This package works around common transient errors and pitfalls in creating a singleton instance of an Actor in Dask. It provides a clean interface for retrieving the singleton instance, and allocating it when necessary.
import dask_actor_singleton
class MyActor:
def __init__(self, arg):
self.value = arg
def inc(self):
self.value += 1
return self.value
client = dask.distributed.Client()
actor = dask_actor_singleton.get('my_actor', create=lambda: MyActor(8))
print(actor.inc().result()) # 9
# Now, on a different computer / dask.distributed.Client, run this script again:
# ...
print(actor.inc().result()) # 10
# If ever the singleton should be deleted, one may call:
dask_actor_singleton.discard('my_actor')
- 2021-08-19 v1.3.3 release. Caching times start after object creation finishes; not at the start. This matters for objects with long load times.
- 2021-08-18 v1.3.2 release. Explicitly forget wrapped future.
- 2021-08-18 v1.3.1 release. Free old actor when re-allocating to potentially reduce memory load.
- 2021-08-02 v1.3.0 release. Priority argument added to improve responsiveness alongside larger processing loads.
- 2021-07-29 v1.2.0 release. Added TTL support for
get
, and better documentation fordiscard
. - 2021-07-29 v1.1.1 release. Fixed issue with "Working holding Actor was lost"
- 2021-07-29 v1.1 release. Supports
discard
to purge a cached singleton. - 2021-07-19 v1.0 release.