This week I’m participating in a virtual social learning experiment run by Phil Eaton / https://twitter.com/eatonphil, called “#wehack PostgreSQL Internals”.

The purpose is to get more people hacking on the internals of PostgreSQL or exploring PostgreSQL topics.

Check out the Announcement blog post: https://eatonphil.com/2023-10-wehack-postgres.html

Since I always have a running list of topics to learn in PostgreSQL, I picked one off my list! I wanted to dive into the new pg_stat_io1 system view in PostgreSQL 16 to understand the information it presents and how I could use that information for performance analysis.

Terminology note: The terms blocks and pages are often used interchangeably2 when discussing physical storage in PostgreSQL. The PostgreSQL Glossary3 uses data pages or pages when talking about storage, so this post will use the term pages. The use of pages in this post though can be thought of as being equivalent to blocks.

What are pages? These are fixed size 8kb (by default) data page structures for storage, inside files in the data directory.

For the purposes of analyzing IO latency, it’s worth noting that PostgreSQL loads the whole page even when just one row is requested. We can use the page size multiplied by the number of buffers accessed to put the data retrieval into kilobytes or megabytes.

In query execution plans (using EXPLAIN (ANALYZE, BUFFERS)) “buffers read” are pages loaded from storage. 100 “buffers read” for example means 100 * 8kb or 800kb.

Let’s get started with pg_stat_io. What is it?

What is it?

pg_stat_io is a system view added in PostgreSQL 16 that adds additional observability to IO operations.

IO operations come from multiple sources, such as client queries like SELECT or INSERT, or other “backend types” like Autovacuum or bulk operations.

Who created it?

Melanie Plageman helped lead the creation, and has given presentations to help educate others on this capability. I met Melanie Plageman at PGConf NYC (check out my recap blog post: PGConf NYC 2023 Recap 🐘), and she was gracious enough to give me an in-person intro to pg_stat_io there.

To understand what the view offers, I watched Melanie’s Citus Con presentation “Additional IO Observability in Postgres with pg_stat_io”.4 As an introduction, I found it very helpful.

While there are many more contributors to pg_stat_io, I wanted to highlight one more person in particular.

Lukas Fittl and the PgAnalyze team helped review and contributed to the capability before PostgreSQL 16 was released, mentioned in the post “Waiting for Postgres 16: Cumulative I/O statistics with pg_stat_io”.5 Lukas explains how IO is split into writes to the WAL stream, or writes to the data directory. Writes don’t happen into a storage device right away, as they can be buffered in PostgreSQL or the Operating System.

Wait a minute, what are writes from the perspective of PostgreSQL?

What “writes” are in the context of files and pages

In PostgreSQL, “writes” are different from what we might think of as client-level operations like INSERT, UPDATE, or DELETE (DML).

Melanie explains this in the Citus Con presentation.4

writes in PostgreSQL relate to files that store data in pages/blocks.

These pages are held in memory. Writes are the process of flushing data to durable storage, or from memory to disk (or more generally, the “storage device”).

When data is stored in PostgreSQL, it’s placed into pages. Existing pages are checked at storage time. When pages contain data they’re considered “dirty”, which means they need to be flushed. After dirty pages are flushed, they’re available for use.

We want to understand when these writes occur. As discussed earlier, writes can occur from different types of operations. What are those?

Distinguishing backend types

Melanie points outs4 these three backend types:

  • client backend
  • background writer
  • checkpointer

These backend types report their IO activity in pg_stat_io. For web applications, we’re mostly interested in client backend types. When analyzing performance, we want to check for excessive or unnecessary writes during client backend queries, because those writes are adding latency.

What other data does the view have?

What does it look like?

Like other system views, this view contains a lot of fields. If you don’t have a PostgreSQL 16 psql client session handy, check out this pgPedia6 page which shows the fields. The field data is related to IO activity across the whole database, showing backend types, operations, and more.

This post “PostgreSQL 16: More I/O statistics”7 describes results:

What you see is one row per backend type, context and object.

You may be interested in querying the view for specific backend type rows.

After initializing PostgreSQL, the view won’t have much data. To generate data, the post above shows how to run pgbench as follows: pgbench -i -s 10 postgres

https://www.postgresql.org/docs/current/monitoring-stats.html#MONITORING-PG-STAT-IO-VIEW

Why does IO matter?

As Lukas Fittl says in “Postgres I/O Basics”:8

PostgreSQL doesn’t load individual rows, it always loads full pages.

For OLTP environments where we want fast queries, IO can add latency to queries.

Often stored data is not stored optimally, densely packed in pages. Pages may be scattered. pg_stat_io helps us analyze the IO happening during client backend operations. We can see how data is placed into and accessed in shared buffers.

Who is this view useful for?

Who is it for?

This view is useful for any PostgreSQL user wishing to gain more insight into the IO of their PostgreSQL installation. Users may wish to explore client backend operations, or other backend operation types like Autovacuum. The statistics presented are from all databases. If you run multiple databases on an instance, keep this in mind.

Let’s look at more use cases.

What are the use cases?

Melanie suggests some use cases for the information in pg_stat_io:

For example, tuning Shared Buffers (Check out: Resource Consumption).

Shared buffers may be undersized. Query performance is optimal when the working data set fits into shared buffers. We could use pg_stat_io to see if there’s a lot of IO from client backends that are outside of shared buffers. This could indicate that more memory should be allocated to shared buffers.

PostgreSQL query execution plans show “shared hits”, which means pages are coming from shared buffers. Shared buffers are filled with what’s currently happening, which could be from other backend types like Autovacuum or bulk operations.

Understanding IO on cloud hosted PostgreSQL could also help with cost analysis. When IO is priced separately, or when usage exceeds IO quotas, it can be helpful to have additional visibility into what operations are contributing to IO.

Extensions

For deeper analysis of the content in shared buffers, enable the pg_buffercache9 extension. pg_buffercache allows you to determine which tables and indexes are in the cache.

Run the sample query on the page to see the tables and indexes with buffers in the cache.

System Wide IO visibility

Lukas points out how prior to pg_stat_io, we could analyze the impact of IO from Vacuum jobs that were logged to postgresql.log on a per-job basis.

With pg_stat_io, we’re able to look at the impact of all Vacuum jobs across the whole system, by querying pg_stat_io for context = vacuum.

The query below shows 3 result rows for backend type = autovacuum worker for contexts bulkread, normal, and vacuum.

SELECT * FROM pg_stat_io
WHERE backend_type = 'autovacuum worker'
OR (context = 'vacuum' AND (
    reads <> 0 OR writes <> 0 OR extends <> 0
    )
);

The post mentions that this information could be used to help analyze an unexplained IO spike.

COPY operations

pg_stat_io gives us visibility for bulk load (COPY) operations.

To do that, query the view for bulkread and bulkwrite values in the io_context column.

Extends

pg_stat_io also shows information on “Extends” operations. What are those?

Melanie describes how Extends are a special kind of write operation. extends are displayed separately from reads and writes.

The following post “PostgreSQL 16: Playing with pg_stat_io (1) – extends”10 has nice examples comparing sizes with an empty table, and showing how the size expands when rows are added. Extend operations happen when there’s no page to store data, and a new page is allocated.

Writebacks

Another concept in pg_stat_io are “Writebacks”. What are they?

Hans-Jürgen Schönig from Cybertec provides some context and an explanation. Normally, “writes” to a storage device go through the OS file system cache. writebacks are when that process is skipped, and writes go directly from PostgreSQL to permanent storage.

Let’s dive in

Let’s explore pg_stat_io locally. Let’s reset the stats and generate some client activity.

To do that, I’ll use the Rideshare Rails app which is set up locally with PostgreSQL 16. I’ll start the Rails Server, then run the simulate_app_activity.rake Rake task which sends queries to the server and database.

Connect to PostgreSQL 16 (or newer) as a superuser:

psql -U postgres -d rideshare_development

Once connected, run the following function to reset the io statistics:

SELECT pg_stat_reset_shared('io');

The stats_reset column shows the timestamp that was just set.

Setting up Rideshare is beyond the scope of this post, but if you’d like to follow along, here’s a brief overview of the steps:

  • Clone the repository. Install Ruby and all the system dependencies. Set up the rideshare_development database on PostgreSQL 16.
  • Start the Rails Server (bin/rails server)
  • Run the Rake task: bin/rails simulate:app_activity

Once the activity simulation task completes, let’s go back to psql and review the client backend rows in pg_stat_io:

SELECT * FROM pg_stat_io
WHERE backend_type = 'client backend';

This is a small 2GB development database with around 20k rows in the largest table.

Below I’ll analyze some of the results:

  • bulkread and bulkwrite are zero, which makes sense since no bulk operations were performed
  • The context value of normal shows 993 read operations, and no write operations. This makes sense since the simulation only ran read only queries.
  • I’m seeing a hits value of 11616. This is good because it means the data was served from shared buffers. Evictions, reuses, and fsyncs were all initially zero.
  • The backend type = checkpointer performed 45 operations. Initially no fsync operations are performed, but later we see them.

In a real system, there will be much data!

Wrap Up

This post was an introduction to the new pg_stat_io system view in PostgreSQL 16. This view adds additional IO visibility into what’s happening with PostgreSQL, as it reads and writes data to physical storage, whether the IO is from client activity or other backend activity. PostgreSQL users can use this information to gain more knowledge on what the contributors are to IO and when the IO operations are happening.

That’s all for now. Please let me know if you spot any technical issues or errors. Leave a comment about how you’re using pg_stat_io in PostgreSQL 16.

Thanks for reading!