[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17795061#comment-17795061 ] wang fanming commented on SPARK-44900: -- Of course, the following is my execution in the spark3.5.1 environment. I only extracted the relevant logs related to the production, storage, and reuse of rdd_1538 partition 1 under the Storage tab of the WEB UI. {quote}23/12/10 18:05:23 INFO MemoryStore: Block rdd_1538_1 stored as values in memory (estimated size 176.7 MiB, free 310.2 MiB) [rdd_1538_1] 23/12/10 18:05:24 INFO BlockManager: Found block rdd_1538_1 locally 23/12/10 18:05:27 INFO BlockManager: Dropping block rdd_1538_1 from memory 23/12/10 18:05:27 INFO BlockManager: Writing block rdd_1538_1 to disk 23/12/10 18:05:34 INFO MemoryStore: Block rdd_1538_1 stored as values in memory (estimated size 176.7 MiB, free 133.5 MiB) 23/12/10 18:05:34 INFO BlockManager: Found block rdd_1538_1 locally 23/12/10 18:05:40 INFO BlockManager: Found block rdd_1538_1 locally 23/12/10 18:05:42 INFO BlockManager: Dropping block rdd_1538_1 from memory 23/12/10 18:05:46 INFO MemoryStore: Block rdd_1538_1 stored as values in memory (estimated size 176.7 MiB, free 133.5 MiB) {quote} Through analysis of the above logs and combined with the UI display situation, the "Size on Disk" displayed on the Storage interface is incomprehensible. If the partitions of this RDD are cached normally, shouldn't the size under "Size on Disk" label change? > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 3.3.0 >Reporter: Varun Nalla >Priority: Major > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17794817#comment-17794817 ] Dongjoon Hyun commented on SPARK-44900: --- Could you try this with Apache Spark 3.5.0, please? > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 3.3.0 >Reporter: Varun Nalla >Priority: Major > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17792872#comment-17792872 ] 王范明 commented on SPARK-44900: - I have analyzed the program execution details in the logs and it seems that there is an issue with the 'org.apache.spark.status.AppStatusListener#updateRDDBlock' method.The method directly calculates the usage of {{rdd.memoryUsed}} and {{{}rdd.diskUsed{}}}, but it does not pay sufficient attention to the {{{}storageLevel{}}}. > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 3.3.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17760700#comment-17760700 ] Varun Nalla commented on SPARK-44900: - [~yxzhang] / [~yao] any update for us ? > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 3.3.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17759821#comment-17759821 ] Yauheni Audzeichyk commented on SPARK-44900: [~yxzhang] looks like it is just disk usage tracking issue as disk space is not used as much. However it affects effectiveness of cached data since Spark spills it to disk as it believes it doesn't fit memory anymore so eventually it becomes 100% stored on disk. > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 3.3.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17759818#comment-17759818 ] Yuexin Zhang commented on SPARK-44900: -- Hi [~varun2807] [~yaud] did you check the actual cached file size on disk, on the yarn node manager local filesystem? Is it really ever growing? > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 3.3.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17759703#comment-17759703 ] Varun Nalla commented on SPARK-44900: - [~yao] hope you got a chance to look into what [~yaud] mentioned. > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 3.3.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17759130#comment-17759130 ] Yauheni Audzeichyk commented on SPARK-44900: [~yao] we tried spark.cleaner.periodicGC.interval=1min but it didn't help. Here are my observations: * this happens even in a very simple scenario - see the example to reproduce below * it happens after join * uncontrollable growth of disk usage starts only when any portion of RDD got spilled to disk * if cached RDD remains 100% in memory this issue doesn't happen * when an executor dies then "Size on Disk" on Storage tab gets reduced by the amount of storage blocks held by that dead executor (makes sense) It looks like some storage block (shuffle blocks?) are being tracked under that cached RDD and never (or at least not in a reasonable time) released until the executor dies. Our worry is whether it is just disk size usage tracking bug or those blocks are actually kept on the disk because our production job disk usage (per Spark UI) grew by 6TB in a span of 10 hours. Here's the code to reproduce: {code:java} val conf = new SparkConf().set("spark.master", "yarn") val spark = SparkSession.builder().config(conf).getOrCreate() import spark.implicits._ val sc = spark.sparkContext val ssc = new StreamingContext(sc, Seconds(10)) // create a pseudo stream val rddQueue = new mutable.Queue[RDD[Long]]() val stream = ssc.queueStream(rddQueue, oneAtATime = true) // create a simple lookup table val lookup = sc.range(start = 0, end = 5000, numSlices = 10) .toDF("id") .withColumn("value", md5(rand().cast(StringType))) .cache() // for every micro-batch perform value lookup via join stream.foreachRDD { rdd => val df = rdd.toDF("id") df.join(lookup, Seq("id"), "leftouter").count() } // run the streaming ssc.start() for (_ <- 1 to 100) { rddQueue.synchronized { val firstId = Random.nextInt(5000) rddQueue += sc.range(start = firstId, end = firstId + 1, numSlices = 4) } Thread.sleep(10) } ssc.stop() {code} Submit parameters (selected to create storage memory deficit and cause cache to be spilled): {code:java} --executor-cores 2 --num-executors 5 --executor-memory 1250m --driver-memory 1g \ --conf spark.dynamicAllocation.enabled=false --conf spark.sql.shuffle.partitions=10 {code} When executed, disk usage of that cached lookup DF grows really fast. > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 3.3.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17758868#comment-17758868 ] Kent Yao commented on SPARK-44900: -- Set spark.cleaner.periodicGC.interval to a smaller value(3min) might help > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 3.3.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17758674#comment-17758674 ] Varun Nalla commented on SPARK-44900: - We do not keep on adding , they are existing cached RDD's which keeps on growing, tested with spark3 as well, seeing same behavior and also changed the persisting strategy to just MEMORY instead of MEMORY_AND_DISK, but no luck. > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 3.3.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17758376#comment-17758376 ] Kent Yao commented on SPARK-44900: -- Please note that there is a storage limit in place. Adding items without removing any may lead to cache evictions and recomputations. In such cases, caching may not be as effective as direct computation, as extra write paths are introduced. Probably, you should optimize your program. > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 2.4.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17758316#comment-17758316 ] Varun Nalla commented on SPARK-44900: - [~yao] Thanks for the comment. However we can't release as cached RDD's are being used in every micro batch that's why could not unpersist. > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 2.4.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17758315#comment-17758315 ] Kent Yao commented on SPARK-44900: -- How about releasing the cached rdds if you never touch it again > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 2.4.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-44900) Cached DataFrame keeps growing
[ https://issues.apache.org/jira/browse/SPARK-44900?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17758161#comment-17758161 ] Varun Nalla commented on SPARK-44900: - [~yao] , is there a way I could prioritize this issue as it's causing us production impact ? > Cached DataFrame keeps growing > -- > > Key: SPARK-44900 > URL: https://issues.apache.org/jira/browse/SPARK-44900 > Project: Spark > Issue Type: Bug > Components: Spark Core >Affects Versions: 2.4.0 >Reporter: Varun Nalla >Priority: Blocker > > Scenario : > We have a kafka streaming application where the data lookups are happening by > joining another DF which is cached, and the caching strategy is > MEMORY_AND_DISK. > However the size of the cached DataFrame keeps on growing for every micro > batch the streaming application process and that's being visible under > storage tab. > A similar stack overflow thread was already raised. > https://stackoverflow.com/questions/55601779/spark-dataframe-cache-keeps-growing -- This message was sent by Atlassian Jira (v8.20.10#820010) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org