# Per-shard compute auto-scaling rules and limits on Atlas > Source-anchored research for Per-shard compute auto-scaling rules and limits on Atlas. Run status: COMPLETED. Parent: [Independent shard scaling and per-shard tiering](https://llms-explorer.com/tree/independent-shard-scaling-and-per-shard-tiering/) · 5 facets · 13 facts · page: https://llms-explorer.com/tree/per-shard-compute-auto-scaling-rules-and-limits-on-atlas/ ## How it works - For Atlas Core sharded clusters with compute auto-scaling enabled, independent shard scaling is enabled by default. Atlas evaluates each shard separately using the replica-set tier-scaling criteria; when a shard's operational or analytics node group meets a trigger, only that group on that shard changes tier. (high confidence) (verified as of 2026-10-20) — [source](https://www.mongodb.com/docs/atlas/cluster-autoscaling-compute-core/) - Terraform per-shard compute scaling uses the new mongodbatlas_advanced_cluster schema: define one unique replication_specs element per shard and omit legacy num_shards. Configure base-node and analytics-node compute scaling separately; compute_enabled defaults to false, a maximum tier is required when it is enabled, and a minimum tier is required when scale-down is enabled. (high confidence) (verified as of 2026-10-20) — [source](https://raw.githubusercontent.com/mongodb/terraform-provider-mongodbatlas/master/docs/guides/advanced-cluster-new-sharding-schema.md) ## Measurements and reference values - Atlas applies reactive scale-up criteria to each shard. For operational nodes on M30+ tiers, average Normalized System CPU or System Memory above 90% for 10 minutes or above 75% for one hour qualifies; queued or rejected operations above zero in every sample for 10 minutes is another trigger for Dedicated M10+ nodes. A brief burst may not scale the cluster because Atlas needs time to prepare resources. (medium confidence) (verified as of 2026-10-19) — [source](https://www.mongodb.com/docs/atlas/cluster-autoscaling-compute-core/) - Reactive scale-down is gated by low use over both 10-minute and four-hour windows: operational-node CPU must be below 45%, WiredTiger cache use below 90% of its maximum, and projected memory on the next tier below 60%. Atlas also blocks scale-down for 24 hours after a scale-down, provisioning, or unpause, and for 12 hours after a node restart. (medium confidence) (verified as of 2026-10-19) — [source](https://www.mongodb.com/docs/atlas/cluster-autoscaling-compute-core/) ## Problems, failure modes and limitations - Independent shard auto-scaling requires matching electable, read-only, and analytics node counts and equal storage capacity across shards; Global clusters can have a different node count per zone. Auto-scaling cannot span different cluster classes, such as M30 General and M40 Low-CPU, and NVMe clusters cannot use independent shard scaling. (high confidence) (verified as of 2026-10-20) — [source](https://www.mongodb.com/docs/atlas/cluster-sharding/) - The Atlas UI can enable compute auto-scaling for independently scaling shards, but manually choosing a different tier per shard requires the Atlas Administration API, Atlas CLI, or Terraform. Changing the tier in the Atlas UI changes every shard, and the Cluster Builder UI cannot manage a cluster after those tools set different shard tiers. (high confidence) (verified as of 2026-10-20) — [source](https://www.mongodb.com/docs/atlas/cluster-sharding/) ## Comparisons and alternatives - MongoDB's current sharding documentation excludes Atlas Infinite clusters during public preview. PointFive's article, updated 2026-09-28, also lists sharding among Flex's missing features, so these per-shard rules apply to Atlas Core dedicated sharded clusters rather than Infinite or Flex. (medium confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/atlas/cluster-sharding/) ## Facts and statements - Compute auto-scaling uses a shared cluster-level minimum and maximum tier range, and the smallest shard can be at most two tier steps below the largest. If scaling up the largest shard would break that spread, Atlas raises smaller shards as needed. (high confidence) (verified as of 2026-10-20) — [source](https://www.mongodb.com/docs/atlas/cluster-sharding/) - Atlas does not allow per-shard auto-scaling across mixed cluster classes. Outside Global clusters, shards must keep the same electable, read-only, and analytics node counts; Global clusters may use different node counts per zone. All shards must keep the same storage capacity and NVMe tier, and NVMe clusters are not compatible with independent shard scaling. (high confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/atlas/cluster-sharding/) - Predictive auto-scaling is narrower than reactive per-shard scaling: eligible clusters must be M30 or larger in the General or Low-CPU class, predictive scaling does not apply to analytics nodes, and a shard added after predictive scaling is active uses reactive scaling until it has two weeks of workload history. (low confidence) (verified as of 2026-10-23) — [source](https://www.mongodb.com/docs/atlas/cluster-autoscaling-compute-core/) - Atlas sharding is supported on M30 or larger Atlas Core clusters. Atlas Infinite does not support sharded clusters during public preview, and Local NVMe SSD clusters are not compatible with independent shard scaling. (high confidence) (verified as of 2026-10-20) — [source](https://www.mongodb.com/docs/atlas/cluster-sharding/) - Atlas enables cluster-tier auto-scaling by default for eligible clusters created in the UI, but API-created clusters require explicit opt-in. The Terraform provider defaults compute_enabled to false, so Terraform-created clusters do not auto-scale until that setting is enabled. (high confidence) (verified as of 2026-10-20) — [source](https://www.mongodb.com/docs/atlas/cluster-autoscaling-compute-core/) - If predictive auto-scaling is active before Atlas adds a shard, that new shard uses reactive auto-scaling until Atlas establishes two weeks of workload patterns. Predictive auto-scaling only scales up; reactive auto-scaling handles downscaling. (medium confidence) (verified as of 2026-10-19) — [source](https://www.mongodb.com/docs/atlas/cluster-autoscaling-compute-core/)