Why shard-level visibility matters
Azure Managed Redis spreads data across shards. A cluster-wide average can look healthy while one shard is running out of memory or has lost its replica. Microsoft's pitch is that the extra detail helps you work out whether workloads are spread unevenly and find potential hotspots worth investigating. It doesn't rebalance anything for you, and it doesn't pinpoint root causes on its own.
The Microsoft Learn reference also fixes an older blind spot. The new metrics give per-shard visibility and come from Redis V2 Prometheus endpoints. The existing Total Keys metric on a clustered cache returns only the key count of the shard with the most keys during the interval. Microsoft's guidance for accurate per-shard key counts is to use the new Shard Key Count metric split by Slots (Range).
The metrics
Core metrics, per Microsoft Learn:
- Shard Memory Used (Bytes): memory used by one shard. On flash-enabled SKUs this includes both DRAM and flash.
- Shard Key Count: the total key count for the shard.
- Shard Replication Link Up: whether a replica is connected to its primary. Only replica shards emit it.
- Shard Memory Clients Normal (Bytes): memory used for input and output buffers of non-replica clients.
- Shard Memory Clients Replica (Bytes): the same buffers for replica clients.
All of these are labeled Preview. Microsoft also documents preview "big keys" metrics. These are counts of collection keys by element-count bucket, such as sets, sorted sets, hashes and lists under 1 million, from 1 million to 8 million, and over 8 million. For strings, only an under-128 MB bucket is exposed.
Dimensions: split by Slots (Range)
Each metric supports four dimensions:
| Dimension | API name | Use |
|---|---|---|
| Instance ID | InstanceId | Identifies the Redis node (VM) |
| Slots (Range) | Slots | Identifies the shard by hash-slot range |
| Shard ID | Shard | The Redis shard UID |
| Shard Role | Role | Primary or replica |
The caveat to remember is that these metrics are published at shard level. If you query without splitting by a dimension, Azure Monitor aggregates across shards with the default aggregation. For most metrics that doesn't give a meaningful cluster-wide total. Microsoft says to always split by Slots (Range). In REST API queries, use the API names rather than portal labels. Names and filter values are case-insensitive.
Diagnostic recipes from Microsoft's guidance
Memory imbalance. Split Shard Memory Used by Slots (Range). A max/min ratio above 2x indicates meaningful imbalance. Then compare Shard Key Count by the same split to see whether the cause is more keys or larger values. The documented remedies include reviewing hash tag usage, breaking up big keys and reviewing TTL policies on high-memory shards.
Replication link failures. Split Shard Replication Link Up by Slots (Range). A 0 means the link is down and a 1 means it is up. A sustained failure removes high-availability protection for that shard and raises data-loss risk if a failover happens first. The metric is noisy, though. Links can drop briefly during failover, shard migration, scaling or maintenance. Microsoft's example of a persistent failure is a link at 0 for 120 minutes or more. It also says to correlate with Shard Memory Used and Shard Memory Clients Replica.
Replication buffer growth. Split Shard Memory Clients Replica by Slots (Range). Look for a sustained upward trend over 15 minutes or more, not a single spike. If growth ends with the link dropping to 0, the replica was probably disconnected and a resynchronization is likely. Microsoft lists big keys, network disruption or latency, and an overloaded primary as common causes.
Alerting limits
Native Azure Monitor metric alerts only test a value against a fixed threshold at a point in time. Two of the suggested alerts fit that model:
- Replication link failure works natively, because the metric aggregates as Minimum and a 0 means the link was down at some point in the window.
- Any non-zero count in an "Over 8M Elements" bucket also works natively.
The others don't fit. Metric alerts can't compare values across dimension values, such as the max/min ratio across slots. They can't compute a ratio between two metrics, and they can't detect a sustained upward trend. Microsoft's workaround is to send metrics to a Log Analytics workspace through diagnostic settings and build log search alerts in KQL.
Limits to know
- Big keys metrics aren't yet supported on active geo-replicated caches or Flash Optimized tier caches.
- Microsoft recommends keeping individual key and value sizes under 512 KB. That is a performance recommendation, not an enforced limit. The bucket boundaries are monitoring thresholds, not endorsed sizes.
- Microsoft's thresholds, such as 2x, 15 minutes and 120 minutes, are starting points from its documentation. Tune them against your own workload.
Bottom line
Shard-level metrics are a practical addition for anyone running clustered Azure Managed Redis. They show where memory, keys and replication health diverge, but they don't fix the divergence. For now, try them in test or staging environments. Build dashboards that split by Slots (Range), and plan on KQL log alerts for the ratio and trend checks that native metric alerts can't do.
References
- (In preview) Public Preview: Shard-level metrics in Azure Managed Redis Azure Updates · 2026-11-17T20:00:32Z
- Monitoring data reference for Azure Managed Redis - Azure Managed Redis learn.microsoft.com
- What’s new in Azure Managed Redis at Microsoft Ignite 2026 techcommunity.microsoft.com