[
https://issues.apache.org/jira/browse/SYSTEMML-1518?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Matthias Boehm updated SYSTEMML-1518:
-------------------------------------
Description:
Both the new and old mlcontext APIs call
{{OptimizerUtils.getUniqueTempFileName()}} to create HDFS filenames for
registered input frames or matrices. This call simply forwards the request to
{{Dag}} for consistency with hdfs filenames of intermediates and to ensure
isolation with regard to concurrently running scripts (from different client
processes on a shared cluster).
However, for this code path the internal scratch space configuration is always
uninitialized leading to corrupt filenames such as
{{/_p1234_1.2.345.678//_t0/temp1_0}}. The missing scratch_space prefix is
problematic because the remainder is interpreted as an absolute file path,
often leading to permission issues because typical users are not granted write
access on HFDS root.
Note that this issue might not be immediately visible in all scenarios because
it only affects input variables that are exported to HDFS (e.g., during guarded
collect or as specific inputs to remote parfor).
was:
Both the new and old mlcontext APIs call
{{OptimizerUtils.getUniqueTempFileName()}} to create HDFS filenames for
registered input frames or matrices. This call simply forwards the request to
{{Dag}} for consistency with hdfs filenames of intermediates and to ensure
isolation with regard to concurrently running scripts (from different client
processes on a shared cluster).
However, for this code path the internal scratch space configuration is always
uninitialized leading to corrupt filenames such as
{{/_p1234_1.2.345.678//_t0/temp1_0}}. The missing scratch_space prefix is
problematic because the remainder is interpreted as an absolute file path,
often leading to permission issues because typical users are not granted write
access on HFDS root.
> Corrupted input file names in old and new mlcontext apis
> --------------------------------------------------------
>
> Key: SYSTEMML-1518
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1518
> Project: SystemML
> Issue Type: Bug
> Reporter: Matthias Boehm
> Priority: Blocker
>
> Both the new and old mlcontext APIs call
> {{OptimizerUtils.getUniqueTempFileName()}} to create HDFS filenames for
> registered input frames or matrices. This call simply forwards the request to
> {{Dag}} for consistency with hdfs filenames of intermediates and to ensure
> isolation with regard to concurrently running scripts (from different client
> processes on a shared cluster).
> However, for this code path the internal scratch space configuration is
> always uninitialized leading to corrupt filenames such as
> {{/_p1234_1.2.345.678//_t0/temp1_0}}. The missing scratch_space prefix is
> problematic because the remainder is interpreted as an absolute file path,
> often leading to permission issues because typical users are not granted
> write access on HFDS root.
> Note that this issue might not be immediately visible in all scenarios
> because it only affects input variables that are exported to HDFS (e.g.,
> during guarded collect or as specific inputs to remote parfor).
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
