2021年3月26日金曜日

Faster Bulk Insertion to Foreign Tables - Introduction to PostgreSQL 14 Committed Features

Introduction

Currently, the global PostgreSQL development team, led by Amit Kapila, continues to contribute to the community as part of Fujitsu's work on PostgreSQL. As a member of this team, I am actively working with talented and passionate community members around the world to advance PostgreSQL.

Our team will publish blog posts focusing on the features and patches that members have worked on. This article discusses "Batch Insert to Foreign Table" committed to PostgreSQL 14, explaining why we've worked on it and the resulting exciting performance improvements.

Why did we work on this?

Our team is focused on improving PostgreSQL's scalability and performance. One big goal is to achieve read-write scale-out. With this, we want to expand the workloads to which PostgreSQL can be applied.

The idea of using postgres_fdw, which is based on the foreign data wrapper (FDW) mechanism, is considered in the community as a promising and practical idea for interactions between multiple servers for scale-out. This is because postgres_fdw has been improved over the years by a number of developers.

At such times, we saw a committer Tomas Vondra try to speed up bulk insertion with postgres_fdw. He says that customers often report slow insertions into databases sharded with FDW.

A user reported a similar problem to the PostgreSQL bug report mailing list. He reported that it took an hour and 50 minutes to insert 20 million rows with an INSERT SELECT statement into a database sharded using postgres_fdw on the Google Cloud. It's 14 times slower than what he got with a non-sharded database, which only took 8 minutes.

The cause of the slowness is the latency of the network and the number of communications. In the current FDW mechanism, the foreign data wrapper inserts rows into the foreign table row by row. Each time, a round trip communication occurs with a remote server.

Tomas Vondra and other people were exploring how to modify only postgres_fdw to send multiple rows together to a foreign server. But while they were looking for a way to avoid changing the executor, development stopped for about three months.

What features have I developed?

I made the following proposal. The idea of sending multiple rows at once to a foreign server is the same as Tomas Vondra's.

  • Add functions for batch row insertion to the FDW interface: BeginForeignBatchInsert, ExecForeignBatchInsert, EndForeignBatchInsert, GetForeignBatchSize.
  • ExecForeignBatchInsert () takes an array of rows and the number of rows and sends them all together to the foreign server.
  • Executor accumulates multiple rows and passes them together to the FDW. The number of rows to accumulate is obtained from the FDW with GetForeignBatchSize ().
  • postgres_fdw uses the row array passed in to assemble "INSERT ... VALUES (row 1), (row 2),..., (row n)" statement and execute it on the foreign server.
  • Add batch_size option to the postgres_fdw's configuration of foreign server and foreign table. This specifies the number of rows to be inserted together for each foreign table.

It's really straightforward and simple, isn't it?

This speeds up INSERT VALUES statements with multiple rows, as well as INSERT SELECT statements. You do not need to change the application.

The above interfaces should be readily implemented in FDWs for other databases, such as oracle_fdw and mysql_fdw.

How much faster?

We first measured the data insertion to a simple table. The target table has one primary key column of type int and one column of type text.

The time it took for an INSERT SELECT statement to insert 1 million rows from another table into that table is as follows. The postgres_fdw batch_size parameter is set to 100. This means sending up to 100 rows at a time to the foreign server:

  • Local table without FDW: 6.1 seconds
  • Remote table with FDW (before improvement): 125.3 seconds
  • Remote table with FDW (after improvement): 11.1 seconds

Great! It's 11 times faster.

Next, split the target table of the same columns into eight hash partitions. Similarly, the time it took to insert 1 million rows with an INSERT SELECT statement is:

  • Local table without FDW: 8.6 seconds
  • Remote table with FDW (before improvement): 113.7 seconds
  • Remote table with FDW (after improvement): 12.5 seconds

This is also good! It's nine times faster.

Note that for this measurement, the foreign server was located on the same host. The network latency measured by "ping localhost "is 34 microseconds. Given the significant performance gains seen with such low latency, this feature will be even more effective in environments with higher network latency, such as the cloud.

Expectations for higher speeds

This feature makes adding data to a single external table dramatically faster. However, this is faster sequential processing using a single CPU.

On the other hand, my colleague Greg Nancarrow is working on making it possible to use multiple CPUs to process INSERT SELECT statements in parallel. Other developers are trying to parallelize the CREATE TABLE AS SELECT (CTAS) statement.

By combining these, I expect to see the potential for even faster data migration to sharded OLTP databases and ETL processing to data warehouses. I am looking forward to delivering good news again.

Acknowledgments

I'd like to express to Tomas Vondra, Amit Langote and other developers my deepest thanks for getting this feature committed relatively quickly. Thanks to their quick and enthusiastic feedback, I believe that we were able to make this feature a reality.