跳至主要内容

Prometheus HA Limitations and How Thanos Handles Them

Core Problem​

Prometheus is stateful and does not replicate its TSDB. Running multiple Prometheus replicas does not give you replicated data — each replica scrapes and stores independently.

Why Replicas Diverge​

Two Prometheus replicas pointing at the same targets will never produce identical data:

  1. Scrape timing: Each replica's scrape loop fires on its own schedule. Sample timestamps differ by milliseconds to seconds.
  2. Crash / restart windows: When replica A is down for rolling restart, replica B keeps scraping. A's TSDB has a gap that B doesn't, and vice versa.
  3. Network blips: A scrape failure on one replica produces a gap that the other may not have.

The result: replica A and replica B each hold a slightly different copy of "the truth" about your targets.

Why a Load Balancer Breaks HA​

The naive HA pattern — put a LB (NLB / ALB / Service ClusterIP with random selection) in front of N Prometheus replicas — does not work.

       LB picks one randomly
            │
   ┌────────┴────────┐
   ▼                 ▼
Replica A         Replica B
(had a gap        (was healthy
 during rolling    during that
 restart 09:00)    window)

If your query hits Replica A for the 09:00 window, you see a gap. If it hits Replica B, you see complete data. Same query, different answers, depending on which backend the LB picked. This is worse than no HA — it makes data quality non-deterministic.

How Thanos Solves It​

Thanos Query connects to all replicas simultaneously (via DNS service discovery — typically dnssrv+_grpc._tcp.<headless-svc>). It then:

  1. Pulls the same time range from every replica
  2. Deduplicates overlapping samples (using replica label)
  3. Fills gaps from one replica with data from another
  4. Returns a single, gap-free merged view to the caller

This requires direct endpoint access to every replica, not a single LB endpoint that hides them. The "many endpoints" model is essential to the algorithm.

Design Implication​

ScenarioRight pattern
1 replica PrometheusLB OK (no dedup needed)
N replica HA Prometheus + ThanosDNS discovery / Headless Service — never a single LB
N replica HA + only a single LBAnti-pattern — breaks Thanos dedup
Thanos Receive (not Sidecar)Receive handles its own replication via RF flag

Key Takeaway​

LB and HA-Prometheus-with-Thanos are mutually exclusive design choices. Choosing HA replicas forces you to expose them via DNS-based service discovery so Thanos Query can reach every replica directly.

Source​

Thanos official documentation: