lib/collection/src/shards/replica_set/mod.rs (1)

1083-1093: Shard-wide activity check is correct; could reuse existing helpers

The implementation correctly reports the shard as active if any peer is in an is_active state, which is exactly what you need to distinguish “all Partial/Dead” shards in custom sharding. Note that this uses ReplicaState::is_active, so shards that are only ActiveRead will be treated as inactive, which aligns with how ActiveRead is treated as unhealthy in other health paths. Based on learnings, this is desirable.

For consistency with the rest of the codebase you could optionally implement this in terms of existing helpers, e.g.:

pub fn shard_is_active(&self) -> bool {
    let state = self.replica_state.read();
    !state.active_peers().is_empty()
    // or: !self.active_shards(false).is_empty()
}

lib/collection/src/shards/shard_holder/mod.rs (1)

588-591: Custom-sharding-only filtering of inactive shards in select_shards(All) looks right

This is_custom_sharding flag plus the !shard.shard_is_active() guard cleanly restricts the “skip inactive shards” behavior to custom sharding, leaving auto sharding unchanged and keeping the existing resharding-up guard intact. That matches the intent to ignore shard keys that currently have no active replicas without masking unexpected full-shard outages in auto mode.

As a tiny style nit, you could write the flag as let is_custom_sharding = matches!(self.sharding_method, ShardingMethod::Custom); but the current match is perfectly fine.

Also applies to: 606-613

tests/consensus_tests/test_tenant_promotion.py (1)

29-42: scroll_points helper correctly exercises the scroll API

The helper is straightforward and matches the existing HTTP-test style: it hits /points/scroll, asserts HTTP OK via assert_http_ok, and returns the points, with shard_key passed through so you can test both global and keyed reads. That should be enough to surface the regression around shards with no active replicas.

If you ever want Ruff F405 to be fully quiet in this file, you could replace the star imports with explicit ones (e.g., import requests and from .assertions import assert_http_ok / from .utils import get_collection_info), but that’s purely a linting concern and consistent with other consensus tests.

Configuration used: CodeRabbit UI

Review profile: CHILL

Plan: Pro

Learnt from: generall
Repo: qdrant/qdrant PR: 7474
File: src/common/health.rs:328-330
Timestamp: 2025-11-10T14:50:02.722Z
Learning: In Qdrant's `ReplicaState`, the `ActiveRead` variant is intentionally considered unhealthy (returns `false` from `is_healthy()`) because it requires a shard transfer operation to finish. This is the correct behavior for health checks.

Learnt from: generall
Repo: qdrant/qdrant PR: 7474
File: src/common/health.rs:328-330
Timestamp: 2025-11-10T14:50:02.722Z
Learning: In Qdrant's `ReplicaState`, the `ActiveRead` variant is intentionally considered unhealthy (returns `false` from `is_healthy()`) because it requires a shard transfer operation to finish. This is the correct behavior for health checks.