It's the only way to properly clean all background tasks,
which the library makes heavy use for in MTProto/Connection
send and receive loops.
Some parts of the code even relied on the fact that it was
asynchronous (it used to return a future so you could await
it and not be breaking).
It's automatically syncified to reduce the damage of being
a breaking change.
Destructors are not guaranteed to run. Despite having good
intentions (saving entities even if the user forgets), it
should be the user's responsability to cleanly close the
client under any circumstances.
Excepting ValueError when creating the SQLiteSession could hide
other errors (e.g. using a newer session file on an older version
of the library). Instead, the original error is raised, as if
sqlite3 was being imported within its __init__ method.
* Create `_NOT_A_REQUEST` when needed. Currently, modifications
in the raised exception would be "global".
* `retries` parameters were actually attempts. This has been fixed
to actually be the amount of retries, so 0 now means don't retry.
* Helper function to deal with retries instead of using a range with
different styles every time.
Things like SQLAlchemy work correctly only for timezone-aware datetimes.
The returned TLObjects all have them, but those that are manually created
were missing them, so serializing the state into SQLAlchemy sessions failed.
Although this commit introduces a race condition since an
event may only be half-resolved. A lock is thus needed,
but it depends on an event-loop to which we don't have
access in the class-level.
This will also fix some bugs where the event type being accessed
would not be available, since it is now built on-demand, without
the need to keep track for the count of each event type.
This implies that export senders will NOT be deleted from
memory once all borrows are returned, thus their auth_key
remains as well. When borrowing them if they existed they
will be connect()ed if it's the first borrow.
This probably fixes#901.
