators below.

rowrecs_to_blocks()

wideTableName <- 'dat'
d <- DBI::dbWriteTable(my_db, wideTableName,
  wrapr::build_frame(
   "ID" , "c1"        , "c2"        , "c3"        , "c4"         |
   "id1", "val_id1_c1", "val_id1_c2", "val_id1_c3", "val_id1_c4" |
   "id2", "val_id2_c1", "val_id2_c2", "val_id2_c3", "val_id2_c4" |
   "id3", "val_id3_c1", "val_id3_c2", "val_id3_c3", "val_id3_c4" ),
             overwrite = TRUE, temporary=TRUE)
controlTable <- wrapr::build_frame(
   "group", "col1", "col2" |
   "aa"   , "c1"  , "c2"   |
   "bb"   , "c3"  , "c4"   )
columnsToCopy <- 'ID'
cdata::rowrecs_to_blocks_q(wideTable = wideTableName,
                         controlTable =  controlTable,
                         my_db = my_db,
                         columnsToCopy = columnsToCopy,
                         tempNameGenerator = tng) %.>%
  DBI::dbReadTable(my_db, .) %.>%
  orderby(., c("ID", "group")) %.>%
  knitr::kable(.)
ID group col1 col2
id1 aa val_id1_c1 val_id1_c2
id1 bb val_id1_c3 val_id1_c4
id2 aa val_id2_c1 val_id2_c2
id2 bb val_id2_c3 val_id2_c4
id3 aa val_id3_c1 val_id3_c2
id3 bb val_id3_c3 val_id3_c4

blocks_to_rowrecs()

tallTableName <- 'dat'
d <- DBI::dbWriteTable(my_db, tallTableName,
  wrapr::build_frame(
   "ID" , "group", "col1"            , "col2"             |
   "id1", "aa"   , "val_id1_gaa_col1", "val_id1_gaa_col2" |
   "id1", "bb"   , "val_id1_gbb_col1", "val_id1_gbb_col2" |
   "id2", "aa"   , "val_id2_gaa_col1", "val_id2_gaa_col2" |
   "id2", "bb"   , "val_id2_gbb_col1", "val_id2_gbb_col2" |
   "id3", "aa"   , "val_id3_gaa_col1", "val_id3_gaa_col2" |
   "id3", "bb"   , "val_id3_gbb_col1", "val_id3_gbb_col2" ),
         overwrite = TRUE, temporary=TRUE)
controlTable <- wrapr::build_frame(
   "group", "col1", "col2" |
   "aa"   , "c1"  , "c2"   |
   "bb"   , "c3"  , "c4"   )
keyColumns <- 'ID'
cdata::blocks_to_rowrecs_q(tallTable = tallTableName,
                            controlTable = controlTable,
                            keyColumns = keyColumns,
                            my_db = my_db,
                            tempNameGenerator = tng) %.>%
  DBI::dbReadTable(my_db, .) %.>%
  orderby(., "ID") %.>%
  knitr::kable(.)
ID c1 c2 c3 c4
id1 val_id1_gaa_col1 val_id1_gaa_col2 val_id1_gbb_col1 val_id1_gbb_col2
id2 val_id2_gaa_col1 val_id2_gaa_col2 val_id2_gbb_col1 val_id2_gbb_col2
id3 val_id3_gaa_col1 val_id3_gaa_col2 val_id3_gbb_col1 val_id3_gbb_col2

Pivot/Un-Pivot

Pivot and un-pivot (or tidyr::spread() and tidyr::gather()) are special cases of the blocks_to_rowrecs() and rowrecs_to_blocks() operators. Pivot/un-pivot are the cases where the control table has two columns.

Pivot

d <- data.frame(
  index = c(1, 2, 3, 1, 2, 3),
  meastype = c('meas1','meas1','meas1','meas2','meas2','meas2'),
  meas = c('m1_1', 'm1_2', 'm1_3', 'm2_1', 'm2_2', 'm2_3'),
  stringsAsFactors = FALSE)
knitr::kable(d)
index meastype meas
1 meas1 m1_1
2 meas1 m1_2
3 meas1 m1_3
1 meas2 m2_1
2 meas2 m2_2
3 meas2 m2_3
# the cdata::pivot_to_rowrecs version
# equivalent to tidyr::spread(d, 'meastype', 'meas')
cdata::pivot_to_rowrecs(d,
                        columnToTakeKeysFrom = 'meastype',
                        columnToTakeValuesFrom = 'meas',
                        rowKeyColumns= 'index',
                        sep= '_') %.>%
  orderby(., "index") %.>%
  knitr::kable(.)
index meastype_meas1 meastype_meas2
1 m1_1 m2_1
2 m1_2 m2_2
3 m1_3 m2_3
# the cdata::blocks_to_rowrecs() version
dtall <- DBI::dbWriteTable(my_db, "dtall", d)
controlTable <- cdata::build_pivot_control_q(tableName = "dtall",
                                             columnToTakeKeysFrom = 'meastype',
                                             columnToTakeValuesFrom = 'meas',
                                             my_db = my_db,
                                             sep = "_")
knitr::kable(controlTable)
meastype meas
meas1 meastype_meas1
meas2 meastype_meas2
blocks_to_rowrecs_q(tallTable = "dtall",
                     controlTable = controlTable,
                     keyColumns = "index",
                     my_db = my_db,
                     tempNameGenerator = tng) %.>%
  DBI::dbReadTable(my_db, .) %.>%
  orderby(., "index") %.>%
  knitr::kable(.)
index meastype_meas1 meastype_meas2
1 m1_1 m2_1
2 m1_2 m2_2
3 m1_3 m2_3

Un-Pivot

d <- data.frame(
  index = c(1, 2, 3),
  info = c('a', 'b', 'c'),
  meas1 = c('m1_1', 'm1_2', 'm1_3'),
  meas2 = c('2.1', '2.2', '2.3'),
  stringsAsFactors = FALSE)
knitr::kable(d)
index info meas1 meas2
1 a m1_1 2.1
2 b m1_2 2.2
3 c m1_3 2.3
# the cdata::unpivot_to_blocks() version
# equivalent to tidyr::gather(d, 'meastype', 'meas', c('meas1','meas2'))
cdata::unpivot_to_blocks(d,
                        nameForNewKeyColumn= 'meastype',
                        nameForNewValueColumn= 'meas',
                        columnsToTakeFrom= c('meas1','meas2')) %.>%
  orderby(., "index") %.>%
  knitr::kable(.)
index info meastype meas
1 a meas1 m1_1
1 a meas2 2.1
2 b meas1 m1_2
2 b meas2 2.2
3 c meas1 m1_3
3 c meas2 2.3
# the cdata::cdata::unpivot_to_blocks() version
dwide <- DBI::dbWriteTable(my_db, "dwide", d)
controlTable <- build_unpivot_control(nameForNewKeyColumn= 'meastype',
                                         nameForNewValueColumn= 'meas',
                                         columnsToTakeFrom= c('meas1','meas2'))
knitr::kable(controlTable)
meastype meas
meas1 meas1
meas2 meas2
keyColumns = c('index', 'info')
rowrecs_to_blocks_q(wideTable = "dwide",
                  controlTable = controlTable,
                  my_db = my_db,
                  columnsToCopy = keyColumns,
                  tempNameGenerator = tng) %.>%
  DBI::dbReadTable(my_db, .) %.>%
  orderby(., "index") %.>%
  knitr::kable(.)
index info meastype meas
1 a meas1 m1_1
1 a meas2 2.1
2 b meas1 m1_2
2 b meas2 2.2
3 c meas1 m1_3
3 c meas2 2.3

Additional Interesting Applications

Interesting applications of cdata::rowrecs_to_blocks() and cdata::blocks_to_rowrecs() include situations where tidyr is not available (databases and Spark) and also when the data transformation is not obviously a single pivot or un-pivot.

Row-parallel dispatch

A particularly interesting application is converting many column operations into a single operation using a row-parallel dispatch.

Suppose we had the following data in the following format in our system of record (but with many more column groups and columns):

purchaseDat <- DBI::dbWriteTable(my_db, 'purchaseDat',
  wrapr::build_frame(
   "ID", "Q1purchases", "Q2purchases", "Q1rebates", "Q2rebates" |
   1   , 20           , 10           , 5          , 3           |
   2   , 5            , 6            , 10         , 12          )
  )
knitr::kable(DBI::dbReadTable(my_db, 'purchaseDat'))
ID Q1purchases Q2purchases Q1rebates Q2rebates
1 20 10 5 3
2 5 6 10 12

Common tasks might include totaling columns and computing rates between columns. However, sometimes that is best done in a row-oriented representation (though outside systems may need column oriented, or more denormalized results).
With fluid data the task is easy:

controlTable <- wrapr::build_frame(
   "group", "purchases"  , "rebates"   |
   "Q1"   , "Q1purchases", "Q1rebates" |
   "Q2"   , "Q2purchases", "Q2rebates" )
controlTable2 <- controlTable %.>%
  extend_nse(., purchasesPerRebate = paste0(group, "PurchPerRebate"))

knitr::kable(controlTable)
group purchases rebates
Q1 Q1purchases Q1rebates
Q2 Q2purchases Q2rebates
purchasesTall <- rowrecs_to_blocks_q(wideTable = "purchaseDat",
                                     columnsToCopy = "ID", 
                                     controlTable = controlTable, 
                                     my_db = my_db,
                                     tempNameGenerator = tng)

print(purchasesTall)
## [1] "fdexample_42473951434910715265_0000000011"
knitr::kable(DBI::dbReadTable(my_db, purchasesTall))
ID group purchases rebates
1 Q1 20 5
1 Q2 10 3
2 Q1 5 10
2 Q2 6 12
# perform the calculation in one easy step

calc <- db_td(my_db, "purchaseDat") %.>% 
  rowrecs_to_blocks(.,
                    columnsToCopy = "ID", 
                    controlTable = controlTable) %.>% 
  extend_nse(., purchasesPerRebate = purchases/rebates) %.>%
  blocks_to_rowrecs(.,
                    keyColumns = "ID",
                    controlTable = controlTable2)

cat(format(calc))
## table(`purchaseDat`; 
##   ID,
##   Q1purchases,
##   Q2purchases,
##   Q1rebates,
##   Q2rebates) %.>%
##  non_sql_node(., rowrecs_to_blocks(.)) %.>%
##  extend(.,
##   purchasesPerRebate := purchases / rebates) %.>%
##  non_sql_node(., blocks_to_rowrecs(.))
execute(my_db, calc) %.>%
  knitr::kable(.)
ID Q1purchases Q2purchases Q1rebates Q2rebates Q1PurchPerRebate Q2PurchPerRebate
1 20 10 5 3 4.0 3.333333
2 5 6 10 12 0.5 0.500000

The point is: the above can work on a large number of rows and columns (especially on a system such as Spark where row operations are performed in parallel).

The above work pattern is particularly powerful on big data systems when the tall table is built up in pieces by appending data (so only the pivot style step is required).

One-hot encoding

Adding indicators or dummy variables (by one-hot encoding, or other methods) are essentially special cases of the pivot flavor of cdata::blocks_to_rowrecs().

Transpose

Transpose is a special case of these operators. In fact the key-columns behave like group specifiers, meaning we can transpose many similarly structured tables at once.

group_by/aggregate

Many operations that look like a complicated pivot in column format are in fact a simple row operation followed a group_by/aggregate (and optional format conversion).

Some fun

The structure of the control table is so similar to the data expected by blocks_to_rowrecs() that you can actually send the control table through blocks_to_rowrecs() to illustrate the kernel of the transformation.

controlTable <- wrapr::build_frame(
   "group", "col1", "col2" |
   "aa"   , "c1"  , "c2"   |
   "bb"   , "c3"  , "c4"   )
tallTableName <- 'dc'
d <- DBI::dbWriteTable(my_db, tallTableName, controlTable)
keyColumns <- NULL
wideTableName <- blocks_to_rowrecs_q(tallTable = tallTableName,
                           controlTable = controlTable,
                           keyColumns = keyColumns,
                           my_db = my_db) 
knitr::kable(DBI::dbReadTable(my_db, wideTableName))
c1 c2 c3 c4
c1 c2 c3 c4

The transformed table is essentially an example row of the wide-form.

And we can, of course, map back.

This “everything maps to a row” form means the control table is essentially a graphical representation of the desired data transform. It also helps make clear that just about any even more general shape to shape transform can be achieved by a blocks_to_rowrecs_q() followed by a rowrecs_to_blocks_q().1 Or by the a rowrecs_to_blocks_q() followed by a blocks_to_rowrecs_q(). One direction is storing all intermediate values in a single denormalized column, the other is storing in many RDF-triple like rows.

rowrecs_to_blocks_q(wideTable = wideTableName,
                  controlTable = controlTable,
                  my_db = my_db) %.>%
  DBI::dbReadTable(my_db, .) %.>%
  orderby(., "group") %.>%
  knitr::kable(.)
group col1 col2
aa c1 c2
bb c3 c4

Conclusion

cdata::rowrecs_to_blocks() and cdata::blocks_to_rowrecs() represent two very general “fluid data” or “coordinatized data” operators that have database scale (via DBI/rquery) and big data scale implementations (via Sparklyr). Some very powerful data transformations can be translated into the above explicit control table terminology. The extra details explicitly managed in the control table notation makes for clear calling interfaces.

for(ti in tng(dumpList = TRUE)) {
  tryCatch(
    DBI::dbRemoveTable(my_db, ti),
    error = function(e) { }
  )
}
DBI::dbDisconnect(my_db)