[SPARK-54367][Connect]Propagate close() from SessionState to CatalogPlugins to prevent leak

[MonkeyCanCode]

What changes were proposed in this pull request?

To fix this, I added the following changes:

1. Apache Iceberg PR(Fix ClassLoader leak in Spark caused by AuthSessionCache and ensure resource cleanup iceberg#14590) to implement closable catalog
2. Make CatalogPlugin interface extends java.io.Closeable
3. Implement a close() method in CatalogManager that iterates through all registered catalogs and calls their close() method
4. Make SessionState implements Closeable and calls catalogManager.close() from its close() method.
5. Invoke session.sessionState.close() from SessionHolder.close() when a Spark Connect session is stopped

Above changes create a clean lifecycle for catalogs when a session ended, a close() call is propagated down the chain which allow each catalog to release its resources.

Why are the changes needed?

Spark Connect server is leaking SparkSession objects each time a client connects and disconnects when dealing with Apache Iceberg Apache Iceberg PR(apache/iceberg#14590).

The SessionHolder.close() method in Spark Connect is responsible for cleaning up a session. It does perform some cleanup such as artifacts and streaming queries but it doesn't perform cleanup on the main SessionState. This is where the CatalogManager lives which holds reference to RESTCatalog and S3FileIO. Since the SessionState is never closed, these Closeable catalogs are never closed and their threads leak.

Does this PR introduce any user-facing change?

N/A

How was this patch tested?

I have a local setup which can easily reproduce this issue. Here is setup details:

REST catalog: Apache Polaris (created the basic polaris entities via getting start example)
Spark Connect server:

1. public released Spark distribution to show this issue is there and we have leaks
2. local build with changes in this PR
   Spark Connect client: install public released apache spark package via pip

Testing config:

1. To make the testing easy, I set spark.connect.session.manager.defaultSessionTimeout from default 60m to 1m

Testing:

1. Check heap dump from spark UI for instance of org.apache.spark.sql.classic.SparkSession and org.apache.spark.sql.internal.SessionState
2. Make a connection to an Iceberg REST catalog and perform close() on spark session implicitly:

import uuid
from pyspark.sql import SparkSession

USER_CLIENT_ID = "xxxxx"
USER_CLIENT_SECRET = "xxxxx"

# Use `create` with defined `session_id` to ensure we will be getting a new session
session_id = str(uuid.uuid4())
spark = (
    SparkSession.builder
    .appName("Iceberg REST Catalog Quickstart")
    .remote(f"sc://localhost/;session_id={session_id}")
    .config("spark.sql.extensions", "org.apache.iceberg.spark.extensions.IcebergSparkSessionExtensions")
    .config("spark.sql.catalog.quickstart_catalog", "org.apache.iceberg.spark.SparkCatalog")
    .config("spark.sql.catalog.quickstart_catalog.catalog-impl", "org.apache.iceberg.rest.RESTCatalog")
    .config("spark.sql.catalog.quickstart_catalog.uri", "http://localhost:8181/api/catalog")
    .config("spark.sql.catalog.quickstart_catalog.warehouse", "quickstart_catalog")
    .config("spark.sql.catalog.quickstart_catalog.header.X-Iceberg-Access-Delegation", "vended-credentials")
    .config("spark.sql.catalog.quickstart_catalog.credential", f"{USER_CLIENT_ID}:{USER_CLIENT_SECRET}")
    .config("spark.sql.catalog.quickstart_catalog.scope", "PRINCIPAL_ROLE:ALL")
    .config("spark.sql.catalog.quickstart_catalog.token-refresh-enabled", "true")
    .config("spark.sql.catalog.quickstart_catalog.client.region", "us-west-2")
    .create()
)

# Setup basic namespace and perform a read/write:
spark.sql("USE quickstart_catalog").show()
spark.sql("CREATE NAMESPACE IF NOT EXISTS quickstart_namespace").show()
spark.sql("CREATE NAMESPACE IF NOT EXISTS quickstart_namespace.schema").show()
spark.sql("USE NAMESPACE quickstart_namespace.schema").show()
spark.sql("CREATE TABLE IF NOT EXISTS quickstart_table (id BIGINT, data STRING) USING ICEBERG").show()
spark.sql("INSERT INTO quickstart_table values (1, 'a')").show()
spark.sql("SELECT * FROM quickstart_table").show()

# Stop the connection
spark.stop()

3. Check heap dump from spark UI for instance of org.apache.spark.sql.classic.SparkSession and org.apache.spark.sql.internal.SessionState again and noticed resources are not getting cleanup
4. To make it more obvious, spin up 300 sessions and let the cleaner context stop them (with overwrite cleaning interval to 1m):

for _ in range (300):
    session_id = str(uuid.uuid4())
    spark = (
        SparkSession.builder
        .appName("Iceberg REST Catalog Quickstart")
        .remote(f"sc://localhost/;session_id={session_id}")
        .config("spark.sql.extensions", "org.apache.iceberg.spark.extensions.IcebergSparkSessionExtensions")
        .config("spark.sql.catalog.quickstart_catalog", "org.apache.iceberg.spark.SparkCatalog")
        .config("spark.sql.catalog.quickstart_catalog.catalog-impl", "org.apache.iceberg.rest.RESTCatalog")
        .config("spark.sql.catalog.quickstart_catalog.uri", "http://localhost:8181/api/catalog")
        .config("spark.sql.catalog.quickstart_catalog.warehouse", "quickstart_catalog")
        .config("spark.sql.catalog.quickstart_catalog.header.X-Iceberg-Access-Delegation", "vended-credentials")
        .config("spark.sql.catalog.quickstart_catalog.credential", f"{USER_CLIENT_ID}:{USER_CLIENT_SECRET}")
        .config("spark.sql.catalog.quickstart_catalog.scope", "PRINCIPAL_ROLE:ALL")
        .config("spark.sql.catalog.quickstart_catalog.token-refresh-enabled", "true")
        .config("spark.sql.catalog.quickstart_catalog.client.region", "us-west-2")
        .create()
    )
    spark.sql("USE quickstart_catalog").show()
    spark.sql("USE NAMESPACE quickstart_namespace.schema").show()
    spark.sql("SELECT * FROM quickstart_table").show()

6. Wait for 2-3 mins (the cleaning should start happen after 1m but it may take an extra 20-30 seconds for the cleaning to complete) then check heap dump from spark UI for instance of org.apache.spark.sql.classic.SparkSession and org.apache.spark.sql.internal.SessionState again. We will noticed the instances of these classes (along with many others) are not getting cleanup with current code. Also, heap usage will stay high and not able to garbage collected.
7. Now to test the fixed, get a local build of iceberg spark runtime jar from the PR above and a local build of apache spark from this PR, and repeat the same tests listed above. This time, we will see resources getting cleanup properly and heap usage decreased after cleanup.

Was this patch authored or co-authored using generative AI tooling?

No

[MonkeyCanCode]

Hello @dongjoon-hyun, I am new to the spark community and I had being using spark connect for a while for work. I found out a memory leak in spark connect while dealing with iceberg. I fixed the memory leak locally (it requires changes from both iceberg and spark). Do you mind take a look to see if this is something we should add for OSS when you get a chance? Thanks.

[hvanhovell]

We need to make sure that the session manager supports cloning.

[MonkeyCanCode]

Thanks for the review and feedback. I just want to make sure I fully understand the requirement here. When you mention this, are you thinking of the spark.newSession() case? My main concern is avoid issues where cloning one session could affected a cloned session as now the catalogs may soon be closable as well (https://github.com/apache/iceberg/pull/14590/files#diff-e263afc81b5b90944281c544005466e8671e7756f52b5b59b67a77e7c63aaf13).

Do you mind point me in the right direction for the implementation you have in mind?

Again, thanks so much for your review and time.

[hvanhovell]

How is this a problem now?

[MonkeyCanCode]

This was added due to failed unit tests with mock object where i was getting class cast exception earlier (sample error can be found in https://github.com/MonkeyCanCode/spark/actions/runs/19394233457/job/55492018508).

Here is the snippet:

[info] org.apache.spark.sql.connect.pipelines.EndToEndAPISuite *** ABORTED *** (81 milliseconds)
[info]   java.lang.ClassCastException: class io.grpc.stub.StreamObserver$MockitoMock$lemjBfIA cannot be cast to class org.apache.spark.sql.connect.execution.ExecuteResponseObserver (io.grpc.stub.StreamObserver$MockitoMock$lemjBfIA and org.apache.spark.sql.connect.execution.ExecuteResponseObserver are in unnamed module of loader 'app')
[info]   at org.apache.spark.sql.connect.service.SparkConnectListenerBusListener.sendResultComplete(SparkConnectListenerBusListener.scala:140)
[info]   at org.apache.spark.sql.connect.service.ServerSideListenerHolder.cleanUp(SparkConnectListenerBusListener.scala:84)
[info]   at org.apache.spark.sql.connect.service.SessionHolder.close(SessionHolder.scala:370)
[info]   at org.apache.spark.sql.connect.service.SparkConnectSessionManager.shutdownSessionHolder(SparkConnectSessionManager.scala:232)
[info]   at org.apache.spark.sql.connect.service.SparkConnectSessionManager.$anonfun$invalidateAllSessions$1(SparkConnectSessionManager.scala:369)
[info]   at java.base/java.util.concurrent.ConcurrentHashMap.forEach(ConcurrentHashMap.java:1603)
[info]   at org.apache.spark.sql.connect.service.SparkConnectSessionManager.invalidateAllSessions(SparkConnectSessionManager.scala:367)
[info]   at org.apache.spark.sql.connect.SparkConnectServerTest.clearAllExecutions(SparkConnectServerTest.scala:83)
[info]   at org.apache.spark.sql.connect.SparkConnectServerTest.clearAllExecutions$(SparkConnectServerTest.scala:80)
[info]   at org.apache.spark.sql.connect.pipelines.EndToEndAPISuite.clearAllExecutions(EndToEndAPISuite.scala:37)
[info]   at org.apache.spark.sql.connect.SparkConnectServerTest.beforeEach(SparkConnectServerTest.scala:72)
[info]   at org.apache.spark.sql.connect.SparkConnectServerTest.beforeEach$(SparkConnectServerTest.scala:70)
[info]   at org.apache.spark.sql.connect.pipelines.EndToEndAPISuite.beforeEach(EndToEndAPISuite.scala:37)
[info]   at org.scalatest.BeforeAndAfterEach.runTest(BeforeAndAfterEach.scala:233)
[info]   at org.scalatest.BeforeAndAfterEach.runTest$(BeforeAndAfterEach.scala:227)
[info]   at org.apache.spark.SparkFunSuite.runTest(SparkFunSuite.scala:68)
[info]   at org.scalatest.funsuite.AnyFunSuiteLike.$anonfun$runTests$1(AnyFunSuiteLike.scala:269)
[info]   at org.scalatest.SuperEngine.$anonfun$runTestsInBranch$1(Engine.scala:413)
[info]   at scala.collection.immutable.List.foreach(List.scala:323)
[info]   at org.scalatest.SuperEngine.traverseSubNodes$1(Engine.scala:401)
[info]   at org.scalatest.SuperEngine.runTestsInBranch(Engine.scala:396)
[info]   at org.scalatest.SuperEngine.runTestsImpl(Engine.scala:475)
[info]   at org.scalatest.funsuite.AnyFunSuiteLike.runTests(AnyFunSuiteLike.scala:269)
[info]   at org.scalatest.funsuite.AnyFunSuiteLike.runTests$(AnyFunSuiteLike.scala:268)
[info]   at org.scalatest.funsuite.AnyFunSuite.runTests(AnyFunSuite.scala:1564)
[info]   at org.scalatest.Suite.run(Suite.scala:1114)
[info]   at org.scalatest.Suite.run$(Suite.scala:1096)
[info]   at org.scalatest.funsuite.AnyFunSuite.org$scalatest$funsuite$AnyFunSuiteLike$$super$run(AnyFunSuite.scala:1564)
[info]   at org.scalatest.funsuite.AnyFunSuiteLike.$anonfun$run$1(AnyFunSuiteLike.scala:273)
[info]   at org.scalatest.SuperEngine.runImpl(Engine.scala:535)
[info]   at org.scalatest.funsuite.AnyFunSuiteLike.run(AnyFunSuiteLike.scala:273)
[info]   at org.scalatest.funsuite.AnyFunSuiteLike.run$(AnyFunSuiteLike.scala:272)
[info]   at org.apache.spark.SparkFunSuite.org$scalatest$BeforeAndAfterAll$$super$run(SparkFunSuite.scala:68)
[info]   at org.scalatest.BeforeAndAfterAll.liftedTree1$1(BeforeAndAfterAll.scala:213)
[info]   at org.scalatest.BeforeAndAfterAll.run(BeforeAndAfterAll.scala:210)
[info]   at org.scalatest.BeforeAndAfterAll.run$(BeforeAndAfterAll.scala:208)
[info]   at org.apache.spark.SparkFunSuite.run(SparkFunSuite.scala:68)
[info]   at org.scalatest.tools.Framework.org$scalatest$tools$Framework$$runSuite(Framework.scala:321)
[info]   at org.scalatest.tools.Framework$ScalaTestTask.execute(Framework.scala:517)
[info]   at sbt.ForkMain$Run.lambda$runTest$1(ForkMain.java:414)
[info]   at java.base/java.util.concurrent.FutureTask.run(FutureTask.java:264)
[info]   at java.base/java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1136)
[info]   at java.base/java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:635)
[info]   at java.base/java.lang.Thread.run(Thread.java:840)

Before this PR, the session cleanup process in the tests was less explicit. Now with proper resources closed, the generic test StreamObserver is causing issue during casting with existed code.

[hvanhovell]

Why is this needed now?

[MonkeyCanCode]

Earlier tests create dummy session that are not yet fully initialized. Now as all sessions got properly closed, calling closed on not initialized sessions was causing failed tests as well.