Getting started!

A comprehensive, fast, pure-Python memcached client library.

Basic Usage

from pymemcache.client.base import Client

client = Client('localhost')
client.set('some_key', 'some_value')
result = client.get('some_key')

The server to connect to can be specified in a number of ways.

If using TCP connections over IPv4 or IPv6, the server parameter can be passed a host string, a host:port string, or a (host, port) 2-tuple. The host part may be a domain name, an IPv4 address, or an IPv6 address. The port may be omitted, in which case it will default to 11211.

ipv4_client = Client('')
ipv4_client_with_port = Client('')
ipv4_client_using_tuple = Client(('', 11211))

ipv6_client = Client('[::1]')
ipv6_client_with_port = Client('[::1]:11211')
ipv6_client_using_tuple = Client(('::1', 11211))

domain_client = Client('localhost')
domain_client_with_port = Client('localhost:11211')
domain_client_using_tuple = Client(('localhost', 11211))

Note that IPv6 may be used in preference to IPv4 when passing a domain name as the host if an IPv6 address can be resolved for that domain.

You can also connect to a local memcached server over a UNIX domain socket by passing the socket’s path to the client’s server parameter. An optional unix: prefix may be used for compatibility in code that uses other client libraries that require it.

client = Client('/run/memcached/memcached.sock')
client_with_prefix = Client('unix:/run/memcached/memcached.sock')

Using a client pool

pymemcache.client.base.PooledClient is a thread-safe client pool that provides the same API as pymemcache.client.base.Client. It’s useful in for cases when you want to maintain a pool of already-connected clients for improved performance.

from pymemcache.client.base import PooledClient

client = PooledClient('', max_pool_size=4)

Using a memcached cluster

This will use a consistent hashing algorithm to choose which server to set/get the values from. It will also automatically rebalance depending on if a server goes down.

from pymemcache.client.hash import HashClient

client = HashClient([
client.set('some_key', 'some value')
result = client.get('some_key')

Key distribution is handled by the hasher argument in the constructor. The default is the built-in pymemcache.client.rendezvous.RendezvousHash hasher. It uses the built-in pymemcache.client.murmur3.murmur3_32 implementation to distribute keys on servers. Overriding these two parts can be used to change how keys are distributed. Changing the hashing algorithm can be done by setting the hash_function argument in the RendezvousHash constructor.

Rebalancing in the pymemcache.client.hash.HashClient functions as follows:

  1. A pymemcache.client.hash.HashClient is created with 3 nodes, node1, node2 and node3.

  2. A number of values are set in the client using set and set_many. Example:

    • key1 -> node2

    • key2 -> node3

    • key3 -> node3

    • key4 -> node1

    • key5 -> node2

  3. Subsequent get calls will hash to the correct server and requests are routed accordingly.

  4. node3 goes down.

  5. The hashclient tries to get("key2") but detects the node as down. This causes it to mark the node as down. Removing it from the hasher. The hasclient can attempt to retry the operation based on the retry_attempts and retry_timeout arguments. If ignore_exc is set, this is treated as a miss, if not, an exception will be raised.

  6. Any get/set for key2 and key3 will now hash differently, example:

    • key2 -> node2

    • key3 -> node1

  7. After the amount of time specified in the dead_timeout argument, node3 is added back into the hasher and will be retried for any future operations.

Using the built-in retrying mechanism

The library comes with retry mechanisms that can be used to wrap all kinds of pymemcache clients. The wrapper allows you to define the exceptions that you want to handle with retries, which exceptions to exclude, how many attempts to make and how long to wait between attempts.

The RetryingClient wraps around any of the other included clients and will have the same methods. For this example, we’re just using the base Client.

from pymemcache.client.base import Client
from pymemcache.client.retrying import RetryingClient
from pymemcache.exceptions import MemcacheUnexpectedCloseError

base_client = Client(("localhost", 11211))
client = RetryingClient(
client.set('some_key', 'some value')
result = client.get('some_key')

The above client will attempt each call three times with a wait of 10ms between each attempt, as long as the exception is a MemcacheUnexpectedCloseError.

Using TLS

Memcached supports authentication and encryption via TLS since version 1.5.13.

A Memcached server running with TLS enabled will only accept TLS connections.

To enable TLS in pymemcache, pass a valid TLS context to the client’s tls_context parameter:

import ssl
from pymemcache.client.base import Client

context = ssl.create_default_context(

client = Client('localhost', tls_context=context)
client.set('some_key', 'some_value')
result = client.get('some_key')


import json
from pymemcache.client.base import Client

class JsonSerde(object):
    def serialize(self, key, value):
        if isinstance(value, str):
            return value, 1
        return json.dumps(value), 2

    def deserialize(self, key, value, flags):
       if flags == 1:
           return value
       if flags == 2:
           return json.loads(value)
       raise Exception("Unknown serialization format")

client = Client('localhost', serde=JsonSerde())
client.set('key', {'a':'b', 'c':'d'})
result = client.get('key')

pymemcache provides a default pickle-based serializer:

from pymemcache.client.base import Client
from pymemcache import serde

class Foo(object):

client = Client('localhost', serde=serde.pickle_serde)
client.set('key', Foo())
result = client.get('key')

The serializer uses the highest pickle protocol available. In order to make sure multiple versions of Python can read the protocol version, you can specify the version by explicitly instantiating pymemcache.serde.PickleSerde:

client = Client('localhost', serde=serde.PickleSerde(pickle_version=2))

Deserialization with Python 3

Values passed to the serde.deserialize() method will be bytestrings. It is therefore necessary to encode and decode them correctly. Here’s a version of the JsonSerde from above which is more careful with encodings:

class JsonSerde(object):
    def serialize(self, key, value):
        if isinstance(value, str):
            return value.encode('utf-8'), 1
        return json.dumps(value).encode('utf-8'), 2

    def deserialize(self, key, value, flags):
       if flags == 1:
           return value.decode('utf-8')
       if flags == 2:
           return json.loads(value.decode('utf-8'))
       raise Exception("Unknown serialization format")

Interacting with pymemcache

For testing purpose pymemcache can be used in an interactive mode by using the python interpreter or again ipython and tools like tox.

One main advantage of using tox to interact with pymemcache is that it comes with its own virtual environments. It will automatically install pymemcache and fetch all the needed requirements at run. See the example below:

$ podman run --publish 11211:11211 -it --rm --name memcached memcached
$ tox -e venv -- python
>>> from pymemcache.client.base import Client
>>> client = Client('')
>>> client.set('some_key', 'some_value')
>>> client.get('some_key')
>>> print(client.get.__doc__)
     The memcached "get" command, but only for one key, as a convenience.
       key: str, see class docs for details.
       default: value that will be returned if the key was not found.
       The value for the key, or default if the key wasn't found.

You can instantiate all the classes and clients offered by pymemcache.

Your client will remain open until you decide to close it or until you decide to quit your interpreter. It can allow you to see what happens if your server is abruptly closed. Below is an example.

Starting your server:

$ podman run --publish 11211:11211 -it --name memcached memcached

Starting your client and set some keys:

$ tox -e venv -- python
>>> from pymemcache.client.base import Client
>>> client = Client('')
>>> client.set('some_key', 'some_value')

Restarting the server:

$ podman restart memcached

The previous client is still open, now try to retrieve some keys:

>>> print(client.get('some_key'))
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/user/pymemcache/pymemcache/client/", line 535, in get
    return self._fetch_cmd(b'get', [key], False).get(key, default)
  File "/home/user/pymemcache/pymemcache/client/", line 910, in _fetch_cmd
    buf, line = _readline(self.sock, buf)
  File "/home/user/pymemcache/pymemcache/client/", line 1305, in _readline
    raise MemcacheUnexpectedCloseError()

We can see that the connection has been closed.

You can also pass a command directly from CLI parameters and get output directly:

$ tox -e venv -- python -c "from pymemcache.client.base import Client; client = Client('127.0.01'); print(client.get('some_key'))"

This kind of usage is useful for debug sessions or to dig manually into your server.

Key Constraints

This client implements the ASCII protocol of memcached. This means keys should not contain any of the following illegal characters:

Keys cannot have spaces, new lines, carriage returns, or null characters. We suggest that if you have unicode characters, or long keys, you use an effective hashing mechanism before calling this client.

At Pinterest, we have found that murmur3 hash is a great candidate for this. Alternatively you can set allow_unicode_keys to support unicode keys, but beware of what unicode encoding you use to make sure multiple clients can find the same key.

Best Practices

  • Always set the connect_timeout and timeout arguments in the pymemcache.client.base.Client constructor to avoid blocking your process when memcached is slow. You might also want to enable the no_delay option, which sets the TCP_NODELAY flag on the connection’s socket.

  • Use the noreply flag for a significant performance boost. The noreply flag is enabled by default for “set”, “add”, “replace”, “append”, “prepend”, and “delete”. It is disabled by default for “cas”, “incr” and “decr”. It obviously doesn’t apply to any get calls.

  • Use pymemcache.client.base.Client.get_many() and pymemcache.client.base.Client.gets_many() whenever possible, as they result in fewer round trip times for fetching multiple keys.

  • Use the ignore_exc flag to treat memcache/network errors as cache misses on calls to the get* methods. This prevents failures in memcache, or network errors, from killing your web requests. Do not use this flag if you need to know about errors from memcache, and make sure you have some other way to detect memcache server failures.

  • Unless you have a known reason to do otherwise, use the provided serializer in pymemcache.serde.pickle_serde for any de/serialization of objects.


noreply will not read errors returned from the memcached server.

If a function with noreply=True causes an error on the server, it will still succeed and your next call which reads a response from memcached may fail unexpectedly.

pymemcached will try to catch and stop you from sending malformed inputs to memcached, but if you are having unexplained errors, setting noreply=False may help you troubleshoot the issue.