[
https://issues.apache.org/jira/browse/FLINK-10518?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Huyen Levan updated FLINK-10518:
--------------------------------
Description:
The ContinuousFileMonitoringFunction class keeps track of the latest file
modification time to rule out all files it has processed in the previous
cycles. For a long-running job, the list of eligible files will be much smaller
than the list of all files in the folder being monitored.
In the current implementation of the getInputSplitsSortedByModTime method, a
(big) list of all available splits are created first, and then every single
split is checked with the list of eligible files.
{quote}for (FileInputSplit split: format.createInputSplits(readerParallelism)) {
FileStatus fileStatus = eligibleFiles.get(split.getPath());
if (fileStatus != null) {
{quote}
The improvement can be done as:
* Listing of all files should be done once in
_ContinuousFileMonitoringFunction.listEligibleFiles()_ (as of now it is done
the 2nd time in _FileInputFormat.createInputSplits()_ )
* The list of file-splits should then be created from the list of paths in
eligibleFiles.
was:
The ContinuousFileMonitoringFunction class keeps track of the latest file
modification time to rule out all file it has processed in the previous cycles.
For a long-running job, the list of eligible files will be much less than the
list of all files in the folder being monitored.
In the current implementation of the getInputSplitsSortedByModTime method, a
list of all available splits are created first, and then every single split is
checked with the list of eligible files.
{quote}for (FileInputSplit split: format.createInputSplits(readerParallelism)) {
FileStatus fileStatus = eligibleFiles.get(split.getPath());
if (fileStatus != null) {
{quote}
The improvement can be done as:
* Listing of all files should be done once in
_ContinuousFileMonitoringFunction.listEligibleFiles()_ (as of now it is done
the 2nd time in _FileInputFormat.createInputSplits()_ )
* The list of file-splits should then be created from the list of paths in
eligibleFiles.
> Inefficient design in ContinuousFileMonitoringFunction
> ------------------------------------------------------
>
> Key: FLINK-10518
> URL: https://issues.apache.org/jira/browse/FLINK-10518
> Project: Flink
> Issue Type: Improvement
> Components: filesystem-connector
> Affects Versions: 1.5.2
> Reporter: Huyen Levan
> Priority: Major
> Labels: Source:FileSystem
>
> The ContinuousFileMonitoringFunction class keeps track of the latest file
> modification time to rule out all files it has processed in the previous
> cycles. For a long-running job, the list of eligible files will be much
> smaller than the list of all files in the folder being monitored.
> In the current implementation of the getInputSplitsSortedByModTime method, a
> (big) list of all available splits are created first, and then every single
> split is checked with the list of eligible files.
> {quote}for (FileInputSplit split:
> format.createInputSplits(readerParallelism)) {
> FileStatus fileStatus = eligibleFiles.get(split.getPath());
> if (fileStatus != null) {
> {quote}
> The improvement can be done as:
> * Listing of all files should be done once in
> _ContinuousFileMonitoringFunction.listEligibleFiles()_ (as of now it is done
> the 2nd time in _FileInputFormat.createInputSplits()_ )
> * The list of file-splits should then be created from the list of paths in
> eligibleFiles.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
