Global sequences in Shardman are implemented on top of regular PostgreSQL sequences with some additional cluster-wide metadata, which among other things holds the interval of globally unused sequence elements.

When CREATE SEQUENCE is issued, an ordinary PostgreSQL sequence with the same name is created on every cluster node. The range of this local sequence is a bounded sub-interval of the global sequence (as defined by MINVALUE and MAXVALUE parameters), and it contains at most block_size elements. The shardman.next_value function returns values from the local sequence until it runs out, then a new sub-interval with block_size elements is allocated from the global sequence using a broadcast query involving all cluster nodes. So, smaller block size values make the generated numbers more monotonic across the cluster, but incur a performance penalty since the broadcast query may be rather expensive. Another way to describe the block size parameter is to say that it controls the size of the second cache level, similarly to how the CACHE parameter works, except at the level of an entire Shardman cluster.

Also note, that every time a new sub-interval is allocated the underlying local sequence is modified (as in ALTER SEQUENCE), which will lock it for the transaction duration, preventing any other local concurrent transactions from obtaining next sequence values.