Skip to content

Sparklyr extension making Flint time series library functionalities (https://github.com/twosigma/flint) easily accessible through R

Notifications You must be signed in to change notification settings

r-spark/sparklyr.flint

Repository files navigation

Sparklyr.flint is a sparklyr extension making Flint time series library functionalities (https://github.com/twosigma/flint) accessible through R.

This extension is currently under active development. It requires the master branch of sparklyr (i.e., the version from devtools::install_github("sparklyr/sparklyr", ref = "master") which is newer that sparklyr 1.2) to be able to load the com.twosigma:sparklyr-flint_* artifacts correctly from the https://dl.bintray.com/yl790/maven repository, along with all transitive dependencies from Maven central.

The com.twosigma:sparklyr-flint_* artifacts contain minor modifications (mostly changes to the build.sbt file) needed to ensure Flint time series functionalities work with Spark 2.4 and Spark 3.0. Artifact names and locations are subject to change. They most likely will be moved to Maven central in future, possibly under a different group ID as well (TBD).

At the moment, Flint time series functionalities are accessible through both Spark 2.x and Spark 3.0 via sparklyr, and some commonly used summarizers such ‘count’ and ‘sum’ are working as expected through a reasonably intuitive R interface (see example below). Meanwhile there are still plenty of other Flint functionalities such as EWMA summarizer, weighted mean, etc that will need similar R interfaces in sparklyr.flint.

Example Usage

First attach sparklyr.flint package and then connect to Spark from sparklyr, e.g.,

library(sparklyr)
library(sparklyr.flint)

spark_version <- "2.4.0"
sc <- spark_connect(master = "local", version = spark_version)

or alternatively,

spark_version <- "3.0.0"
sc <- spark_connect(master = "local", version = spark_version)

since this extension also works with Spark 3.0.

For the purpose of this illustration, we shall create some simple data points such that verifying the correctness of summarized results in all examples below will be an easy exercise for the reader.

df <- tibble::tibble(
  t = c(1, 3, 4, 6, 7, 10, 15, 16, 18, 19),
  v = c(4, -2, NA, 5, NA, 1, -4, 5, NA, 3)
)
sdf <- copy_to(sc, df, overwrite = TRUE)

Next, we shall copy data points from above from a Spark data frame into a TimeSeriesRDD so that Flint can analyze them:

ts_rdd <- fromSDF(sdf, is_sorted = TRUE, time_unit = "SECONDS", time_column = "t")

Alternatively, one can also create a builder object if the same (is_sorted, time_unit, time_column) settings need to be applied to multiple Spark data frames or RDDs:

builder <- ts_rdd_builder(sc, is_sorted = TRUE, time_unit = "SECONDS", time_column = "t")
ts_rdd_1 <- builder$fromSDF(sdf_1)
ts_rdd_2 <- builder$fromSDF(sdf_2)
ts_rdd_3 <- builder$fromRDD(rdd_3, schema_of_rdd_3)

Let’s say for each time point t (in seconds), we are interested in a summary all rows within the time span of [t - 3, t], then we can specify the desired time window in R as in_past("3s"), and apply various summarizers with this time window on the TimeSeriesRDD from above.

ts_count <- summarize_count(ts_rdd, window = in_past("3s"))
ts_count %>% collect()

should output the total number of rows within each time window:

## # A tibble: 10 x 3
##    time                    v count
##    <dttm>              <dbl> <dbl>
##  1 1970-01-01 00:00:01     4     1
##  2 1970-01-01 00:00:03    -2     2
##  3 1970-01-01 00:00:04   NaN     3
##  4 1970-01-01 00:00:06     5     3
##  5 1970-01-01 00:00:07   NaN     3
##  6 1970-01-01 00:00:10     1     2
##  7 1970-01-01 00:00:15    -4     1
##  8 1970-01-01 00:00:16     5     2
##  9 1970-01-01 00:00:18   NaN     3
## 10 1970-01-01 00:00:19     3     3
ts_count <- summarize_count(ts_rdd, column = "v", window = in_past("3s"))
ts_count %>% collect()

should output the total number of values from column v that are not NULL or NaN within each time window:

## # A tibble: 10 x 3
##    time                    v v_count
##    <dttm>              <dbl>   <dbl>
##  1 1970-01-01 00:00:01     4       1
##  2 1970-01-01 00:00:03    -2       2
##  3 1970-01-01 00:00:04   NaN       2
##  4 1970-01-01 00:00:06     5       2
##  5 1970-01-01 00:00:07   NaN       1
##  6 1970-01-01 00:00:10     1       1
##  7 1970-01-01 00:00:15    -4       1
##  8 1970-01-01 00:00:16     5       2
##  9 1970-01-01 00:00:18   NaN       2
## 10 1970-01-01 00:00:19     3       2

and

ts_sum <- summarize_sum(ts_rdd, column = "v", window = in_past("3s"))
ts_sum %>% collect()

should output the sum of values from column v within each time window, ignoring NULL or NaN values:

## # A tibble: 10 x 3
##    time                    v v_sum
##    <dttm>              <dbl> <dbl>
##  1 1970-01-01 00:00:01     4     4
##  2 1970-01-01 00:00:03    -2     2
##  3 1970-01-01 00:00:04   NaN     2
##  4 1970-01-01 00:00:06     5     3
##  5 1970-01-01 00:00:07   NaN     5
##  6 1970-01-01 00:00:10     1     1
##  7 1970-01-01 00:00:15    -4    -4
##  8 1970-01-01 00:00:16     5     1
##  9 1970-01-01 00:00:18   NaN     1
## 10 1970-01-01 00:00:19     3     8