Hi All,
Apologies for the noise .. Still getting my bearings on how to contribute
(realized too late that gh issues were sourced from the jira project).
I want to bring to your attention a small proposed feature as described below:
The current CachedModeledFrameworkImpl doesn't manage cache initialization for
you. A perfect example of the kind of code you can expect to see in the wild is
in the CachedModeledFramework tests themselves, i.e. blocking on semaphores
that pin the reading thread to prevent it from certain operations that are
cache dependent (but as far as I can tell exactly which operations are cache
dependent is not really guaranteed so this is arguably cumbersome). Either way,
this is fine in a lot of cases...
However, I propose an additional InitializedCachedModeledFramework
implementation which asynchronously waits for the cache initialization trigger
and only then proceeds to read from the cache. I implemented something similar
for my personal use-case where I couldn't rely on, i.e. readThrough, to handle
the uninitialized case because readThrough cannot disambiguate between a znode
that is missing because it truly is missing from zk vs one that is missing
because the cache hasn't initialized. In my case the znode wouldn't always
exist in zk and so using readThrough would result in a lot of wasted calls to
zk, greatly reducing the benefit of the cache in the first place.
To reiterate InitializedCachedModeledFramework has a couple benefits over the
existing implementation:
1)No more possibility of misleading NoNodeException in
the case of reading from CachedModeledFramework before the cache has warmed. (I
say misleading because the node may exist .. just not in the cache)
2)No more temptation to add blocking semaphores in
front of this non-blocking interface.
3)IMO generally less confusion about how to properly
use this otherwise great(!) feature.
A first pass at an implementation is here
https://github.com/apache/curator/pull/1247/files ..
But I just wanted to gauge the community's sense of whether they think this is
a good idea and, assuming it is, if a new CMF implementation is the right way
to go about it.
I'd appreciate any feedback either way.
Thanks,
Luke
