Hi Jason,
I read your notes and the code simulating the problem as link
https://issues.apache.org/jira/browse/SPARK-38388 and the specific
repartition issue (SPARK-38388) that this code aims to demonstrate
The code below from the above link Jira
import scala.sys.process._
import org.apache.spark.TaskContext
case class TestObject(id: Long, value: Double)
val ds = spark.range(0, 1000 * 1000, 1).repartition(100,
$"id").withColumn("val", rand()).repartition(100).map {
row => if (TaskContext.get.stageAttemptNumber == 0 &&
TaskContext.get.attemptNumber == 0 && TaskContext.get.partitionId > 97) {
throw new Exception("pkill -f java".!!)
}
TestObject(row.getLong(0), row.getDouble(1))
}
ds.toDF("id", "value").write.mode("overwrite").saveAsTable("tmp.test_table")
spark.sql("select count(distinct id) from tmp.test_table").show
*contains a potential security risk* by using scala.sys.process to execute
the pkill -f java command. While the code aims to demonstrate the
repartition issue, using pkill is IMO unnecessary and risky. This could
potentially terminate critical processes on the cluster as well. Instead of
throwing an exception based on partition ID, you can try to filter out
unwanted partitions before applying the map transformation like below
val filteredDS = ds.filter($"id".lt(98)) // Filter out partitions with ID
>= 98
filteredDS.map { row => TestObject(row.getLong(0), row.getDouble(1)) }
By using filteredDS for subsequent transformations or actions, you avoid
redundant processing and potential complications from the conditional logic
in the original map transformation. This approach is a safer simulation of
the repartition issue by only working with the filtered dataset, representing
the partitions that would have hypothetically succeeded.
HTH
Mich Talebzadeh,
Dad | Technologist | Solutions Architect | Engineer
London
United Kingdom
view my Linkedin profile
<https://www.linkedin.com/in/mich-talebzadeh-ph-d-5205b2/>
https://en.everybodywiki.com/Mich_Talebzadeh
*Disclaimer:* The information provided is correct to the best of my
knowledge but of course cannot be guaranteed . It is essential to note
that, as with any advice, quote "one test result is worth one-thousand
expert opinions (Werner <https://en.wikipedia.org/wiki/Wernher_von_Braun>Von
Braun <https://en.wikipedia.org/wiki/Wernher_von_Braun>)".
On Mon, 4 Mar 2024 at 18:26, Jason Xu <jasonxu.sp...@gmail.com> wrote:
> Hi Prem,
>
> From the symptom of shuffle fetch failure and few duplicate data and few
> missing data, I think you might run into this correctness bug:
> https://issues.apache.org/jira/browse/SPARK-38388.
>
> Node/shuffle failure is hard to avoid, I wonder if you have
> non-deterministic logic and calling repartition() (round robin
> partitioning) in your code? If you can avoid either of these, you can avoid
> the issue from happening for now. To root fix the issue, it requires a
> non-trivial effort, I don't think there's a solution available yet.
>
> I have heard that there are community efforts to solve this issue, but I
> lack detailed information. Hopefully, someone with more knowledge can
> provide further insight.
>
> Best,
> Jason
>
> On Mon, Mar 4, 2024 at 9:41 AM Prem Sahoo <prem.re...@gmail.com> wrote:
>
>> super :(
>>
>> On Mon, Mar 4, 2024 at 6:19 AM Mich Talebzadeh <mich.talebza...@gmail.com>
>> wrote:
>>
>>> "... in a nutshell if fetchFailedException occurs due to data node
>>> reboot then it can create duplicate / missing data . so this is more of
>>> hardware(env issue ) rather than spark issue ."
>>>
>>> As an overall conclusion your point is correct but again the answer is
>>> not binary.
>>>
>>> Spark core relies on a distributed file system to store data across data
>>> nodes. When Spark needs to process data, it fetches the required blocks
>>> from the data nodes.* FetchFailedException*: means that Spark
>>> encountered an error while fetching data blocks from a data node. If a data
>>> node reboots unexpectedly, it becomes unavailable to Spark for a
>>> period. During this time, Spark might attempt to fetch data blocks from the
>>> unavailable node, resulting in the FetchFailedException.. Depending on the
>>> timing and nature of the reboot and data access, this exception can lead
>>> to:the following:
>>>
>>> - Duplicate Data: If Spark retries the fetch operation successfully
>>> after the reboot, it might end up processing the same data twice, leading
>>> to duplicates.
>>> - Missing Data: If Spark cannot fetch all required data blocks due
>>> to the unavailable data node, some data might be missing from the
>>> processing results.
>>>
>>> The root cause of this issue lies in the data node reboot itself. So we
>>> can conclude that it is not a problem with Spark core functionality but
>>> rather an environmental issue within the distributed storage systemL You
>>> need to ensure that your nodes are stable and minimise unexpected reboots
>>> for whatever reason. Look at the host logs or run /usr/bin/dmesg to see
>>> what happened..
>>>
>>> Good luck
>>>
>>> Mich Talebzadeh,
>>> Dad | Technologist | Solutions Architect | Engineer
>>> London
>>> United Kingdom
>>>
>>>
>>> view my Linkedin profile
>>> <https://www.linkedin.com/in/mich-talebzadeh-ph-d-5205b2/>
>>>
>>>
>>> https://en.everybodywiki.com/Mich_Talebzadeh
>>>
>>>
>>>
>>> *Disclaimer:* The information provided is correct to the best of my
>>> knowledge but of course cannot be guaranteed . It is essential to note
>>> that, as with any advice, quote "one test result is worth one-thousand
>>> expert opinions (Werner
>>> <https://en.wikipedia.org/wiki/Wernher_von_Braun>Von Braun
>>> <https://en.wikipedia.org/wiki/Wernher_von_Braun>)".
>>>
>>>
>>> On Mon, 4 Mar 2024 at 01:30, Prem Sahoo <prem.re...@gmail.com> wrote:
>>>
>>>> thanks Mich, in a nutshell if fetchFailedException occurs due to data
>>>> node reboot then it can create duplicate / missing data . so this is
>>>> more of hardware(env issue ) rather than spark issue .
>>>>
>>>>
>>>>
>>>> On Sat, Mar 2, 2024 at 7:45 AM Mich Talebzadeh <
>>>> mich.talebza...@gmail.com> wrote:
>>>>
>>>>> Hi,
>>>>>
>>>>> It seems to me that there are issues related to below
>>>>>
>>>>> *<Prem> I think when a task failed in between and retry task started
>>>>> and completed it may create duplicate as failed task has some data + retry
>>>>> task has full data. but my question is why spark keeps delta data or
>>>>> according to you if speculative and original task completes generally
>>>>> spark
>>>>> kills one of the tasks to get rid of dups data. when a data node is
>>>>> rebooted then spark fault tolerant should go to other nodes isn't it ?
>>>>> then
>>>>> why it has missing data.*
>>>>>
>>>>> Spark is designed to be fault-tolerant through lineage and
>>>>> recomputation. However, there are scenarios where speculative execution or
>>>>> task retries might lead to duplicated or missing data. So what are these?
>>>>>
>>>>> - Task Failure and Retry: You are correct that a failed task might
>>>>> have processed some data before encountering the FetchFailedException. If
>>>>> a
>>>>> retry succeeds, it would process the entire data partition again, leading
>>>>> to duplicates. When a task fails, Spark may recompute the lost data by
>>>>> recomputing the lost task on another node. The output of the retried task
>>>>> is typically combined with the output of the original task during the
>>>>> final
>>>>> stage of the computation. This combination is done to handle scenarios
>>>>> where the original task partially completed and generated some output
>>>>> before failing. Spark does not intentionally store partially processed
>>>>> data. However, due to retries and speculative execution, duplicate
>>>>> processing can occur. To the best of my knowledge, Spark itself doesn't
>>>>> have a mechanism to identify and eliminate duplicates automatically. While
>>>>> Spark might sometimes kill speculative tasks if the original one finishes,
>>>>> it is not a guaranteed behavior. This depends on various factors like
>>>>> scheduling and task dependencies.
>>>>>
>>>>> - Speculative Execution: Spark supports speculative execution, where
>>>>> the same task is launched on multiple executors simultaneously. The result
>>>>> of the first completed task is used, and the others are usually killed to
>>>>> avoid duplicated results. However, speculative execution might introduce
>>>>> some duplication in the final output if tasks on different executors
>>>>> complete successfully.
>>>>>
>>>>> - Node Reboots and Fault Tolerance: If the data node reboot leads to
>>>>> data corruption or loss, that data might be unavailable to Spark. Even
>>>>> with
>>>>> fault tolerance, Spark cannot recover completely missing data. Fault
>>>>> tolerance focuses on recovering from issues like executor failures, not
>>>>> data loss on storage nodes. Overall, Spark's fault tolerance is designed
>>>>> to
>>>>> handle executor failures by rescheduling tasks on other available
>>>>> executors
>>>>> and temporary network issues by retrying fetches based on configuration.
>>>>>
>>>>> Here are some stuff to consider:
>>>>>
>>>>> - Minimize retries: Adjust spark.shuffle.io.maxRetries to a lower
>>>>> value such as 1 or 2 to reduce the chance of duplicate processing
>>>>> attempts, if retries are suspected to be a source.
>>>>> - Disable speculative execution if needed: Consider disabling
>>>>> speculative execution (spark.speculation=false) if duplicates are a major
>>>>> concern. However, this might impact performance.
>>>>> - Data persistence: As mentioned in the previous reply, persist
>>>>> intermediate data to reliable storage (HDFS, GCS, etc.) if data integrity
>>>>> is critical. This ensures data availability even during node failures.
>>>>> - Data validation checks: Implement data validation checks after
>>>>> processing to identify potential duplicates or missing data.
>>>>> HTH
>>>>> Mich Talebzadeh,
>>>>> Dad | Technologist | Solutions Architect | Engineer
>>>>> London
>>>>> United Kingdom
>>>>>
>>>>>
>>>>> view my Linkedin profile
>>>>> <https://www.linkedin.com/in/mich-talebzadeh-ph-d-5205b2/>
>>>>>
>>>>>
>>>>> https://en.everybodywiki.com/Mich_Talebzadeh
>>>>>
>>>>>
>>>>>
>>>>> *Disclaimer:* The information provided is correct to the best of my
>>>>> knowledge but of course cannot be guaranteed . It is essential to note
>>>>> that, as with any advice, quote "one test result is worth one-thousand
>>>>> expert opinions (Werner
>>>>> <https://en.wikipedia.org/wiki/Wernher_von_Braun>Von Braun
>>>>> <https://en.wikipedia.org/wiki/Wernher_von_Braun>)".
>>>>>
>>>>>
>>>>> On Sat, 2 Mar 2024 at 01:43, Prem Sahoo <prem.re...@gmail.com> wrote:
>>>>>
>>>>>> Hello Mich,
>>>>>> thanks for your reply.
>>>>>>
>>>>>> As an engineer I can chip in. You may have partial execution and
>>>>>> retries meaning when spark encounters a *FetchFailedException*, it
>>>>>> may retry fetching the data from the unavailable (the one being rebooted)
>>>>>> node a few times before marking it permanently unavailable. However, if
>>>>>> the
>>>>>> rebooted node recovers quickly within this retry window, some executors
>>>>>> might successfully fetch the data after a retry. *This leads to
>>>>>> duplicate processing of the same data partition*.
>>>>>>
>>>>>> <Prem> data node reboot is taking more than 20 mins and our config
>>>>>> spark.network.timeout=300s so we don't have dupls for the above reason.
>>>>>> I am not sure this one applies to your spark version but spark may
>>>>>> speculatively execute tasks on different executors to improve
>>>>>> performance. If a task fails due to the *FetchFailedException*, a
>>>>>> speculative task might be launched on another executor. This is where fun
>>>>>> and games start. If the unavailable node recovers before the speculative
>>>>>> task finishes, both the original and speculative tasks might complete
>>>>>> successfully,* resulting in duplicates*. With regard to missing
>>>>>> data, if the data node reboot leads to data corruption or loss, some data
>>>>>> partitions might be completely unavailable. In this case, spark may skip
>>>>>> processing that missing data, leading to missing data in the final
>>>>>> output.
>>>>>>
>>>>>> <Prem> I think when a task failed in between and retry task started
>>>>>> and completed it may create duplicate as failed task has some data +
>>>>>> retry
>>>>>> task has full data. but my question is why spark keeps delta data or
>>>>>> according to you if speculative and original task completes generally
>>>>>> spark
>>>>>> kills one of the tasks to get rid of dups data. when a data node is
>>>>>> rebooted then spark fault tolerant should go to other nodes isn't it ?
>>>>>> then
>>>>>> why it has missing data.
>>>>>> Potential remedies: Spark offers some features to mitigate these
>>>>>> issues, but it might not guarantee complete elimination of duplicates or
>>>>>> data loss:. You can adjust parameters like *spark.shuffle.retry.wa*it
>>>>>> and *spark.speculation* to control retry attempts and speculative
>>>>>> execution behavior. Lineage tracking is there to help. Spark can track
>>>>>> data
>>>>>> lineage, allowing you to identify potentially corrupted or missing data
>>>>>> in
>>>>>> some cases. You can consider persisting intermediate data results to a
>>>>>> reliable storage (like HDFS or GCS or another cloud storage) to avoid
>>>>>> data
>>>>>> loss in case of node failures. Your mileage varies as it adds additional
>>>>>> processing overhead but can ensure data integrity.
>>>>>>
>>>>>> <Prem> How spark will handle these without a checkpoint as it will
>>>>>> slow down the process . I have data loss or duplication is due to
>>>>>> fetchFailedException as a part of data node reboot.
>>>>>> I have few config to minimize fetchFailedException
>>>>>> spark.network.timeout=300s
>>>>>> spark.reducer.maxReqsInFlight=4
>>>>>> spark.shuffle.io.retryWait=30s
>>>>>> spark.shuffle.io.maxRetries=3
>>>>>>
>>>>>> When we get a fetchFailedException due to data node reboot then spark
>>>>>> should handle it gracefully isn't it ?
>>>>>> or how to handle it ?
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Fri, Mar 1, 2024 at 5:35 PM Mich Talebzadeh <
>>>>>> mich.talebza...@gmail.com> wrote:
>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> Your point -> "When Spark job shows FetchFailedException it creates
>>>>>>> few duplicate data and we see few data also missing , please explain
>>>>>>> why.
>>>>>>> We have scenario when spark job complains *FetchFailedException as
>>>>>>> one of the data node got ** rebooted middle of job running ."*
>>>>>>>
>>>>>>> As an engineer I can chip in. You may have partial execution and
>>>>>>> retries meaning when spark encounters a *FetchFailedException*,
>>>>>>> it may retry fetching the data from the unavailable (the one being
>>>>>>> rebooted) node a few times before marking it permanently unavailable.
>>>>>>> However, if the rebooted node recovers quickly within this retry window,
>>>>>>> some executors might successfully fetch the data after a retry. *This
>>>>>>> leads to duplicate processing of the same data partition*.
>>>>>>>
>>>>>>> I am not sure this one applies to your spark version but spark may
>>>>>>> speculatively execute tasks on different executors to improve
>>>>>>> performance. If a task fails due to the *FetchFailedException*, a
>>>>>>> speculative task might be launched on another executor. This is where
>>>>>>> fun
>>>>>>> and games start. If the unavailable node recovers before the speculative
>>>>>>> task finishes, both the original and speculative tasks might complete
>>>>>>> successfully,* resulting in duplicates*. With regard to missing
>>>>>>> data, if the data node reboot leads to data corruption or loss, some
>>>>>>> data
>>>>>>> partitions might be completely unavailable. In this case, spark may skip
>>>>>>> processing that missing data, leading to missing data in the final
>>>>>>> output.
>>>>>>>
>>>>>>> Potential remedies: Spark offers some features to mitigate these
>>>>>>> issues, but it might not guarantee complete elimination of duplicates or
>>>>>>> data loss:. You can adjust parameters like *spark.shuffle.retry.wa*it
>>>>>>> and *spark.speculation* to control retry attempts and speculative
>>>>>>> execution behavior. Lineage tracking is there to help. Spark can track
>>>>>>> data
>>>>>>> lineage, allowing you to identify potentially corrupted or missing data
>>>>>>> in
>>>>>>> some cases. You can consider persisting intermediate data results to a
>>>>>>> reliable storage (like HDFS or GCS or another cloud storage) to avoid
>>>>>>> data
>>>>>>> loss in case of node failures. Your mileage varies as it adds
>>>>>>> additional
>>>>>>> processing overhead but can ensure data integrity.
>>>>>>>
>>>>>>> HTH
>>>>>>>
>>>>>>> Mich Talebzadeh,
>>>>>>> Dad | Technologist
>>>>>>> London
>>>>>>> United Kingdom
>>>>>>>
>>>>>>>
>>>>>>> view my Linkedin profile
>>>>>>> <https://www.linkedin.com/in/mich-talebzadeh-ph-d-5205b2/>
>>>>>>>
>>>>>>>
>>>>>>> https://en.everybodywiki.com/Mich_Talebzadeh
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> *Disclaimer:* The information provided is correct to the best of my
>>>>>>> knowledge but of course cannot be guaranteed . It is essential to note
>>>>>>> that, as with any advice, quote "one test result is worth one-thousand
>>>>>>> expert opinions (Werner
>>>>>>> <https://en.wikipedia.org/wiki/Wernher_von_Braun>Von Braun
>>>>>>> <https://en.wikipedia.org/wiki/Wernher_von_Braun>)".
>>>>>>>
>>>>>>>
>>>>>>> On Fri, 1 Mar 2024 at 20:56, Prem Sahoo <prem.re...@gmail.com>
>>>>>>> wrote:
>>>>>>>
>>>>>>>> Hello All,
>>>>>>>> in the list of JIRAs i didn't find anything related to
>>>>>>>> fetchFailedException.
>>>>>>>>
>>>>>>>> as mentioned above
>>>>>>>>
>>>>>>>> "When Spark job shows FetchFailedException it creates few duplicate
>>>>>>>> data and we see few data also missing , please explain why. We have a
>>>>>>>> scenario when spark job complains FetchFailedException as one of the
>>>>>>>> data
>>>>>>>> nodes got rebooted in the middle of job running .
>>>>>>>> Now due to this we have few duplicate data and few missing data .
>>>>>>>> Why is spark not handling this scenario correctly ? kind of we
>>>>>>>> shouldn't
>>>>>>>> miss any data and we shouldn't create duplicate data . "
>>>>>>>>
>>>>>>>> We have to rerun the job again to fix this data quality issue .
>>>>>>>> Please let me know why this case is not handled properly by Spark ?
>>>>>>>>
>>>>>>>> On Thu, Feb 29, 2024 at 9:50 PM Dongjoon Hyun <
>>>>>>>> dongjoon.h...@gmail.com> wrote:
>>>>>>>>
>>>>>>>>> Please use the url as thr full string including '()' part.
>>>>>>>>>
>>>>>>>>> Or you can seach directly at ASF Jira with 'Spark' project and
>>>>>>>>> three labels, 'Correctness', 'correctness' and 'data-loss'.
>>>>>>>>>
>>>>>>>>> Dongjoon
>>>>>>>>>
>>>>>>>>> On Thu, Feb 29, 2024 at 11:54 Prem Sahoo <prem.re...@gmail.com>
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>>> Hello Dongjoon,
>>>>>>>>>> Thanks for emailing me.
>>>>>>>>>> Could you please share a list of fixes as the link provided by
>>>>>>>>>> you is not working.
>>>>>>>>>>
>>>>>>>>>> On Thu, Feb 29, 2024 at 11:27 AM Dongjoon Hyun <
>>>>>>>>>> dongj...@apache.org> wrote:
>>>>>>>>>>
>>>>>>>>>>> Hi,
>>>>>>>>>>>
>>>>>>>>>>> If you are observing correctness issues, you may hit some old
>>>>>>>>>>> (and fixed) correctness issues.
>>>>>>>>>>>
>>>>>>>>>>> For example, from Apache Spark 3.2.1 to 3.2.4, we fixed 31
>>>>>>>>>>> correctness issues.
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> https://issues.apache.org/jira/issues/?filter=12345390&jql=project%20%3D%20SPARK%20AND%20fixVersion%20in%20(3.2.1%2C%203.2.2%2C%203.2.3%2C%203.2.4)%20AND%20labels%20in%20(Correctness%2C%20correctness%2C%20data-loss)
>>>>>>>>>>>
>>>>>>>>>>> There are more fixes in 3.3 and 3.4 and 3.5, too.
>>>>>>>>>>>
>>>>>>>>>>> Please use the latest version, Apache Spark 3.5.1, because
>>>>>>>>>>> Apache Spark 3.2 and 3.3 are in the End-Of-Support status of the
>>>>>>>>>>> community.
>>>>>>>>>>>
>>>>>>>>>>> It would be help if you can report any correctness issues with
>>>>>>>>>>> Apache Spark 3.5.1.
>>>>>>>>>>>
>>>>>>>>>>> Thanks,
>>>>>>>>>>> Dongjoon.
>>>>>>>>>>>
>>>>>>>>>>> On 2024/02/29 15:04:41 Prem Sahoo wrote:
>>>>>>>>>>> > When Spark job shows FetchFailedException it creates few
>>>>>>>>>>> duplicate data and
>>>>>>>>>>> > we see few data also missing , please explain why. We have
>>>>>>>>>>> scenario when
>>>>>>>>>>> > spark job complains FetchFailedException as one of the data
>>>>>>>>>>> node got
>>>>>>>>>>> > rebooted middle of job running .
>>>>>>>>>>> >
>>>>>>>>>>> > Now due to this we have few duplicate data and few missing
>>>>>>>>>>> data . Why spark
>>>>>>>>>>> > is not handling this scenario correctly ? kind of we shouldn't
>>>>>>>>>>> miss any
>>>>>>>>>>> > data and we shouldn't create duplicate data .
>>>>>>>>>>> >
>>>>>>>>>>> >
>>>>>>>>>>> >
>>>>>>>>>>> > I am using spark3.2.0 version.
>>>>>>>>>>> >
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> ---------------------------------------------------------------------
>>>>>>>>>>> To unsubscribe e-mail: dev-unsubscr...@spark.apache.org
>>>>>>>>>>>
>>>>>>>>>>>
