UltraDict2

Sychronized, streaming Python dictionary that uses shared memory as a backend

This is a maintained fork of ronny-rentner/UltraDict, published on PyPI as UltraDict2. Install with pip install UltraDict2 and import with from UltraDict2 import UltraDict.

Warning: This is an early hack. There are only few unit tests and so on. Maybe not stable!

Features:

• Fast (compared to other sharing solutions)
• No running manager processes
• Works in spawn and fork context
• Safe locking between independent processes
• Tested with Python 3.11 - 3.14 on Linux, Windows and Mac
• Convenient, no setter or getters necessary
• Optional recursion for nested dicts

Atomic operations via atomics2

Shared locking is built on atomics2, a maintained fork of the abandoned atomics package, rebuilt on patomic v1.1.0 with prebuilt wheels for Python 3.11 - 3.14 across Linux, Windows, and macOS (including arm64). It is installed automatically as a dependency, so shared_lock=True works out of the box.

General Concept

UltraDict uses multiprocessing.shared_memory to synchronize a dict between multiple processes.

It does so by using a stream of updates in a shared memory buffer. This is efficient because only changes have to be serialized and transferred.

If the buffer is full, UltraDict will automatically do a full dump to a new shared memory space, reset the streaming buffer and continue to stream further updates. All users of the UltraDict will automatically load full dumps and continue using streaming updates afterwards.

What lives where

The dict itself is not stored in shared memory. Every process keeps its own complete copy as an ordinary Python dict; shared memory carries the update stream and the snapshots used to keep those copies in step.

flowchart LR
    subgraph proc1["Process 1 (writer)"]
        d1["self.data<br/>full local copy"]
    end
    subgraph proc2["Process 2"]
        d2["self.data<br/>full local copy"]
    end
    subgraph proc3["Process 3"]
        d3["self.data<br/>full local copy"]
    end

    subgraph shm["Shared memory"]
        control["control<br/>1000 bytes<br/>position + dump counter"]
        buffer["update buffer<br/>buffer_size<br/>stream of changes"]
        dump["full dump<br/>snapshot of the whole dict"]
    end

    d1 -- "append frame" --> buffer
    d1 -- "snapshot when buffer is full" --> dump
    buffer -- "replay frames" --> d2
    buffer -- "replay frames" --> d3
    dump -- "reload everything" --> d2
    dump -- "reload everything" --> d3
    control -.- buffer
    control -.- dump

Two consequences worth planning for:

• Reads are local. A read compares two integers in the control block and then hits the local dict. Nothing is deserialized unless something actually changed.
• Memory scales with the number of processes. A 500 MB dict shared by 8 processes costs roughly 8 x 500 MB of RAM plus the shared segments, not 500 MB. Shared memory here is a transport, not a way to store one copy.

Writing

flowchart TD
    w["d[key] = value"] --> ser["serialize the change"]
    ser --> fits{"does the frame fit<br/>in the update buffer?"}
    fits -- yes --> append["append frame, advance position"]
    append --> cheap["cost: size of the change"]
    fits -- no --> full["serialize the whole dict into a full dump"]
    full --> reset["publish it, bump the dump counter, reset the buffer"]
    reset --> pricey["cost: size of the entire dict,<br/>paid again by every reader"]

A write that does not fit is never streamed at all: it reaches other processes only inside the snapshot. That is why buffer_size matters far more than it looks.

Issues

On Windows, if no process has any handles on the shared memory, the OS will gc all of the shared memory making it inaccessible for future processes. To work around this issue you can currently set full_dump_size which will cause the creator of the dict to set a static full dump memory of the requested size. This full dump memory will live as long as the creator lives. This approach has the downside that you need to plan ahead for your data size and if it does not fit into the full dump memory, it will break.

Security

UltraDict uses pickle to serialize data by default. Unpickling attacker-controlled data leads to arbitrary code execution, and any local process that can guess or enumerate the shared memory names can write to the buffers. Only use UltraDict between processes that already trust each other and run under user accounts you control; do not use it as a boundary against untrusted local processes. If you need to share data with less trusted processes, pass a safe serializer (e.g. one based on JSON) instead of pickle.

Alternatives

There are many alternatives:

• multiprocessing.Manager
• shared-memory-dict
• mpdict
• Redis
• Memcached

How to use?

Simple

In one Python REPL:

Python 3.11 on linux
>>>
>>> from UltraDict2 import UltraDict
>>> ultra = UltraDict({ 1:1 }, some_key='some_value')
>>> ultra
{1: 1, 'some_key': 'some_value'}
>>>
>>> # We need the shared memory name in the other process.
>>> ultra.name
'psm_ad73da69'

In another Python REPL:

Python 3.11 on linux
>>>
>>> from UltraDict2 import UltraDict
>>> # Connect to the shared memory with the name above
>>> other = UltraDict(name='psm_ad73da69')
>>> other
{1: 1, 'some_key': 'some_value'}
>>> other[2] = 2

Back in the first Python REPL:

>>> ultra[2]
2

Nested

In one Python REPL:

Python 3.11 on linux
>>>
>>> from UltraDict2 import UltraDict
>>> ultra = UltraDict(recurse=True)
>>> ultra['nested'] = { 'counter': 0 }
>>> type(ultra['nested'])
<class 'UltraDict2.UltraDict2.UltraDict'>
>>> ultra.name
'psm_0a2713e4'

In another Python REPL:

Python 3.11 on linux
>>>
>>> from UltraDict2 import UltraDict
>>> other = UltraDict(name='psm_0a2713e4')
>>> other['nested']['counter'] += 1

Back in the first Python REPL:

>>> ultra['nested']['counter']
1

Performance comparison

Lets compare a classical Python dict, UltraDict, multiprocessing.Manager and Redis.

Median across Python 3.11 - 3.14 on Debian 12 bare metal, 10,000 keys:

Reading through UltraDict costs about 11x a plain dict, and is 82x faster than a Manager dict and 227x faster than a local Redis. Against a Redis on another machine, which is the more usual deployment, it is roughly 1900x faster. Reading ultra.data directly is within ~20% of a plain dict, at the cost of not seeing updates.

The gap in the middle of the chart is the point: there is nothing between reading local memory and crossing a socket. A read that finds nothing new never leaves the process, which is what buys the two orders of magnitude.

Writes cost more than reads because the change is serialized and published, and because a write that does not fit the buffer forces a full dump of the whole dict. See Buffer sizes and read performance for how to size it.

Python version

Across 3.11 - 3.14 everything lands within 1.1 - 1.4x, which is close enough to call flat:

No version is worth picking over another, and that was already true before the control block counters stopped being read through int.from_bytes(): on this machine the spread was 1.14x then and 1.11x now.

One caveat that cost us a wrong conclusion, recorded here so nobody repeats it: the same benchmark inside a virtual machine reported UltraDict as 2.1x slower on 3.13 than on 3.11, reproducibly. On bare metal that gap is 1.1x. Virtualized IPC and memory access inflate this workload unevenly across versions, so benchmark UltraDict on the kind of machine you will actually deploy on.

Method

Debian 12 bare metal, 12 cores, each Python version in its official Docker image, Redis 7 on the same host over loopback, 10,000 keys, best of three timing repeats. Every candidate gets its own iteration count and results are reported per operation, so a 11 ns operation and a 25 µs one are directly comparable. The LAN row was measured from a second machine on a 1GbE network, where the four Python versions agree to within 1.02x because the wire dominates.

This is not a real life workload and it is single process: it measures the cost of the operations, not what happens when several processes contend.

Reproduce it with tests/performance/compare.py. There is also a throughput test in tests/performance/performance.py, and tests/performance/buffer_size_sweep.py for choosing buffer_size.

I am interested in extending the performance testing to other solutions (like sqlite, memcached, etc.) and to more complex use cases with multiple processes working in parallel.

Parameters

Ultradict(*arg, name=None, create=None, buffer_size=10000, serializer=pickle, shared_lock=False, full_dump_size=None, auto_unlink=None, recurse=False, recurse_register=None, **kwargs)

name: Name of the shared memory. A random name will be chosen if not set. By default, if a name is given a new shared memory space is created if it does not exist yet. Otherwise the existing shared memory space is attached.

create: Can be either True or False or None. If set to True, a new UltraDict will be created and an exception is thrown if one exists already with the given name. If kept at the default value None, either a new UltraDict will be created if the name is not taken or an existing UltraDict will be attached.

Setting create=True does ensure not accidentally attaching to an existing UltraDict that might be left over.

buffer_size: Size of the shared memory buffer used for streaming changes of the dict. The buffer size limits the biggest change that can be streamed, so when you use large values or deeply nested dicts you might need a bigger buffer. Otherwise, if the buffer is too small, it will fall back to a full dump. Creating full dumps can be slow, depending on the size of your dict.

Whenever the buffer is full, a full dump will be created. A new shared memory is allocated just big enough for the full dump. Afterwards the streaming buffer is reset. All other users of the dict will automatically load the full dump and continue streaming updates.

(Also see the section Memory management below!)

serializer: Use a different serialized from the default pickle, e. g. marshal, dill, jsons. The module or object provided must support the methods loads() and dumps()

shared_lock: When writing to the same dict at the same time from multiple, independent processes, they need a shared lock to synchronize and not overwrite each other's changes. Shared locks are slow. They rely on the atomics2 package for atomic locks. By default, UltraDict will use a multiprocessing.RLock() instead which works well in fork context and is much faster.

(Also see the section Locking below!)

full_dump_size: If set, uses a static full dump memory instead of dynamically creating it. This might be necessary on Windows depending on your write behaviour. On Windows, the full dump memory goes away if the process goes away that had created the full dump. Thus you must plan ahead which processes might be writing to the dict and therefore creating full dumps.

auto_unlink: If True, the creator of the shared memory will automatically unlink the handle at exit so it is not visible or accessible to new processes. All existing, still connected processes can continue to use the dict.

recurse: If True, any nested dict objects will be automaticall wrapped in an UltraDict allowing transparent nested updates.

recurse_register: Has to be either the name of an UltraDict or an UltraDict instance itself. Will be used internally to keep track of dynamically created, recursive UltraDicts for proper cleanup when using recurse=True. Usually does not have to be set by the user.

Memory management

UltraDict uses shared memory buffers and those usually live is RAM. UltraDict does not use any management processes to keep track of buffers. Also it cannot know when to free those shared memory buffers again because you might want the buffers to outlive the process that has created them.

By convention you should set the parameter auto_unlink to True for exactly one of the processes that is using the UltraDict. The first process that is creating a certain UltraDict will automatically get the flag auto_unlink=True unless you explicitly set it to False. When this process with the auto_unlink=True flag ends, it will try to unlink (free) all shared memory buffers.

A special case is the recursive mode using recurse=True parameter. This mode will use an additional internal UltraDict to keep track of recursively nested UltraDict instances. All child UltraDicts will write to this register the names of the shared memory buffers they are creating. This allows the buffers to outlive the processes and still being correctly cleanup up by at the end of the program.

Buffer sizes and read performance:

There are 3 cases that can occur when you read from an `UltraDict:

1. No new updates: This is the fastes cases. UltraDict was optimized for this case to find out as quickly as possible if there are no updates on the stream and then just return the desired data. If you want even better read perforamance you can directly access the underlying data attribute of your UltraDict, though at the cost of not getting real time updates anymore.

2. Streaming update: This is usually fast, depending on the size and amount of that data that was changed but not depending on the size of the whole UltraDict. Only the data that was actually changed has to be unserialized.

3. Full dump load: This can be slow, depending on the total size of your data. If your UltraDict is big it might take long to unserialize it.

flowchart TD
    r["d[key]"] --> chk{"dump counter<br/>changed?"}
    chk -- yes --> load["load the full dump"]
    load --> c3["case 3: cost is the whole dict"]
    chk -- no --> pos{"stream position<br/>moved?"}
    pos -- yes --> replay["replay only the new frames"]
    replay --> c2["case 2: cost is the changes"]
    pos -- no --> hit["return from the local copy"]
    hit --> c1["case 1: two integer comparisons"]

Note that synchronization is pull based and lazy: there is no background thread. A process only catches up when it touches the dict, so an idle process pays for everything it missed on its next access.

Given the above 3 cases, you need to balance the size of your data and your write patterns with the streaming buffer_size of your UltraDict. If the streaming buffer is full, a full dump has to be created. Thus, if your full dumps are expensive due to their size, try to find a good buffer_size to avoid creating too many full dumps.

On the other hand, if for example you only change back and forth the value of one single key in your UltraDict, it might be useless to process a stream of all these back and forth changes. It might be much more efficient to simply do one full dump which might be very small because it only contains one key.

Locking

Every UltraDict instance has a lock attribute which is either a multiprocessing.RLock or an UltraDict.SharedLock if you set shared_lock=True when creating the UltraDict.

RLock is the fastest locking method that is used by default but you can only use it if you fork your child processes. Forking is the default on Linux systems.

In contrast, on Windows systems, forking is not available and Python will automatically use the spawn method when creating child processes. You should then use the parameter shared_lock=True when using UltraDict. This uses the atomics2 package, which is installed automatically as a dependency.

How to use the locking?

ultra = UltraDict(shared_lock=True)

with ultra.lock:
	ultra['counter']++

# The same as above with all default parameters
with ultra.lock(timeout=None, block=True, steal=False, sleep_time=0.000001):
	ultra['counter']++

# Busy wait, will result in 99 % CPU usage, fastest option
# Ideally number of processes using the UltraDict should be < number of CPUs
with ultra.lock(sleep_time=0):
	ultra['counter']++

try:
	result = ultra.lock.acquire(block=False)
	ultra.lock.release()
except UltraDict.Exceptions.CannotAcquireLock as e:
	print(f'Process with PID {e.blocking_pid} is holding the lock')

try:
	with ultra.lock(timeout=1.5):
		ultra['counter']++
except UltraDict.Exceptions.CannotAcquireLockTimeout:
	print('Stale lock?')

with ultra.lock(timeout=1.5, steal_after_timeout=True):
	ultra['counter']++

Explicit cleanup

Sometimes, when your program crashes, no cleanup happens and you might have a corrupted shared memeory buffer that only goes away if you manually delete it.

On Linux/Unix systems, those buffers usually live in a memory based filesystem in the folder /dev/shm. You can simply delete the files there.

Another way to do this in code is like this:

# Unlink both shared memory buffers possibly used by UltraDict
name = 'my-dict-name'
UltraDict.unlink_by_name(name, ignore_errors=True)
UltraDict.unlink_by_name(f'{name}_memory', ignore_errors=True)

Advanced usage

See examples folder

>>> ultra = UltraDict({ 'init': 'some initial data' }, name='my-name', buffer_size=100_000)
>>> # Let's use a value with 100k bytes length.
>>> # This will not fit into our 100k bytes buffer due to the serialization overhead.
>>> ultra[0] = ' ' * 100_000
>>> ultra.print_status()
{'buffer': SharedMemory('my-name_memory', size=100000),
 'buffer_size': 100000,
 'control': SharedMemory('my-name', size=1000),
 'full_dump_counter': 1,
 'full_dump_counter_remote': 1,
 'full_dump_memory': SharedMemory('psm_765691cd', size=100057),
 'full_dump_memory_name_remote': 'psm_765691cd',
 'full_dump_size': None,
 'full_dump_static_size_remote': <memory at 0x7fcbf5ca6580>,
 'lock': <RLock(None, 0)>,
 'lock_pid_remote': 0,
 'lock_remote': 0,
 'name': 'my-name',
 'recurse': False,
 'recurse_remote': <memory at 0x7fcbf5ca6700>,
 'serializer': <module 'pickle' from '/usr/lib/python3.11/pickle.py'>,
 'shared_lock_remote': <memory at 0x7fcbf5ca6640>,
 'update_stream_position': 0,
 'update_stream_position_remote': 0}

Note: All status keys ending with _remote are stored in the control shared memory space and shared across processes.

Other things you can do:

>>> # Create a full dump
>>> ultra.dump()

>>> # Load latest full dump if one is available
>>> ultra.load()

>>> # Show statistics
>>> ultra.print_status()

>>> # Force load of latest full dump, even if we had already processed it.
>>> # There might also be streaming updates available after loading the full dump.
>>> ultra.load(force=True)

>>> # Apply full dump and stream updates to
>>> # underlying local dict, this is automatically
>>> # called by accessing the UltraDict in any usual way,
>>> # but can be useful to call after a forced load.
>>> ultra.apply_update()

>>> # Access underlying local dict directly for maximum performance
>>> ultra.data

>>> # Use any serializer you like, given it supports the loads() and dumps() methods
>>> import jsons
>>> ultra = UltraDict(serializer=jsons)

>>> # Close connection to shared memory; will return the data as a dict
>>> ultra.close()

>>> # Unlink all shared memory, it will not be visible to new processes afterwards
>>> ultra.unlink()

Metrics

get_metrics() returns a plain dataclass of scalars, so you can feed a monitoring system without UltraDict pulling in a monitoring dependency.

>>> ultra = UltraDict()
>>> ultra['some key'] = 'some value'
>>> ultra.get_metrics()
Metrics(item_count=1, item_size_bytes_min=39, item_size_bytes_max=39, item_size_bytes_sum=39,
        item_size_observations_total=1, buffer_size_bytes=10000, buffer_used_bytes=45,
        buffer_used_fraction=0.0045, full_dump_size_bytes=None, full_dump_last_bytes=None,
        full_dump_total=0, buffer_full_forced_dump_total=0, full_dump_memory_full_total=0,
        full_dump_too_fast_total=0, shm_total_bytes=67108864, shm_used_bytes=1052672,
        shm_free_bytes=66056192)

The item size fields describe the serialized size of the writes this instance has made, not a measurement of the current contents. Measuring the latter would mean re-serializing every key on every call. The mean is item_size_bytes_sum / item_size_observations_total.

The shm_* fields report the whole /dev/shm filesystem, shared by every process on the machine. They are None on platforms without /dev/shm, such as Windows and macOS.

Counters are per instance, are never shared through the control memory, and reset when the process restarts. get_metrics() takes no lock, so a value can be one write stale.

Exporting to Prometheus

Register a collector that scrapes on demand:

from prometheus_client.core import CounterMetricFamily, GaugeMetricFamily, REGISTRY

class UltraDictCollector:
    def __init__(self, ultra):
        self.ultra = ultra

    def collect(self):
        metrics = self.ultra.get_metrics()
        labels  = [self.ultra.name]

        gauge = GaugeMetricFamily('ultradict_items', 'Items in the dict', labels=['name'])
        gauge.add_metric(labels, metrics.item_count)
        yield gauge

        gauge = GaugeMetricFamily('ultradict_buffer_used_fraction', 'Update buffer fill level', labels=['name'])
        gauge.add_metric(labels, metrics.buffer_used_fraction)
        yield gauge

        counter = CounterMetricFamily('ultradict_buffer_full_forced_dumps', 'Full dumps caused by a full buffer', labels=['name'])
        counter.add_metric(labels, metrics.buffer_full_forced_dump_total)
        yield counter

REGISTRY.register(UltraDictCollector(ultra))

What to watch:

Contributing

Contributions are always welcome!