If there is a PHP fatal error and destructors do not run (this can
happen in out-of-memory errors, and maybe if there is an error in a
previous destructor) then Postgres cursors may be left open.
Usually this does not cause a problem because the connection is
closed anyway, but if using persistent connections, a future
request may reuse the connection with a cursor open. It then gets
errors when it tries to create a new cursor with the same name.
This change closes all cursors at the start of a persistent
connection.
It's more memory efficient to use `pg_fetch_assoc` for each row than to
call `pg_fetch_all` and release memory immediately. This is because we
can treat the assoc fetch like an iterator and it only fetches the
current record into memory one at a time, whilst the all fetch fetches
all records and never unsets them. Attempting to unset them is extremely
time consuming.
- Clumsy fallback only when there is no full-text search support
- Mimic solr tests
- pgsql tokenization using simple configuration
- workaround for mysql '*' search issue
- total results proper calculation
- SQL server FTS support
- Standarize dml full-text search checkings
- Upgrade note about the new dml method
- Set search_simpledb as default engine if no solr config
On Postgres, get_recordset_sql loads all the results into memory
(within the Postgres library, which doesn't count towards the PHP
memory limit, but does count towards making your server run out of
memory) as soon as the query completes.
This commit changes the code to use cursors, which in Postgres
allow the results to be returned in smaller chunks (by default
100,000 rows).
pgbouncer doesn't accept the PostgreSQL options command as it can't
process it. If you are using pgBouncer you need to make sure your
database is configured to set client encoding and standard_conforming_strings
correctly on each connection. pgBouncer can do this, as can ALTER
commands.
Also using sockets without a filename did not allow different ports to work.
This is because you either specify a filename which includes a port or
you just specify a port and libpq works out the correct socket location.
PostgreSQL 9.1 allows hex formating for binary which is handled better
by pg_query_params().
Getting bytea isn't required on connection, it can be used as pg_field_type()
when binary needs to be checked.
When insert_record is called and it doesn't find any columns
to insert data into, that is a dml_exception as you need at least
one column in the database to insert into it. This can happen
when you have created a zero column database table, or the table
does not exist.
Includes following fixes:
* support for MS SQL Server
* optional trimming of of oversized VARCHAR fields
* conversion to forms library
* full localisation
* other cleanup
List of changes:
* New OOP API using PHP namespace \core\session\.
* All handlers now update the sessions table consistently.
* Experimental DB session support in Oracle.
* Full support for session file handler (filesystem locking required).
* New option for alternative session directory.
* Official memcached session handler support.
* Workaround for memcached version with non-functional gc.
* Improved security - forced session id regeneration.
* Improved compatibility with recent PHP releases.
* Fixed borked CSS during install in debug mode.
* Switched to file based sessions in new installs.
* DB session setting disappears if DB does not support sessions.
* DB session setting disappears if session handler specified in config.php.
* Fast purging of sessions used in request only.
* No legacy distinction - file, database and memcached support the same functionality.
* Session handler name included in performance info.
* Fixed user_loggedin and user_loggedout event triggering.
* Other minor bugfixing and improvements.
* Fixed database session segfault if MUC disposed before $DB.
Limitations:
* Session access time is now updated right after session start.
* Support for $CFG->sessionlockloggedinonly was removed.
* First request does not update userid in sessions table.
* The timeouts may break badly if server hosting forces PHP.ini session settings.
* The session GC is a lot slower, we do not rely on external session timeouts.
* There cannot be any hooks triggered at the session write time.
* File and memcached handlers do not support session lock acquire timeouts.
* Some low level PHP session functions can not be used directly in Moodle code.
The query of the information_schema view seems to be painfully slow on
older versions of postgres (2000ms instead of 20ms). Therefore, it is
worth detecting that case, and using a more postgres specific feedback.
This is particularly important for 2.3+ unit tests.
Thanks to Petr Skoda for his help with this.
