[
https://issues.apache.org/jira/browse/BEAM-5428?focusedWorklogId=306400&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-306400
]
ASF GitHub Bot logged work on BEAM-5428:
----------------------------------------
Author: ASF GitHub Bot
Created on: 04/Sep/19 16:07
Start Date: 04/Sep/19 16:07
Worklog Time Spent: 10m
Work Description: mxm commented on pull request #9418: [BEAM-5428]
Implement cross-bundle user state caching in the Python SDK
URL: https://github.com/apache/beam/pull/9418#discussion_r320845993
##########
File path: sdks/python/apache_beam/runners/worker/sdk_worker.py
##########
@@ -620,6 +628,106 @@ def _next_id(self):
return str(self._last_id)
+class CachingMaterializingStateHandler(object):
+ """ A State handler which retrieves and caches state. """
+
+ def __init__(self, global_state_cache, underlying_state):
+ self._underlying = underlying_state
+ self._state_cache = global_state_cache
+ self._context = threading.local()
+
+ @contextlib.contextmanager
+ def process_instruction_id(self, bundle_id, cache_tokens):
+ if getattr(self._context, 'cache_token', None) is not None:
+ raise RuntimeError(
+ 'Cache tokens already set to %s' % self._context.cache_token)
+ # TODO Also handle cache tokens for side input, if present
+ user_state_cache_token = None
+ for cache_token_struct in cache_tokens:
+ if cache_token_struct.HasField("user_state"):
+ # There should only be one user state token present
+ assert not user_state_cache_token
+ user_state_cache_token = cache_token_struct.token
+ try:
+ self._context.cache_token = user_state_cache_token
+ with self._underlying.process_instruction_id(bundle_id):
+ yield
+ finally:
+ self._context.cache_token = None
+
+ def blocking_get(self, state_key, coder, is_cached=False):
+ if not self._should_be_cached(is_cached):
+ # no cache / tokens, can't do a lookup/store in the cache
+ return self._materialize_iter(state_key, coder)
+ # Cache lookup
+ cache_state_key = self._convert_to_cache_key(state_key)
+ cached_value = self._state_cache.get(cache_state_key,
+ self._context.cache_token)
+ if cached_value is None:
+ # Cache miss, need to retrieve from the Runner
+ materialized = cached_value = []
Review comment:
For the cache to work effectively, we want to cache the whole value. Note
that if the cache is disabled, we still retain the old behavior of only
materializing a page at a time.
Yes, it has the drawback of the value to fit into memory. The design
currently doesn't do memory estimation but keeps a fixed upper limit of items.
If we move to a memory estimation, then also reflecting this here would make
sense.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
Issue Time Tracking
-------------------
Worklog Id: (was: 306400)
Time Spent: 15h (was: 14h 50m)
> Implement cross-bundle state caching.
> -------------------------------------
>
> Key: BEAM-5428
> URL: https://issues.apache.org/jira/browse/BEAM-5428
> Project: Beam
> Issue Type: Improvement
> Components: sdk-py-harness
> Reporter: Robert Bradshaw
> Assignee: Rakesh Kumar
> Priority: Major
> Time Spent: 15h
> Remaining Estimate: 0h
>
> Tech spec:
> [https://docs.google.com/document/d/1BOozW0bzBuz4oHJEuZNDOHdzaV5Y56ix58Ozrqm2jFg/edit#heading=h.7ghoih5aig5m]
> Relevant document:
> [https://docs.google.com/document/d/1ltVqIW0XxUXI6grp17TgeyIybk3-nDF8a0-Nqw-s9mY/edit#|https://docs.google.com/document/d/1ltVqIW0XxUXI6grp17TgeyIybk3-nDF8a0-Nqw-s9mY/edit]
> Mailing list link:
> [https://lists.apache.org/thread.html/caa8d9bc6ca871d13de2c5e6ba07fdc76f85d26497d95d90893aa1f6@%3Cdev.beam.apache.org%3E]
--
This message was sent by Atlassian Jira
(v8.3.2#803003)
