[
https://issues.apache.org/jira/browse/IGNITE-3877?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15744890#comment-15744890
]
Ivan Veselovsky commented on IGNITE-3877:
-----------------------------------------
The problem is in mixing notions GroupBlockSize and BlockSize.
As comment to class
{code}org.apache.ignite.igfs.IgfsGroupDataBlocksKeyMapper{code} states, when
{code}org.apache.hadoop.fs.FileSystem{code} lies upon IGFS, the Hadoop Fs block
size equals to underlying IGFS group block size (which, in turn, is the block
size multiplied by groupSize).
When we have IGFS over Hadoop FileSystem, we also use the group block size as
the block size of the created secondary Fs files (see
{code}org.apache.ignite.internal.processors.igfs.IgfsSecondaryFileSystemCreateContext#create{code}).
This way, when reading files we should follow the same logic: IGFS *group*
block size == Hadoop block size, and IGFS block size is just the configuration
value
({code}org.apache.ignite.configuration.FileSystemConfiguration#getBlockSize{code})
.
This way we make the logic consistent, and fix the assertion issue described
above.
> Clarify if IgfsFile -> FileStatus conversion should treat groupBlockSize as
> blockSize
> -------------------------------------------------------------------------------------
>
> Key: IGNITE-3877
> URL: https://issues.apache.org/jira/browse/IGNITE-3877
> Project: Ignite
> Issue Type: Bug
> Components: IGFS
> Affects Versions: 1.6
> Reporter: Ivan Veselovsky
> Assignee: Vladimir Ozerov
> Fix For: 2.0
>
>
> During Metrics tests repairing test
> org.apache.ignite.igfs.Hadoop1DualAbstractTest#testMetricsBlock revealed the
> following problem:
> org.apache.ignite.hadoop.fs.v1.IgniteHadoopFileSystem#convert(org.apache.ignite.igfs.IgfsFile)
> method treats groupBlockSize as blockSize for Hadoop FileStatus.
> groupBlockSize can be several times larger than blockSize, so blockSize in
> status gets different to that in original IgfsFile .
> changing file.groupBlockSize() to file.blockSize() fixes problem in metrics
> tests, but creates problems in Hadoop tests that are bound to splits
> calculation, since split calculation related to blockSizes.
> Need to
> 1) clarify if the treatment of groupBlcokSize was intentional.
> 2) fix either metrics tests or Hadoop tests.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
