[ideaship]

ansible-core 2.19 changed how cache plugins store entries. Before 2.19 the redis cache plugin stores a host's facts as plain JSON under ansible_facts<host>. From 2.19 on they land under ansible_factss1_<host> as {"__payload__": "<JSON>"}, and values set with set_fact: cacheable keep data tags ({"value": ..., "tags": [...], "__ansible_type": ...}).

kolla-ansible (2.20) already writes the new key, osism-ansible (2.18) and ceph-ansible (2.16) write the old one, and both coexist on a running cluster. Every reader in python-osism used only the old key:

• osism get facts / osism dump facts and osism get states (commands/get.py)
• GET /v1/inventory/hosts/{host}/facts and .../facts/{fact} (api.py)
• the RabbitMQ node address lookup (utils/rabbitmq.py), used by the RabbitMQ commands in commands/migrate.py and commands/status.py
• osism reset facts --limit (commands/reset.py), which deleted only the old key
• the facts freshness check run by osism apply (check_ansible_facts), which derived the host name s1_<host> from new keys and then skipped those entries as malformed

Change

New module osism/utils/fact_cache.py, and all of the readers above go through it:

• decode() parses an entry in either schema, unwraps __payload__ and strips data tags itself rather than with ansible-core's private decoder, so it does not depend on which ansible-core the reading image ships.
• get_host_facts() reads both keys and returns the newer entry by ansible_date_time.epoch.
• scan_host_keys() groups both schemas' keys by host; the freshness check judges each host by its newest entry.
• delete_host_facts() deletes both keys and removes <host> / s1_<host> from the ansible_cache_keys sorted set the redis cache plugin maintains. osism reset facts without --limit also deletes that set and reports distinct hosts.

The API facts endpoints catch ValueError (which decode() also raises for a non-mapping entry). The freshness check logs "unusable ansible_date_time.epoch" for a well-formed entry with a non-numeric epoch instead of calling it malformed; the host is still skipped.

Testing

• tests/unit/utils/test_fact_cache.py (fakeredis): both schemas alone and together, either one newer, entries without a timestamp, tagged values, corrupt entries, the sorted-set cleanup. New freshness-check and API tests with new-format entries; existing tests that pinned the exact redis calls now expect both keys. Unit suite: 3539 passed, 3 xfailed.
• OSISM 11.0.3 as released (osism-ansible 2.18, kolla-ansible 2.20; both formats present, written ~20 s apart), read-only with this branch mounted into a container of the osism image: the newer entry is returned for every host, osism dump facts shows it where the current code shows the older one, and the freshness check names each host once.
• An OSISM 11.0.3 deploy with a locally built osism-ansible image on ansible-core 2.20.10 and the osism image, both carrying this change (Ceph via cephadm, so no runtime writes the old key): redis held 0 old-format and 8 new-format keys; osism dump facts returned current facts and GET /v1/inventory/hosts/<host>/facts/<fact> answered 200.

🤖 Generated with Claude Code