ceilometer/releasenotes/notes/manager-based-ipc-queues-85e3bf59ffdfb0ac.yaml
gord chung 911b973d70 queues per manager rather than per pipeline
currently we create queue per pipeline which is not necessary. it
creates more memory usage and doesn't necessarily distribute
work more effectively. this hashes data to queues based on
manager but still internally, the data is destined to specific
pipeline based on event_type. this will minimise queue usage while
keeping internal code path the same.

Change-Id: I0ccd51f13457f208fe2ccedb6e680c91e132f78f
2017-11-16 14:43:46 -05:00

18 lines
865 B
YAML

---
features:
- |
Workload partitioning of notification agent is now split into queues
based on pipeline type (sample, event, etc...) rather than per individual
pipeline. This will save some memory usage specifically for pipeline
definitions with many source/sink combinations.
upgrade:
- |
If workload partitioning of the notification agent is enabled, the
notification agent should not run alongside pre-Queens agents. Doing so
may result in missed samples when leveraging transformations. To upgrade
without loss of data, set `notification_control_exchanges` option to
empty so only existing `ceilometer-pipe-*` queues are processed. Once
cleared, reset `notification_control_exchanges` option and launch the new
notification agent(s). If `workload_partitioning` is not enabled, no
special steps are required.