Point-in-time recovery (PITR) restores MongoDB to a specific cluster timestamp — not just the snapshot you took at midnight, but 14:37:22 yesterday when the ill-fated migration ran. The mechanic is the same regardless of the platform: a base snapshot plus the oplog entries between the snapshot and the target time, replayed in order. Atlas and Ops Manager package this; if you self-host you build it yourself out of mongodump --oplog and mongorestore --oplogReplay. This article is the runbook.
How to verify
Before you need it: prove your snapshot + oplog window is actually wide enough to span the maximum tolerable recovery window:
mongosh --host primary.example.internal:27017
use local
db.oplog.rs.find().sort({ $natural: 1 }).limit(1).next().ts
db.oplog.rs.find().sort({ $natural: -1 }).limit(1).next().ts
rs.printReplicationInfo()
aws s3 ls s3://backups/mongo/ --recursive | tail -5
oplog first and oplog last define the PITR window for your most recent snapshot. If the difference is smaller than your “go back N hours” objective, the window must be expanded before you need to use it.
What’s happening
A base snapshot is a complete copy of the dataset at some ts. The oplog records every write since ts. To restore to ts + Δ, you restore the snapshot, then replay oplog entries with ts in the range (snapshot_ts, target_ts]. mongorestore --oplogReplay does exactly this when you pass it both a dump directory and the oplog file from a mongodump --oplog.
For continuous PITR (as opposed to “PITR within the last snapshot’s oplog window”), you need an out-of-band oplog archive — Atlas and Ops Manager run a backup agent that tails the oplog continuously and ships entries to durable storage. The PITR window then equals the oplog archive retention, not the database’s local oplog size. In self-hosted setups, the equivalent is a custom oplog-tailer that writes entries to S3 or another durable store.
The base snapshot can be a filesystem-level snapshot of the WiredTiger files (LVM snapshot, EBS snapshot, ZFS snapshot) or a mongodump --archive --oplog. The filesystem-level form restores faster but requires the same storage engine and version. The dump form is portable across versions but restores slower.
The procedure
Standard recovery (within local oplog window)
-
Determine the target timestamp. From application logs, an incident report, or
db.system.profile.Timestamp(1717272000, 1) // unix epoch seconds, increment -
Find the most recent snapshot before the target time. Restoring forward is the only direction
--oplogReplayruns. -
Restore the snapshot to a clean target. Do not restore over the live cluster — restore to a side-by-side environment, validate, then cut over.
mongorestore --archive=backup-20260601-0000.archive.gz --gzip --oplogReplay \ --uri="mongodb://restore:secret@restore-target:27017" -
--oplogReplayconsumes theoplog.bsonfrom the dump. It stops at the end of that file’s entries. -
For ranges that exceed the dump’s oplog, supply additional oplog as
--oplogFile. Eachmongorestoreinvocation can replay one oplog file; chain them in order.mongorestore --oplogReplay --oplogFile=oplog-20260601-0100.bson \ --oplogLimit 1717272000:1 \ --uri="mongodb://restore:secret@restore-target:27017" -
--oplogLimit ts[:inc]stops replay at the target. Without this, replay continues to the end of the oplog file.
Continuous PITR (Atlas / Ops Manager)
-
From Atlas UI or API, request a PITR restore with the target timestamp.
atlas backups restores create pointInTime --clusterName prod \ --pointInTimeUTC 2026-06-01T14:37:22Z --targetClusterName prod-restore -
Atlas restores the latest snapshot before the target, then replays archived oplog entries to the target, into a new cluster — never in place over the existing one.
-
Validate the restored cluster against application-side invariants — record counts, recent transactions, foreign-key-like consistency.
-
Cut over by repointing application connection strings to the restored cluster, or by re-importing the affected collections into the live cluster.
Operational notes
- PITR is per-replica-set on sharded clusters. Atlas handles cross-shard restore by snapshotting all shards at the same cluster time — self-hosted PITR across shards requires coordinating restores at the same
clusterTime, which is non-trivial. - The oplog window must exceed the recovery objective by a safety margin. A 24-hour RPO with a 24-hour oplog window leaves no buffer for finding the bug and deciding to restore. Aim for 2-3× the recovery objective.
- Snapshot consistency matters. A filesystem snapshot taken while
mongodis writing is recoverable thanks to journaling, but quiescing the database (db.fsyncLock()) produces a cleaner snapshot. - Restore to a new cluster, validate, then cut over. Restoring over production prevents rollback and obscures the original state if the restore itself fails.
- Test the runbook quarterly. PITR is the kind of operation that nobody remembers under pressure. Run a full drill — pick a random snapshot, restore to a sandbox, prove the data is correct — every quarter.
Stack Harbor maintains PITR runbooks for clients running self-hosted MongoDB and exercises them on a quarterly cadence. We pair filesystem snapshots from a backup secondary with a continuous oplog archive in S3, sized for 2-3× the contractual RPO. See /en/services/managed-operations/ for the engagement model and /en/knowledge-base/mongodb-mongodump-streaming-large/ for the streaming-dump techniques the base snapshots rely on.