Steven Dolg wrote:
Sylvain Wallez schrieb:
Reinhard Pötz wrote:
<snip/>
But let me broaden the picture: Based on our work from about two weeks
ago, I created another aspect which implements the support for
conditional GET requests and also takes care that a pipeline isn't
executed unless it is really necessary. I was also able to fix all
failing test cases. I created an issue that contains a patch:
https://issues.apache.org/jira/browse/COCOON3-47
Additionally there is also another feature that I would like to add:
The
current patch only takes care of 'If-Modified-Since' requests. I also
want to support 'If-None-Match' requests that are based on the 'ETag'
response header. (see http://en.wikipedia.org/wiki/HTTP_ETag).
Using ETag has the advantage that we could support conditional GET
requests also in the case where we can't use a timestamp based approach
(e.g. when using o.a.c.pipeline.caching.ParameterCacheKey) or to
provide conditional GET support in REST controllers.
As an ETag value we could use the hash code of a pipeline's cache key.
I don't fully get the context of this conversation, but this last
sentence triggered a question to me: how can we validate a cache
entry with its _key_? Looking at the code, I see that CacheKey holds
both the identifier information (the actual key) and the validity
information.
There is a naming issue here which leads to some confusion between
key and key-and-validity that we can see it in the code:
ExpiresCacheKey doesn't include the validity information in
hashcode() and equals() whereas ParameterCacheKey does. What is the
right contract?
I'm not sure I understand what you mean.
The implementations of hashCode() in ExpiresCacheKey and
ParameterCacheKey are as similar in both the code and actual behaviour
as they can be.
Neither of them performs any operations necessary to check their
validity in the hashCode() or equals() methods.
Hmm... ok, so parameter values are part of the key. When reading the
code, and because of this mixing of validity and key (and lack of docs)
I thought the parameter keys were defining the identity and their values
were defining the validity.
Your confusion might arrise from the point that ParameterCacheKey
cannot become invalid because the same parameter value means always
the same parameter value, there is no way this can become invalid (as
opposed to a cache file contents which can become invalid when the
file is changed, even if it is still the same file)
So the isValid() method basically performs the equals check, since
this is a required condition for being valid. (valid = equal & not
expired; since expired = false here: valid = equal).
Ok. So with the definition of parameter keys and values being part of
the identity, ExpiresCacheKey and ParameterCacheKey effectively behave
consistenly.
Now you'll understand that this becomes really confusing, and if there's
not a very well defined contract for equals and hashcode (and even with
that) there's a big opportunity for people to implement wrongly their
CacheKey.
The ExpiresCacheKey performs an additional check in its isValid()
method, namely checking the expiresTimestamp.
This is not done in either the hashcode() or equals() method. So here
valid := equals & not expired.
This principle holds true for each and every CacheKey currently
implemented (unless there is faulty implementation).
And this is also the answer to your question:
CacheKey contains information to check its validity, but this
information is not used for identifying (iow, equals() and hashCode()
methods) CacheKeys.
Which means frequently invalidated CacheKeys will not fill the cache
but instead overwrite each other.
This is only true for cache implementations that rely on hashcode and
equals(), i.e. that keep an index in a Map or a Set (ehcache does this
for its DiskStore).
But if you use a non-java cache or persistent store, you have no other
solution than serializing the key and its validity information. This is
for example the case with memcache which requires the key to be a
String. And this is where the problem arises if you don't want or can't
keep an in-memory index, e.g. because of size or distribution.
As a side note, both classes include the class' hashcode in the
instance's hash code, which means hash codes will be different a
every JVM restart, or across JVM instances in a cluster, and is
likely to break persistent and distributed caches.
That is a good hint.
We will want to look into that and amend things if necessary.
Well, I would call it a bug that needs fixing, because it's basically
equivalent to clearing persistent caches at every JVM restart or class
reloading.
That being said, I'm wondering if this aggregation of key and
validity won't cause other kinds of problems with distributed cache
implementations. For example, Java memcached clients serialize the
cache key and use this result as the memcache key. If the key
includes validity information, the memcache key will change every
time the underlying data changes (e.g. a file's timestamp).
At first sight, this can sound good as it means we will have a cache
miss when the validity has changed, and will even avoid having to
compare the validity of cached content. But this can have a
desastrous impact on the cache efficiency in situations where you
have some often requested content that changes frequently: the cache
will quickly fill up with obsolete versions of this content under
different key values, that will lead older content to be evicted,
reducing the overall cache efficiency. Whereas a key that's only an
indentifier will lead the entry to be _replaced_ and not a new one
being added.
So in the end, my feeling is that key and validity information really
should be separated.
Now going back to the ETag discussion, using the pipeline's cache key
won't work IMHO because of the implementation of some key's
hashcode() using only the identifier part of the key and not the
validity. Confusion, I told you ;-)
We (intend to) use a layer for integrating caches since we don't want
to compile directly against the API of one specific provider and then
have to stick with that provider till the end of time (Avalon, anyone?)
Avalon certainly had its problems, but never mandated a particular cache
API beyond the one that we, Cocoon devs, defined in Excalibur
(org.apache.excalibur.store.Store). But even the Excalibur Store wasn't
a requirement, since it was just the store abstraction used by a
particular implementation of org.apache.cocoon.caching.Cache.
This additional layer is used to perform validity checks when
necessary and/or desired and not check the validity if not.
The intention here is to reuse the abstraction layer and not have this
kind of (critical) logic scattered in the individual cache provider
adaptors.
Agree. A cache storage should not have to do much more than get(key),
put(key, value) and maybe delete(key) and clear().
So it is possible to check if a CacheKey is pointing to the same
resource *and* if that cached data is still valid - even tho the
underlying cache provider has no means of performing the second check
(validity).
I totally agree. Now this doesn't solve the issue: to implement put(key,
value) on an arbitrary non-java store, I wouldn't trust the CacheKey's
hashcode() method to produce a uniform distribution that would avoid
conflicts (same hashcode for different CacheKeys). The solution would be
to to serialize the CacheKey and either use this as the store key if
it's not too long, or use a strong hash (e.g. MD5 or FNV) of this
serialized representation otherwise.
Mixing key and validity in a single CacheKey object means having several
store keys (which is different from _cache_ keys in this case) for
CacheKeys that are equal(), leading to the problems I outlined.
And BTW, what is the "jmxGroupName" property on CacheKey used for?
The jmxGroupName is used for making them accessible via JMX, no?
Well, I guessed this was somehow related to JMX ;-) Did my homework and
found its use in cocoon-monitoring's CacheEntrysMonitorInitializer. Nice
stuff, but I'm wondering if exposing the full key set of a big cache to
JMX actually scales.
And there are some cache implementations (again, memcached) that don't
expose their key set. But since this is mostly used for monitoring
AFAIU, returning an empty set in these cases should be acceptable.
Sylvain
--
Sylvain Wallez - http://bluxte.net
