Choose a Shard Key

Shard Key Cardinality

The cardinality of a shard key determines the maximum number of chunks the balancer can create. Where possible, choose a shard key with high cardinality. A shard key with low cardinality reduces the effectiveness of horizontal scaling in the cluster.

Each unique shard key value can exist on no more than a single chunk at any given time. Consider a dataset that contains user data with a continent field. If you chose to shard on continent, the shard key would have a cardinality of 7. A cardinality of 7 means there can be no more than 7 chunks within the sharded cluster, each storing one unique shard key value. This constrains the number of effective shards in the cluster to 7 as well - adding more than seven shards would not provide any benefit.

The following image illustrates a sharded cluster using the field X as the shard key. If X has low cardinality, the distribution of inserts may look similar to the following:

click to enlarge

If your data model requires sharding on a key that has low cardinality, consider using an indexed compound of fields to increase cardinality.

A shard key with high cardinality does not, on its own, guarantee even distribution of data across the sharded cluster. The frequency of the shard key and the potential for monotonically changing shard key values also contribute to the distribution of the data.

Shard Key Frequency

The frequency of the shard key represents how often a given shard key value occurs in the data. If the majority of documents contain only a subset of the possible shard key values, then the chunks storing the documents with those values can become a bottleneck within the cluster. Furthermore, as those chunks grow, they may become indivisible chunks as they cannot be split any further. This reduces the effectiveness of horizontal scaling within the cluster.

The following image illustrates a sharded cluster using the field X as the shard key. If a subset of values for X occur with high frequency, the distribution of inserts may look similar to the following:

click to enlarge

If your data model requires sharding on a key that has high frequency values, consider using a compound index using a unique or low frequency value.

A shard key with low frequency does not, on its own, guarantee even distribution of data across the sharded cluster. The cardinality of the shard key and the potential for monotonically changing shard key values also contribute to the distribution of the data.

Monotonically Changing Shard Keys

A shard key on a value that increases or decreases monotonically is more likely to distribute inserts to a single chunk within the cluster.

This occurs because every cluster has a chunk that captures a range with an upper bound of MaxKey. maxKey always compares as higher than all other values. Similarly, there is a chunk that captures a range with a lower bound of MinKey. minKey always compares as lower than all other values.

If the shard key value is always increasing, all new inserts are routed to the chunk with maxKey as the upper bound. If the shard key value is always decreasing, all new inserts are routed to the chunk with minKey as the lower bound. The shard containing that chunk becomes the bottleneck for write operations.

To optimize data distribution, the chunks that contain the global maxKey (or minKey) do not stay on the same shard. When a chunk is split, the new chunk with the maxKey (or minKey) chunk is located on a different shard.

The following image illustrates a sharded cluster using the field X as the shard key. If the values for X are monotonically increasing, the distribution of inserts may look similar to the following:

click to enlarge

If the shard key value was monotonically decreasing, then all inserts would route to Chunk A instead.

If your data model requires sharding on a key that changes monotonically, consider using Hashed Sharding.

A shard key that does not change monotonically does not, on its own, guarantee even distribution of data across the sharded cluster. The cardinality and frequency of the shard key also contribute to the distribution of the data.

Sharding Query Patterns

The ideal shard key distributes data evenly across the sharded cluster while also facilitating common query patterns. When you choose a shard key, consider your most common query patterns and whether a given shard key covers them.

In a sharded cluster, the mongos routes queries to only the shards that contain the relevant data if the queries contain the shard key. When the queries do not contain the shard key, the queries are broadcast to all shards for evaluation. These types of queries are called scatter-gather queries. Queries that involve multiple shards for each request are less efficient and do not scale linearly when more shards are added to the cluster.

This does not apply for aggregation queries that operate on a large amount of data. In these cases, scatter-gather can be a useful approach that allows the query to run in parallel on all shards.