…sts metrics (#23957)

* add mapping for updated metric name litellm_remaining_tokens_metric

* changelog

* add sister metric litellm_remaining_requests_metric

* changelog update

…ent events (#23778)

* kafka_consumer: collect Schema Registry compatibility, drop legacy agent events

Add global compatibility (GET /config) and per-subject effective
compatibility (GET /config/{subject}?defaultToGlobal=true) to the
data-streams-message payload for schemas. Per-subject results are
cached alongside version metadata; a compatibility change is included
in the cache key so it triggers re-emission like a schema change.

Stop emitting agent-side events for broker, topic, and schema
configurations — the same payloads continue to flow to the
data-streams intake, which is the canonical consumer.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Add changelog entries for #23778

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Refresh per-subject schema compatibility on the configs cadence

Previously per-subject compatibility was only fetched alongside a Tier 2
schema fetch — so a customer flipping a subject's compatibility without
publishing a new version was invisible until the next version bump.

Move compatibility onto its own TTL-driven cache, reusing the same
broker/topic configs refresh interval (default 3 minutes, with the same
jitter). The per-run batch size mirrors the version-check tier (200).
Subjects with a new version this run are still bundled into the same
parallel wave, so we avoid a separate HTTP burst when the two coincide.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Fix two_tier_fetch test to account for compatibility in cache entry

The new compatibility-refresh logic populates the compatibility field on
both the changed subject (via the full-fetch wave) and the unchanged
subject (via the standalone compat-refresh path). Add the mocks and
update the cache assertions to match.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Include global_compatibility in the schema emission cache key

A global compatibility change with no per-subject change would otherwise
leave subjects with explicit overrides emitting stale global_compatibility
in their payload until the TTL eviction kicked in.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Address review feedback: TypedDict, docstrings, test isolation, compat-flip test

- Add `compatibility: NotRequired[str | None]` to SubjectVersionInfo so the
  TypedDict reflects the actual cache shape after this PR
- Collapse multi-line docstring on private _get_schema_registry_subject_compatibility
  to one-liner per project style guide
- Trim 3-line comment block to one line
- Fix `if subject_compat:` → `if subject_compat is not None:` to avoid
  silently dropping an empty-string API response
- Remove dead test variables (expected_broker_config/topic_config/schema)
- Add compatibility mocks to three existing tests that relied on exception
  handlers swallowing live HTTP calls
- Add test_schema_registry_compatibility_flip_triggers_reemission covering the
  headline behaviour: a compat change without a version bump triggers re-emission
- Consolidate changelog into a single 23778.changed entry

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Fix NameError: restore expected_* dicts used by DS assertions in test_collect_cluster_metadata

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Fix global_compatibility is-None guard and add missing docstring

- `if global_compatibility is not None:` to match the subject_compat fix and
  avoid silently dropping an empty-string API response
- Add one-liner docstring to _get_schema_registry_global_compatibility for
  consistency with its sibling method

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Wrap over-length list comprehension in compatibility-flip test

The schema-event filter exceeded the 120-char line limit. Split the
data-streams payload extraction and the config_type filter into two
readable comprehensions.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Address all AAraKKe suggestions: helper, log level, compat fallbacks, batch clamp, test cleanup

- Extract _schema_registry_get helper; all five GET helpers become one-liners
- Raise per-subject compat fetch log level from debug to warning
- Persist global_compatibility on success; restore last known value on failure
- Clamp compat_due to max(0, BATCH_SIZE - len(subjects_needing_full_fetch)) to keep total bounded
- Standalone loop skips subject in schema_responses (not subjects_needing_full_fetch) so a failed schema fetch no longer silently discards the compat value
- Fall back to cached compatibility when version-bump path gets None from a failed fetch
- Parameterise mock_schema_registry_methods with global_compat/subject_compat; remove call-site overrides
- Move expected_* dicts adjacent to their consumers in test_collect_cluster_metadata
- Add three failure-path tests: subject compat failure preserves cached value, global compat failure uses last known, None compat omits field from payload

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Address review: global-compat None fallback, parallel-fetch helper, emit-loop extraction, test dedup

- Fall back to cached global compatibility on a None /config response, not just on error
- Extract _parallel_fetch helper to collapse the triplicated ThreadPoolExecutor blocks
- Extract _emit_schema_registry_events from _collect_schema_registry_info
- Document the TTL-reset-on-failure, defaultToGlobal None semantics, and post-upgrade re-emission
- Add schema_ds_events and mock_compatibility_methods test helpers; cover the global-compat-only flip

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* Address round-2 review: cache-key ordering comment, _wire_cache test helper, relocate _make_schema_registry_check

- Document the ordering invariant between the two latest_version_cache write sites
- Note the per-subject compatibility-flip cadence latency in the inline comment
- Add a _wire_cache helper and adopt it across the schema-registry tests, removing the
  duplicated cache_storage/mock_read/mock_write scaffolding
- Move _make_schema_registry_check up near the other shared helpers and migrate the
  inline instance/client wiring in the older schema-registry tests to use it

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* Address round-3 review: type-annotate _schema_registry_get, extract compat/schema-info helpers, test compat-skip

- Annotate _schema_registry_get as (**kwargs: Any) -> Any
- Extract _collect_subject_compatibilities from _collect_schema_registry_info, documenting that
  the batch-size clamp is not a hard ceiling on /config/{subject} calls
- Extract _build_schema_info to share the canonical payload dict between the fetched and
  cache-reconstruction paths
- Add a test asserting compatibility is not refetched when its cadence cache is fresh

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* Address round-4 review: lift compat cache mutation to call site, add SchemaInfo TypedDict, parametrize flip tests

- Remove the hidden side effect from _collect_subject_compatibilities: it now only fetches and
  returns, with the standalone-flip cache mutation applied explicitly at the call site
- Add a SchemaInfo TypedDict and annotate _build_schema_info's return with it
- Merge the two near-identical compatibility-flip tests into one parametrized test (subject/global)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* Fix ruff formatting after rebase conflict resolution

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Rename changelog entry from changed to added for PR #23778

Not a breaking change — existing events are deduplicated, not removed.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…#23902)

* Emit connection_error DSM event when Kafka connection fails

When cluster monitoring is enabled and request_metadata_update() fails,
emit a data-streams-message with config_type=connection_error and a
reason field carrying the exception message, so consumers of the
heartbeat pipeline can distinguish unreachable clusters from healthy ones.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Add changelog entry for PR #23902

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Address review comments on connection_error heartbeat

- Guard emission call in try/except so errors from event_platform_event
  never mask the AdminClient connection error message
- Extract _emit_cluster_monitoring_event helper to share collection_timestamp,
  bootstrap_servers, and event_platform_event call across both senders
- Add match= to pytest.raises in all three new tests
- Extract _connection_error_events helper to deduplicate test extraction logic
- Move import json to stdlib block (isort)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Apply ruff format

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* Add regression test: sink failure must not mask AdminClient error

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* update vSphere dashboard

* remove tag

* filter correctly