6:06 a.m.
I was looking at the cache issue reported by customer. I found the cause
of it and couple of other related issues:
KEYCLOAK-3857 - Bad performance with clustered invalidation cache when
updating object
KEYCLOAK-3858 - Removing model object send lots of invalidation messages
across cluster
KEYCLOAK-3859 - Lots of userCache invalidation messages when
invalidating realm
KEYCLOAK-3860 - All realm users are invalidated from cache when removing
some realm object
In shortcut, our cache works fine in local mode. But in cluster, there
are issues with the invalidation caches . We don't have issues with
stale entries, but this is purchased but lots of various performance
issues like:
- There are lots of invalidation messages sent across the cluster
- Eviction on the node, which received invalidation event, is also very
uneffective. For example evicting realm with 1000 roles needs to call
1000 predicates, which iterates the cache 1000 times.
- Invalidation cache doesn't allow to differ between the context why the
object was invalidated. For example when I update realm settings on
node1, I need to invalidate just the CachedRealm object, but not all the
other objects dependent on the realm. However the invalidation event
received on node2 doesn't know, if I invalidated CachedRealm because of
realm update or because of realm removal. So for more safety, it assumes
removal, which evicts all realm objects! See
https://issues.jboss.org/browse/KEYCLOAK-3857 for details.
- Finally we have the workaround with the "invalidation.key" objects in
our invalidation caches. This is currently needed because when
invalidating object on node1, the invalidation event is NOT received on
node2 unless the object is there. Hence the workaround with the
"invalidation.key" records just to avoid this limitation of invalidation
cache.
For solve all these issues, I propose:
- Instead of relying on invalidation caches, we will send notification
across cluster what happened (eg. message "realm XY was updated"). All
the nodes will receive this notification and will evict all their
locally cached objects accordingly and bump their revisions locally.
This would be much more stable, performant and will allow us to remove
some workarounds.
Some details:
- The caches "realms" and "users" won't be
"invalidation" caches, but
they will be "local" caches.
- When any object needs to be removed from cache because of some reason
(eg. updating realm), the notification message will be sent from node1
to all other cluster nodes. We will use the replicated cache for that.
Node1 will send the notification message like "realm XY was updated" .
- Other cluster nodes will receive this message and they will locally
trigger evictions of all the objects dependent on particular realm. In
case of realm update, it's just the CachedRealm object itself. In case
of realm removal, it is all realm objects etc.
- Note message will contain also context "realm XY was updated" or
"realm XY was removed" . Not just "invalidate realm XY". This allows
much more flexibility and in particular avoid the issues like
https://issues.jboss.org/browse/KEYCLOAK-3857 .
- We already have replicated cache called "work", which we are using to
notify other cluster nodes about various events. So we will just use
this one. No need to add another replicated cache, we will probably just
need to configure LRU eviction for the existing one.
- Also note that messages will be always received. We won't need
workaround with "invalidation.key" objects anymore.
- Also we don't need recursive evictions (which has very poor
performance. See https://issues.jboss.org/browse/KEYCLOAK-3857 ),
because receiving node will know exactly what happened. It will remove
objects just the same way like the "sender" node.
- Finally the amount of traffic sent across the cluster will be much lower.
This sounds like the big step, but IMO it's not that bad :) Note that we
already have all the predicates in place for individual objects. The
only change will be about sending/receiving notifications across
cluster. I think I am able to prototype something by tomorrow to
doublecheck this approach working and then finish it somewhen middle
next week. WDYT?
Marek