Dogpile Usage Guide
At its core, Dogpile provides a locking interface around a "value creation" function.
The interface supports several levels of usage, starting from one that is very rudimentary, then providing more intricate usage patterns to deal with certain scenarios. The documentation here will attempt to provide examples that use successively more and more of these features, as we approach how a fully featured caching system might be constructed around Dogpile.
Note that when using the dogpile.cache package, the constructs here provide the internal implementation for that system, and users of that system don't need to access these APIs directly (though understanding the general patterns is a terrific idea in any case). Using the core Dogpile APIs described here directly implies you're building your own resource-usage system outside, or in addition to, the one dogpile.cache provides.
A simple example:
from dogpile import Dogpile # store a reference to a "resource", some # object that is expensive to create. the_resource = [None] def some_creation_function(): # create the resource here the_resource = create_some_resource() def use_the_resource(): # some function that uses # the resource. Won't reach # here until some_creation_function() # has completed at least once. the_resource.do_something() # create Dogpile with 3600 second # expiry time dogpile = Dogpile(3600) with dogpile.acquire(some_creation_function): use_the_resource()
Above, some_creation_function() will be called when :meth:`.Dogpile.acquire` is first called. The remainder of the with block then proceeds. Concurrent threads which call :meth:`.Dogpile.acquire` during this initial period will be blocked until some_creation_function() completes.
Once the creation function has completed successfully the first time, new calls to :meth:`.Dogpile.acquire` will call some_creation_function() each time the "expiretime" has been reached, allowing only a single thread to call the function. Concurrent threads which call :meth:`.Dogpile.acquire` during this period will fall through, and not be blocked. It is expected that the "stale" version of the resource remain available at this time while the new one is generated.
By default, :class:`.Dogpile` uses Python's threading.Lock() to synchronize among threads within a process. This can be altered to support any kind of locking as we'll see in a later section.
Using a Value Function with a Cache Backend
The dogpile lock includes a more intricate mode of usage to optimize the usage of a cache like Memcached. The difficulties Dogpile addresses in this mode are:
- Values can disappear from the cache at any time, before our expiration time is reached. Dogpile needs to be made aware of this and possibly call the creation function ahead of schedule.
- There's no function in a Memcached-like system to "check" for a key without actually retrieving it. If we need to "check" for a key each time, we'd like to use that value instead of calling it twice.
- If we did end up generating the value on this get, we should return that value instead of doing a cache round-trip.
To use this mode, the steps are as follows:
- Create the Dogpile lock with init=True, to skip the initial "force" of the creation function. This is assuming you'd like to rely upon the "check the value" function for the initial generation. Leave it at False if you'd like the application to regenerate the value unconditionally when the dogpile lock is first created (i.e. typically application startup).
- The "creation" function should return the value it creates.
- An additional "getter" function is passed to acquire() which should return the value to be passed to the context block. If the value isn't available, raise NeedRegenerationException.
from dogpile import Dogpile, NeedRegenerationException def get_value_from_cache(): value = my_cache.get("some key") if value is None: raise NeedRegenerationException() return value def create_and_cache_value(): value = my_expensive_resource.create_value() my_cache.put("some key", value) return value dogpile = Dogpile(3600, init=True) with dogpile.acquire(create_and_cache_value, get_value_from_cache) as value: return value
Note that get_value_from_cache() should not raise :class:`.NeedRegenerationException` a second time directly after create_and_cache_value() has been called.
Using Dogpile for Caching
Dogpile is part of an effort to "break up" the Beaker package into smaller, simpler components (which also work better). Here, we illustrate how to approximate Beaker's "cache decoration" function, to decorate any function and store the value in Memcached. We create a Python decorator function called cached() which will provide caching for the output of a single function. It's given the "key" which we'd like to use in Memcached, and internally it makes usage of its own :class:`.Dogpile` object that is dedicated to managing this one function/key:
import pylibmc mc_pool = pylibmc.ThreadMappedPool(pylibmc.Client("localhost")) from dogpile import Dogpile, NeedRegenerationException def cached(key, expiration_time): """A decorator that will cache the return value of a function in memcached given a key.""" def get_value(): with mc_pool.reserve() as mc: value = mc.get(key) if value is None: raise NeedRegenerationException() return value dogpile = Dogpile(expiration_time, init=True) def decorate(fn): def gen_cached(): value = fn() with mc_pool.reserve() as mc: mc.put(key, value) return value def invoke(): with dogpile.acquire(gen_cached, get_value) as value: return value return invoke return decorate
Above we can decorate any function as:
@cached("some key", 3600) def generate_my_expensive_value(): return slow_database.lookup("stuff")
The Dogpile lock will ensure that only one thread at a time performs slow_database.lookup(), and only every 3600 seconds, unless Memcached has removed the value in which case it will be called again as needed.
In particular, Dogpile's system allows us to call the memcached get() function at most once per access, instead of Beaker's system which calls it twice, and doesn't make us call get() when we just created the value.
Scaling Dogpile against Many Keys
The patterns so far have illustrated how to use a single, persistently held :class:`.Dogpile` object which maintains a thread-based lock for the lifespan of some particular value. The :class:`.Dogpile` also is responsible for maintaining the last known "creation time" of the value; this is available from a given :class:`.Dogpile` object from the :attr:`.Dogpile.createdtime` attribute.
For an application that may deal with an arbitrary number of cache keys retrieved from a remote service, this approach must be revised so that we don't need to store a :class:`.Dogpile` object for every possible key in our application's memory.
The two challenges here are:
- We need to create new :class:`.Dogpile` objects as needed, ideally sharing the object for a given key with all concurrent threads, but then not hold onto it afterwards.
- Since we aren't holding the :class:`.Dogpile` persistently, we need to store the last known "creation time" of the value somewhere else, i.e. in the cache itself, and ensure :class:`.Dogpile` uses it.
The approach is another one derived from Beaker, where we will use a registry that can provide a unique :class:`.Dogpile` object given a particular key, ensuring that all concurrent threads use the same object, but then releasing the object to the Python garbage collector when this usage is complete. The :class:`.NameRegistry` object provides this functionality, again constructed around the notion of a creation function that is only invoked as needed. We also will instruct the :meth:`.Dogpile.acquire` method to use a "creation time" value that we retrieve from the cache, via the value_and_created_fn parameter, which supercedes the value_fn we used earlier. value_and_created_fn expects a function that will return a tuple of (value, created_at), where it's assumed both have been retrieved from the cache backend:
import pylibmc import pickle import time from dogpile import Dogpile, NeedRegenerationException, NameRegistry mc_pool = pylibmc.ThreadMappedPool(pylibmc.Client("localhost")) def create_dogpile(key, expiration_time): return Dogpile(expiration_time) dogpile_registry = NameRegistry(create_dogpile) def get_or_create(key, expiration_time, creation_function): def get_value(): with mc_pool.reserve() as mc: value = mc.get(key) if value is None: raise NeedRegenerationException() # deserialize a tuple # (value, createdtime) return pickle.loads(value) def gen_cached(): value = creation_function() with mc_pool.reserve() as mc: # serialize a tuple # (value, createdtime) value = (value, time.time()) mc.put(mangled_key, pickle.dumps(value)) return value dogpile = dogpile_registry.get(key, expiration_time) with dogpile.acquire(gen_cached, value_and_created_fn=get_value) as value: return value
Stepping through the above code:
- After the imports, we set up the memcached backend using the pylibmc library's recommended pattern for thread-safe access.
- We create a Python function that will, given a cache key and an expiration time, produce a :class:`.Dogpile` object which will produce the dogpile mutex on an as-needed basis. The function here doesn't actually need the key, even though the :class:`.NameRegistry` will be passing it in. Later, we'll see the scenario for which we'll need this value.
- We construct a :class:`.NameRegistry`, using our dogpile creator function, that will generate for us new :class:`.Dogpile` locks for individual keys as needed.
- We define the get_or_create() function. This function will accept the cache key, an expiration time value, and a function that is used to create a new value if one does not exist or the current value is expired.
- The get_or_create() function defines two callables, get_value() and gen_cached(). These two functions are exactly analogous to the the functions of the same name in :ref:`caching_decorator` - get_value() retrieves the value from the cache, raising :class:`.NeedRegenerationException` if not present; gen_cached() calls the creation function to generate a new value, stores it in the cache, and returns it. The only difference here is that instead of storing and retrieving the value alone from the cache, the value is stored along with its creation time; when we make a new value, we set this to time.time(). While the value and creation time pair are stored here as a pickled tuple, it doesn't actually matter how the two are persisted; only that the tuple value is returned from both functions.
- We acquire a new or existing :class:`.Dogpile` object from the registry using :meth:`.NameRegistry.get`. We pass the identifying key as well as the expiration time. A new :class:`.Dogpile` is created for the given key if one does not exist. If a :class:`.Dogpile` lock already exists in memory for the given key, we get that one back.
- We then call :meth:`.Dogpile.acquire` as we did in the previous cache examples, except we use the value_and_created_fn keyword for our get_value() function. :class:`.Dogpile` uses the "created time" value we pull from our cache to determine when the value was last created.
An example usage of the completed function:
import urllib2 def get_some_value(key): """retrieve a datafile from a slow site based on the given key.""" def get_data(): return urllib2.urlopen( "http://someslowsite.com/some_important_datafile_%s.json" % key ).read() return get_or_create(key, 3600, get_data) my_data = get_some_value("somekey")
Using a File or Distributed Lock with Dogpile
The final twist on the caching pattern is to fix the issue of the Dogpile mutex itself being local to the current process. When a handful of threads all go to access some key in our cache, they will access the same :class:`.Dogpile` object which internally can synchronize their activity using a Python threading.Lock. But in this example we're talking to a Memcached cache. What if we have many servers which all access this cache? We'd like all of these servers to coordinate together so that we don't just prevent the dogpile problem within a single process, we prevent it across all servers.
To accomplish this, we need an object that can coordinate processes. In this example we'll use a file-based lock as provided by the lockfile package, which uses a unix-symlink concept to provide a filesystem-level lock (which also has been made threadsafe). Another strategy may base itself directly off the Unix os.flock() call, and still another approach is to lock within Memcached itself, using a recipe such as that described at Using Memcached as a Distributed Locking Service. The type of lock chosen here is based on a tradeoff between global availability and reliable performance. The file-based lock will perform more reliably than the memcached lock, but may be difficult to make accessible to multiple servers (with NFS being the most likely option, which would eliminate the possibility of the os.flock() call). The memcached lock on the other hand will provide the perfect scope, being available from the same memcached server that the cached value itself comes from; however the lock may vanish in some cases, which means we still could get a cache-regeneration pileup in that case.
What all of these locking schemes have in common is that unlike the Python threading.Lock object, they all need access to an actual key which acts as the symbol that all processes will coordinate upon. This is where the key argument to our create_dogpile() function introduced in :ref:`scaling_on_keys` comes in. The example can remain the same, except for the changes below to just that function:
import lockfile import os from hashlib import sha1 # ... other imports and setup from the previous example def create_dogpile(key, expiration_time): lock_path = os.path.join("/tmp", "%s.lock" % sha1(key).hexdigest()) return Dogpile( expiration_time, lock=lockfile.FileLock(path) ) # ... everything else from the previous example
Where above,the only change is the lock argument passed to the constructor of :class:`.Dogpile`. For a given key "some_key", we generate a hex digest of it first as a quick way to remove any filesystem-unfriendly characters, we then use lockfile.FileLock() to create a lock against the file /tmp/53def077a4264bd3183d4eb21b1f56f883e1b572.lock. Any number of :class:`.Dogpile` objects in various processes will now coordinate with each other, using this common filename as the "baton" against which creation of a new value proceeds.
Locking the "write" phase against the "readers"
A less prominent feature of Dogpile ported from Beaker is the ability to provide a mutex against the actual resource being read and created, so that the creation function can perform certain tasks only after all reader threads have finished. The example of this is when the creation function has prepared a new datafile to replace the old one, and would like to switch in the new file only when other threads have finished using it.
from dogpile import SyncReaderDogpile dogpile = SyncReaderDogpile(3600) def some_creation_function(dogpile): create_expensive_datafile() with dogpile.acquire_write_lock(): replace_old_datafile_with_new() # usage: with dogpile.acquire(some_creation_function): read_datafile()
With the above pattern, :class:`.SyncReaderDogpile` will allow concurrent readers to read from the current version of the datafile as the create_expensive_datafile() function proceeds with its job of generating the information for a new version. When the data is ready to be written, the :meth:`.SyncReaderDogpile.acquire_write_lock` call will block until all current readers of the datafile have completed (that is, they've finished their own :meth:`.Dogpile.acquire` blocks). The some_creation_function() function then proceeds, as new readers are blocked until this function finishes its work of rewriting the datafile.
Note that the :class:`.SyncReaderDogpile` approach is useful for when working with a resource that itself does not support concurent access while being written, namely flat files, possibly some forms of DBM file. It is not needed when dealing with a datasource that already provides a high level of concurrency, such as a relational database, Memcached, or NoSQL store. Currently, the :class:`.SyncReaderDogpile` object only synchronizes within the current process among multiple threads; it won't at this time protect from concurrent access by multiple processes. Beaker did support this behavior however using lock files, and this functionality may be re-added in a future release.