New information sources and spark_apply() capabilities, higher interfaces for sparklyr extensions, and extra!


Sparklyr 1.7 is now accessible on CRAN!

To put in sparklyr 1.7 from CRAN, run

On this weblog put up, we want to current the next highlights from the sparklyr 1.7 launch:

Picture and binary information sources

As a unified analytics engine for large-scale information processing, Apache Spark is well-known for its skill to deal with challenges related to the amount, velocity, and final however not least, the number of large information. Due to this fact it’s hardly shocking to see that – in response to current advances in deep studying frameworks – Apache Spark has launched built-in help for picture information sources and binary information sources (in releases 2.4 and three.0, respectively). The corresponding R interfaces for each information sources, particularly, spark_read_image() and spark_read_binary(), have been shipped not too long ago as a part of sparklyr 1.7.

The usefulness of knowledge supply functionalities equivalent to spark_read_image() is maybe greatest illustrated by a fast demo under, the place spark_read_image(), via the usual Apache Spark ImageSchema, helps connecting uncooked picture inputs to a complicated characteristic extractor and a classifier, forming a strong Spark utility for picture classifications.

The demo

Picture by Daniel Tuttle on Unsplash

On this demo, we will assemble a scalable Spark ML pipeline able to classifying pictures of cats and canine precisely and effectively, utilizing spark_read_image() and a pre-trained convolutional neural community code-named Inception (Szegedy et al. (2015)).

Step one to constructing such a demo with most portability and repeatability is to create a sparklyr extension that accomplishes the next:

A reference implementation of such a sparklyr extension might be present in right here.

The second step, in fact, is to utilize the above-mentioned sparklyr extension to carry out some characteristic engineering. We are going to see very high-level options being extracted intelligently from every cat/canine picture based mostly on what the pre-built Inception-V3 convolutional neural community has already discovered from classifying a much wider assortment of pictures:


# NOTE: the right spark_home path to make use of relies on the configuration of the
# Spark cluster you're working with.
spark_home <- "/usr/lib/spark"
sc <- spark_connect(grasp = "yarn", spark_home = spark_home)

data_dir <- copy_images_to_hdfs()

# extract options from train- and test-data
image_data <- checklist()
for (x in c("prepare", "check")) {
  # import
  image_data[[x]] <- c("canine", "cats") %>%
      operate(label) {
        numeric_label <- ifelse(equivalent(label, "canine"), 1L, 0L)
          sc, dir = file.path(data_dir, x, label, fsep = "/")
        ) %>%
          dplyr::mutate(label = numeric_label)
    ) %>%, .)

  dl_featurizer <- invoke_new(
    random_string("dl_featurizer") # uid
  ) %>%
    invoke("setModelName", "InceptionV3") %>%
    invoke("setInputCol", "picture") %>%
    invoke("setOutputCol", "options")
  image_data[[x]] <-
    dl_featurizer %>%
    invoke("rework", spark_dataframe(image_data[[x]])) %>%

Third step: geared up with options that summarize the content material of every picture effectively, we are able to construct a Spark ML pipeline that acknowledges cats and canine utilizing solely logistic regression

label_col <- "label"
prediction_col <- "prediction"
pipeline <- ml_pipeline(sc) %>%
    features_col = "options",
    label_col = label_col,
    prediction_col = prediction_col
mannequin <- pipeline %>% ml_fit(image_data$prepare)

Lastly, we are able to consider the accuracy of this mannequin on the check pictures:

predictions <- mannequin %>%
  ml_transform(image_data$check) %>%

cat("Predictions vs. labels:n")
predictions %>%
  dplyr::choose(!!label_col, !!prediction_col) %>%
  print(n = sdf_nrow(predictions))

cat("nAccuracy of predictions:n")
predictions %>%
    label_col = label_col,
    prediction_col = prediction_col,
    metric_name = "accuracy"
  ) %>%
## Predictions vs. labels:
## # Supply: spark<?> [?? x 2]
##    label prediction
##    <int>      <dbl>
##  1     1          1
##  2     1          1
##  3     1          1
##  4     1          1
##  5     1          1
##  6     1          1
##  7     1          1
##  8     1          1
##  9     1          1
## 10     1          1
## 11     0          0
## 12     0          0
## 13     0          0
## 14     0          0
## 15     0          0
## 16     0          0
## 17     0          0
## 18     0          0
## 19     0          0
## 20     0          0
## Accuracy of predictions:
## [1] 1

New spark_apply() capabilities

Optimizations & customized serializers

Many sparklyr customers who’ve tried to run spark_apply() or doSpark to parallelize R computations amongst Spark staff have most likely encountered some challenges arising from the serialization of R closures. In some eventualities, the serialized dimension of the R closure can turn out to be too massive, usually because of the dimension of the enclosing R setting required by the closure. In different eventualities, the serialization itself could take an excessive amount of time, partially offsetting the efficiency achieve from parallelization. Just lately, a number of optimizations went into sparklyr to deal with these challenges. One of many optimizations was to make good use of the broadcast variable assemble in Apache Spark to scale back the overhead of distributing shared and immutable activity states throughout all Spark staff. In sparklyr 1.7, there may be additionally help for customized spark_apply() serializers, which presents extra fine-grained management over the trade-off between pace and compression stage of serialization algorithms. For instance, one can specify

choices(sparklyr.spark_apply.serializer = "qs")


which can apply the default choices of qs::qserialize() to realize a excessive compression stage, or

choices(sparklyr.spark_apply.serializer = operate(x) qs::qserialize(x, preset = "quick"))
choices(sparklyr.spark_apply.deserializer = operate(x) qs::qdeserialize(x))


which can goal for sooner serialization pace with much less compression.

Inferring dependencies robotically

In sparklyr 1.7, spark_apply() additionally supplies the experimental auto_deps = TRUE possibility. With auto_deps enabled, spark_apply() will study the R closure being utilized, infer the checklist of required R packages, and solely copy the required R packages and their transitive dependencies to Spark staff. In lots of eventualities, the auto_deps = TRUE possibility might be a considerably higher various in comparison with the default packages = TRUE conduct, which is to ship every part inside .libPaths() to Spark employee nodes, or the superior packages = <package deal config> possibility, which requires customers to provide the checklist of required R packages or manually create a spark_apply() bundle.

Higher integration with sparklyr extensions

Substantial effort went into sparklyr 1.7 to make lives simpler for sparklyr extension authors. Expertise suggests two areas the place any sparklyr extension can undergo a frictional and non-straightforward path integrating with sparklyr are the next:

We are going to elaborate on current progress in each areas within the sub-sections under.

Customizing the dbplyr SQL translation setting

sparklyr extensions can now customise sparklyr’s dbplyr SQL translations via the spark_dependency() specification returned from spark_dependencies() callbacks. Such a flexibility turns into helpful, as an example, in eventualities the place a sparklyr extension must insert sort casts for inputs to customized Spark UDFs. We are able to discover a concrete instance of this in sparklyr.sedona, a sparklyr extension to facilitate geo-spatial analyses utilizing Apache Sedona. Geo-spatial UDFs supported by Apache Sedona equivalent to ST_Point() and ST_PolygonFromEnvelope() require all inputs to be DECIMAL(24, 20) portions somewhat than DOUBLEs. With none customization to sparklyr’s dbplyr SQL variant, the one method for a dplyr question involving ST_Point() to really work in sparklyr could be to explicitly implement any sort solid wanted by the question utilizing dplyr::sql(), e.g.,

my_geospatial_sdf <- my_geospatial_sdf %>%
    x = dplyr::sql("CAST(`x` AS DECIMAL(24, 20))"),
    y = dplyr::sql("CAST(`y` AS DECIMAL(24, 20))")
  ) %>%
  dplyr::mutate(pt = ST_Point(x, y))


This might, to some extent, be antithetical to dplyr’s purpose of liberating R customers from laboriously spelling out SQL queries. Whereas by customizing sparklyr’s dplyr SQL translations (as applied in right here and right here ), sparklyr.sedona permits customers to easily write

my_geospatial_sdf <- my_geospatial_sdf %>% dplyr::mutate(pt = ST_Point(x, y))

as an alternative, and the required Spark SQL sort casts are generated robotically.

Improved interface for invoking Java/Scala features

In sparklyr 1.7, the R interface for Java/Scala invocations noticed quite a lot of enhancements.

With earlier variations of sparklyr, many sparklyr extension authors would run into hassle when trying to invoke Java/Scala features accepting an Array[T] as one in every of their parameters, the place T is any sort certain extra particular than java.lang.Object / AnyRef. This was as a result of any array of objects handed via sparklyr’s Java/Scala invocation interface might be interpreted as merely an array of java.lang.Objects in absence of extra sort data. Because of this, a helper operate jarray() was applied as a part of sparklyr 1.7 as a solution to overcome the aforementioned drawback. For instance, executing

sc <- spark_connect(...)

arr <- jarray(
  seq(5) %>% lapply(operate(x) invoke_new(sc, "MyClass", x)),
  element_type = "MyClass"

will assign to arr a reference to an Array[MyClass] of size 5, somewhat than an Array[AnyRef]. Subsequently, arr turns into appropriate to be handed as a parameter to features accepting solely Array[MyClass]s as inputs. Beforehand, some potential workarounds of this sparklyr limitation included altering operate signatures to just accept Array[AnyRef]s as an alternative of Array[MyClass]s, or implementing a “wrapped” model of every operate accepting Array[AnyRef] inputs and changing them to Array[MyClass] earlier than the precise invocation. None of such workarounds was a great resolution to the issue.

One other comparable hurdle that was addressed in sparklyr 1.7 as effectively includes operate parameters that have to be single-precision floating level numbers or arrays of single-precision floating level numbers. For these eventualities, jfloat() and jfloat_array() are the helper features that enable numeric portions in R to be handed to sparklyr’s Java/Scala invocation interface as parameters with desired sorts.

As well as, whereas earlier verisons of sparklyr didn’t serialize parameters with NaN values accurately, sparklyr 1.7 preserves NaNs as anticipated in its Java/Scala invocation interface.

Different thrilling information

There are quite a few different new options, enhancements, and bug fixes made to sparklyr 1.7, all listed within the file of the sparklyr repo and documented in sparklyr’s HTML reference pages. Within the curiosity of brevity, we is not going to describe all of them in nice element inside this weblog put up.


In chronological order, we want to thank the next people who’ve authored or co-authored pull requests that have been a part of the sparklyr 1.7 launch:

We’re additionally extraordinarily grateful to everybody who has submitted characteristic requests or bug experiences, lots of which have been tremendously useful in shaping sparklyr into what it’s right this moment.

Moreover, the writer of this weblog put up is indebted to @skeydan for her superior editorial solutions. With out her insights about good writing and story-telling, expositions like this one would have been much less readable.

In case you want to be taught extra about sparklyr, we advocate visiting,, and likewise studying some earlier sparklyr launch posts equivalent to sparklyr 1.6 and sparklyr 1.5.

That’s all. Thanks for studying!

Databricks, Inc. 2019. Deep Studying Pipelines for Apache Spark (model 1.5.0). deal/databricks/spark-deep-learning.
Elson, Jeremy, John (JD) Douceur, Jon Howell, and Jared Saul. 2007. “Asirra: A CAPTCHA That Exploits Curiosity-Aligned Guide Picture Categorization.” In Proceedings of 14th ACM Convention on Pc and Communications Safety (CCS), Proceedings of 14th ACM Convention on Pc and Communications Safety (CCS). Affiliation for Computing Equipment, Inc.
Szegedy, Christian, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. 2015. “Going Deeper with Convolutions.” In Pc Imaginative and prescient and Sample Recognition (CVPR).


Leave a Reply

Your email address will not be published. Required fields are marked *