MongoDB

MongoDB sources are read with change streams, which tail the replica set's oplog for every insert, update, and delete. Initial load is a chunked, resumable scan sorted by _id, and schema is inferred by sampling live documents — MongoDB has no catalog to read.

Supported deployments

The same connector serves self-hosted MongoDB replica sets and MongoDB Atlas. An Atlas cluster is a replica set, so change streams work unchanged; the connection string and the user's privileges are the only differences.

Prerequisites

A replica set is required

Change streams do not work on a standalone mongod. Fabric refuses a deployment with no replica set name. The one exception is a documented single-node development deployment, enabled with the Direct connection checkbox — never use that against a production topology.

User privileges

The connection's user needs three actions on the target database: find, changeStream, and listCollections. A read role alone is not enough — changeStream is a separate action.

// mongosh, as a user with userAdmin on the target db
db.getSiblingDB("myapp").createRole({
  role: "fabricCdc",
  privileges: [{
    resource: { db: "myapp", collection: "" },
    actions: ["find", "changeStream", "listCollections"]
  }],
  roles: []
})
db.getSiblingDB("myapp").createUser({
  user: "fabric",
  pwd: "<password>",
  roles: [{ role: "fabricCdc", db: "myapp" }]
})

Oplog retention

A change-stream resume token can expire if the oplog rolls over. Keep at least 24 hours of oplog history:

db.adminCommand({ setParameter: 1, oplogMinRetentionHours: 24 })

Network

The driver must reach every replica-set member, not just one host, unless direct connection is enabled. A mongodb+srv:// URI also needs DNS SRV resolution from the Fabric host.

Connection fields

How it replicates

• Each collection is scanned find().sort(_id, 1) in chunks, checkpointing the last key after every chunk. The _id index makes this always correct.
• The current clusterTime is captured before the first snapshot row and streaming starts strictly after it, so a crash mid-load never skips intervening writes.
• Streaming resumes from an opaque resume token. If it expires (ChangeStreamHistoryLost) Fabric raises and requires a re-snapshot rather than silently restarting from now.

Gotchas