Skip to main content
Card class: Hero • Category: Nerve Centre

​
At a glance

An alert card that fires when the Elasticsearch cluster health status leaves green and stays at yellow or red for a sustained window. This is the single most Elasticsearch-defining signal on the board. green means every primary and every replica shard is allocated. yellow means all primaries are allocated but one or more replicas are missing, so you are running without redundancy. red means at least one primary shard is unallocated, which means data on that shard is unavailable to reads and writes right now. A red cluster is a page-the-on-call event because part of your index is simply not answering.
What it tracksThe cluster-level status field returned by the Elasticsearch cluster health API, surfaced as an alert when it is not green. The card lists the affected indices, the count of unassigned shards, and the duration the cluster has been off-green.
Data sourceGET /_cluster/health (and GET /_cluster/health?level=indices for the per-index breakdown). The status field is the canonical traffic-light value Elasticsearch computes from shard allocation state. Detail: “Elasticsearch-distinctive, RED = data unavailable on affected indexes. Page on-call.”
Time windowRT (real-time, polled on the standard fast cadence for alert cards).
Alert triggerstatus in (yellow, red) sustained for 5 minutes. The 5-minute sustain is deliberate: brief yellow flaps during a rolling restart or a routine relocation are normal and should not page anyone.
Severity splityellow is a warning (redundancy lost, data still served). red is critical (data unavailable, page on-call immediately). The card renders red status outlined in red.
What does NOT trigger itA green cluster with high heap, high disk, or slow queries. Those are separate cards. This card is purely the shard-allocation traffic light.
Rolesplatform, SRE, DBA, on-call

​
Calculation

Elasticsearch computes the cluster status itself; Vortex IQ does not derive it. The engine polls GET /_cluster/health and reads the status string, which Elasticsearch sets according to these rules:
  • green: all primary shards and all replica shards are allocated.
  • yellow: all primary shards are allocated, but at least one replica is unallocated. Reads and writes still succeed; you have simply lost redundancy for the affected indices.
  • red: at least one primary shard is unallocated. Any document that lives on that shard cannot be read or written until the shard recovers.
The cluster status is the worst status of any index in the cluster: a single red index makes the whole cluster red, a single yellow index (with no red) makes it yellow. The alert evaluates the status on every poll and starts a timer when it first leaves green. If the status is still non-green after 5 continuous minutes, the alert fires. If the status returns to green inside the window, the timer resets and nothing pages. To find which indices are responsible, the engine also calls GET /_cluster/health?level=indices and surfaces the indices whose own status is yellow or red, along with unassigned_shards per index. That is what turns “the cluster is red” into “the products-v7 primary is unallocated”.

​
Worked example

A platform team runs a 6-node Elasticsearch cluster (3 data, 3 master-eligible) backing product search and a logging pipeline for a mid-size retailer. Snapshot taken on 14 Apr 26 at 02:18 BST, during an overnight maintenance window. A data node (es-data-02) ran out of disk after a log-index rollover went wrong and was evicted from the cluster. The cluster health API now returns: 
{
  "cluster_name": "prod-search",
  "status": "red",
  "number_of_nodes": 5,
  "number_of_data_nodes": 2,
  "active_primary_shards": 411,
  "active_shards": 798,
  "unassigned_shards": 14
}
The per-index drill (?level=indices) shows:
IndexStatusUnassigned shardsNotes
products-v7red2One primary + one replica unallocated; the primary lived on es-data-02 and has no in-sync copy elsewhere
logs-2026.04.14yellow8Primaries intact on the surviving nodes; replicas gone
categories-v3yellow4Replicas gone, primaries fine
The Nerve Centre headline reads RED, 1 index red, 2 yellow, off-green for 6m, outlined in red, and the on-call engineer is paged. The story is clear from the card alone:
  1. products-v7 is red, so product search is partly broken. Any query hitting the missing primary’s shard returns partial results or a shard-failure error. Storefront search for the SKUs on that shard is degraded. Pair with Unassigned Shards to see the 14 unassigned shards and with Active Node Count, which dropped from 6 to 5 at 02:12.
  2. The yellow indices are not the emergency. logs-2026.04.14 and categories-v3 lost replicas only; their data is still fully served. They will go green again automatically once es-data-02 rejoins and shards re-replicate.
  3. Root cause is disk, not the shard layer. Storage Usage % on es-data-02 had crossed the flood-stage watermark before the node dropped. The fix is to free disk or add capacity, after which the unallocated primary recovers from its translog.
Decision path for the on-call engineer:
  1. Is any index RED? Yes (products-v7) -> data is unavailable -> this is the priority.
  2. Why is the primary unallocated? Use GET /_cluster/allocation/explain.
     -> "node left, no in-sync replica" because replica was on the same lost node.
  3. Can the lost node come back? es-data-02 is disk-full -> free disk, restart node.
  4. Once node rejoins, primary recovers from translog; replicas re-replicate.
  5. Cluster goes yellow (recovering), then green. Alert clears automatically.
The actionable lesson: a red cluster is about shard allocation, but the cause is almost always one layer down (a lost node, a full disk, a watermark breach). The card tells you data is unavailable; the sibling cards tell you why.

​
Sibling cards

CardWhy pair it with Cluster Not GreenWhat the combination tells you
Cluster Status (green / yellow / red)The always-on KPI version of the same status field.This alert is the paging wrapper around that KPI; the KPI shows the steady-state colour.
Unassigned ShardsThe direct cause of yellow/red.Any unassigned shard explains the colour; zero unassigned with non-green means a transient relocation.
Active Node CountA lost node is the most common trigger.Node count dropping at the same instant the cluster went red points straight at the missing node.
Storage Usage %A full disk evicts nodes and blocks shard allocation.A watermark breach just before red means disk is the root cause.
Initializing / Relocating ShardsRecovery from red passes through initialising shards.High initialising count after a red event means the cluster is healing; falling to zero means recovery is done.
Elasticsearch Health ScoreThe composite that weights cluster status heavily.A red cluster alone drags the health score well below the 70 alert line.
Pending Cluster TasksMaster-node backlog during mass reallocation.A red event plus a high pending-task queue means the master is busy reassigning shards.

​
Reconciling against the source

Where to look in Elasticsearch’s own tooling:
GET /_cluster/health returns the authoritative status field. This is the exact value the card reads. GET /_cluster/health?level=indices breaks the status down per index so you can see which index is responsible. GET /_cat/shards?v&h=index,shard,prirep,state,unassigned.reason lists every shard and the reason any are unassigned, in human-readable rows. GET /_cluster/allocation/explain explains in plain terms why a specific shard cannot be allocated.
On a managed service, the same value appears in the console: Elastic Cloud surfaces it under the deployment Health view, and AWS OpenSearch Service / managed Elasticsearch offerings expose it as the ClusterStatus.green / .yellow / .red CloudWatch metrics. The console may lag the live API by a minute or two because it samples on its own schedule. Why our number may legitimately differ from a manual _cluster/health call:
ReasonDirectionWhy
Sustain windowCard lags a raw API callA bare _cluster/health shows yellow the instant a replica drops; the card only fires after 5 minutes of sustained non-green, so a brief flap shows in the API but not as an alert.
Poll cadenceUp to one poll intervalThe card samples on a fixed interval; a status that flips green between polls may never be captured.
Per-index drill timingIndex list can lag the colourThe headline colour and the per-index breakdown are two API calls; during fast recovery the colour can update before the index list refreshes.
Managed-console samplingConsole can lag bothElastic Cloud and CloudWatch sample on their own cadence, so the vendor console can show green while the live API still reports yellow for a few seconds.

​
Known limitations / FAQs

Why did the cluster go yellow during a routine node restart, and is that a problem? When you stop a node for a rolling restart, the replicas that lived on it become unallocated, so the cluster goes yellow until those replicas re-replicate elsewhere or the node returns. This is expected and harmless: all primaries are still allocated, so reads and writes succeed. The 5-minute sustain window is designed to swallow short restarts; if you do slow rolling restarts, consider raising the sustain window or using cluster.routing.allocation.enable: primaries during the maintenance so replicas are not reshuffled needlessly. The cluster is yellow but everything works fine. Why page anyone? Yellow does not page on-call; only red does in the default configuration. Yellow is a warning that you have lost redundancy: if a second node fails while you are yellow, you risk going red. Treat yellow as “fix soon”, red as “fix now”. A cluster that sits yellow for days usually has a real allocation problem (a disk watermark, a stuck shard, or index.number_of_replicas set higher than the number of available nodes). Can a single-node cluster ever be green? No, not for any index with replicas. A replica shard can never be allocated to the same node as its primary, so a single-node cluster with the default 1 replica is permanently yellow. This is by design, not a fault. For a deliberate single-node setup, set index.number_of_replicas: 0 and the cluster can be green, but then you have no redundancy at all. What is the difference between this alert and the Unassigned Shards card? This card reads the cluster status traffic light directly. The Unassigned Shards card counts the shards behind that colour. They move together: any unassigned replica makes the cluster yellow, any unassigned primary makes it red. Use this card to know “is data unavailable?”, use Unassigned Shards to know “how many shards and which ones?”. The cluster went red but search still returns results. How? Search against a red cluster returns partial results: shards that are allocated answer, the unallocated primary’s shard returns a shard-failure, and Elasticsearch merges what it can unless you set allow_partial_search_results: false. So the storefront may look mostly fine while silently dropping the SKUs on the missing shard. That silent partiality is exactly why red is a paging event even when the site “looks okay”. Does this card cover index-level red, or only cluster-level? The headline reflects the cluster status, which is the worst of any single index. The per-index drill (sourced from ?level=indices) names the specific red or yellow indices so you are not left guessing. A single red index makes the whole cluster red even if 500 other indices are green. How long does recovery from red usually take? It depends entirely on the cause. If a node rejoins and its primary recovers from the translog, recovery can be seconds to a few minutes. If a primary has to be rebuilt from a replica on another node, it scales with shard size and network throughput, often tens of minutes for large shards. Watch Initializing / Relocating Shards trend to zero as the signal that recovery is finishing.

​
Tracked live in Vortex IQ Nerve Centre

Cluster Not Green (yellow or red) is one of hundreds of KPI pulses Vortex IQ tracks across Elasticsearch and 70+ other ecommerce connectors. Nerve Centre runs the detection layer; Vortex Mind investigates the cause when something moves; Ask Viq lets you interrogate any number in plain English. Start for free or book a demo to see this metric running on your own data.