About an index merge policy

An index merge policy determines how frequently the Dgraph merges incremental update generations in its index files.

The data layer stores the Dgraph index files as a series of internal files with versions. As a result:

Old versions can be accessed while new versions are created.
Old versions are garbage-collected when no longer needed.

A version of the index is stored as a sequence of generation files. A new version appends a new generation file to the sequence. Query latency depends, in part, on the number and size of generation files used to store the index files.

Generation files are combined through a process called merging. Merging is a background task that does not affect the Dgraph request processing, but may affect its performance. Because of this, you can set a merge policy that dictates the aggressiveness of the merges.

The merge policy has two settings:

Balanced: This policy strikes a balance between low latency and high throughput. This is the default policy that is recommended for the majority of deployments.
Note: Under normal conditions, you do not need to change the default balanced policy.
Aggressive: This policy merges index generations frequently and completely to keep query latency low at the expense of average throughput. Aggressive merge policy may help deployments where query latency is the primary concern.

In addition to setting the merge policy, you can also force a one-time index merge, without changing the overall policy that will be used in all other instances. See Manually forcing a merge.

You can configure the index merge policy (or force a merge) using one of these two mechanisms:

The Enterprise Manager plug-in (if you use it for the Big Data Discovery). For forcing a merge or setting the merge policy, see Merging index updates in the Dgraph.
Options in the dgraph-admin command of the bdd-admin.sh script. To force a merge, see merge. To set the merge policy, see merge.

In a clustered environment, a request to set the merge policy is automatically routed to the leader Dgraph node.

Merging is not affected by the Dgraph's memory limit.