[
https://issues.apache.org/jira/browse/BEAM-2398?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Antony Mayi updated BEAM-2398:
------------------------------
Description:
Over the time the end-to-end latency of a pipeline running on DirectRunner is
significantly increasing.
This is caused by ever growing sets of:
* {{WatermarkManager.TransformWatermarks.inputWatermark.pendingElements}}
*
{{WatermarkManager.TransformWatermarks.synchronizedProcessingInputWatermark.pendingBundles}}
That means calls to {{WatermarkManager.TransformWatermarks.refresh()}} which
need to iterate through that collections take longer and longer and the latency
is growing.
I believe it is the line {{WaterMark.updatePending()}} line:
{quote}
if (input != null) {
// Add the unprocessed inputs
completedTransform.addPending(result.getUnprocessedInputs());
{quote}
that's adding the items that are never removed.
See attached [^LatencyTest.java] demo code showing the increasing latency.
was:
Over the time the end-to-end latency of a pipeline running on DirectRunner is
significantly increasing.
This is caused by ever growing sets of:
* {{WatermarkManager.TransformWatermarks.inputWatermark.pendingElements}}
*
{{WatermarkManager.TransformWatermarks.synchronizedProcessingInputWatermark.pendingBundles}}
That means calls to {{WatermarkManager.TransformWatermarks.refresh()}} which
need to iterate through that collections take longer and longer and the latency
is growing.
I believe it is the line {{WaterMark.updatePending()}} line:
{quote}
if (input != null) {
// Add the unprocessed inputs
completedTransform.addPending(result.getUnprocessedInputs());
{quote}
that's adding the items that are never removed.
See attached demo code showing the increasing latency.
> Increasing latency within DirectRunner caused by cumulated TransformWatermarks
> ------------------------------------------------------------------------------
>
> Key: BEAM-2398
> URL: https://issues.apache.org/jira/browse/BEAM-2398
> Project: Beam
> Issue Type: Bug
> Components: runner-direct
> Affects Versions: 2.0.0
> Reporter: Antony Mayi
> Assignee: Thomas Groh
> Attachments: LatencyTest.java
>
>
> Over the time the end-to-end latency of a pipeline running on DirectRunner is
> significantly increasing.
> This is caused by ever growing sets of:
> * {{WatermarkManager.TransformWatermarks.inputWatermark.pendingElements}}
> *
> {{WatermarkManager.TransformWatermarks.synchronizedProcessingInputWatermark.pendingBundles}}
> That means calls to {{WatermarkManager.TransformWatermarks.refresh()}} which
> need to iterate through that collections take longer and longer and the
> latency is growing.
> I believe it is the line {{WaterMark.updatePending()}} line:
> {quote}
> if (input != null) {
> // Add the unprocessed inputs
> completedTransform.addPending(result.getUnprocessedInputs());
> {quote}
> that's adding the items that are never removed.
> See attached [^LatencyTest.java] demo code showing the increasing latency.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
