`build_key(key, namespace)` is missing from some modules/classes

[padraic-shafer]

There are some parts of aiocache that do not honor namespaces. This means that BaseCache.build_key(key, namespace) is never called to join the namespace with the key. The affected modules/classes are:

• decorators.cached
• plugins.HitMissRatioPlugin
• backends.memory.SimpleMemoryCache

This affects sharing an aliased cache across multiple @cached-decorated callables. This currently works only when namespace==None.

A PR will be submitted shortly.

[Dreamsorcerer]

Needs test/reproducer.

[padraic-shafer]

OK, I'll add these tests during the next few days.

[padraic-shafer]

Here is a start...

I've added two new tests to "tests/acceptance/test_decorators.py::TestCached" in the test-namespace-key branch of this fork.

1. test_cached_without_namespace passes because the cache key is created with the default format when no namespace is provided.
2. test_cached_with_namespace fails because the default key for a decorated function/coroutine does not include the namespace.

Run these tests:

pytest tests/acceptance/test_decorators.py::TestCached

[padraic-shafer]

The other changes included in https://github.com/aio-libs/aiocache/pull/570 are related to using cache keys that are not strings (notably tuple keys), so I will suggest those separately in a new issue.

@Dreamsorcerer if you agree that the ~tests above are~ test above is worth resolving, I'll submit a new PR with just those changes to "decorators.py" and their corresponding tests.

[Dreamsorcerer]

Tests look reasonable, go ahead and complete the PR.

The other changes included in #570 are related to using cache keys that are not strings (notably tuple keys), so I will suggest those separately in a new issue.

I was under the impression that cache keys are expected to be strings always..

[padraic-shafer]

Tests look reasonable, go ahead and complete the PR.

Thanks. Will do.

I was under the impression that cache keys are expected to be strings always..

Yeah, in retrospect I'm seeing that now more than when I started using aiocache some months ago. FWIW, Using non-string keys works quite well with the existing package when using a custom key_builder and the key_filter that I suggested in https://github.com/aio-libs/aiocache/pull/570.

[padraic-shafer]

...but I only use the SimpleMemory cache at the moment; not redis or memcached.

[Dreamsorcerer]

Yeah, I suspect it will work in some/many cases, but I don't think it's something we want to actively support. The code seems to assume str keys in many places and I think we'll keep typing support simple if we stick to str.

[padraic-shafer]

I've updated the tests to correctly pass the namespace to cache.exists(). This is consistent with cached.get_cache_key() returning a "raw" key without a namespace as prefix.

[padraic-shafer]

I needed to update the tests again to use the namespace from the cache fixtures. Otherwise these always overwrite the arbitrary namespace I originally chose for the test.

[padraic-shafer]

After reviewing the motivation for submitting this issue, I think it might be helpful to break this into multiple pieces.

1. Observation: the key_builder argument to BaseCache and its subclasses joins a namespace and a "raw" key; whereas the key_builder argument to decorator classes builds a "raw" key from the callable and its arguments when it is called.

2. Using a custom key_builder with a cache would be better supported after some simple updates:

   a. Include examples of using a custom key_builder when creating a cache. b. BaseCache._build_key uses namespace argument if provided, otherwise it uses self.namespace. A custom key_builder breaks this expectation because it cannot access self.namespace -- there is no self for the externally provided key_builder.

   • RECOMMENDATION: BaseCache public members that accept a namespace parameter (add/set/get/etc.) should pass self.namespace to their private equivalents (add/set/get/etc.) if the namespace argument is None.
   • This enables consistent "fallback" behavior for BaseCache.build_key, regardless of whether it delegates to _build_key or key_builder;
   • it enables decorator classes to use alias caches that were created with a namespace and a key_builder;
   • and it enables cache locks to use client caches that were created with a a custom key_builder. c. Cache locks do not generate the correct lock key when using a client cache that was created with a a custom key_builder.
   • RECOMMENDATION: Cache locks should call client.build_key rather than client._build_key.

3. Clashes between cache key_builder and decorator class key_builder. a. Decorators cannot be used with ad hoc caches that need a custom key_builder. The decorator class key_builder parameter shadows the parameter with the same name that would have otherwise been passed to the underlying cache via **kwargs.

   • RECOMMENDATION: Add a parameter to [multi_]cached.__init__ that enables a user to pass a key builder to the underlying cache. For example, cache_key_builder. In general **kwargs could be filtered to pop any keyword arguments that start with "cache_", strip the prefix and pass it to the underlying cache.
   • RECOMMENDATION: This changes the interface and probably merits more discussion. It could be considered separately for a future release. Meanwhile, decorators can use an aliased cache with a key builder because of the changes in Item 2 (above). b. Make namespace an explicit parameter to decorator caches, to better indicate that the namespace inclusion does not happen in the key_builder argument passed to the decorator. Instead, this is performed by the underlying cache, in self.cache.build_key().

4. A custom key_builder passed to a decorator class has no access to self.noself and so must hardcode or introspect the decorated callable to determine whether to use the first argument (self) in the key that is generated.

   • RECCOMMENDATION: key_builder should include a noself parameter that could be passed explicitly by the user or implicitly by decorator.get_cache_key() a la Item 2b (above).
   • RECOMMENDATION: This changes the interface and probably merits more discussion. It could be considered separately for a future release.

5. It would be helpful to issue a warning when a decorator class is instantiated with both alias and other parameters that do not get used because "alias takes precedence".

6. Included in the original issue here was an enhancement to use namespace keys without requiring ns_key.startswith(namespace). I proposed a key_filter(ns_key, namespace) member for BaseCache and _key_filter(ns_key, namespace) for backends that returns True if ns_key is in the namespace, namespace; otherwise it returns False. This would be used by clear()/_clear() and other cache key inspections.

   • RECOMMENDATION: This changes the interface and probably merits more discussion. It could be considered separately for a future release.

7. The issue with plugins.HitMissRatioPlugin that was originally referenced here was due to an old version that did not use **kwargs in post_get(). It was resolved by #507.

[padraic-shafer]

All changes to the test suite were included in the test-namespace-key branch of this fork before merging them into the PR. This makes it simple to run the tests on the codebase from the main branch.

[Dreamsorcerer]

I think this mostly resolved now, with improvements to come in PRs for v1.

Feel free to open new issues to track specific problems, but I suspect you'll cover it all in another PR anyway.