By resetting the replay state the transaction replay can start again on a
new server. This allows the replay process work when a master server is
shutting down.
By delaying the replay for a second, we give the monitor a small chance to
adap to master failures. It'll also prevent rapid re-querying if multiple
transaction replays are supported.
A transaction that just completed will go through the start_trx_replay
function as from the client protocol's point of view the transaction is
still open. The debug assertion did not take this into account and would
fail if a successful commit was the last thing done on master that failed.
Also fixed the formatting.
When a server is stopping, it'll send an error to the client before
terminating the TCP connection. The code in readwritesplit would detect
this error and create a hangup event on the DCB. This would cause it to
appear as if the TCP connection was broken and the router would
immediately try to reconnect to the same server.
By ignoring the error and allowing the connection to die on its own, we
avoid immediately reconnecting and retrying any transactions on the
stopping server. This increases the chances that the monitor will see it
first and assign the server states correctly before the transaction replay
is attempted.
If the server where a query is being executed is shutting down,
readwritesplit should treat it as an error to make retrying of the query
possible.
By treating server shutdowns as network errors, the same code path that is
used for actual network errors can be taken. This removes the need for any
extra retrying logic for this particular case.
The transaction replay could get mixed up with new queries if the client
managed to perform one while the delayed routing was taking place. A
proper way to solve this would be to cork the client DCB until the
transaction is fully replayed. As this change would be relatively more
complex compared to simply labeling queries that are being retried the
corking implementation is left for later when a more complete solution can
be designed.
This commit also adds some of the missing info logging for the transaction
replaying which makes analysis of failures easier.
As the router is the only one that knows what backends a particular
statement has been sent to, it is the responsibility of the router
to keep the session bookkeeping up to date. If it doesn't we will
know what statements a session has received (provided at least some
component in the routing chain has RCAP_TYPE_STMT_INPUT capability),
but not how long their processing took. Currently only readwritesplit
does that.
All queries are stored and not just COM_QUERY as that makes the
overall bookkeeping simpler; at clientReply() time we do not need to
know whether or not to bookkeep information, we can just do it.
When session information is queried for, we report as much information
we have available.
With causal_reads enabled, the query would return with an error if the
slave was not able to catch up to the master fast enough. By automatically
retrying the query on the master, we're guaranteed that a valid result is
always returned to the client.
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.
By splitting the processing and state querying into two separate
functions, the result can be inspected multiple times without triggering
the result processing.
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.
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 main piece of code, slave selection (backend_cmp_response_time), uses the available
method of pair-wise comparison of slaves. This will be changed to selection using all
available slaves, along with removal of hard coded values.
The configuration doesn't need to be contained in shared pointer as each
session holds its own version of it. This removes most of the overhead in
configuration reloading. The only thing that's left is any overhead added
by the use of thread-local storage.
The configuration used the wrong parameter name. The test also did not
explicitly enable tracking of the last_gtid variable which caused it to
fail if it wasn't already on.
The transaction migration in the case of a changed master never worked as
transaction replay would only be triggered when the master fails. To cover
this case, the transaction replay just needs to be started when the need
for a transaction migration is detected.
To help diagnose the behavior, the Trx class no longer logs a message when
a transaction is closed. This is now done by readwritesplit which has more
knowledge of the context in which the transaction is closed.
If a transaction is replayed, queued commands must not be processed. The
exception to this rule is when pending session commands are executed
before the first statement in the replayed transaction is executed.
If transaction replaying was enabled and a result was returned in more
than one call to clientReply, a NULL value would be added to the statement
which in turn would trigger a debug assertion.
Similarly any following statements in the transaction would be executed
regardless of whether the result was complete.
Renamed the statement execution function to better describe what it does.
Extended the basic functional test case to cover this.
By using a shared pointer instead of a plain object, we can replace the
router configuration without it affecting existing sessions. This is a
change that is required to enable runtime reconfiguration of
readwritesplit.
All executed session commands were logged in the RWSplitSession
destructor. This is not really necessary and shouldn't have been placed
there in the first place.
When the `optimistic_trx` mode is enabled, all transactions are started on
a slave server. If the client executes a query inside the transaction that
is not of a read-only nature, the transaction is rolled back and replayed
on the master.
Unconditionally update the previous target on each routed query. This
allows routing to the previous server in case it is needed. One example of
this is a new type of hint that allows routing to the same server where
the previous query was sent.
Also added a minor clarifying comment to the resetting of the
current_query.
Readwritesplit would crash with the following transaction:
BEGIN;
SET @a = 1; -- This is where it would crash
COMMIT;
When a session command was a part of the transaction, empty queries
(i.e. NULL GWBUFs) would be added to the transaction. If the transaction
were to be replayed, MaxScale would crash when these NULL queries were
executed.
Once the empty responses were fixed, the replaying of the transaction
would fail with a checksum mismatch. This was caused by the wrong order of
processing in RWSplitSession::clientReply. The response processing for
session commands was done after the response processing for replayed
transactions. This would trigger a checksum comparison too early for the
transaction in question.
Backend::execute_session_command would use the overridden write method
instead of the Backend::write method that it intended to use. This caused
session commands that did not expect a response to be in a state that
expected a result.
Also fixed RWBackend::write pass the response_type value to
Backend::write.
The state of the backend needs to be checked before any pending session
commands are executed on it.
Added debug assertions to catch invalid use of the status functions of
closed backends.
Allowing calls to select_connect_backend_servers even when all slaves are
connected solves the debug assertion in select_connect_backend_servers
that happens when the execution of a queued query causes a new connection
to be created.
If the starting of a transaction was interrupted by a server failure, the
query needs to be retried. This needs to be done as a transaction replay
to keep the routing logic consistent and simple.
When a non-autocommit transaction is interrupted, there will be no query
in progress and no replaying is needed. To handle this case, the replay
initialization logic needed to be altered to treat truly empty
transactions as a success case.
