[r387] Revert "HA dedup on every sample (#13665)" (#15023)

Backport 49c3084 from #15018

<!-- CURSOR_SUMMARY -->
---

> [!NOTE]
> **Medium Risk**
> Changes HA deduplication behavior in the distributor by no longer
evaluating cluster/replica labels per-series within a write request,
which can alter ingestion outcomes for mixed-label batches. Risk is
mitigated by updated docs but could impact non-standard Prometheus
setups (federation/proxies) that send heterogeneous series in one
request.
> 
> **Overview**
> Reverts the distributor HA deduplication optimization that evaluated
HA state per-series, and replaces it with a single HA check based on the
*first* timeseries’ cluster/replica labels, applying the decision
uniformly across the whole request.
> 
> This removes the per-replica state tracking/sorting logic and
associated tests/benchmarks, updates `findHALabels` to return `(cluster,
replica)` strings, and updates docs/CHANGELOG to reflect the new
assumption that all series in a request share the same HA labels.
> 
> <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit
1c41453. Bugbot is set up for automated
code reviews on this repo. Configure
[here](https://www.cursor.com/dashboard/bugbot).</sup>
<!-- /CURSOR_SUMMARY -->

Author: pr00se

CHANGELOG.md

@@ -154,7 +154,6 @@
 * [ENHANCEMENT] Query-frontend: Extend query blocking to optionally only apply a blocking rule if the query is an unaligned range query. Set `unaligned_range_queries: true` to enable. #14643
 * [ENHANCEMENT] Store-gateway: Add experimental flag `blocks-storage.bucket-store.partitioner-max-gap-bytes-chunks` to specify the gap size for the chunks partitioner. #14649
 * [ENHANCEMENT] Compactor: Add expermental `-compactor.first-level-compaction-ooo-wait-period` to configure a separate compaction wait period for out-of-order blocks. It's an analogue of `-compactor.first-level-compaction-wait-period`, which currently ignores out-of-order blocks. #14627
-* [ENHANCEMENT] HA: Deduplicate per sample instead of per batch. #13665
 * [ENHANCEMENT] Usage-tracker: Improve performance of TrackSeriesBatch by preprocessing input data. #14702 #14734
 * [ENHANCEMENT] MQE: Improve per-query memory consumption limit enforcement in histogram function evaluations. #14691
 * [ENHANCEMENT] MQE: Improve per-query memory consumption limit enforcement within aggregation operations. #14735

docs/sources/mimir/configure/configure-high-availability-deduplication.md

@@ -45,14 +45,7 @@ Incoming samples are considered duplicated (and thus dropped) if they are receiv
 
 If the HA tracker is enabled but incoming samples contain only one or none of the cluster and replica labels, these samples are accepted by default and never deduplicated.
 
-> Note: the HA tracker checks the cluster and replica label of every series in the request to determine whether each series in the request should be deduplicated.
-
-### Error responses
-
-When the HA tracker drops samples, Mimir returns one of the following errors depending on the reason:
-
-- **Replicas did not match**: When samples are received from a non-elected replica, Mimir returns an HTTP `202 Accepted` response with the message `replicas did not match, rejecting sample: replica=<replica>, elected=<elected>`. This indicates that the samples were successfully deduplicated and can be safely ignored.
-- **Too many HA clusters**: When the number of HA clusters for a tenant exceeds the configured limit, Mimir returns an HTTP `400 Bad Request` response with the error ID `err-mimir-tenant-too-many-ha-clusters`. To adjust this limit, configure `-distributor.ha-tracker.max-clusters` or contact your service administrator.
+> Note: for performance reasons, the HA tracker only checks the cluster and replica label of the first series in the request to determine whether all series in the request should be deduplicated. This assumes that all series inside the request have the same cluster and replica labels, which is typically true when Prometheus is configured with external labels. Ensure this requirement is honored if you have a non-standard Prometheus setup (for example, you're using Prometheus federation or have a metrics proxy in between).
 
 ## Configuration