ASOF Joins, OLS Regression, and extra summarizers

Since sparklyr.flint, a sparklyr extension for leveraging Flint time collection functionalities via sparklyr, was introduced in September, we have now made quite a lot of enhancements to it, and have efficiently submitted sparklyr.flint 0.2 to CRAN.

On this weblog submit, we spotlight the next new options and enhancements from sparklyr.flint 0.2:

ASOF Joins

For these unfamiliar with the time period, ASOF joins are temporal be a part of operations based mostly on inexact matching of timestamps. Throughout the context of Apache Spark, a be a part of operation, loosely talking, matches data from two knowledge frames (let’s name them left and proper) based mostly on some standards. A temporal be a part of implies matching data in left and proper based mostly on timestamps, and with inexact matching of timestamps permitted, it’s sometimes helpful to hitch left and proper alongside one of many following temporal instructions:

  1. Wanting behind: if a document from left has timestamp t, then it will get matched with ones from proper having the latest timestamp lower than or equal to t.
  2. Wanting forward: if a document from left has timestamp t, then it will get matched with ones from proper having the smallest timestamp higher than or equal to (or alternatively, strictly higher than) t.

Nevertheless, oftentimes it isn’t helpful to contemplate two timestamps as “matching” if they’re too far aside. Due to this fact, a further constraint on the utmost period of time to look behind or look forward is normally additionally a part of an ASOF be a part of operation.

In sparklyr.flint 0.2, all ASOF be a part of functionalities of Flint are accessible through the asof_join() technique. For instance, given 2 timeseries RDDs left and proper:


sc <- spark_connect(grasp = "native")
left <- copy_to(sc, tibble::tibble(t = seq(10), u = seq(10))) %>%
  from_sdf(is_sorted = TRUE, time_unit = "SECONDS", time_column = "t")
proper <- copy_to(sc, tibble::tibble(t = seq(10) + 1, v = seq(10) + 1L)) %>%
  from_sdf(is_sorted = TRUE, time_unit = "SECONDS", time_column = "t")

The next prints the results of matching every document from left with the latest document(s) from proper which might be at most 1 second behind.

print(asof_join(left, proper, tol = "1s", path = ">=") %>% to_sdf())

## # Supply: spark<?> [?? x 3]
##    time                    u     v
##    <dttm>              <int> <int>
##  1 1970-01-01 00:00:01     1    NA
##  2 1970-01-01 00:00:02     2     2
##  3 1970-01-01 00:00:03     3     3
##  4 1970-01-01 00:00:04     4     4
##  5 1970-01-01 00:00:05     5     5
##  6 1970-01-01 00:00:06     6     6
##  7 1970-01-01 00:00:07     7     7
##  8 1970-01-01 00:00:08     8     8
##  9 1970-01-01 00:00:09     9     9
## 10 1970-01-01 00:00:10    10    10

Whereas if we alter the temporal path to “<”, then every document from left shall be matched with any document(s) from proper that’s strictly sooner or later and is at most 1 second forward of the present document from left:

print(asof_join(left, proper, tol = "1s", path = "<") %>% to_sdf())

## # Supply: spark<?> [?? x 3]
##    time                    u     v
##    <dttm>              <int> <int>
##  1 1970-01-01 00:00:01     1     2
##  2 1970-01-01 00:00:02     2     3
##  3 1970-01-01 00:00:03     3     4
##  4 1970-01-01 00:00:04     4     5
##  5 1970-01-01 00:00:05     5     6
##  6 1970-01-01 00:00:06     6     7
##  7 1970-01-01 00:00:07     7     8
##  8 1970-01-01 00:00:08     8     9
##  9 1970-01-01 00:00:09     9    10
## 10 1970-01-01 00:00:10    10    11

Discover no matter which temporal path is chosen, an outer-left be a part of is all the time carried out (i.e., all timestamp values and u values of left from above will all the time be current within the output, and the v column within the output will include NA at any time when there isn’t a document from proper that meets the matching standards).

OLS Regression

You is perhaps questioning whether or not the model of this performance in Flint is kind of equivalent to lm() in R. Seems it has far more to supply than lm() does. An OLS regression in Flint will compute helpful metrics reminiscent of Akaike information criterion and Bayesian information criterion, each of that are helpful for mannequin choice functions, and the calculations of each are parallelized by Flint to totally make the most of computational energy accessible in a Spark cluster. As well as, Flint helps ignoring regressors which might be fixed or practically fixed, which turns into helpful when an intercept time period is included. To see why that is the case, we have to briefly look at the aim of the OLS regression, which is to search out some column vector of coefficients (mathbf{beta}) that minimizes (|mathbf{y} – mathbf{X} mathbf{beta}|^2), the place (mathbf{y}) is the column vector of response variables, and (mathbf{X}) is a matrix consisting of columns of regressors plus a whole column of (1)s representing the intercept phrases. The answer to this downside is (mathbf{beta} = (mathbf{X}^intercalmathbf{X})^{-1}mathbf{X}^intercalmathbf{y}), assuming the Gram matrix (mathbf{X}^intercalmathbf{X}) is non-singular. Nevertheless, if (mathbf{X}) accommodates a column of all (1)s of intercept phrases, and one other column shaped by a regressor that’s fixed (or practically so), then columns of (mathbf{X}) shall be linearly dependent (or practically so) and (mathbf{X}^intercalmathbf{X}) shall be singular (or practically so), which presents a difficulty computation-wise. Nevertheless, if a regressor is fixed, then it basically performs the identical function because the intercept phrases do. So merely excluding such a continuing regressor in (mathbf{X}) solves the issue. Additionally, talking of inverting the Gram matrix, readers remembering the idea of “situation quantity” from numerical evaluation should be considering to themselves how computing (mathbf{beta} = (mathbf{X}^intercalmathbf{X})^{-1}mathbf{X}^intercalmathbf{y}) may very well be numerically unstable if (mathbf{X}^intercalmathbf{X}) has a big situation quantity. Because of this Flint additionally outputs the situation variety of the Gram matrix within the OLS regression outcome, in order that one can sanity-check the underlying quadratic minimization downside being solved is well-conditioned.

So, to summarize, the OLS regression performance applied in Flint not solely outputs the answer to the issue, but in addition calculates helpful metrics that assist knowledge scientists assess the sanity and predictive high quality of the ensuing mannequin.

To see OLS regression in motion with sparklyr.flint, one can run the next instance:

mtcars_sdf <- copy_to(sc, mtcars, overwrite = TRUE) %>%
  dplyr::mutate(time = 0L)
mtcars_ts <- from_sdf(mtcars_sdf, is_sorted = TRUE, time_unit = "SECONDS")
mannequin <- ols_regression(mtcars_ts, mpg ~ hp + wt) %>% to_sdf()

print(mannequin %>% dplyr::choose(akaikeIC, bayesIC, cond))

## # Supply: spark<?> [?? x 3]
##   akaikeIC bayesIC    cond
##      <dbl>   <dbl>   <dbl>
## 1     155.    159. 345403.

# ^ output says situation variety of the Gram matrix was inside purpose

and procure (mathbf{beta}), the vector of optimum coefficients, with the next:

print(mannequin %>% dplyr::pull(beta))

## [[1]]
## [1] -0.03177295 -3.87783074

Further Summarizers

The EWMA (Exponential Weighted Transferring Common), EMA half-life, and the standardized second summarizers (specifically, skewness and kurtosis) together with a couple of others which have been lacking in sparklyr.flint 0.1 are actually totally supported in sparklyr.flint 0.2.

Higher Integration With sparklyr

Whereas sparklyr.flint 0.1 included a acquire() technique for exporting knowledge from a Flint time-series RDD to an R knowledge body, it didn’t have an identical technique for extracting the underlying Spark knowledge body from a Flint time-series RDD. This was clearly an oversight. In sparklyr.flint 0.2, one can name to_sdf() on a timeseries RDD to get again a Spark knowledge body that’s usable in sparklyr (e.g., as proven by mannequin %>% to_sdf() %>% dplyr::choose(...) examples from above). One may get to the underlying Spark knowledge body JVM object reference by calling spark_dataframe() on a Flint time-series RDD (that is normally pointless in overwhelming majority of sparklyr use instances although).


We have now offered quite a lot of new options and enhancements launched in sparklyr.flint 0.2 and deep-dived into a few of them on this weblog submit. We hope you might be as enthusiastic about them as we’re.

Thanks for studying!


The writer want to thank Mara (@batpigandme), Sigrid (@skeydan), and Javier (@javierluraschi) for his or her unbelievable editorial inputs on this weblog submit!

Posit AI Weblog: Classifying photographs with torch

Posit AI Weblog: Optimizers in torch