fix(async): honor reset_connection when AsyncMilvusClient rotates a password

[Anai-Guo]

What

AsyncMilvusClient.update_password() has no reset_connection flag, so an async client that rotates its own password keeps sending the old one on every subsequent call.

The sync client has had this since the beginning:

# pymilvus/milvus_client/milvus_client.py
def update_password(self, user_name, old_password, new_password,
                    reset_connection: Optional[bool] = False, ...):
    ...
    if reset_connection:
        conn._setup_authorization_interceptor(user_name, new_password, None)
        conn._setup_grpc_channel()

AsyncGrpcHandler even carries a mirrored _setup_authorization_interceptor() — but it has zero call sites; the async client never wires it up. So today the only way for an async caller to keep working after update_password() is to throw the client away and build a new one.

Why adding the flag alone isn't enough

Porting the sync two-liner verbatim does not work, because the two handlers install interceptors differently:

• sync _setup_authorization_interceptor() only records self._authorization_interceptor; _setup_grpc_channel() then rebuilds _final_channel from the raw channel, so the old interceptor is dropped.
• async _setup_authorization_interceptor() appends to the live _final_channel._unary_unary_interceptors, and _build_stub() appends again. Nothing is ever removed.

Each authorization interceptor appends its own authorization metadata entry (async_header_adder_interceptor), so the superseded credential stays on the wire — and, being first in the list, is the one the server sees. Measured against the current master with a naive port of the sync code:

init      : ['root:old_pw']
1st reset : ['root:old_pw', 'root:new_pw', 'root:new_pw']
2nd reset : ['root:old_pw', 'root:new_pw', 'root:new_pw', 'root:newer_pw', 'root:newer_pw']

(_setup_grpc_channel() carries the comment # avoid to add duplicate headers. — _build_stub does not currently honor it.)

What this PR changes

1. AsyncMilvusClient.update_password(..., reset_connection=False) — same semantics as the sync client.
2. AsyncGrpcHandler._uninstall_authorization_interceptor() (new) — detaches the interceptor being replaced from the live channel; _setup_authorization_interceptor() calls it before installing the new one.
3. AsyncGrpcHandler._build_stub() — no longer re-appends an authorization interceptor that is already installed on that channel, honoring its own comment. Building a stub on a fresh channel is unchanged.

After the change, the same measurement:

init      : ['root:old_pw']
1st reset : ['root:new_pw']
2nd reset : ['root:newer_pw']

Signature note: reset_connection is appended at the end of the async signature rather than placed after new_password as in the sync client, so existing positional calls like update_password(u, old, new, 30) keep binding timeout. Happy to mirror the sync ordering instead if you'd prefer exact positional parity.

Tests

New in tests/unit/async_grpc_handler/test_async_auth.py (5 tests): credential rotation replaces the stale interceptor and installs exactly one; the uninstall helper is a no-op when nothing is installed; reset_connection=True re-authorizes the connection and the default leaves it alone; and the sync/async signatures carry the same parameter set.

The rotation tests assert on the decoded authorization headers the installed interceptors would emit, not just list length, so they fail loudly if a stale credential survives.

Verification (Windows, Python 3.12, no server required — grpc.aio channels are constructed but never connected):

• new code + new tests → 174 passed in tests/unit/async_grpc_handler/

• unmodified source + new tests → 5 failed, 169 passed, the key one being

  AssertionError: assert ['root:old_pa...oot:new_pass'] == ['root:new_pass']
  

No existing test was modified or weakened. An earlier draft moved interceptor installation out of _setup_authorization_interceptor and broke test_setup_authorization_interceptor_appends_header; the design in this PR keeps that method's contract intact (it still appends to the channel) and instead removes the interceptor it supersedes.

Full unit suite: 3377 passed. The two unrelated failures on my machine (test_check.py::test_get_commit, plus collection errors in test_version.py / the bulk-writer modules) are missing optional dev dependencies (setuptools_scm) and reproduce identically on unmodified master.

ruff check and ruff format --check clean on all three files (tests checked with tests/ruff.toml).

---

This was found by diffing MilvusClient against AsyncMilvusClient method signatures. The scan also surfaced two naming drifts I have deliberately not touched, since they may be intentional and changing them would be breaking: get_load_state(partition_name=) vs (partition_names=), and transfer_replica(source_group, target_group, num_replicas) vs (source, target, num_replica). Happy to file those separately if they're unintended.

🤖 Generated with Claude Code

[sre-ci-robot]

[APPROVALNOTIFIER] This PR is NOT APPROVED

This pull-request has been approved by: Anai-Guo
To complete the pull request process, please assign longjiquan after the PR has been reviewed.
You can assign the PR to them by writing /assign @longjiquan in a comment when ready.

The full list of commands accepted by this bot can be found here.

Details

Needs approval from an approver in each of these files:

• OWNERS

Approvers can indicate their approval by writing /approve in a comment
Approvers can cancel approval by writing /approve cancel in a comment

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request