State management extensions for Apache Spark to keep data across micro-batches during stateful stream processing.
Out of the box, Apache Spark has only one implementation of state store providers. It's HDFSBackedStateStoreProvider
which stores all of the data in memory, what is a very memory consuming approach. To avoid OutOfMemory
errors, this repository and custom state store providers were created.
To use the custom state store provider for your pipelines use the following additional configuration for the submit script/ SparkConf:
--conf spark.sql.streaming.stateStore.providerClass="ru.chermenin.spark.sql.execution.streaming.state.RocksDbStateStoreProvider"
Here is some more information about it: https://docs.databricks.com/spark/latest/structured-streaming/production.html
Alternatively, you can use the useRocksDBStateStore()
helper method in your application while creating the SparkSession,
import ru.chermenin.spark.sql.execution.streaming.state.implicits._
val spark = SparkSession.builder().master(...).useRocksDBStateStore().getOrCreate()
Note: For the helper methods to be available, you must import the implicits as shown above.
With semantics similar to those of GroupState
/ FlatMapGroupWithState
, state timeout features have been built directly into the custom state store.
Important points to note when using State Timeouts,
- Timeouts can be set differently for each streaming query. This relies on
queryName
and itscheckpointLocation
. - The poll trigger set on a streaming query may or may not be set to a different value than the state expiration.
- Timeouts are currently based on processing time
- The timeout will occur once
- a fixed duration has elapsed after the entry's creation, or
- the most recent replacement (update) of its value, or
- its last access
- Unlike
GroupState
, the timeout is not eventual as it is independent from query progress - Since the processing time timeout is based on the clock time, it is affected by the variations in the system clock (i.e. time zone changes, clock skew, etc.)
- Timeout may or may not be set to strict expiration at the slight cost of memory. More info here.
There are 2 different ways configure state timeout:
-
Via additional configuration on SparkConf:
To set a processing time timeout for all streaming queries in strict mode.
--conf spark.sql.streaming.stateStore.stateExpirySecs=5 --conf spark.sql.streaming.stateStore.strictExpire=true
To configure state timeout differently for each query the above configs can be modified to,
--conf spark.sql.streaming.stateStore.stateExpirySecs.queryName1=5 --conf spark.sql.streaming.stateStore.stateExpirySecs.queryName2=10 ... ... --conf spark.sql.streaming.stateStore.strictExpire=true
-
Via
stateTimeout()
helper method (recommended way):import ru.chermenin.spark.sql.execution.streaming.state.implicits._ val spark: SparkSession = ... val streamingDF: DataFrame = ... streamingDF.writeStream .format(...) .outputMode(...) .trigger(Trigger.ProcessingTime(1000L)) .queryName("myQuery1") .option("checkpointLocation", "chkpntloc") .stateTimeout(spark.conf, expirySecs = 5) .start() spark.streams.awaitAnyTermination()
Preferably, the
queryName
andcheckpointLocation
can be set directly via thestateTimeout()
method, as below:streamingDF.writeStream .format(...) .outputMode(...) .trigger(Trigger.ProcessingTime(1000L)) .stateTimeout(spark.conf, queryName="myQuery1", expirySecs = 5, checkpointLocation ="chkpntloc") .start()
Note: If queryName
is invalid/ unavailable, the streaming query will be tagged as UNNAMED
and timeout applicable will be as per the value of spark.sql.streaming.stateStore.stateExpirySecs
(which defaults to -1, but can be overridden via SparkConf)
Other state timeout related points (applicable on global and query level),
- For no timeout, i.e. infinite state, set
spark.sql.streaming.stateStore.stateExpirySecs=-1
- For stateless processing, i.e. no state, set
spark.sql.streaming.stateStore.stateExpirySecs=0
You're welcome to submit pull requests with any changes for this repository at any time. I'll be very glad to see any contributions.
The standard Apache 2.0 license is used for this project.