[
https://issues.apache.org/jira/browse/UIMA-1970?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15470717#comment-15470717
]
Marshall Schor commented on UIMA-1970:
--------------------------------------
These are interesting ideas. I think Jörn was mentioning UIMA-AS just as an
example... so I'm not sure this belongs there?
The general rule on whether or not some feature / capability gets added to UIMA
framework, or left for users to grapple with, has been some judgement on
whether or not the feature / capability would be generally useful for a wider
audience, and therefore, merit a careful / robust (insert more adjectives here)
implementation that might be individually done.
The use case Jörn mentions, of "resources" changing more frequently than
annotators, seems common enough. To implement such a thing seems to require:
* a "trigger"
* a way to update
The trigger could be some (user) code run every so often; the way to update
could be some user-code called when the trigger fires. These could be
additional APIs the framework would call in a resource.
Very common cases (if there are any) could be handled by framework code: e.g.
getting some signal from a file system when a file was updated.
Resources are typically run on potentially multiple threads, so this would need
to be taken into account.
> Reload updated resources while an analysis pipeline is running
> --------------------------------------------------------------
>
> Key: UIMA-1970
> URL: https://issues.apache.org/jira/browse/UIMA-1970
> Project: UIMA
> Issue Type: New Feature
> Components: Async Scaleout
> Reporter: Joern Kottmann
> Labels: Resources
>
> In many use cases resources used by Analysis Engine are updated more
> frequently than the AE implementation.
> Samples of such resources are huge dictionaries which are extended with new
> names on a daily basis
> or statistical models which are re-trained frequently to be adjusted to data
> from the past minutes.
> For a resource updates an analysis pipeline has to be stopped and started
> manually to pick up the updated resource.
> It would be nice if the framework could automate this procedure. In order to
> do that it must have a capability to detect
> updated resources and perform the re-initialization of affected Analysis
> Engines.
> If the pipelines run multiple instances of these Analysis Engines it would be
> nice if a rolling
> update could be performed in a way that the processing of CASes never comes
> to an halt.
> That should be handled in different issues for the CPE and UIMA-AS.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
