When a read-only transaction fails due to a connection error, no message
would be logged. Also added an info level message for the case when a
backend connection would get closed before the session is in the correct
state and a debug assertion that the router session should never be closed
when the handleError method is called.
The binlogrouter uses buffers across worker threads which is no longer OK
in 2.3. The correct solution would be to store data in something other
than a GWBUF (e.g. std::vector) and protect the sharing with a mutex. The
current solution simply works around the assertions by using macros
instead of functions.
By biasing the values of all counter type scores to positive integers, the
server weights are always taken into use.
This fixes the case when weights were ignored until all score base values
were larger than zero (the mxs922_server test).
If the DCB is closed in handleError, it would be NULL in closeSession. To
only close the DCB in one place, the handleError can be reduced to writing
an error to the client and marking the failure as a fatal one.
The check for a closed session should never be needed as the core
correctly orders the calls to the module functions. The version numbering
was also not used.
Fixed the incrementation of the query counter so that it uses relaxed
atomic operations.
Replaced SPINLOCK with std::mutex where possible, leaving out the more
complex cases. The big offenders remaining are the binlogrouter and the
gateway.cc OpenSSL locks.
Removed the almost equal comparison and subsequent selection based on historical number of connections.
The effect of it was this: Select the server that has historically, weights or not, been slower. Tested this with 2.2
with maxscale on one server and mariadb:s on two servers with different network lags. The tests with historical
selects were clearly slower.
The debug assertion is wrong as the code was changed to prioritize hints
over the router target selection. Also removed the superficial check for
master, slave and relay master states as they are implied by the fact that
the connection is in use.
The readwritesplit transaction management was a large part of the
clientReply function. Moving it into a separate function clarifies the
clientReply function by hiding the comments and details of the transaction
management.
The schemarouter now uses the RWBackend to track the response states. This
fixes the debug assertions that happened with the mxs1113_schemarouter_ps
test.
By splitting the processing and state querying into two separate
functions, the result can be inspected multiple times without triggering
the result processing.
This column will be used to store the relative path of the file
where a particular event can be found.
Unless the path is stored, BLR will not be able to find an event based
on the gtid if BRL is connected to a node in a Galera cluster and updates
are made to nodes other than that node as in that case, the GTID domain id
and server id, will not identify the correct directory.
The full path is not stored in the column binlog_file as the path
would in that case (without other modifications) be visible to the
client.
The collection of resultsets needs to be disabled by default when a
response is received to cover the cases where an error is returned.
The collection of results should also not be set for queries that do not
generate any responses.
The read-write distribution in readwritesplit is now stored in a map
partitioned by the servers that the router has used. Currently, the
statistics for removed servers aren't dropped so some filtering still
needs to be added.
This only deactivates some of the more heavy-handed features. All files have
been formatted with the new settings, which seems to have only affected lines
which were not formatted before.
When the setting up of filters for a session fails, the DCB is closed and
the client DCB's session pointer is set to NULL. This needs to be checked
in the schemarouter before the `m_client->session` pointer is used.
The act of setting the session pointer to NULL should not be necessary as
the session is freed once the reference count drops down to zero. Due to
the fact that changing this would require moderate changes in session code
means that it should not be done in a patch release as the risks are too
high.
The additions into the server.h header used C++ language which caused C
programs to fail to compile. Moved the implementation of the EMAverage
class into the private Server class in the server.hh header and exposed it
via functions in the server.h header. Also temporarily moved
almost_equal_server_scores into the public server.hh as there is no
service.hh header.
Changes to the ChangeMasterOptions made it non-trivial so it cannot be
safely memset anymore. The maxavrocheck was missing the linkage to the
maxscale-common library.
See script directory for method. The script to run in the top level
MaxScale directory is called maxscale-uncrustify.sh, which uses
another script, list-src, from the same directory (so you need to set
your PATH). The uncrustify version was 0.66.
Changes that allow slow or new servers to quickly apply samples towards the
server average. The most important changes are to not ignore the first N samples,
and apply an average to the server as soon as there is one available.
The new ResponseStat::make_valid() will use filter samples to add an average,
if no averages have yet been added, even if the number of filter samples is less
than the filter limit.
The math becomes simpler when the weight is inverted, i.e. a simple multiplication
to get the (inverse) score. Inverse weights are normalized to the range [0..1] where a lower
number is a higher weight,
The enum select_criteria_t is used to provide a std::function that takes the backends
as vector (rather than the prior pairwise compares) and returns the best backend.
