As Krinkle points out, it always writes to DB_MASTER, so it doesn't make
much sense to offer an option to read from DB_REPLICA. This also
resolves the race condition that I believe arises from the one caller
that passes in DB_REPLICA.
Change-Id: I6976e5479debc3f4a8f28d53b1616c01475772be
To use WANObjectCache correctly in a multi-DC-safe way, we need to use
getWithSetCallback() to read data, and call delete() when it changes.
NotifUser's caching of notification counts and timestamps relied
heavily on set() calls, and so wasn't multi-DC-safe.
Changes in this commit:
* Rather than caching counts/timestamps in separate cache keys, and
using separate cache keys for each section (alert/message/all), put
all this data in an array and store that in a single cache key.
This reduces the number of cache keys per user per wiki from 6 to 1.
* Similarly, use a single global cache key per user. The global check
key for the last updated timestamp is retained, so we now have
2 global cache keys per user (down from 7)
* Remove preloading using getMulti(), no longer needed
* Move computation of counts and timestamps into separate compute
functions (one for local, one for global), and wrap them with
a getter that uses getWithSetCallback().
* Use TS_MW strings instead of MWTimestamp objects internally, to
simplify comparisons and max() operations.
* Make existing getters wrap around this new getter. They now ignore
their $cached and $dbSource parameters, and we should deprecate/change
these function signatures.
* In resetNotificationCounts(), just delete the cache keys. In global
mode, also recompute the notification counts and put them in the
echo_unread_wikis table. We could also set() the data into the cache
at this point, but don't, because you're not supposed to mix set() and
getWithSetCallback() calls and I don't want to find out what happens
if you do.
Bug: T164860
Change-Id: I4f86aab11d50d20280a33e0504ba8ad0c6c01842
We almost always call it with DB_MASTER, and I'm pretty sure that the
one call that uses DB_REPLICA introduces a race condition. I didn't want
to change that quite yet, though, so I left it in for now.
Change-Id: Ia5a59fdda357b799e327b8ed224f3ccb09509a8a
In principle, if $alertCount > 0, $alertTime should never be null,
but it feels much safer to check $alertTime itself instead.
Change-Id: Ifb107e257fbcd65e8181e2a214028f1b3733ac30
So that classes that get a UserNotificationGateway dependency-injected
into them can get something to call Database::getCacheSetOptions() on.
Change-Id: Ie1c98f09d1581ea17809a4dfc68adec1e2805580
The call of wfGetLangObj() is not necessary anymore.
Depends-On: I36834dbd119f4299f84b3dfe5e085eb0f6b8499c
Change-Id: I4e047393eee337432dd5edd0b4e3d650b6c7d393
NotificationJob used to include an instance
of EchoEvent, which is not serializable.
With this change, it only includes the
event Id (int) and retrieve the event
instance in its run() function.
Bug: T192945
Change-Id: I00950ddfa37717c7dfc19efdca9701693622da5d
