Frequent invocation of javax.xml.stream.XMLOutputFactory::newInstance negatively impacts performance

[robsman]

While profiling a client application, it showed that there are lots of javax.xml.stream.XMLOutputFactory.newInstance() invocations, specifically when a client retrieves extracted document data from search results. In our use case, this almost totaled to the same amount of CPU time as the actual search.

[robsman]

I have applied a hotfix in the client project but still need to beef it up into a proper pull request.

The problematic code in our use case is in Utilities::writeEvents. However, there are other places that retrieve XML factories, e.g. SearchHandle::receiveContent, StructuredQueryBuilder::makeSerializer and SearchResponseImpl::handleMetrics.

My suggestion is to consolidate retrieval and caching of expensive factories into a utility class (maybe XmlFactoriesAccess ?) and adjust client code accordingly.

[sammefford]

Thanks Rob!

[sammefford]

We'll let you submit your patch first, then we'll use that as a model for the rest.

[robsman]

Hi Sam, after some further research it turned out the various XML factories are not guaranteed to be thread-safe. To address this I have reworked pull request #477 to use instance-per-thread caching.

The code is a bit verbose as the develop-3.0 branch is still Java 6. But nevertheless, it should be relatively straightforward.
We're running a copy of this in our project for some days now, so far, so good.

[sammefford]

Thanks Rob!

[sammefford]

@georgeajit as long as this didn't create any issues in unit tests, I think we're good on this

[georgeajit]

Don't see any regressions as of now (01/25/2018) on the develop branch. Shipping this issue.