# Atlas Infinite oplog and storage-layer replication > Source-anchored research for Atlas Infinite oplog and storage-layer replication. Run status: COMPLETED. Parent: [Oplog sizing under storage auto-scaling](https://llms-explorer.com/tree/oplog-sizing-under-storage-auto-scaling/) · 4 facets · 17 facts · page: https://llms-explorer.com/tree/atlas-infinite-oplog-and-storage-layer-replication/ ## Structure and components - Atlas Infinite stores and replicates data in a managed storage layer shared by its compute nodes; scaling or replacing compute does not require copying the complete dataset onto each new compute node. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/infinite/atlas-infinite-architecture.md) - On Infinite, majority write concern refers to a majority within the storage layer; compute nodes do not elect the primary, because storage coordinates standby promotion. (low confidence) — [source](https://www.mongodb.com/docs/atlas/infinite/atlas-infinite-architecture.md) ## Problems, failure modes and limitations - A compute node joining or rejoining Atlas Infinite cannot fall behind the local.oplog.rs window because nodes read shared storage. Inference: enlarging that collection to protect a slow secondary from falling off its window is not the applicable replication-sizing model for Infinite. (low confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/infinite/atlas-infinite-architecture.md) ## Comparisons and alternatives - The local.oplog.rs collection supports change streams and limited internal history rather than replication; nodes joining or rejoining cannot fall behind its retention window. (medium confidence) — [source](https://www.mongodb.com/docs/atlas/infinite/atlas-infinite-architecture.md) - On Atlas Infinite, local.oplog.rs serves change streams and limited internal history rather than replica-set data replication. The collection remains accessible to users with the appropriate local-database read privilege. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/infinite/atlas-infinite-architecture.md) - On Atlas Infinite, majority write concern refers to a majority inside the storage layer rather than a majority of compute nodes. Compute-node count must not be used to infer the durability quorum. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/infinite/atlas-infinite-architecture.md) ## Facts and statements - Atlas Infinite permits reading local.oplog.rs but does not permit setting an oplog size. (low confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/reference/atlas-oplog.md) - The collection retention window bounds interrupted change-stream resumption, not point-in-time restore history; retained oplog entries count toward cluster storage. (low confidence) — [source](https://www.mongodb.com/docs/atlas/infinite/atlas-infinite-architecture.md) - The documented default minimum oplog retention window is 24 hours; users can configure the minimum retention window in Additional Settings. (low confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/infinite/atlas-infinite-architecture.md) - MongoDB's October 8 engineering account describes a logical oplog entry replicated by the Log Service across three zones, with acknowledgment after durable majority storage. A separate phylog carries checkpoint-produced page images and deltas; the page materializer distributes those records and uploads both logs to object storage. (low confidence) (verified as of 2026-10-08) — [source](https://www.mongodb.com/company/blog/engineering/ground-beneath-database) - MongoDB's engineering account describes standby/read-replica reads as a composition of checkpoint-consistent shared pages and a local ingest table containing the recent logical-log tail. This design avoids waiting for the next checkpoint to expose recent writes; the post does not establish a universal measured lag guarantee. (low confidence) (verified as of 2026-10-08) — [source](https://www.mongodb.com/company/blog/engineering/ground-beneath-database) - Atlas alert documentation explicitly scopes the replication-oplog interpretation: Infinite replicates through storage, so the general alert category must not be read as proof that local.oplog.rs drives replication. (medium confidence) — [source](https://www.mongodb.com/docs/atlas/reference/alert-conditions.md) - The implementation ticket SERVER-129897 explicitly connects an optional per-document oplog hash and a PersistenceProvider capability to Atlas Infinite, giving a concrete lead for deeper implementation investigation. (low confidence) — [source](https://jira.mongodb.org/browse/SERVER-129897) - SERVER-136893 requests use of disagg_storage.step_up.write_unavailability_duration in Infinite replication dashboards; the ticket is an observability lead, not measured failover performance. (low confidence) (verified as of 2026-10-13) — [source](https://jira.mongodb.org/browse/SERVER-136893) - Atlas does not allow customers to set an oplog size on Atlas Infinite. Generic instructions to increase configured oplog size therefore need an edition check before use. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/reference/atlas-oplog.md) - The architecture documentation specifies a default minimum oplog retention window of 24 hours, configurable in Additional Settings. Atlas keeps entries for at least that window, then deletes older entries. The window limits interrupted change-stream resumability, not cluster point-in-time restore history. (low confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/infinite/atlas-infinite-architecture.md) - The local.oplog.rs collection counts toward Infinite cluster storage and storage charges. Storage is metered as uncompressed logical data in GB-hours; retained continuous-backup history is accounted for separately as backup. Inference: retention still has a storage-cost consequence even though the collection does not drive replication. (medium confidence) (verified as of 2026-10-14) — [source](https://www.mongodb.com/docs/atlas/infinite/atlas-infinite-architecture.md)