The existing implementation was exposed to very poor performance when matching
with multiple delimiters against a large buffer with many delimiters. In that
case all matchers are invoked many times (as many as the number of delimiters)
even though some of them found no match at all on the first pass.
The revised implementation uses a single index and advances all matchers
together, checking one byte a time, and not letting any one of them search to
the end of the entire buffer on a single pass.
Closes gh-25915
Simplify and optimize the processing of the input stream.
The existing implementation was using bufferUntil and creating a List
for every line along with an EndFrameBuffer inserted for the
bufferUntil predicate. So the larger the input buffer and the more
lines it contained, the greater the overhead.
The new implementation avoids bufferUntil for all lines and
instead uses concatMapIterable to aggregate the lines from a buffer
into a single list. So the larger the input buffer and the more
lines it contains, the better the throughput. The only buffering
used then is for partial chunks and those are accumulated in a list.
See gh-25915
Prior to this commit, references to `JsonGenerator` and
`ByteArrayBuilder` were not closed/released within codecs calls.
This prevents Jackson from reusing more efficiently shared memory
resources.
This commit properly closes/releases Jackson resources in Spring MVC,
Spring WebFlux and Spring Messaging codecs.
A benchmark on WebFlux codecs (in both single value/streaming mode)
shows significant throughput and allocation improvements for small
payloads.
Closes gh-25910
Before this commit, ReadCompletionHandler delayed closing the channel
to allow an ongoing read to complete/fail so we can get a hold of the
associated DataBuffer and release it. This can be problematic since the
read take time to complete but even more importantly there was a race
condition where we didn't check if we've been disposed concurrently
while releasing the read flag.
This commit removes the delay and closes the channel immediately from
cancel() and that should in turn fail any ongoing read operation,
according to AsynchronousChannel, and the DataBuffer is released.
Further improvements include:
- combining the "reading" and "disposed" AtomicBoolean's into a
single "state" AtomicReference.
- an optimistic check to remain in READING mode and avoid state
switches when there is demand to continue reading.
Closes gh-25831
Issues gh-25038 and gh-25618 collectively introduced a regression for
thread-scoped and transaction-scoped beans.
For example, given a thread-scoped bean X that depends on another
thread-scoped bean Y, if the names of the beans (when used as map keys)
end up in the same bucket within a ConcurrentHashMap AND an attempt is
made to retrieve bean X from the ApplicationContext prior to retrieving
bean Y, then the use of Map::computeIfAbsent in SimpleThreadScope
results in recursive access to the same internal bucket in the map.
On Java 8, that scenario simply hangs. On Java 9 and higher,
ConcurrentHashMap throws an IllegalStateException pointing out that a
"Recursive update" was attempted.
In light of these findings, we are reverting the changes made to
SimpleThreadScope and SimpleTransactionScope in commits 50a4fdac6e and
148dc95eb1.
Closes gh-25801
Prior to this commit, the `docsZip` task would not reference the new
output locations for the `asciidoctor` and `asciidoctorPdf` tasks.
This results with missing reference docs in the docs zip.
This commit updates the input locations of the Zip task to include the
produced reference docs.
Fixes gh-25783
Prior to this commit, the Asciidoctor Gradle tasks generated top-level
HTML and PDF documents for AsciiDoc files that are included in other
top-level documents. This causes slower builds and results in each
include-file being published twice:
1) inline in the including document (as intended)
2) as a top-level document but missing surrounding context (unintended)
The reason these include-files are generated as top-level documents is
that the asciidoctor and asciidoctorPdf Gradle tasks are configured to
use '*.adoc' as the input source files.
This commit addresses this issue by moving the following include-files
to new subdirectories. Locating the include-files in the subdirectories
causes them to be ignored in the '*.adoc' pattern used to identify
input source files.
- data-access-appendix.adoc -> data-access/data-access-appendix.adoc
- integration-appendix.adoc -> integration/integration-appendix.adoc
- testing-webtestclient.adoc -> testing/testing-webtestclient.adoc
Closes gh-25783
Prior to this commit, the asciidoctor Gradle task was configured to
generate both the HTML5 and PDF backends. Unfortunately, this resulted
in resources such as HTML, JavaScript, CSS, and images being published
alongside the generated PDF documents.
This commit addresses this issue by introducing the use of a dedicated
asciidoctorPdf Gradle task. The existing asciidoctor Gradle task has
been modified to only generate HTML5 output. Both of these tasks now
share common configuration supplied by the updated asciidoctorj Gradle
task.
In addition, the asciidoctor task now depends on the asciidoctorPdf
task. Thus, invoking `./gradlew asciidoctor` will still generate both
the HTML5 and PDF outputs; whereas, `./gradlew asciidoctorPdf` will
generate only the PDF outputs. We may later decide to rework the tasks
to introduce a dedicated asciidoctorHtml task so that we can generate
the HTML outputs without having to generate the PDF outputs (which are
more time consuming).
See gh-25783
