curator-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jordan Zimmerman <jor...@jordanzimmerman.com>
Subject Re: Persistent locks
Date Fri, 17 Jan 2014 16:48:01 GMT
I’m not sure. You can give it a try. I can’t guarantee that we’ll accept the patch. That
code is already very complicated. But, if it’s reasonable we’ll take it.

-JZ

From: Jozef Vilcek Jozef Vilcek
Reply: Jozef Vilcek jozo.vilcek@gmail.com
Date: January 16, 2014 at 11:20:58 PM
To: Jordan Zimmerman jordan@jordanzimmerman.com
Subject:  Re: Persistent locks  
Jordan,

so, do you think it is feasible to do such changes into Curator locks? Via changes into LockInternalsDriver?
If yes, should I go ahead and create a JIRA and draft a patch?

Best,
Jozef


On Wed, Jan 15, 2014 at 12:06 PM, Jozef Vilcek <jozo.vilcek@gmail.com> wrote:
Yes. I was digging a bit in curator code and thinking about how can I implement this. If I
would be able to somehow intercept creation of lock nodes, I could build GC/heartbeat logic
around the lock by "installing" appropriate hooks.

One idea is that lock would accept a kind of LockNodeFactory, which would be responsible for
creating actual zookeeper nodes. LockInternals.attemptLock() would delegate construction to
the factory. There, in my own factory, I could add posibility to listen to created lock nodes
and apply GC/heartbeat logic.

As you suggested, same could be done if node creation would be moved to the LockInternalsDriver
and made public/reusable. I like this even better. It could be tricky how to reuse e.g. read-write
lock, but it would not be so painful to "re-implement" only that part of curator code on my
end.


Jozo


On Tue, Jan 14, 2014 at 6:51 PM, Jordan Zimmerman <jordan@jordanzimmerman.com> wrote:
You might be able to re-use the code in the LockInternals class. It could easily be modified
to do what you want via the LockInternalsDriver. Of course it would need to be made public.
The gc/heartbeat stuff would have to be coded fresh, though.

-Jordan

From: Jozef Vilcek Jozef Vilcek
Reply: Jozef Vilcek jozo.vilcek@gmail.com
Date: January 14, 2014 at 12:18:15 AM
To: Jordan Zimmerman jordan@jordanzimmerman.com
Subject:  Re: Persistent locks
Yes, you are correct. I tried to sketch this in my post. There will be possibility to create
persistent lock, which will cause deadlock if holder crashes and one with possibility to to
be gc collected after some time if holder crashes. The mechanism for this has to be implemented.

Off course that I use "classical" ephemeral locks where possible. But I have few cases where
it is not feasible.


On Tue, Jan 14, 2014 at 2:42 AM, Jordan Zimmerman <jordan@jordanzimmerman.com> wrote:
How will you avoid dead locks? You’ll need to write some kind of heartbeat/gc mechanism
to the lock. Otherwise I don’t see how it works if the lock holder crashes.

-JZ

From: Jozef Vilcek Jozef Vilcek
Reply: user@curator.apache.org user@curator.apache.org
Date: January 13, 2014 at 7:24:06 AM
To: user@curator.apache.org user@curator.apache.org
Subject:  Persistent locks
Hi

I have a question about curator locks. I see that locks are implemented via znode type EPHEMERAL_SEQUENTIAL.
I am thinking about having an implementation via PERSISTENT_SEQUENTIAL. 

Main reason for this are processes with critical sections, where we can not afford to loose
a lock due to session expiration. In such case, others might acquire a lock and kick in while
the previous process is still running but e.g. experiencing connection issues. To kill this
temporally detached process in favor of others would be too costly.

My thoughts are to have:

* Persistent lock - if something go south and client code does not release lock, it will stay
there until removed by manual or some other intervention

* Persistent ephemeral lock - this would be ephemeral implemented by persistent lock. For
not so much critical stuff but e.g. for unstable environments. This would install kind of
refresh hook on a created lock node. Other clients waiting to acquire lock could garbage collect
locks which does not received refresh for reasonable long time (scaling beyond session timeouts).

What I would like to know:

* any wisdom, if this does make sense or if there is a better way out there
* support from curator ... There is a lot of good code in curator I would have to copy to
make this work. I want to avoid this. Would it be possible of provide either path for making
locks core extensible/reusable (or to contribute implementation of locks if considered worth
for framework) ?


Best,
Jozo




Mime
View raw message