curator-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Nick Hill <>
Subject Re: TreeCache question
Date Thu, 28 Jan 2016 15:21:51 GMT
Thanks Cameron, Scott,

> I'm not aware of any
> concurrency bugs in TreeCache right now, I don't think I relied on the
> atomic refs for safety.

I guess I wasn't aware in general that atomic refs had a purpose other  
than this.

It just seems that not guaranteeing consistency of the stat and data  
in the returned ChildData objects restricts the usefulness of the  
cache, since you would typically have to get the data/stat again  
directly before doing any conditional update to a znode.

Ignoring the less-important suggestion for simplifying ChildData  
itself, wouldn't there still be significant value in a simple change  
to move the TreeNode data and stat atomic refs into a single ChildData  
atomic ref?


Quoting Scott Blum <>:

> HI Nick,
> TreeCache came later, and literally the only reason for reusing ChildData
> was to not create additional API surface area, and make for an easier
> transition / drop-in replacement for NodeCache & PathChildrenCache.
> Ordinarily, I'm a big fan of immutable objects.  I'm not aware of any
> concurrency bugs in TreeCache right now, I don't think I relied on the
> atomic refs for safety.
> The main issue with changing it would just be additional surface area or
> breaking old client code.
> HTH!
> Scott
> On Wed, Jan 27, 2016 at 10:16 PM, Cameron McKenzie <>
> wrote:
>> hey Nick,
>> Sounds like a reasonable suggestion to me. I'll let Scott chime in though
>> as he wrote that code originally and may have had some reason for
>> structuring it as he has.
>> cheers
>> On Thu, Jan 28, 2016 at 1:42 PM, Nick Hill <> wrote:
>> > Hi, I have been looking at the TreeCache impl and have some questions.
>> >
>> > It doesn't look right to me that there's separate atomic refs for a
>> node's
>> > data and stat. It seems the stat in a ChildData object obtained from
>> > getCurrentData() might not correspond to the data that it's with. This
>> > could be problematic when doing conditional state changes given
>> assumptions
>> > about that data.
>> >
>> > An obvious and simple solution to this would be to have a single
>> > AtomicReference<ChildData> field instead, which would have the additional
>> > significant benefit of eliminating ChildData obj creation on every cache
>> > access. PathChildrenCache works this way, but my understanding was that
>> > TreeCache is intended to be a (more flexible) replacement.
>> >
>> > Furthermore I'd propose that the data field of ChildData be just a final
>> > byte[] instead of an AtomicReference. This would avoid needing two
>> volatile
>> > reads to get to the data, and mean that "sharing" these (per above) is a
>> > bit safer. The ChildData byte[] AtomicReference is only used by
>> > PathChildrenCache.clearDataBytes() (not currently used by TreeCache at
>> > all), and that capability could be easily maintained by having
>> > PathChildrenCache use it's own simple subclass of ChildData containing
>> the
>> > atomic reference.
>> > If similar capability were to be added to TreeCache, I'd suggest it would
>> > be better to just replace the node's ChildData object with a copy that
>> has
>> > the byte[] field nulled out (but same stat ref).
>> >
>> > I'm fairly new to the code so apologies if there's something I've
>> > missed/misunderstood! But if there is agreement, I'd also be happy to
>> > prepare a PR.
>> >
>> > Regards,
>> > Nick
>> >

View raw message