Chapter 5. Performance Tuning
Table of Contents
Performanсe tuning should be done during application development and include an accurate choice of hardware (for example, estimating the number of CPUs and memory per Shardman cluster node or tuning your storage), OS tuning (for example, tuning the swappiness parameter or network-related behavior) and DBMS tuning (choosing efficient configuration). But first of all, an application should be tested and tuned for distributed DBMS. This includes designing a distributed schema (or converting an existing schema to a distributed one), tuning queries, using connection poolers, caching and even checking performance issues related to possible serialization errors or Shardman node outage. The design of the schema should include accurate selection of a sharding key and a decision which tables should become global. Usually you select a sharding key so that:
Most of the queries filter out most of sharded table partitions.
Sharded tables are colocated and all joins of sharded tables are equi-joins on the sharding key.
These rules allow Shardman to efficiently exclude unused shards from queries and to push down joins to shards where the required data resides.
Each Shardman node operates as a usual DBMS server, so all standard recommendations for tuning PostgreSQL for production load remain in place. You should select shared_buffers, work_mem, effective_cache_size depending on resources available to DBMS. Keep in mind that if the cluster topology is set to cross, Repfactor instances run on a single node w. When all cluster nodes are online, replicas should not utilize a lot of CPUs. However, in case of node failure, masters for Repfactor replication groups can become running on one server, which can create significant load on it. While tuning the max_connections parameter, note that each transaction can initiate n-1 connections, where n is the number of replication groups in the cluster. When Silk is enabled, it is still true for transactions containing DML operations. When Silk is disabled, it is also true for read-only transactions.
Other parameters, which you perhaps would like to tune, are foreign server options. They can be set in FDWOptions section of Shardman configuration file. Parameters that significantly affect Shardman performance are fetch_size, batch_size and async_capable. When Silk transport is not enabled, fetch_size determines the number of records that are fetched from a remote server at once. When Silk transport is enabled, fetch_size currently does not have significant impact on the query execution. batch_size specifies how many rows can be combined in a single remote INSERT operation for a sharded table. async_capable allows asynchronous execution and should always be turned on (which is the default).
The shardman.gt_batch_size configuration parameter allows you to optimize the size of an intermediate buffer for INSERT and DELETE operations on global tables.