6:49 p.m.
On 23 Feb 2019, at 02:46, Mark Reynolds <mreynolds(a)redhat.com>
wrote:
I want to start a brief discussion about a major problem we have backend transaction
plugins and the entry caches. I'm finding that when we get into a nested state of be
txn plugins and one of the later plugins that is called fails then while we don't
commit the disk changes (they are aborted/rolled back) we DO keep the entry cache
changes!
For example, a modrdn operation triggers the referential integrity plugin which renames
the member attribute in some group and changes that group's entry cache entry, but
then later on the memberOf plugin fails for some reason. The database transaction is
aborted, but the entry cache changes that RI plugin did are still present :-( I have also
found other entry cache issues with modrdn and BE TXN plugins, and we know of other
currently non-reproducible entry cache crashes as well related to mishandling of cache
entries after failed operations.
It's time to rework how we use the entry cache. We basically need a transaction
style caching mechanism - we should not commit any entry cache changes until the original
operation is fully successful. Unfortunately the way the entry cache is currently
designed and used it will be a major change to try to change it.
William wrote up this doc: http://www.port389.org/docs/389ds/design/cache_redesign.html
But this also does not currently cover the nested plugin scenario either (not yet). I do
know how how difficult it would be to implement William's proposal, or how difficult
it would be to incorporate the txn style caching into his design. What kind of time frame
could this even be implemented in? William what are your thoughts?
I like coffee? How cool are planes? My thoughts are simple :)
I think there is a pretty simple mental simplification we can make here though. Nested
transactions “don’t really exist”. We just have *recursive* operations inside of one
transaction.
Once reframed like that, the entire situation becomes simpler. We have one thread in a
write transaction that can have recursive/batched operations as required, which means that
either “all operations succeed” or “none do”. Really, this is the behaviour we want
anyway, and it’s the transaction model of LMDB and other kv stores that we could consider
(wired tiger, sled in the future).
If William's design is too huge of a change that will take too long to safely
implement then perhaps we need to look into revising the existing cache design where we
use "cache_add_tentative" style functions and only apply them at the end of the
op. This is also not a trivial change.
It’s pretty massive as a change - if we want to do it right. I’d say we need:
* development and testing of a MVCC/COW cache implementation (proof that it really really
works transactionally)
* allow “disable/disconnect” of the entry cache, but with the higher level txn’s so that
we can prove the txn semantics are correct
* re-architect our transaction calls so that they are “higher” up. An example is that
internal_modify shouldn’t start a txn, it should be given the current txn state as an arg.
Combined with the above, we can prove we haven’t corrupted our server transaction
guarantees.
* integrate the transactional cache.
I don’t know if I would still write a transactional cache the same way as I proposed in
that design, but I think the ideas are on the right path.
And what impact would changing the entry cache have on Ludwig's plugable backend
work?
Should be none, it’s seperate layers. If anything this change is going to make Ludwig’s
work better because our current model won’t really take good advantage of the MVCC nature
of modern kv stores.
Anyway we need to start thinking about redesigning the entry cache - no matter what
approach we want to take. If anyone has any ideas or comments please share them, but I
think due to the severity of this flaw redesigning the entry cache should be one of our
next major goals in DS (1.4.1?).
Thanks,
Mark
_______________________________________________
389-devel mailing list -- 389-devel(a)lists.fedoraproject.org
To unsubscribe send an email to 389-devel-leave(a)lists.fedoraproject.org
Fedora Code of Conduct: https://getfedora.org/code-of-conduct.html
List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines
List Archives:
https://lists.fedoraproject.org/archives/list/389-devel@lists.fedoraproje...
—
Sincerely,
William Brown
Software Engineer, 389 Directory Server
SUSE Labs