Previously, the second mtime check for new files usually happened right
after the first one, which could have caused consumption of docs that
were still being modified.
We're now waiting for at least FILES_MIN_UNMODIFIED_DURATION (0.5s).
This also cleans up the logic by eliminating the consumer.stats attribute
and the weird double call to consumer.run().
Additionally, this a fixes memory leak in consumer.stats where paths could be
added but never removed if the corresponding files disappeared from
the consumer dir before being considered ready.
I've broken out the OCR-specific code from the consumers and dumped it
all into its own app, `paperless_tesseract`. This new app should serve
as a sample of how to create one's own consumer for different file
types.
Documentation for how to do this isn't ready yet, but for the impatient:
* Create a new app
* containing a `parsers.py` for your parser modelled after
`paperless_tesseract.parsers.RasterisedDocumentParser`
* containing a `signals.py` with a handler moddelled after
`paperless_tesseract.signals.ConsumerDeclaration`
* connect the signal handler to
`documents.signals.document_consumer_declaration` in
`your_app.apps`
* Install the app into Paperless by declaring
`PAPERLESS_INSTALLED_APPS=your_app`. Additional apps should be
separated with commas.
* Restart the consumer
Creating the scratch-files in `_get_grayscale` using a random integer is
for one inherently unsafe and can cause a collision. On the other hand,
it should be unnecessary given that the files will be cleaned up after
the OCR run.
Since we don't know if OCR runs might be parallel in the future, this
commit implements thread-safe and deterministic directory-creation.
Additionally it fixes the call to `_cleanup` by `consume`. In the
current implementation `_cleanup` will not be called if the last
consumed document failed with an `OCRError`, this commit fixes this.
