[
https://issues.apache.org/jira/browse/SPARK-30666?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Enrico Minack updated SPARK-30666:
----------------------------------
Description:
This proposes a pragmatic improvement to allow for reliable single-stage
accumulators. Under the assumption that a given stage / partition / rdd
produces identical results, non-deterministic code incrementing accumulators
also produces identical accumulator increments on success. Rerunning partitions
for any reason should always produce the same increments on success.
With this pragmatic approach, increments from individual partitions / tasks are
compared to earlier increments. Depending on the strategy of how a new
increment updates over an earlier increment from the same partition, different
semantics of accumulators (here called accumulator modes) can be implemented:
- ALL sums over all increments of each partition: this represents the current
implementation of accumulators
- MAX over all increments of each partition: assuming accumulators only
increment while a partition is processed, a successful task provides an
accumulator value that is always larger than any value of failed tasks, hence
it paramounts any failed task's value. This produces reliable accumulator
values. This should only be used in a single stage.
- LAST increment: allows to retrieve the latest increment for each partition
only.
The implementation for MAX and LAST requires extra memory that scales with the
number of partitions. The current ALL implementation does not require extra
memory.
was:
This proposes a pragmatic improvement to allow for reliable single-stage
accumulators. Under the assumption that a given stage / partition / rdd
produces identical results, non-deterministic code incrementing accumulators
also produces identical accumulator increments on success. Rerunning partitions
for any reason should always produce the same increments on success.
With this pragmatic approach, increments from individual partitions / tasks are
compared to earlier increments. Depending on the strategy of how a new
increment updates over an earlier increment from the same partition, different
semantics of accumulators (here called accumulator modes) can be implemented:
- SUM over all increments of each partition: this represents the current
implementation of accumulators
- MAX over all increments of each partition: assuming accumulators only
increment while a partition is processed, a successful task provides an
accumulator value that is always larger than any value of failed tasks, hence
it paramounts any failed task's value. This produces reliable accumulator
values. This should only be used in a single stage.
- LAST increment: allows to retrieve the latest increment for each partition
only.
The implementation for MAX and LAST requires extra memory that scales with the
number of partitions. The current SUM implementation does not require extra
memory.
> Reliable single-stage accumulators
> ----------------------------------
>
> Key: SPARK-30666
> URL: https://issues.apache.org/jira/browse/SPARK-30666
> Project: Spark
> Issue Type: Improvement
> Components: SQL
> Affects Versions: 3.0.0
> Reporter: Enrico Minack
> Priority: Major
>
> This proposes a pragmatic improvement to allow for reliable single-stage
> accumulators. Under the assumption that a given stage / partition / rdd
> produces identical results, non-deterministic code incrementing accumulators
> also produces identical accumulator increments on success. Rerunning
> partitions for any reason should always produce the same increments on
> success.
> With this pragmatic approach, increments from individual partitions / tasks
> are compared to earlier increments. Depending on the strategy of how a new
> increment updates over an earlier increment from the same partition,
> different semantics of accumulators (here called accumulator modes) can be
> implemented:
> - ALL sums over all increments of each partition: this represents the
> current implementation of accumulators
> - MAX over all increments of each partition: assuming accumulators only
> increment while a partition is processed, a successful task provides an
> accumulator value that is always larger than any value of failed tasks, hence
> it paramounts any failed task's value. This produces reliable accumulator
> values. This should only be used in a single stage.
> - LAST increment: allows to retrieve the latest increment for each partition
> only.
> The implementation for MAX and LAST requires extra memory that scales with
> the number of partitions. The current ALL implementation does not require
> extra memory.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
