Skip to main content

Advanced Redis plugins for `dogpile.cache`.

Project description

Python package

This package supports Python2 and Python3

dogpile_backend_redis_advanced

This is a plugin for the dogpile.cache system that offers some alternatives to the standard Redis datastore implementation.

Two new backends are offered:

backend description
dogpile_backend_redis_advanced extends the dogpile.cache.redis backend and allows for custom pickling overrides
dogpile_backend_redis_advanced_hstore extends dogpile_backend_redis_advanced and allows for some specific hstore operations

There is a negligible performance hit in dogpile_backend_redis_advanced_hstore, as cache keys must be inspected to determine if they are an hstore or not -- and there are some operations involved to coordinate values.

Additionally, some behavior is changed:

  • The constructor now accepts a lock_class argument, which can be used to wrap a mutex and alter how releases are handled. This can be necessary if you have a distributed lock and timeout or flush issues (via LRU or otherwise). A lock disappearing in Redis will raise a fatal exception under the standard redis backend.

  • The constructor now accepts a lock_prefix argument, which can be used to alter the prefix used for locks. The standard redis backend uses _lock as the prefix -- which can be hard to read or isolate for tests. One might want to use "_" as the lock prefix (so that keys "\_*" will show all locks).

purpose:

Mike Bayer's dogpile.cache is an excellent package for general purpose development.

The system offers 3 key features:

  1. elegant read-through caching functionality
  2. a locking mechanism that ensures only the first request of a cache-miss will create the resource (turning the rest into consumers of the first-requestor's creation)
  3. integrated cache expiry against time and library versions

Unfortunately, the integrated cache expiry feature comes at a cost -- objects are wrapped into a tuple with some metadata and pickled before hitting the datastore.

The additional metadata or pickle format may not be needed or wanted. Look how the size of 'a' grows by the time it becomes something passed off to Redis:

type example
string a
pickle(string) S'a'\np0\n.
CachedValue(string) ('a', {'ct': 1471113698.76127, 'v': 1})
pickle(CachedValue(string) ) cdogpile.cache.api\nCachedValue\np0\n(S'a'\np1\n(dp2\nS'ct'\np3\nF1471113698.76127\nsS'v'\np4\nI1\nstp5\nRp6\n.

By adding in hooks for custom serializers, this backend lets developers choose better ways to cache data.

You may want a serializer that doesn't care about the expiry of cached data, so just uses simpler strings.

type example 1 example 2
string a mellifluous
json.dumps(string) "a" "mellifluous"
msgpack.packb(string) \xa1a \xabmellifluous

or you may want to fool dogpile.cache by manipulating what the cached is. instead of using a python dict, of time and API version, you might just track the time but only to the second.

type example 1 example 2
AltCachedValue(string) ('a', 1471113698) ('mellifluous', 1471113698)
json.dumps(AltCachedValue(string)) '["a", 1471113698]' '["mellifluous", 1471113698]'
msgpack.packb(AltCachedValue(string)) '\x92\xa1a\xceW\xafi\xe2' '\x92\xabmellifluous\xceW\xafi\xe2'

This is how dogpile.cache stores "a"

cdogpile.cache.api\nCachedValue\np0\n(S'a'\np1\n(dp2\nS'ct'\np3\nF1471113698.76127\nsS'v'\np4\nI1\nstp5\nRp6\n.

This package lets us cache a raw string and trick dogpile.cache into thinking it is timely

a

or include a simpler version fo the time, along with a different serializer

["a", 1471113698] \x92\xa1a\xceW\xafi\xe2

If you cache lots of big objects, dogpile.cache's overhead is minimal -- but if you have a cache that works for mapping short bits of text, like ids to usernames (and vice-versa) you will see considerable savings.

Another way to make Redis more efficient is to use hash storage

Let's say you have a lot of keys that look like this:

region.set("user-15|posts", x)
region.set("user-15|friends", y)
region.set("user-15|profile", z)
region.set("user-15|username", z1)

You could make Redis a bit more efficient by using hash storage, in which you have 1 key with multiple fields:

region.hset("user-15", {'posts': x,
                        'friends', y,
                        'profile', z,
                        'username', z1,
                        })

Redis tends to operate much more efficiently in this situation (more below), but you can also save some bytes by not repeating the key prefix. Instagram's engineering team has a great article on this (http://instagram-engineering.tumblr.com/post/12202313862/storing-hundreds-of-millions-of-simple-key-value)

90% of dogpile.cache users who choose Redis will never need this package. A decent number of other users with large datasets have been trying to squeeze every last bit of memory and performance out of their machines -- and this package is designed to facilitate that.

usage:

myfile.py

# importing will register the plugins
import dogpile_backend_redis_advanced

then simply configure dogpile.cache with dogpile_backend_redis_advanced or dogpile_backend_redis_advanced_hstore as the backend.

RedisAdvancedBackend

Two new configuration options are offered to specify custom serializers via loads and dumps. The default selection is to use dogpile.cache's choice of pickle.

This option was designed to support msgpack as the serializer:

import msgpack
from dogpile.cache.api import CachedValue

def msgpack_loads(value):
    """pickle maintained the `CachedValue` wrapper of the tuple
       msgpack does not, so it must be added back in.
       """
    value = msgpack.unpackb(value, use_list=False)
    return CachedValue(*value)

region = make_region().configure(
    arguments= {'loads': msgpack_loads,
                'dumps': msgpack.packb,
                }
    )

One can also abuse/misuse dogpile.cache and defer all cache expiry to Redis using this serializer hook.

dogpile.cache doesn't cache your value as-is, but wraps it in a CachedValue object which contains an API version and a timestamp for the expiry.

This format is necessary for most cache backends, but Redis offers the ability to handle expiry in the cloud. By using the slim msgpack format and only storing the payload, you can drastically cut down the bytes needed to store this information.

This approach SHOULD NOT BE USED by 99% of users. However, if you do aggressive caching, this will allow you to leverage dogpile.cache's excellent locking mechanism for handling read-through caching while slimming down your cache size and the traffic on-the-wire.

import time
from dogpile.cache.api import CachedValue
from dogpile.cache.region import value_version
import msgpack

def raw_dumps(value):
    ''''pull the payload out of the CachedValue and serialize that
    '''
    value = value.payload
    value = msgpack.packb(value)
    return value

def raw_loads(value):
    ''''unpack the value and return a CachedValue with the current time
    '''
    value = msgpack.unpackb(value, use_list=False)
    return CachedValue(
        value,
        {
            "ct": time.time(),
            "v": value_version
        })

region = make_region().configure(
    arguments= {'loads': msgpack_loads,
                'dumps': msgpack.packb,
                'redis_expiration_time': 1,
                }
    )

RedisAdvancedHstoreBackend

This backend extends RedisAdvancedBackend with drop-in support for Hash storage under Redis.

  • if key names are tuples, they will be treated as hash operations on Redis.
  • by setting redis_expiration_time_hash to a boolean value, you can control how expiry times work within redis

This backend has a slight, negligible, overhead:

  • all key operations (get/get_multi/set/set_multi/delete) require an inspection of keys.
  • get_multi requires the order of keys to be tracked, and results from multiple get/hget operations are then correlated
  • set_multi requires the mapping to be analyzed and bucketed into different hmsets

redis_expiration_time_hash allows some extended management of expiry in Redis. by default it is set to None

  • False - ignore hash expiry. (never set a TTL in redis)
  • None - set redis_expiration_time on new hash creation only. this requires a check to the Redis key before a set.
  • True - unconditionally set redis_expiration_time on every hash key set/update.

Please note the following:

  • Redis manages the expiry of hashes on the key, making it global for all fields in the hash
  • Redis does not support setting a ttl on hashes while doing another operation. ttl must be set via another request
  • if redis_expiration_time_hash is set to True, there will be 2 calls to the Redis API for every key: hset or hmset then expires
  • if redis_expiration_time_hash is set to None, there will be 2-3 calls to the Redis API for every key: exists, hset or hmset, and possibly expires

Memory Savings and Suggested Usage

Redis is an in-memory datastore that offers persistence -- optimizing storage is incredibly important because the entire set must be held in-memory.

Example Demo

The attached demo.py (results in demo.txt) shows some potential approaches to caching and hashing by priming a Redis datastore with some possible strategies of a single dataset.

It's worth looking at demo.txt to see how the different serializesr encode the data -- sample keys are pulled for each format.

test memory bytes memory human relative ttl on redis? ttl in dogpile? backend encoder
region_redis 249399504 237.85M 0% Y Y dogpile.cache.redis pickle
region_json 222924496 212.60M 89.38% Y Y dogpile_backend_redis_advanced json
region_msgpack 188472048 179.74M 75.57% Y Y dogpile_backend_redis_advanced msgpack
region_redis_local 181501200 173.09M 72.78% - Y dogpile.cache.redis pickle
region_json_raw 171554880 163.61M 68.79% Y - dogpile_backend_redis_advanced json
region_msgpack_raw 170765872 162.86M 68.47% Y - dogpile_backend_redis_advanced msgpack
region_json_local 162612752 155.08M 65.20% - Y dogpile_backend_redis_advanced json
region_json_local_int 128648576 122.69M 57.71% - Y, int(time) dogpile_backend_redis_advanced json
region_msgpack_local 128160048 122.22M 51.39% - Y dogpile_backend_redis_advanced msgpack
region_msgpack_local_int 126938576 121.06M 50.89% - Y, int(time) dogpile_backend_redis_advanced msgpack
region_json_raw_local 111241280 106.09M 44.60% - - dogpile_backend_redis_advanced json
region_msgpack_raw_local 110455968 105.34M 44.29% - - dogpile_backend_redis_advanced msgpack
region_msgpack_raw_hash 28518864 27.20M 11.44% Y, only keys - dogpile_backend_redis_advanced_hstore msgpack
region_json_raw_hash 24836160 23.69M 9.96% Y, only keys - dogpile_backend_redis_advanced_hstore json

Notes:

  • the _local variants do not set a TTL on Redis
  • the _raw variants strip out the dogpile CachedValue wrapper and only store the payload
  • the _msgpack variants use msgpack instead of pickle
  • the _json variants use json instead of pickle
  • the _int variant applies int() to the dogpile timestamp, dropping a few bytes per entry

Wait WHAT? LOOK AT region_msgpack_raw_hash and region_json_raw_hash - that's a HUGE savings!

Yes.

The HSTORE has considerable savings due to 2 reasons:

  • Redis internally manages a hash much more effectively than keys.
  • Redis will only put an expiry on the keys (buckets), not the hash fields

HSTORE ends up being a much tighter memory usage for this example set, as we're setting 100 fields in each key. The savings would not be so severe if you were setting 5-10 fields per key

Note that region_msgpack_raw_local and region_json_raw_local should not be used unless you're running a LRU -- they have no expiry.

Assumptions

This demo is assuming a few things that are not tested here (but there are plenty of benchmarks on the internet showing this):

msgpack is the fastest encoder for serializing and deserializing data. json outperforms cpickle on serializing; cpickle outperforms json on deserializing data.

here are some benchmarks and links:

Caveats

In the examples above, we deal with (de)serializing simple, native, datatypes: string, int, bool, list, dict, tuple. For these datatypes, msgpack is both the smallest datastore and the fastest performer.

If you need to store more complex types, you will need to provide a custom encoder/decoder and will likely suffer a performance hit on the speed of (de)serialization. Unfortunately, the more complex data types that require custom encoding/decoding include standard datetime objects, which can be annoying.

The file custom_serializer.py shows an example class for handling (de)serialization -- MsgpackSerializer. Some common datetime formats are supported; they are encoded as a specially formatted dict, and decoded correspondingly. A few tricks are used to shave off time and make it roughly comparable to the speed of pickle.

Key Takeaways

  • this was surprising - while the differences are negligible on small datasets, using Redis to track expiry on long data-sets is generally not a good idea(!). dogpile.cache tracks this data much more efficiently. you can enable an LRU policy in Redis to aid in expiry.
  • msgpack and json are usually fairly comparable in size [remember the assumption that msgpack is better for speed]
  • reformatting the dogpile.cache metadata (replacing a dict an int() of the expiry) saves a lot of space under JSON when you have small payloads. the strings are a fraction of the size.
  • msgpack is really good with nested data structures

The following payloads for 1 are strings:

region_json_local =        '[10, {"v": 1, "ct": 1471113698.76127}]'
region_json_local_int =    '[10, 1471113753]'
region_msgpack_local =     '\x92\n\x82\xa1v\x01\xa2ct\xcbA\xd5\xeb\x92\x83\xe9\x97\x9a'
region_msgpack_local_int = '\x92\n\xceW\xafct'

So what should you use?

There are several tradeoffs and concepts to consider:

  1. Do you want to access information outside of dogpile.cache (in Python scripts, or even in another language)
  2. Are you worried about the time to serialize/deserialize? are you write-heavy or read-heavy?
  3. Do you want the TTL to be handled by Redis or within Python?
  4. What are your expiry needs? what do your keys look like? there may not be any savings possible. but if you have a lot of recycled prefixes, there could be.
  5. What do your values look like? How many are there?

This is test uses a particular dataset, and differences are inherent to the types of data and keys. Using the strategies from the region_msgpack_raw_hash on our production data has consistently dropped a 300MB Redis imprint to the 60-80MB range.

The Redis configuration file is also enclosed. the above tests are done with Redis compression turned on (which is why memory size fluctuates in the full demo reporting).

Custom Lock Classes

If your redis db gets flushed the lock will disappear. This will cause the redis backend to raise an exception EVEN THOUGH you have succeeded in generating your data.

By using a lock_class, you can catch the exception and decide what to do -- log it?, continue on, raise an error? It's up to you!

import redis.exceptions

class RedisDistributedLockProxy(object):
	"""example lock wrapper
	this will silently pass if a LockError is encountered
	"""
	mutex = None

	def __init__(self, mutex):
		self.mutex = mutex

	def acquire(self, *_args, **_kwargs):
		return self.mutex.acquire(*_args, **_kwargs)

	def release(self):
		# defer imports until backend is used
		global redis
		import redis  # noqa
		try:
			self.mutex.release()
		except redis.exceptions.LockError, e:
			# log.debug("safe lock timeout")
			pass
		except Exception as e:
			raise

To Do

I've been experimenting with handling the TTL within a hash bucket (instead of using the Redis or dogpile.cache methods). This looks promising. The rationale is that it is easier for Redis to get/set an extra field from the same hash, than it is to do separate calls to TTL/EXPIRES.

in code:

- hset('example', 'foo', 'bar')
- expires('example', 3600)
+ hmset('example', {'foo': 'bar',
                    'expires': time.time() + 3600,
                    }

I've also been experimenting with blessing the result into a subclass of dict that would do the object pair decoding lazily as-needed. That would speed up most use cases.

Maturity

This package is pre-release. I've been using these strategies in production via a custom fork of dogpile.cache for several years, but am currently migrating it to a plugin.

Maintenance and Upstream Compatibility

Some files in /tests are entirely from dogpile.cache as-is:

  • /tests/conftest.py
  • /tests/cache/_init_.py
  • /tests/cache/_fixtures.py

They are versions from dogpile.cache 0.6.2

The core file, /cache/backends/redis_advanced.py inherits from dogpile.cache's /cache/backends/redis.py

Testing

This ships with full tests.

Much of the core package and test fixtures are from dogpile.cache and copyright from that project, which is available under the MIT license.

Tests are handled through tox

Examples:

tox
tox -e py27 -- tests/cache/test_redis_backend.py
tox -e py27 -- tests/cache/test_redis_backend.py::RedisAdvanced_SerializedRaw_Test
tox -e py27 -- tests/cache/test_redis_backend.py::HstoreTest

Tests pass on the enclosed redis.conf file:

/usr/local/Cellar/redis/3.0.7/redis-server ./redis-server--6379.conf

License

This project is available under the same MIT license as dogpile.cache.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for dogpile-backend-redis-advanced, version 0.3.1
Filename, size File type Python version Upload date Hashes
Filename, size dogpile_backend_redis_advanced-0.3.1.tar.gz (14.6 kB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page