automatic_parallel_replicas_max_duplicated_read_ratio
How much of a query’s read volume may be read by every replica instead of being split between them, as a fraction of the whole, before automatic parallel replicas declines the query. Only the read parallel replicas coordinate is split; every other read of the same subtree runs on each replica in full. Those reads take the same wall-clock time either way, so the cost-model comparison is blind to them, but the cluster performsmax_parallel_replicas times as much work for them. This setting is the
limit on that waste: at the default 0.5 a query whose coordinated read is less than half of what it reads
keeps running on one node, however the time comparison turns out.
Set to 1 to accept any amount of duplicated reading, which restores the behaviour of only comparing times.
automatic_parallel_replicas_min_bytes_per_replica
Threshold of bytes to read per replica to enable parallel replicas automatically (applies only whenautomatic_parallel_replicas_mode=1). 0 means no threshold.
The total number of bytes to read is estimated based on the collected statistics.
automatic_parallel_replicas_mode
Enable automatic switching to execution with parallel replicas based on collected statistics. Requiresenable_parallel_replicas != 0, parallel_replicas_local_plan = 1 and providing cluster_for_parallel_replicas.
0 - disabled, 1 - enabled, 2 - only statistics collection is enabled (switching to execution with parallel replicas is disabled).