This adds a requirement for all definitions that have foreign keys to also
require indices for all source (local) attributes mentioned in foreign keys.
MariaDB/MySQL creates indices for source attributes automatically, so this
serves as a way to get rid of those automatic indices and create clean explicit
ones instead.
In PostgreSQL, most of the time, indices on the source are necessary to
decrease performance penalty of foreign keys (like in MariaDB), but they aren't
created automatically, so this serves to remove that difference between
PostgreSQL and MariaDB.
- getMessages() is now fetching from the Notice table as supposed
- every show{format}* method is properly updated to use Notice objects
- json and xml responses retrieve multi-recipients without compromising
backwards compatibility
Instead of relying on the MariaDB's ON UPDATE CURRENT_TIMESTAMP trigger update
"modified" attributes in Managed_DataObject. Every raw query that needs
adjusting is adjusted, as they won't update "modified" automatically anymore.
The main goal behind this change is to fix "modified" updates on PostgreSQL.
Password hashes are now stored in a TEXT attribute, not limited to 199 symbols.
That limitation makes no sense as password hashes are not the kind of
information to be indexed.
Actually replace crypt() with password_verify() for password checking, current
code left password_verify() unused.
Only update passwords when they use a different algorithm from the current
default. Previously "overwrite" meant rehashing every login.
Replace the "argon" boolean option with "algorithm" and "algorithm_options" for
better configurability.
The default remains whichever is default for PHP's password_hash.
This way UNKNOWN (NULL) explicitly turns to FALSE when three-valued logic is
reduced to binary.
In pgsqlschema, however, use "IS FALSE" as boolean attributes in pg_index are
non-nullable, there is no outer join and there's no clear preference for NULL
reduction.
Over-complicated constructions in TagCloud queries have been simplified, which
should not affect their performance.
Additionally, in TagCloud's lib/subscriptionspeopleselftagcloudsection.php
a typing mistake in an equi-join of "profile_tag" and "profile_list" on
"tagger" was fixed.
That regression was introduced in f446db8e2a
Argument 3 passed to htmloutputter::input() must be of the type string or null, array given, called in /srv/gnusocial/plugins/Bookmark/forms/bookmark.php on line 166
Argument 1 passed to xmloutputter::text() must be of the type string, null given, called in /srv/gnusocial/plugins/ExtendedProfile/lib/extendedprofilewidget.php on line 556
After adding a verb condition there, MariaDB now prefers the
("created", "id", "is_local") and ("profile_id", "verb", "created", "id")
indices for that query, even though they are slow for the job.
So replace them with ("is_local", "created", "id") and
("profile_id", "verb", "created", "id") respectively.
Also fix the naming of the ("profile_id", "created", "id") index.
There are only three possible visible notice verbs: POST, SHARE and DELETE.
What including all verbs does is it makes limiting (FETCH FIRST) unreliable as
the query will fetch invisible stuff (like favourites) and count it in, but
nothing will be displayed.
NoticeStream only allows POST and SHARE, so this effectively removes tombstones
from the profile page like in 78a111b57d
Introduce a new property in CachingNoticeStream for always checking if there
are any new elements in the stream.
It would be extremely hard to blow InboxNoticeStream, so instead the database
hit will still occur, but it is be much faster than starting fresh.
This fixes a regression introduced in 36a55d8436
There was no checking of attributedTo, actors and referent object IDs to make
sure they exist in the same domain. Therefore, one could spoof messages from
people by doing attributedTo: whoever-i-want-to-spoof
The notice.created sort forced the notice_tag by notice join plain to employ
materialisation, which can have a serious performance penalty depending on the
size of the database.
Sort by notice_tag.created instead, which should be exactly the same.
All verbs for not visible notices are filtered out, so this should not
break the timeline.
Additionally, filter by profile outside of the derived relation as that shows
better performance in PostgreSQL and MariaDB both.
It appears this was added to display "tombstones" of deleted notices.
However, it has other side-effects and the concept of keeping them visible has
not been adopted by the wider fediverse.
Avoid the use of deprecated MariaDB "zero dates" globally. If they're present
as attribute defaults somewhere, they will be replaced with NULL implicitly.
The existing "zero dates" in MariaDB storage will be left intact and this
should not present any issues.
The "timestamp" type in table definitions now corresponds to DATETIME in
MariaDB with "DEFAULT CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP", which
should be close enough to the original behaviour for compatibility purposes.
It is now the recommended type for "modified" attributes, because of the
update trigger on MariaDB. But there is no such trigger implemented on
PostgreSQL as of this moment.
