Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

I found a thread that breaks into two in production. Ironically the
problem isn't reproducible in isolation here, but we don't have tests
where both references and in-reply-to info is used (duplicate
reference) so I figured we can keep it :)

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

In attempt to debug a threading problem I wrote this test so I could
find the bug. But it shows that the threading algorithm itself is fine.

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Just like a CLI priority inbox model training gives all the details, we
want to have the same to diagnose slow/faulty account syncs. This
changes the console logger adapter for the PSR logger and adds it to the
sync process.

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

They clutter the output.

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

We cache the HTTP response for when no avatar is available, so the same
browser won't ask for this for the next few hours. However, this doesn't
work across devices, so a tiny bit of work is spent on a task that can
be cached.

This is most noticeable in dev setup where the HTTP cache is disabled.
The server had to always check if avatars are available. This is quite
slow and causes lots of requests to queue.

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Julius Härtl <jus@bitgrid.net>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

* ID is supposed to be the database ID
* UID is the IMAP UID
* UUID is the globally unique identifier used in the front-end

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: GretaD <gretadoci@gmail.com>
Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

When looking for changed and vanished UIDs we typically have to send the
list of the known UIDs to the IMAP server to know what went missing
meanwhile. With the number of UIDs sent, the size of the IMAP command
grows. At some point the server will just refuse the command due to its
size.
To circumvent this scenario, this patch splits the sync process into
new, changed and vanished. New is cheap in terms of data sent, so it
stays untouched. Changed and vanished will now split the known UIDs list
into chunks and retrieve partial sync results that way.

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

We need good training data to create a high quality classifier from the
user's data. If the data set does not contain enough important messages,
the ML would learn that few things are important and thus always develop
a bias towards non-important classification. To solve this we fall back
to a lower quality rule-based classifier that uses the same features as
the ML classifier. The importance is determined by hard-coded thresholds
of the individual features.

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

Signed-off-by: Holger Dehnhardt <holger@dehnhardt.org>
Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>

This will allow using kitinerary-extractor when it's installed on the
system but invoking our binary and flatpak fail (due to platform
incompatibilities).

Signed-off-by: Christoph Wurst <christoph@winzerhof-wurst.at>