Since fork is not available on some platforms (aka. Windows), we
avoid it and use normal Threads instead. We have fail saves in place to
ensure that tests are still valid even on MRI with its GIL.
Each flow now has an associated executor which performs all actions
(writing events, closing, reopening) asynchronously by default using a
Concurrent::SingleThreadExecutor.
This improves the responsiveness of the application by performing the
(usually) IO-bound task of writing the logs to a background thread.
By creating a flow with `synchronous: true`, all actions are run in the
calling thread as before, making the flow blocking.
With this, we also drop the ability to define conditions in the filter
itself. When adding a filter, users can still setup a condition using
the common functionality of all filters.
We now support two different modes of file rotation at the same time:
* auto_reopen can be used to automatically reopen a logfile at the
original location if the file was moved or deleted from the filesystem
* rotate can be used to write to a rotate file which can be reopened /
created based on Date pattern.
The user can now decide whether they want to use an external logrotate
command or use internal rotation with Rackstash instead.
Instead of defining the specific buffering behavior on a Buffer, we can
now mark individual flows as auto_flushing or now. An auto_flushing Flow
with a buffering Buffer behaves the same as a Buffer with `buffering:
:data` would before.
This allows us to simplify the buffering logic on the Buffer. Also, we
can now use "normal" flows and auto_flushing flows on the same logger in
parallel. Each of them behaves as expected with the same unchanged logger
code.
It is thus easier to define behavior for a development or production
environment of an app since the necessary changes can all be defined on
the logger itself (through the defined flows) without having to adapt
the code which creates suitable Buffers with the Logger#with_buffer
method in any way.
With that, we can lazy-transform the Buffer to the event hash. B using
the common `.to_h` protocol, we can also support various other objects
here instead of just Buffers (including actual raw Hashes).
Previously, we would only setup a `::Rack::BodyProxy` to eventually log
the request once the response is deivered to the client. We did this
mainly in order to have a more "accurate" duration in the logs, similar
to what the `::Rack::CommonLogger` middleware does.
This has some significant disadvantages though:
* If an exception is raised by an outer middleware after we returned,
the response is never delivered and the log is lost.
* If a timeout occured (and e.g. a Unicorn worker is killed or a
`Rack::Timeout` stroke) before we return the full response, the log is
lost.
* If the client goes away before the request is finished, the app server
might decide no just throw away the response.
Besides addressing these issues, logging directly after the request
leaves our middleware scope makes the whole process much easier to
reason about.
Most notably, we want to use the plus character in URI schemes. By only
checking the first character of the scheme registration, we can better
fullfil our contract and better distinguish between schemes and class
names of registered Adapters.
If we were to fall through, undefined schemes would always end up as a
file adapter. This is generally undesired since it hides the fact that
we have not found a suitable adapter. Most of the time, this will be a
configuration error which should be reported early.
If a user still wants to create a file adapter with a filename that
looks like a URI, they can create a Rackstash::Adapter::File object
manually.
