[
https://issues.apache.org/jira/browse/PARQUET-520?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Wes McKinney updated PARQUET-520:
---------------------------------
Description: Noted this while working on PARQUET-497. If we are using a
memory-mapped file, then copying data into a {{ScopedInMemoryInputStream}} as
we are now is unnecessary and will yield improved performance. Perhaps this
should be made a property of the {{InputStream}} (i.e. indicate whether it
support zero-copy reads, and the returned buffer does not become invalid after
future reads as long as the stream -- the memory map specifically in this
example -- is alive) (was: Noted this while working on PARQUET-497. If we are
using a memory-mapped file, then copying data into a
{{ScopedInMemoryInputStream}} as we are not is unnecessary and will yield
improved performance. Perhaps this should be made a property of the
{{InputStream}} (i.e. that it support zero-copy reads, and the returned buffer
does not become invalid after future reads as long as the stream -- the memory
map specifically in this example -- is alive))
> Add support for zero-copy InputStreams on memory-mapped files
> -------------------------------------------------------------
>
> Key: PARQUET-520
> URL: https://issues.apache.org/jira/browse/PARQUET-520
> Project: Parquet
> Issue Type: Improvement
> Components: parquet-cpp
> Reporter: Wes McKinney
>
> Noted this while working on PARQUET-497. If we are using a memory-mapped
> file, then copying data into a {{ScopedInMemoryInputStream}} as we are now is
> unnecessary and will yield improved performance. Perhaps this should be made
> a property of the {{InputStream}} (i.e. indicate whether it support zero-copy
> reads, and the returned buffer does not become invalid after future reads as
> long as the stream -- the memory map specifically in this example -- is alive)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
