Apache Spark Connect Client for Rust

This project houses the experimental client for Spark Connect for Apache Spark written in Rust

Current State of the Project

Currently, the Spark Connect client for Rust is highly experimental and should not be used in any production setting. This is currently a "proof of concept" to identify the methods of interacting with Spark cluster from rust.

The spark-connect-rs aims to provide an entrypoint to Spark Connect, and provide similar DataFrame API interactions.

Project Layout

├── core       <- core implementation in Rust
│   └─ spark   <- git submodule for apache/spark
├── rust       <- shim for 'spark-connect-rs' from core
├── examples   <- examples of using different aspects of the crate
├── datasets   <- sample files from the main spark repo

Future state would be to have additional bindings for other languages along side the top level rust folder.

Getting Started

This section explains how run Spark Connect Rust locally starting from 0.

Step 1: Install rust via rustup: https://www.rust-lang.org/tools/install

Step 2: Ensure you have a cmake and protobuf installed on your machine

Step 3: Run the following commands to clone the repo

git clone https://github.com/sjrusso8/spark-connect-rs.git
git submodule update --init --recursive

cargo build

Step 4: Setup the Spark Driver on localhost either by downloading spark or with docker.

With local spark:

1. Download Spark distribution (3.5.1 recommended), unzip the package.

2. Set your SPARK_HOME environment variable to the location where spark was extracted to,

3. Start the Spark Connect server with the following command (make sure to use a package version that matches your Spark distribution):

$ $SPARK_HOME/sbin/start-connect-server.sh --packages "org.apache.spark:spark-connect_2.12:3.5.1,io.delta:delta-spark_2.12:3.0.0" \
      --conf "spark.driver.extraJavaOptions=-Divy.cache.dir=/tmp -Divy.home=/tmp" \
      --conf "spark.sql.extensions=io.delta.sql.DeltaSparkSessionExtension" \
      --conf "spark.sql.catalog.spark_catalog=org.apache.spark.sql.delta.catalog.DeltaCatalog"

With docker:

1. Start the Spark Connect server by leveraging the created docker-compose.yml in this repo. This will start a Spark Connect Server running on port 15002

$ docker compose up --build -d

Step 5: Run an example from the repo under /examples

Features

The following section outlines some of the larger functionality that are not yet working with this Spark Connect implementation.

• TLS authentication & Databricks compatability via the feature flag feature = 'tls'
• StreamingQueryManager
• UDFs or any type of functionality that takes a closure (foreach, foreachBatch, etc.)

SparkSession

Spark Session type object and its implemented traits

SparkSession	API	Comment
active
addArtifact(s)
addTag
clearTags
copyFromLocalToFs
createDataFrame		Partial. Only works for RecordBatch
getActiveSessions
getTags
interruptAll		splitn
interruptOperation
interruptTag
newSession
range
removeTag
sql
stop
table
catalog		Catalog
client		unstable developer api for testing only
conf		Conf
read		DataFrameReader
readStream		DataStreamReader
streams		Streams
udf		Udf - may not be possible
udtf		Udtf - may not be possible
version

SparkSessionBuilder

SparkSessionBuilder	API	Comment
appName
config
master
remote		Validate using spark connection string

StreamingQueryManager

StreamingQueryManager	API	Comment
awaitAnyTermination
get
resetTerminated
active

StreamingQuery

StreamingQuery	API	Comment
awaitTermination
exception
explain
processAllAvailable
stop
id
isActive
lastProgress
name
recentProgress
runId
status

DataStreamReader

DataStreamReader	API	Comment
csv
format
json
load
option
options
orc
parquet
schema
table
text

DataFrameReader

DataFrameReader	API	Comment
csv
format
json
load
option
options
orc
parquet
schema
table
text

DataStreamWriter

Start a streaming job and return a StreamingQuery object to handle the stream operations.

DataStreamWriter	API	Comment
foreach
foreachBatch
format
option
options
outputMode		Uses an Enum for OutputMode
partitionBy
queryName
start
toTable
trigger		Uses an Enum for TriggerMode

StreamingQueryListener

StreamingQueryListener	API	Comment
onQueryIdle
onQueryProgress
onQueryStarted
onQueryTerminated

UdfRegistration (may not be possible)

UDFRegistration	API	Comment
register
registerJavaFunction
registerJavaUDAF

UdtfRegistration (may not be possible)

UDTFRegistration	API	Comment
register

RuntimeConfig

RuntimeConfig	API	Comment
get
isModifiable
set
unset

Catalog

Catalog	API	Comment
cacheTable
clearCache
createExternalTale
createTable
currentCatalog
currentDatabase
databaseExists
dropGlobalTempView
dropTempView
functionExists
getDatabase
getFunction
getTable
isCached
listCatalogs
listDatabases
listFunctions
listTables
recoverPartitions
refreshByPath
refreshTable
registerFunction
setCurrentCatalog
setCurrentDatabase
tableExists
uncacheTable

DataFrame

Spark DataFrame type object and its implemented traits.

DataFrame	API	Comment
agg
alias
approxQuantile
cache
checkpoint
coalesce
colRegex
collect
columns
corr
count
cov
createGlobalTempView
createOrReplaceGlobalTempView
createOrReplaceTempView
createTempView
crossJoin
crosstab
cube
describe
distinct
drop
dropDuplicates
dropDuplicatesWithinWatermark		Windowing functions are currently in progress
drop_duplicates
dropna
dtypes
exceptAll
explain
fillna
filter
first
foreach
foreachPartition
freqItems
groupBy
head
hint
inputFiles
intersect
intersectAll
isEmpty
isLocal
isStreaming
join
limit
localCheckpoint
mapInPandas		TBD on this exact implementation
mapInArrow		TBD on this exact implementation
melt
na
observe
offset
orderBy
persist
printSchema
randomSplit
registerTempTable
repartition
repartitionByRange
replace
rollup
sameSemantics
sample
sampleBy
schema
select
selectExpr
semanticHash
show
sort
sortWithinPartitions
sparkSession
stat
storageLevel
subtract
summary
tail
take
to
toDF
toJSON		Does not return an RDD but a long JSON formatted String
toLocalIterator
toPandas to_polars & toPolars		Convert to a polars::frame::DataFrame
new to_datafusion & toDataFusion		Convert to a datafusion::dataframe::DataFrame
transform
union
unionAll
unionByName
unpersist
unpivot
where		use filter instead, where is a keyword for rust
withColumn
withColumns
withColumnRenamed
withColumnsRenamed
withMetadata
withWatermark
write
writeStream
writeTo

DataFrameWriter

Spark Connect should respect the format as long as your cluster supports the specified type and has the required jars

DataFrameWriter	API	Comment
bucketBy
csv
format
insertInto
jdbc
json
mode
option
options
orc
parquet
partitionBy
save
saveAsTable
sortBy
text

DataFrameWriterV2

DataFrameWriterV2	API	Comment
append
create
createOrReplace
option
options
overwrite
overwritePartitions
partitionedBy
replace
tableProperty
using

Column

Spark Column type object and its implemented traits

Column	API	Comment
alias
asc
asc_nulls_first
asc_nulls_last
astype
between
cast
contains
desc
desc_nulls_first
desc_nulls_last
dropFields
endswith
eqNullSafe
getField		This is depreciated but will need to be implemented
getItem		This is depreciated but will need to be implemented
ilike
isNotNull
isNull
isin
like
name
otherwise
over		Refer to Window for creating window specifications
rlike
startswith
substr
when
withField
eq ==		Rust does not like when you try to overload == and return something other than a bool. Currently implemented column equality like col('name').eq(col('id')). Not the best, but it works for now
addition +
subtration -
multiplication *
division /
OR |
AND &
XOR ^
Negate ~

Functions

Only a few of the functions are covered by unit tests.

Functions	API	Comment
abs
acos
acosh
add_months
aggregate
approxCountDistinct
approx_count_distinct
array
array_append
array_compact
array_contains
array_distinct
array_except
array_insert
array_intersect
array_join
array_max
array_min
array_position
array_remove
array_repeat
array_sort
array_union
arrays_overlap
arrays_zip
asc
asc_nulls_first
asc_nulls_last
ascii
asin
asinh
assert_true
atan
atan2
atanh
avg
base64
bin
bit_length
bitwiseNOT
bitwise_not
broadcast
bround
bucket
call_udf
cbrt
ceil
coalesce
col
collect_list
collect_set
column
concat
concat_ws
conv
corr
cos
cosh
cot
count
countDistinct
count_distinct
covar_pop
covar_samp
crc32
create_map
csc
cume_dist
current_date
current_timestamp
date_add
date_format
date_sub
date_trunc
datediff
dayofmonth
dayofweek
dayofyear
days
decode
degrees
dense_rank
desc
desc_nulls_first
desc_nulls_last
element_at
encode
exists
exp
explode
explode_outer
expm1
expr
factorial
filter
first
flatten
floor
forall
format_number
format_string
from_csv
from_json
from_unixtime
from_utc_timestamp
functools
get
get_active_spark_context
get_json_object
greatest
grouping
grouping_id
has_numpy
hash
hex
hour
hours
hypot
initcap
inline
inline_outer
input_file_name
inspect
instr
isnan
isnull
json_tuple
kurtosis
lag
last
last_day
lead
least
length
levenshtein
lit
localtimestamp
locate
log
log10
log1p
log2
lower
lpad
ltrim
make_date
map_concat
map_contains_key
map_entries
map_filter
map_from_arrays
map_from_entries
map_keys
map_values
map_zip_with
max
max_by
md5
mean
median
min
min_by
minute
mode
monotonically_increasing_id
month
months
months_between
nanvl
next_day
np
nth_value
ntile
octet_length
overlay
overload
pandas_udf
percent_rank
percentile_approx
pmod
posexplode
posexplode_outer
pow
product
quarter
radians
raise_error
rand
randn
rank
regexp_extract
regexp_replace
repeat
reverse
rint
round
row_number
rpad
rtrim
schema_of_csv
schema_of_json
sec
second
sentences
sequence
session_window
sha1
sha2
shiftLeft
shiftRight
shiftRightUnsigned
shiftleft
shiftright
shiftrightunsigned
shuffle
signum
sin
sinh
size
skewness
slice
sort_array
soundex
spark_partition_id
split
sqrt
stddev
stddev_pop
stddev_samp
struct
substring
substring_index
sum
sumDistinct
sum_distinct
sys
tan
tanh
timestamp_seconds
toDegrees
toRadians
to_csv
to_date
to_json
to_str
to_timestamp
to_utc_timestamp
transform
transform_keys
transform_values
translate
trim
trunc
try_remote_functions
udf
unbase64
unhex
unix_timestamp
unwrap_udt
upper
var_pop
var_samp
variance
warnings
weekofyear
when
window
window_time
xxhash64
year
years
zip_with

Data Types

Data types are used for creating schemas and for casting columns to specific types

Column	API	Comment
ArrayType
BinaryType
BooleanType
ByteType
DateType
DecimalType
DoubleType
FloatType
IntegerType
LongType
MapType
NullType
ShortType
StringType
CharType
VarcharType
StructField
StructType
TimestampType
TimestampNTZType
DayTimeIntervalType
YearMonthIntervalType

Literal Types

Create Spark literal types from these rust types. E.g. lit(1_i64) would be a LongType() in the schema.

An array can be made like lit([1_i16,2_i16,3_i16]) would result in an ArrayType(Short) since all the values of the slice can be translated into literal type.

Spark Literal Type	Rust Type	Status
Null
Binary	&[u8]
Boolean	bool
Byte
Short	i16
Integer	i32
Long	i64
Float	f32
Double	f64
Decimal
String	&str / String
Date	chrono::NaiveDate
Timestamp	chrono::DateTime<Tz>
TimestampNtz	chrono::NaiveDateTime
CalendarInterval
YearMonthInterval
DayTimeInterval
Array	slice / Vec
Map	Create with the function create_map
Struct	Create with the function struct_col or named_struct

Window & WindowSpec

For ease of use it's recommended to use Window to create the WindowSpec.

Window	API	Comment
currentRow
orderBy
partitionBy
rangeBetween
rowsBetween
unboundedFollowing
unboundedPreceding
WindowSpec.orderBy
WindowSpec.partitionBy
WindowSpec.rangeBetween
WindowSpec.rowsBetween