wassname/ray
1
0
Fork 0
mirror of https://github.com/wassname/ray.git synced 2026-07-13 17:45:08 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Serialization] API for deregistering serializers; code & doc cleanup (#13471)

Browse Source 
* make methods private, remove confusion brackets and usages

* unregister serializer; fix doc

* Cleanup doc

* rename unregister -> deregister
This commit is contained in:
Siyuan (Ryans) Zhuang
2021-01-29 10:27:05 -08:00
committed by GitHub
parent b20a38febb
commit 0b598c0f05
6 changed files with 85 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files
doc/source/serialization.rst
+55 -46
View File
@@ -5,15 +5,6 @@ Serialization
Since Ray processes do not share memory space, data transferred between workers and nodes will need to **serialized** and **deserialized**. Ray uses the `Plasma object store <https://arrow.apache.org/docs/python/plasma.html>`_ to efficiently transfer objects across different processes and different nodes. Numpy arrays in the object store are shared between workers on the same node (zero-copy deserialization).
.. _plasma-store:
Plasma Object Store
-------------------
Plasma is an in-memory object store that is being developed as part of Apache Arrow. Ray uses Plasma to efficiently transfer objects across different processes and different nodes. All objects in Plasma object store are **immutable** and held in shared memory. This is so that they can be accessed efficiently by many workers on the same node.
Each node has its own object store. When data is put into the object store, it does not get automatically broadcasted to other nodes. Data remains local to the writer until requested by another task or actor on another node.
Overview
--------
@@ -21,8 +12,17 @@ Ray has decided to use a customized `Pickle protocol version 5 <https://www.pyth
Ray is currently compatible with Pickle protocol version 5, while Ray supports serialization of a wider range of objects (e.g. lambda & nested functions, dynamic classes) with the help of cloudpickle.
.. _plasma-store:
Plasma Object Store
~~~~~~~~~~~~~~~~~~~
Plasma is an in-memory object store that is being developed as part of Apache Arrow. Ray uses Plasma to efficiently transfer objects across different processes and different nodes. All objects in Plasma object store are **immutable** and held in shared memory. This is so that they can be accessed efficiently by many workers on the same node.
Each node has its own object store. When data is put into the object store, it does not get automatically broadcasted to other nodes. Data remains local to the writer until requested by another task or actor on another node.
Numpy Arrays
------------
~~~~~~~~~~~~
Ray optimizes for numpy arrays by using Pickle protocol 5 with out-of-band data.
The numpy array is stored as a read-only object, and all Ray workers on the same node can read the numpy array in the object store without copying (zero-copy reads). Each numpy array object in the worker process holds a pointer to the relevant array held in shared memory. Any writes to the read-only object will require the user to first copy it into the local process memory.
@@ -48,7 +48,7 @@ Serialization notes
- Lock objects are mostly unserializable, because copying a lock is meaningless and could cause serious concurrency problems. You may have to come up with a workaround if your object contains a lock.
Customized Serialization
________________________
------------------------
Sometimes you may want to customize your serialization process because
the default serializer used by Ray (pickle5 + cloudpickle) does
@@ -61,29 +61,29 @@ There are at least 3 ways to define your custom serialization process:
function inside the corresponding class. This is commonly done
by most Python libraries. Example code:
.. code-block:: python
.. code-block:: python
import ray
import sqlite3
import ray
import sqlite3
ray.init()
ray.init()
class DBConnection:
def __init__(self, path):
self.path = path
self.conn = sqlite3.connect(path)
class DBConnection:
def __init__(self, path):
self.path = path
self.conn = sqlite3.connect(path)
# without '__reduce__', the instance is unserializable.
def __reduce__(self):
deserializer = DBConnection
serialized_data = (self.path,)
return deserializer, serialized_data
# without '__reduce__', the instance is unserializable.
def __reduce__(self):
deserializer = DBConnection
serialized_data = (self.path,)
return deserializer, serialized_data
original = DBConnection("/tmp/db")
print(original.conn)
original = DBConnection("/tmp/db")
print(original.conn)
copied = ray.get(ray.put(original))
print(copied.conn)
copied = ray.get(ray.put(original))
print(copied.conn)
2. If you want to customize the serialization of a type of objects,
but you cannot access or modify the corresponding class, you can
@@ -112,8 +112,17 @@ There are at least 3 ways to define your custom serialization process:
A, serializer=custom_serializer, deserializer=custom_deserializer)
ray.get(ray.put(A(1))) # success!
# You can deregister the serializer at any time.
ray.util.deregister_serializer(A)
ray.get(ray.put(A(1))) # fail!
# Nothing happens when deregister an unavailable serializer.
ray.util.deregister_serializer(A)
NOTE: Serializers are managed locally for each Ray worker. So for every Ray worker,
if you want to use the serializer, you need to register the serializer.
if you want to use the serializer, you need to register the serializer. Deregister
a serializer also only applies locally.
If you register a new serializer for a class, the new serializer would replace
the old serializer immediately in the worker. This API is also idempotent, there are
no side effects caused by re-registering the same serializer.
@@ -121,29 +130,29 @@ There are at least 3 ways to define your custom serialization process:
3. We also provide you an example, if you want to customize the serialization
of a specific object:
.. code-block:: python
.. code-block:: python
import threading
import threading
class A:
def __init__(self, x):
self.x = x
self.lock = threading.Lock() # could not serialize!
class A:
def __init__(self, x):
self.x = x
self.lock = threading.Lock() # could not serialize!
ray.get(ray.put(A(1))) # fail!
ray.get(ray.put(A(1))) # fail!
class SerializationHelperForA:
"""A helper class for serialization."""
def __init__(self, a):
self.a = a
class SerializationHelperForA:
"""A helper class for serialization."""
def __init__(self, a):
self.a = a
def __reduce__(self):
return A, (self.a.x,)
def __reduce__(self):
return A, (self.a.x,)
ray.get(ray.put(SerializationHelperForA(A(1)))) # success!
# the serializer only works for a specific object, not all A
# instances, so we still expect failure here.
ray.get(ray.put(A(1))) # still fail!
ray.get(ray.put(SerializationHelperForA(A(1)))) # success!
# the serializer only works for a specific object, not all A
# instances, so we still expect failure here.
ray.get(ray.put(A(1))) # still fail!
Troubleshooting
python/ray/actor.py
+1 -1
View File
@@ -937,7 +937,7 @@ class ActorHandle:
def __reduce__(self):
"""This code path is used by pickling but not by Ray forking."""
state = self._serialization_helper()
return ActorHandle._deserialization_helper, (state)
return ActorHandle._deserialization_helper, state
def modify_class(cls):
python/ray/serialization.py
+9 -8
View File
@@ -31,7 +31,7 @@ class DeserializationError(Exception):
pass
def object_ref_deserializer(reduced_obj_ref, owner_address):
def _object_ref_deserializer(binary, owner_address):
# NOTE(suquark): This function should be a global function so
# cloudpickle can access it directly. Otherwise couldpickle
# has to dump the whole function definition, which is inefficient.
@@ -40,9 +40,7 @@ def object_ref_deserializer(reduced_obj_ref, owner_address):
# the core worker to resolve the value. This is to make sure
# that the ref count for the ObjectRef is greater than 0 by the
# time the core worker resolves the value of the object.
# UniqueIDs are serialized as (class name, (unique bytes,)).
obj_ref = reduced_obj_ref[0](*reduced_obj_ref[1])
obj_ref = ray.ObjectRef(binary)
# TODO(edoakes): we should be able to just capture a reference
# to 'self' here instead, but this function is itself pickled
@@ -61,7 +59,7 @@ def object_ref_deserializer(reduced_obj_ref, owner_address):
return obj_ref
def actor_handle_deserializer(serialized_obj):
def _actor_handle_deserializer(serialized_obj):
# If this actor handle was stored in another object, then tell the
# core worker.
context = ray.worker.global_worker.get_serialization_context()
@@ -85,7 +83,7 @@ class SerializationContext:
serialized, actor_handle_id = obj._serialization_helper()
# Update ref counting for the actor handle
self.add_contained_object_ref(actor_handle_id)
return actor_handle_deserializer, (serialized, )
return _actor_handle_deserializer, (serialized, )
self._register_cloudpickle_reducer(ray.actor.ActorHandle,
actor_handle_reducer)
@@ -96,13 +94,16 @@ class SerializationContext:
worker.check_connected()
obj, owner_address = (
worker.core_worker.serialize_and_promote_object_ref(obj))
return object_ref_deserializer, (obj.__reduce__(), owner_address)
return _object_ref_deserializer, (obj.binary(), owner_address)
self._register_cloudpickle_reducer(ray.ObjectRef, object_ref_reducer)
def _register_cloudpickle_reducer(self, cls, reducer):
pickle.CloudPickler.dispatch[cls] = reducer
def _unregister_cloudpickle_reducer(self, cls):
pickle.CloudPickler.dispatch.pop(cls, None)
def _register_cloudpickle_serializer(self, cls, custom_serializer,
custom_deserializer):
def _CloudPicklerReducer(obj):
@@ -198,7 +199,7 @@ class SerializationContext:
elif metadata_fields[
0] == ray_constants.OBJECT_METADATA_TYPE_ACTOR_HANDLE:
obj = self._deserialize_msgpack_data(data, metadata_fields)
return actor_handle_deserializer(obj)
return _actor_handle_deserializer(obj)
# Otherwise, return an exception object based on
# the error type.
try:
python/ray/tests/test_serialization.py
+7
View File
@@ -616,6 +616,13 @@ def test_custom_serializer(ray_start_shared_local_modes):
A, serializer=custom_serializer, deserializer=custom_deserializer)
ray.get(ray.put(A(1)))
ray.util.deregister_serializer(A)
with pytest.raises(Exception):
ray.get(ray.put(A(1)))
# deregister again takes no effects
ray.util.deregister_serializer(A)
if __name__ == "__main__":
import pytest
python/ray/util/__init__.py
+2 -1
View File
@@ -6,7 +6,7 @@ from ray.util.debug import log_once, disable_log_once_globally, \
from ray.util.placement_group import (placement_group, placement_group_table,
remove_placement_group)
from ray.util import rpdb as pdb
from ray.util.serialization import register_serializer
from ray.util.serialization import register_serializer, deregister_serializer
from ray.util.client_connect import connect, disconnect
@@ -25,4 +25,5 @@ __all__ = [
"connect",
"disconnect",
"register_serializer",
"deregister_serializer",
]
python/ray/util/serialization.py
+11
View File
@@ -16,3 +16,14 @@ def register_serializer(cls, *, serializer, deserializer):
"""
context = ray.worker.global_worker.get_serialization_context()
context._register_cloudpickle_serializer(cls, serializer, deserializer)
def deregister_serializer(cls):
"""Deregister the serializer associated with the type ``cls``.
There is no effect if the serializer is unavailable.
Args:
cls: A Python class/type.
"""
context = ray.worker.global_worker.get_serialization_context()
context._unregister_cloudpickle_reducer(cls)