© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

PGConf 2018

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

System Overview (simplified version)

One Database usually serves multiple types

of clients with different access patterns and

IO requirements.

To satisfy all these requirements we need to

understand the nature and priorities of the

different clients.

Browsers B2B Clients

DB

Scripts

API Server ReportsWeb Server

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

System Overview (less simplified)

Browsers B2B Clients

DB

Scripts

API Server ReportsWeb Server

Most of the clients have multiple threads or

processes.

This can lead to dozens if not hundreds of

long living database connections.

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

System Overview (closer to the truth)

Always expect to find some “legacy code”.Browsers B2B Clients

DB

Old Zoo

API Server ReportsWeb Server

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

System Overview (even closer to the truth)

Many of those, who draw these types of

system diagrams, don’t really know what a

database is or does.

Browsers B2B Clients Old Zoo

API Server ReportsWeb Server

Magical Bit

Container

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

System Overview

Today we focus on the IO of PostgreSQL

PostgreSQL

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

Buffering

PostgreSQL does not operate on raw

devices. It relies on a regular file system and

the OS buffer cache.

Because of that tuning must not be limited to

changing parameters in the file

postgresql.conf. The OS kernel parameters

are not “off limits”.Backend

OS Buffers

Disk

Shared Bufs

Data

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

Normal IO for Data

Under normal circumstances the backend

should seldom to never write data (heap and

index blocks).

postgresql.conf:• shared_buffers

• checkpoint_timeout

• checkpoint_completion_target

Kernel parameters:• dirty_background_bytes

• dirty_ratio

• deadline IO scheduler

Backend

OS Buffers

Disk

Shared Bufs

Data

Checkpointer

BG-Writer

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

Temporary Files

PostgreSQL can spill large sort operations to

disk.

This is done on purpose because glibc’s

malloc()/free() clings to memory.

postgresql.conf:• work_mem

• maintenance_work_mem

Backend

OS Buffers

Disk

Shared Bufs

Checkpointer

BG-Writer

Data

TMP

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

Add WAL to the mix

WAL (redo log) has a very different IO

pattern compared to data.

WAL is written sequentially (more or less)

and has a high fdatasync() rate for OLTP.

postgresql.conf• wal_buffers

• max_wal_size

• synchronous_commit

Backend

OS Buffers

Disk

Shared Bufs WAL Buffers

Checkpointer

BG-WriterWAL-Writer

Data

TMP

WAL

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

Separate the WAL

Due to its high fdatasync() rate (O_DIRECT

on Linux) it is best to put WAL on a separate

IO subsystem.

Even if that is not possible, having it on its

own filesystem still has advantages (like that

the database will start up when $PGDATA

throws “no space left on device”).Backend

OS Buffers

Disk

Shared Bufs WAL Buffers

Checkpointer

BG-WriterWAL-Writer

Data

TMP

WAL

Disk

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.

And all the rest …

PostgreSQL IO (simplified)

Backend

OS Buffers

Disk

Shared Bufs WAL Buffers

Checkpointer

BG-WriterWAL-Writer

Data

TMP

WAL

WAL-Sender

Backup

Disk