Prior to this commit, the `ConcurrentReferenceHashMap#Segment` would
recalculate its element count during the restructure phase by
subtracting the number of polled elements from the queue to the current
count.
Later, the newly restructured `references` array would only contain
references that 1) are not in the set of elements to purge and
2) that hold a non-null value.
The issues with this approach are multiple. The newly calculated count
is only an estimate, as some situations can make it invalid, even
temporarily.
* since we initially collected all elements to be purged from the queue,
the GC might have collected new values. This means that we might
filter out more references that we initially intended to
* because the restructure operation re-creates new references for all
elements in the original array, we might later get references from the
queue that are not in the array anymore. This could lead to
"duplicate" removals for the same value
Because several methods in the Segment class have special no-op behavior
when `count == 0`, an invalid count can lead to keys appearing missing
when they are actually still present. In some scenarios, this can
decrease the performance of the cache since values need to be
recalculated.
This commit fixes this inconsistency count issue by first using an
estimate in order to decide whether the array needs a resize and then by
counting the actual numbers of elements inserted in the restructured
array as the new count.
Fixes gh-31373
This commit adds a warning in the reference guide to address the
use cases where users might be tempted to use several property
placeholder configurers.
Closes gh-14623
Prior to this commit, HTTP server observations for Spring WebFlux could
be recorded twice for a single request in some cases. The "COMPLETE" and
"CANCEL" signals would race in the reactive pipeline and would trigger
both the `doOnComplete()` and ` `doOnCancel()` operators, each calling
`observation.stop()` on the current observation.
This would in fact publish two different observations for the same
request.
This commit ensures that the instrumentation uses the `Mono#tap`
operator to guard against this case and only call `Observation#stop`
once for each request.
Fixes gh-31417
Prior to this commit, a cancelled exchange would always result in an
`"status":"UNKNOWN"` KeyValue. This only applied to reactive variants,
as cancelled exchanges are not currently detected for Servlet
implementations.
In some cases, exchanges can be cancelled by clients before they are
completed, but the response was actually received by the client. The
response status information has been set by the application and the
response has been committed. For those cases, we shouldn't assume an
"UNKNOWN" value.
This commit assumes that committed responses have a response status set
by the application and that the observations should reflect that. From
now on, we only assume an "UNKNOWN" status if the response has not been
commited.
Fixes gh-31388
