Change Streams Production Recommendations

If you drop or rename a collection or database with change streams opened against it, the change stream cursors close when they advance to that point in the oplog. Change stream cursors with the fullDocument : updateLookup option may return null for the lookup document.

Attempting to resume a change stream against a dropped collection results in an error. Any data changes that occurred on the collection between the last event the change stream captured and the collection drop event are lost.

Change stream response documents must adhere to the 16MB BSON document limit. Depending on the size of documents in the collection against which you open a change stream, notifications may fail if the resulting notification document exceeds the 16MB limit. For example, update operations on change streams configured to return the full updated document, or insert/replace operations with a document that is at or just below the limit.

Replica Sets

For replica sets with arbiter members, change streams may remain idle if enough data-bearing members are unavailable such that operations cannot be majority committed.

For example, consider a 3-member replica set with two data-bearing nodes and an arbiter. If the secondary goes down, such as due to failure or an upgrade, writes cannot be majority committed. The change stream remains open, but does not send any notifications.

In this scenario, the application can catch up to all operations that occurred during the downtime so long as the last operation the application received is still in the oplog of the replica set.

If significant downtime is estimated, such as for an upgrade or a significant disaster, consider increasing the size of the oplog such that operations are retained for a duration of time greater than the estimated downtime. Use rs.printReplicationInfo() to retrieve information on the oplog status, including the size of the oplog and the time range of operations.

Sharded Clusters

Change streams provide a total ordering of changes across shards by utilizing a global logical clock. MongoDB guarantees the order of changes are preserved and change stream notifications can be safely interpreted in the order received. For example, a change stream cursor opened against a 3-shard sharded cluster returns change notifications respecting the total order of those changes across all three shards.

To guarantee total ordering of changes, for each change notification the mongos checks with each shard to see if the shard has seen more recent changes. Sharded clusters with one or more shards that have little or no activity for the collection, or are “cold”, can negatively affect the response time of the change stream as the mongos must still check with those cold shards to guarantee total ordering of changes. This effect may be more apparent with geographically distributed shards, or workloads where the majority of operations occur on a subset of shards in the cluster.

If a sharded collection has high levels of activity, the mongos may not be able to keep up with the changes across all of the shards. Consider utilizing notification filters for these types of collections. For example, passing a $match pipeline configured to filter only insert operations.

For sharded collections, update operations with multi : true may cause any change streams opened against that collection to send notifications for orphaned documents.

From the moment an unsharded collection is sharded until the time the change stream catches up to the first chunk migration, the documentKey in the change stream notification document only includes the _id of the document, not the full shard key.