Greater-order Features, Avro and Customized Serializers


sparklyr 1.3 is now obtainable on CRAN, with the next main new options:

To put in sparklyr 1.3 from CRAN, run

On this publish, we will spotlight some main new options launched in sparklyr 1.3, and showcase eventualities the place such options come in useful. Whereas a variety of enhancements and bug fixes (particularly these associated to spark_apply(), Apache Arrow, and secondary Spark connections) have been additionally an necessary a part of this launch, they won’t be the subject of this publish, and it is going to be a simple train for the reader to search out out extra about them from the sparklyr NEWS file.

Greater-order Features

Greater-order features are built-in Spark SQL constructs that enable user-defined lambda expressions to be utilized effectively to advanced knowledge sorts equivalent to arrays and structs. As a fast demo to see why higher-order features are helpful, let’s say at some point Scrooge McDuck dove into his enormous vault of cash and located giant portions of pennies, nickels, dimes, and quarters. Having an impeccable style in knowledge constructions, he determined to retailer the portions and face values of every little thing into two Spark SQL array columns:


sc <- spark_connect(grasp = "native", model = "2.4.5")
coins_tbl <- copy_to(
    portions = listing(c(4000, 3000, 2000, 1000)),
    values = listing(c(1, 5, 10, 25))

Thus declaring his internet price of 4k pennies, 3k nickels, 2k dimes, and 1k quarters. To assist Scrooge McDuck calculate the whole worth of every sort of coin in sparklyr 1.3 or above, we are able to apply hof_zip_with(), the sparklyr equal of ZIP_WITH, to portions column and values column, combining pairs of components from arrays in each columns. As you may need guessed, we additionally have to specify the best way to mix these components, and what higher strategy to accomplish that than a concise one-sided components   ~ .x * .y   in R, which says we wish (amount * worth) for every sort of coin? So, we have now the next:

result_tbl <- coins_tbl %>%
  hof_zip_with(~ .x * .y, dest_col = total_values) %>%

result_tbl %>% dplyr::pull(total_values)
[1]  4000 15000 20000 25000

With the end result 4000 15000 20000 25000 telling us there are in whole $40 {dollars} price of pennies, $150 {dollars} price of nickels, $200 {dollars} price of dimes, and $250 {dollars} price of quarters, as anticipated.

Utilizing one other sparklyr perform named hof_aggregate(), which performs an AGGREGATE operation in Spark, we are able to then compute the online price of Scrooge McDuck primarily based on result_tbl, storing the end in a brand new column named whole. Discover for this mixture operation to work, we have to make sure the beginning worth of aggregation has knowledge sort (specifically, BIGINT) that’s per the information sort of total_values (which is ARRAY<BIGINT>), as proven beneath:

result_tbl %>%
  dplyr::mutate(zero = dplyr::sql("CAST (0 AS BIGINT)")) %>%
  hof_aggregate(begin = zero, ~ .x + .y, expr = total_values, dest_col = whole) %>%
  dplyr::choose(whole) %>%
[1] 64000

So Scrooge McDuck’s internet price is $640 {dollars}.

Different higher-order features supported by Spark SQL up to now embody remodel, filter, and exists, as documented in right here, and just like the instance above, their counterparts (specifically, hof_transform(), hof_filter(), and hof_exists()) all exist in sparklyr 1.3, in order that they are often built-in with different dplyr verbs in an idiomatic method in R.


One other spotlight of the sparklyr 1.3 launch is its built-in assist for Avro knowledge sources. Apache Avro is a broadly used knowledge serialization protocol that mixes the effectivity of a binary knowledge format with the flexibleness of JSON schema definitions. To make working with Avro knowledge sources easier, in sparklyr 1.3, as quickly as a Spark connection is instantiated with spark_connect(..., package deal = "avro"), sparklyr will robotically work out which model of spark-avro package deal to make use of with that connection, saving loads of potential complications for sparklyr customers attempting to find out the proper model of spark-avro by themselves. Much like how spark_read_csv() and spark_write_csv() are in place to work with CSV knowledge, spark_read_avro() and spark_write_avro() strategies have been carried out in sparklyr 1.3 to facilitate studying and writing Avro recordsdata by an Avro-capable Spark connection, as illustrated within the instance beneath:


# The `package deal = "avro"` possibility is just supported in Spark 2.4 or larger
sc <- spark_connect(grasp = "native", model = "2.4.5", package deal = "avro")

sdf <- sdf_copy_to(
    a = c(1, NaN, 3, 4, NaN),
    b = c(-2L, 0L, 1L, 3L, 2L),
    c = c("a", "b", "c", "", "d")

# This instance Avro schema is a JSON string that basically says all columns
# ("a", "b", "c") of `sdf` are nullable.
avro_schema <- jsonlite::toJSON(listing(
  sort = "file",
  identify = "topLevelRecord",
  fields = listing(
    listing(identify = "a", sort = listing("double", "null")),
    listing(identify = "b", sort = listing("int", "null")),
    listing(identify = "c", sort = listing("string", "null"))
), auto_unbox = TRUE)

# persist the Spark knowledge body from above in Avro format
spark_write_avro(sdf, "/tmp/knowledge.avro", as.character(avro_schema))

# after which learn the identical knowledge body again
spark_read_avro(sc, "/tmp/knowledge.avro")
# Supply: spark<knowledge> [?? x 3]
      a     b c
  <dbl> <int> <chr>
  1     1    -2 "a"
  2   NaN     0 "b"
  3     3     1 "c"
  4     4     3 ""
  5   NaN     2 "d"

Customized Serialization

Along with generally used knowledge serialization codecs equivalent to CSV, JSON, Parquet, and Avro, ranging from sparklyr 1.3, personalized knowledge body serialization and deserialization procedures carried out in R may also be run on Spark employees through the newly carried out spark_read() and spark_write() strategies. We will see each of them in motion by a fast instance beneath, the place saveRDS() is known as from a user-defined author perform to save lots of all rows inside a Spark knowledge body into 2 RDS recordsdata on disk, and readRDS() is known as from a user-defined reader perform to learn the information from the RDS recordsdata again to Spark:


sc <- spark_connect(grasp = "native")
sdf <- sdf_len(sc, 7)
paths <- c("/tmp/file1.RDS", "/tmp/file2.RDS")

spark_write(sdf, author = perform(df, path) saveRDS(df, path), paths = paths)
spark_read(sc, paths, reader = perform(path) readRDS(path), columns = c(id = "integer"))
# Supply: spark<?> [?? x 1]
1     1
2     2
3     3
4     4
5     5
6     6
7     7

Different Enhancements


Sparklyr.flint is a sparklyr extension that goals to make functionalities from the Flint time-series library simply accessible from R. It’s at the moment underneath lively growth. One piece of excellent information is that, whereas the unique Flint library was designed to work with Spark 2.x, a barely modified fork of it should work properly with Spark 3.0, and inside the present sparklyr extension framework. sparklyr.flint can robotically decide which model of the Flint library to load primarily based on the model of Spark it’s related to. One other bit of excellent information is, as beforehand talked about, sparklyr.flint doesn’t know an excessive amount of about its personal future but. Perhaps you possibly can play an lively half in shaping its future!

EMR 6.0

This launch additionally encompasses a small however necessary change that permits sparklyr to appropriately hook up with the model of Spark 2.4 that’s included in Amazon EMR 6.0.

Beforehand, sparklyr robotically assumed any Spark 2.x it was connecting to was constructed with Scala 2.11 and tried to load any required Scala artifacts constructed with Scala 2.11 as properly. This turned problematic when connecting to Spark 2.4 from Amazon EMR 6.0, which is constructed with Scala 2.12. Ranging from sparklyr 1.3, such drawback may be mounted by merely specifying scala_version = "2.12" when calling spark_connect() (e.g., spark_connect(grasp = "yarn-client", scala_version = "2.12")).

Spark 3.0

Final however not least, it’s worthwhile to say sparklyr 1.3.0 is understood to be absolutely appropriate with the just lately launched Spark 3.0. We extremely suggest upgrading your copy of sparklyr to 1.3.0 if you happen to plan to have Spark 3.0 as a part of your knowledge workflow in future.


In chronological order, we wish to thank the next people for submitting pull requests in direction of sparklyr 1.3:

We’re additionally grateful for precious enter on the sparklyr 1.3 roadmap, #2434, and #2551 from [@javierluraschi](, and nice non secular recommendation on #1773 and #2514 from @mattpollock and @benmwhite.

Please be aware if you happen to imagine you’re lacking from the acknowledgement above, it could be as a result of your contribution has been thought of a part of the following sparklyr launch reasonably than half of the present launch. We do make each effort to make sure all contributors are talked about on this part. In case you imagine there’s a mistake, please be happy to contact the writer of this weblog publish through e-mail (yitao at rstudio dot com) and request a correction.

When you want to study extra about sparklyr, we suggest visiting,, and a number of the earlier launch posts equivalent to sparklyr 1.2 and sparklyr 1.1.

Thanks for studying!


Leave a Reply

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