If there are queued request when connection closes we may try to process
those requests (by trying to send them to already terminated db process,
and waiting until we hit timeout) before we see that driver is not longer
alive. This change adds check for driver exit before we process next
queued sql request, and immediately switch to connection state if we have
one.
This can be used for all upsert expressions (where REPLACE INTO used
previously were only possible to use for subset of queries), and may
potentially help with deadlocks reported by mysql when we issues multiple
querier for same key in quick succession.
Don't forget the push session ID if the client resumes a stream
management session. This makes sure no push notifications are generated
for a connected client with a resumed stream management session.
Thanks to Thilo Molitor for spotting the issue.
The fact that mod_push uses a timestamp as a push session ID is just an
implementation detail. Rename things accordingly to avoid confusion.
However, stick to "timestamp" in the storage modules, as it corresponds
to the database column names/types, there.
Fix the problem that no push notifications were triggered in the case
where all of the following conditions applied:
- The XEP-0198 session has been resumed at least once.
- The XEP-0198 session is pending (i.e., the client is disconnected).
- The message wasn't stored in the user's MAM archive (e.g., because it
was of type "groupchat").
The problem was introduced in cd336369a5.
Many thanks to Friedrich Altheide for reporting the issue and to Thilo
Molitor for his help with tracking it down.
Allow for specifying the 'type' of 'domain_balancing' without specifying
the 'component_number' (as per the example in the documentation). The
balancing 'type' is then applied to the dynamic number of component
instances.