Instrumentations for Apache HttpComponents do not meter errors and leak memory

[cachescrubber]

Describe the bug
An HttpAsyncClient instrumented with MicrometerHttpClientInterceptor does not meter IO errors.
For example, requests resulting in a ConnectTimeoutException, SocketTimeoutException are not metered.

MicrometerHttpRequestExecutor however does meter IO errors by setting the status key to IO_ERROR.

Environment

• Micrometer version [1.11.0-RC1]
• Micrometer registry [prometheus]
• Java version: [17]

To Reproduce
How to reproduce the bug:
Make an instrumented client to run into some sort of IO Exception, ConnectTimeoutException, SocketTimeoutException, etc.

Expected behavior

• The request is counted, and elapsed time is measured.
• The Tag status has the value IO_ERROR
• The Tag outcome is UNKNOWN.

Additional context
The Apache Http Components instrumentation does not support the tag exception, which is quite commonly used by other http client instrumentations.

[cachescrubber]

I dig a bit deeper into this. As it stands right now, The httpcomponents binder is not able to meter connection related IO errors like connection-refused or connection-timeout, TLS handshake related errors, etc. This affects both the classic and the async client. I committed test cases to #3801 to highlight the issue. Bummer.

Implemented in #3801 is a measurement of IO errors during an active http request, for example socket read timeouts. This brings the async implementation on par with classic.

[cachescrubber]

As it stands right now, the httpcomponents binder does not only miss connection related errors - the time necessary to establish a connection is not metered at all. #3801 has changed the instrumentation to use ExecChainHandler instead of Request/Response interceptor to address this issue.

[cachescrubber]

Backport to 4.x would require to Subclass HttpClientBuilder and override its #decorateMainExec or #decorateProtocolExec methods.