[
https://issues.apache.org/jira/browse/BEAM-1206?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ryan Skraba updated BEAM-1206:
------------------------------
Description:
The SparkRunner keeps track of "leaf" transforms in the job graph and ensures
that they have been executed when the pipeline is run.
However, when a PCollection is used as a sideInput, it is evaluated *once* to
get the materialized values but is still considered a leaf node. It is
evaluated a second time at the end of the Spark job, which is unnecessary (and
can cause unexpected behaviour).
One of the symptoms of this bug is that {{Sink}}s will create spurious writers
that execute, but are never finalized. Specifically the HDFSFileSink will
_always_ die with a {{Writer results and output files do not match}} error.
was:
The SparkRunner keeps track of "leaf" transforms in the job graph and ensures
that they have been executed when the pipeline is run.
However, when a PCollection is used as a sideInput, it is evaluated *once* to
get the materialized values but is still considered a leaf node. It is
evaluated a second time at the end of the Spark job, which is unnecessary (and
can cause unexpected behaviour).
One of the symptoms of this bug is that `Sink`s will create spurious writers
that execute, but are never finalized. Specifically the HDFSFileSink will
_always_ die with a `Writer results and output files do not match` error.
> PCollections used as a sideInput are unnecessarily re-evaluated.
> ----------------------------------------------------------------
>
> Key: BEAM-1206
> URL: https://issues.apache.org/jira/browse/BEAM-1206
> Project: Beam
> Issue Type: Bug
> Components: runner-spark
> Affects Versions: 0.4.0-incubating
> Reporter: Ryan Skraba
> Assignee: Jean-Baptiste Onofré
>
> The SparkRunner keeps track of "leaf" transforms in the job graph and ensures
> that they have been executed when the pipeline is run.
> However, when a PCollection is used as a sideInput, it is evaluated *once* to
> get the materialized values but is still considered a leaf node. It is
> evaluated a second time at the end of the Spark job, which is unnecessary
> (and can cause unexpected behaviour).
> One of the symptoms of this bug is that {{Sink}}s will create spurious
> writers that execute, but are never finalized. Specifically the HDFSFileSink
> will _always_ die with a {{Writer results and output files do not match}}
> error.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
