Versus

PostgreSQL vs MariaDB Performance comparison

PostgreSQL and MariaDB are great open source databases. Both systems will work well as a general-purpose RDBMS. However, in my opinion, PostgreSQL is better than MariaDB, and you should use it unless you have a specific use case that requires MariaDB.

The reasons I would choose PostgreSQL over MariaDB for data storage include:

  • Faster performance
  • Community size
  • MariaDB has to maintain some compatibility with MySQL
  • Many new databases are compatible with PostgreSQL clients

Let’s get right to the MariaDB vs PostgreSQL performance comparison.

PostgreSQL, MariaDB, and HammerDB deployment

To compare the performance characteristics of these SQL databases, I deployed two AWS RDS instances, along with an EC2 instance running HammerDB.

RDS is Amazon’s service for deploying managed relational databases. This service is a great way to get the most out of a database without being an expert on replication strategies, memory allocation/configuration, or other details required to maintain high availability, high performance, and data integrity.

Deploying the databases in this way allowed me to compare their data access efficiency, with as little variance coming from configuration and hardware differences as possible.

HammerDB is a popular, open source database performance testing software. It implements a test called TPROC-C, which is loosely based on the TPC-C benchmark. You can read more about TPC’s other measures on TPC’s official page.

PostgreSQL benchmark setup

  • Version: PostgreSQL 13.3
  • Instance type: DB.m5.large
  • Two x86 cores
  • 8 GB of RAM
  • 100 GB of storage
  • gp2 storage type (no provisioned IOPS)
  • No replication of any kind (no multi-AZ or horizontal partitioning)
  • No data encryption at rest

MariaDB benchmark setup

  • Version: MariaDB 10.5.13
  • Instance type: DB.m5.large
  • Two x86 cores
  • 8 GB of RAM
  • 100 GB of storage
  • gp2 storage type (no provisioned IOPS)
  • No replication of any kind (no multi-AZ or horizontal partitioning)
  • No data encryption at rest

HammerDB benchmark setup

  • Version: HammerDB 4.3
  • Operating system: Debian 11
  • Instance type: t3.medium
  • Two x86 cores
  • 4 GB of RAM

Benchmark Setup

Now that we have the hardware and software deployed, we can give HammerDB the database credentials and issue a command to set up the test schemas. HammerDB provisions the following structured data schema in both PostgreSQL and MariaDB.

Picture from HammerDB docs

This schema approximates a simple online transactional processing (OLTP) database. Most “real” databases will have more tables than this, but it is a good starting point with enough variety to simulate real-world usage.

According to the official HammerDB docs:

… the workload is defined by a mix of 5 transactions selected at random according to the balance of the percentage value shown as follows:

  • New-order: receive a new order from a customer: 45%
  • Payment: update the customer’s balance to record a payment: 43%
  • Delivery: deliver orders asynchronously: 4%
  • Order-status: retrieve the status of customers most recent order: 4%
  • Stock-level: return the status of the warehouse’s inventory: 4%

In our opinion, this database structure and query mix will properly test indexes, read-write performance, and general ability for storing data. It will not require advanced features, complex queries, or data warehouse-style queries, but that is not the point of a transactional test.

PostgreSQL setup

For PostgreSQL, we simply connected as the superuser postgres, on port 5432 and built ten “warehouses”, which are essentially duplicate schemas. This duplication of warehouses may somewhat overcome the fact that such a limited schema would only be possible in smaller databases.

PostgreSQL HammerDB settings
PostgreSQL HammerDB create schema

We then created ten virtual users and logged their output. We setup the test in this way to approximate ten backend servers querying the database simultaneously, which is not an unreasonable load to expect for a busy database. PostgreSQL supports many more concurrent users, but 10 is a nice round number to start.

PostgreSQL HammerDB create users

MariaDB setup

As you can see, the setup is primarily the same for MariaDB. We used the root user, the InnoDB storage engine, and the default port to connect to the database and build ten warehouses.

MariaDB HammerDB settings

Once again, we created ten users to simulate ten backend servers querying the database.

MariaDB HammerDB settings

Running the tests

And now, we will show the configuration we used to run the tests and show what HammerDB looked like during the run.

PostgreSQL

To start the test for PostgreSQL, we selected “Timed driver script” and left everything else on default, except checking the “Use all warehouses” box.

PostgreSQL HammerDB start test

These screenshots taken at the beginning of the run shows the ten users “ramping up” for two minutes.

PostgreSQL HammerDB ramp up

After less than a minute, PostgreSQL’s transactions per minute plateaued around 100,000 to 120,000 transactions per minute, with occasional dips.

PostgreSQL HammerDB transactions per minute

The AWS console shows that PostgreSQL, on average, used about 90% of its available CPU resources. Note: This chart shows me making three different runs (trying to get the logging right), which is why there are two significant dips in CPU usage.

PostgreSQL CPU usage

MariaDB

Now we will run the same test with MariaDB. As you can see here, the settings are all left on default, except for “Use all warehouses,” just like PostgreSQL.

MariaDB HammerDB test spinup

Interestingly, MariaDB seemed to ramp up to its maximum transactions per minute faster than PostgreSQL. However, it plateaued at a lower level.


MariaDB initial transactions per minute

More interestingly, it showed a lot less variance in transactions per minute – it was rock solid, around 60,000 to 70,000 transactions per minute.


MariaDB HammerDB transactions per minute

Looking at CPU usage on the AWS console shows similar results to PostgreSQL. It averaged around 90% usage of CPU resources.

MariaDB CPU usage

MariaDB vs PostgreSQL benchmark results


MariaDB vs PostgreSQL performance comparison

The chart above shows the transactions per minute performed by both systems after the initial 2-minute ramp-up period was complete. Two things stand out to me in the chart:

  • Overall, PostgreSQL is faster than MariaDB by a wide margin.
  • MariaDB’s transactions per minute was a lot more stable than PostgreSQL’s.

Why is MariaDB so much more consistent? Does PostgreSQL make some optimistic assumptions that allow it to perform better overall, at the expense of lost time? Is it faster at certain operations but slower at others? I will look into the possibilities more soon when doing more benchmarks.

The chart below shows HammerDB’s orders per minute metric. As stated in the docs, insertions into the order table is 45% of the work done by the benchmark


MariaDB vs PostgreSQL orders per minute performance comparison

The number of transactions required to complete a given action can differ based on the DB system’s design. Measuring the number of insertions the database completes may offer a more accurate measure of the system’s performance. As such, we consider the above chart to be a more accurate representation of each database’s performance level.

In the end, PostgreSQL inserted 84% more orders than MariaDB while only needing 39% more transactions. This is an impressive performance win for PostgreSQL.

For transparency, let’s look at each database’s performance logs to see how we calculated results and allow you to audit our results.

  • NOPM is the number of orders per minute.
  • TPM is transactions per minute.

PostgreSQL performance benchmark logs

Here are the logs from the PostgreSQL test.

PostgreSQL general test logs

Hammerdb Log @ Fri Jan 07 16:18:01 UTC 2022
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-
Timestamp 1 @ Fri Jan 07 16:18:09 UTC 2022
Vuser 1:Beginning rampup time of 2 minutes
Timestamp 2 @ Fri Jan 07 16:18:10 UTC 2022
Vuser 2:VU 2 : Assigning WID=1 based on VU count 10, Warehouses = 10 (1 out of 1)
Timestamp 2 @ Fri Jan 07 16:18:10 UTC 2022
Vuser 2:Processing 10000000 transactions with output suppressed…

… Many logs that look exactly the same as above

Timestamp 1 @ Fri Jan 07 16:19:09 UTC 2022
Vuser 1:Rampup 1 minutes complete ...
Timestamp 1 @ Fri Jan 07 16:20:09 UTC 2022
Vuser 1:Rampup 2 minutes complete ...
Timestamp 1 @ Fri Jan 07 16:20:09 UTC 2022
Vuser 1:Rampup complete, Taking start Transaction Count.
Timestamp 1 @ Fri Jan 07 16:20:09 UTC 2022
Vuser 1:Timing test period of 5 in minutes
Timestamp 1 @ Fri Jan 07 16:21:09 UTC 2022
Vuser 1:1 ...,

… Many logs that looks exactly the same as above

Vuser 1:Test complete, Taking end Transaction Count.
Timestamp 1 @ Fri Jan 07 16:25:10 UTC 2022
Vuser 1:10 Active Virtual Users configured
Timestamp 1 @ Fri Jan 07 16:25:10 UTC 2022
Vuser 1:TEST RESULT : System achieved 41396 NOPM from 95427 PostgreSQL TPM

PostgreSQL transactions log

Hammerdb Transaction Counter Log @ Fri Jan 07 16:18:05 UTC 2022
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
0 PostgreSQL tpm @ Fri Jan 07 16:18:07 UTC 2022
23514 PostgreSQL tpm @ Fri Jan 07 16:18:17 UTC 2022
65520 PostgreSQL tpm @ Fri Jan 07 16:18:27 UTC 2022
75858 PostgreSQL tpm @ Fri Jan 07 16:18:37 UTC 2022
51966 PostgreSQL tpm @ Fri Jan 07 16:18:47 UTC 2022
78126 PostgreSQL tpm @ Fri Jan 07 16:18:57 UTC 2022
84552 PostgreSQL tpm @ Fri Jan 07 16:19:07 UTC 2022
102498 PostgreSQL tpm @ Fri Jan 07 16:19:17 UTC 2022
129360 PostgreSQL tpm @ Fri Jan 07 16:19:28 UTC 2022
76722 PostgreSQL tpm @ Fri Jan 07 16:19:38 UTC 2022
65472 PostgreSQL tpm @ Fri Jan 07 16:19:48 UTC 2022
77298 PostgreSQL tpm @ Fri Jan 07 16:19:58 UTC 2022
79362 PostgreSQL tpm @ Fri Jan 07 16:20:08 UTC 2022
119226 PostgreSQL tpm @ Fri Jan 07 16:20:18 UTC 2022
82638 PostgreSQL tpm @ Fri Jan 07 16:20:28 UTC 2022
69834 PostgreSQL tpm @ Fri Jan 07 16:20:38 UTC 2022
94422 PostgreSQL tpm @ Fri Jan 07 16:20:48 UTC 2022
123588 PostgreSQL tpm @ Fri Jan 07 16:20:58 UTC 2022
110622 PostgreSQL tpm @ Fri Jan 07 16:21:08 UTC 2022
116982 PostgreSQL tpm @ Fri Jan 07 16:21:18 UTC 2022
53394 PostgreSQL tpm @ Fri Jan 07 16:21:28 UTC 2022
56808 PostgreSQL tpm @ Fri Jan 07 16:21:38 UTC 2022
87126 PostgreSQL tpm @ Fri Jan 07 16:21:48 UTC 2022
80850 PostgreSQL tpm @ Fri Jan 07 16:21:58 UTC 2022
111102 PostgreSQL tpm @ Fri Jan 07 16:22:08 UTC 2022
100014 PostgreSQL tpm @ Fri Jan 07 16:22:18 UTC 2022
57012 PostgreSQL tpm @ Fri Jan 07 16:22:28 UTC 2022
75600 PostgreSQL tpm @ Fri Jan 07 16:22:38 UTC 2022
109644 PostgreSQL tpm @ Fri Jan 07 16:22:48 UTC 2022
103122 PostgreSQL tpm @ Fri Jan 07 16:22:58 UTC 2022
69336 PostgreSQL tpm @ Fri Jan 07 16:23:08 UTC 2022
64176 PostgreSQL tpm @ Fri Jan 07 16:23:18 UTC 2022
105996 PostgreSQL tpm @ Fri Jan 07 16:23:28 UTC 2022
111750 PostgreSQL tpm @ Fri Jan 07 16:23:38 UTC 2022
127266 PostgreSQL tpm @ Fri Jan 07 16:23:48 UTC 2022
130368 PostgreSQL tpm @ Fri Jan 07 16:23:58 UTC 2022
112146 PostgreSQL tpm @ Fri Jan 07 16:24:08 UTC 2022
47106 PostgreSQL tpm @ Fri Jan 07 16:24:18 UTC 2022
85182 PostgreSQL tpm @ Fri Jan 07 16:24:28 UTC 2022
101190 PostgreSQL tpm @ Fri Jan 07 16:24:38 UTC 2022
129216 PostgreSQL tpm @ Fri Jan 07 16:24:48 UTC 2022
133782 PostgreSQL tpm @ Fri Jan 07 16:24:58 UTC 2022
106734 PostgreSQL tpm @ Fri Jan 07 16:25:08 UTC 2022
33378 PostgreSQL tpm @ Fri Jan 07 16:25:18 UTC 2022
126 PostgreSQL tpm @ Fri Jan 07 16:25:28 UTC 2022
102 PostgreSQL tpm @ Fri Jan 07 16:25:38 UTC 2022

MariaDB performance benchmark logs

Here are the logs from the MariaDB test.

MariaDB general test logs

Hammerdb Log @ Fri Jan 07 17:07:07 UTC 2022
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-
Timestamp 1 @ Fri Jan 07 17:07:09 UTC 2022
Vuser 1:Beginning rampup time of 2 minutes
Timestamp 2 @ Fri Jan 07 17:07:10 UTC 2022
Vuser 2:VU 2 : Assigning WID=1 based on VU count 10, Warehouses = 10 (1 out of 1)
Timestamp 2 @ Fri Jan 07 17:07:10 UTC 2022
Vuser 2:Processing 10000000 transactions with output suppressed...

… Many logs that looks exactly the same as above

Timestamp 1 @ Fri Jan 07 17:08:09 UTC 2022
Vuser 1:Rampup 1 minutes complete ...
Timestamp 1 @ Fri Jan 07 17:09:09 UTC 2022
Vuser 1:Rampup 2 minutes complete ...
Timestamp 1 @ Fri Jan 07 17:09:09 UTC 2022
Vuser 1:Rampup complete, Taking start Transaction Count.
Timestamp 1 @ Fri Jan 07 17:09:09 UTC 2022
Vuser 1:Timing test period of 5 in minutes
Timestamp 1 @ Fri Jan 07 17:10:09 UTC 2022
Vuser 1:1 ...,

… Many logs that looks exactly the same as above

Vuser 1:Test complete, Taking end Transaction Count.
Timestamp 1 @ Fri Jan 07 17:14:09 UTC 2022
Vuser 1:10 Active Virtual Users configured
Timestamp 1 @ Fri Jan 07 17:14:09 UTC 2022
Vuser 1:TEST RESULT : System achieved 22496 NOPM from 68574 MariaDB TPM

MariaDB transactions log

Hammerdb Transaction Counter Log @ Fri Jan 07 17:07:03 UTC 2022
+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+
0 MariaDB tpm @ Fri Jan 07 17:07:05 UTC 2022
23058 MariaDB tpm @ Fri Jan 07 17:07:15 UTC 2022
62820 MariaDB tpm @ Fri Jan 07 17:07:25 UTC 2022
62058 MariaDB tpm @ Fri Jan 07 17:07:35 UTC 2022
63774 MariaDB tpm @ Fri Jan 07 17:07:45 UTC 2022
64536 MariaDB tpm @ Fri Jan 07 17:07:55 UTC 2022
60996 MariaDB tpm @ Fri Jan 07 17:08:05 UTC 2022
63810 MariaDB tpm @ Fri Jan 07 17:08:15 UTC 2022
63924 MariaDB tpm @ Fri Jan 07 17:08:25 UTC 2022
62310 MariaDB tpm @ Fri Jan 07 17:08:35 UTC 2022
59196 MariaDB tpm @ Fri Jan 07 17:08:45 UTC 2022
58752 MariaDB tpm @ Fri Jan 07 17:08:55 UTC 2022
57858 MariaDB tpm @ Fri Jan 07 17:09:05 UTC 2022
62622 MariaDB tpm @ Fri Jan 07 17:09:15 UTC 2022
64134 MariaDB tpm @ Fri Jan 07 17:09:25 UTC 2022
63606 MariaDB tpm @ Fri Jan 07 17:09:35 UTC 2022
63804 MariaDB tpm @ Fri Jan 07 17:09:45 UTC 2022
63378 MariaDB tpm @ Fri Jan 07 17:09:55 UTC 2022
59142 MariaDB tpm @ Fri Jan 07 17:10:05 UTC 2022
63600 MariaDB tpm @ Fri Jan 07 17:10:15 UTC 2022
64578 MariaDB tpm @ Fri Jan 07 17:10:25 UTC 2022
71754 MariaDB tpm @ Fri Jan 07 17:10:35 UTC 2022
72336 MariaDB tpm @ Fri Jan 07 17:10:45 UTC 2022
71700 MariaDB tpm @ Fri Jan 07 17:10:55 UTC 2022
69858 MariaDB tpm @ Fri Jan 07 17:11:05 UTC 2022
70170 MariaDB tpm @ Fri Jan 07 17:11:15 UTC 2022
71172 MariaDB tpm @ Fri Jan 07 17:11:25 UTC 2022
72072 MariaDB tpm @ Fri Jan 07 17:11:35 UTC 2022
71784 MariaDB tpm @ Fri Jan 07 17:11:45 UTC 2022
72210 MariaDB tpm @ Fri Jan 07 17:11:55 UTC 2022
69306 MariaDB tpm @ Fri Jan 07 17:12:05 UTC 2022
69576 MariaDB tpm @ Fri Jan 07 17:12:15 UTC 2022
70050 MariaDB tpm @ Fri Jan 07 17:12:25 UTC 2022
70878 MariaDB tpm @ Fri Jan 07 17:12:35 UTC 2022
69552 MariaDB tpm @ Fri Jan 07 17:12:45 UTC 2022
66912 MariaDB tpm @ Fri Jan 07 17:12:55 UTC 2022
64632 MariaDB tpm @ Fri Jan 07 17:13:05 UTC 2022
70638 MariaDB tpm @ Fri Jan 07 17:13:15 UTC 2022
71598 MariaDB tpm @ Fri Jan 07 17:13:25 UTC 2022
71298 MariaDB tpm @ Fri Jan 07 17:13:35 UTC 2022
72540 MariaDB tpm @ Fri Jan 07 17:13:45 UTC 2022
71292 MariaDB tpm @ Fri Jan 07 17:13:55 UTC 2022
69846 MariaDB tpm @ Fri Jan 07 17:14:05 UTC 2022
59406 MariaDB tpm @ Fri Jan 07 17:14:15 UTC 2022

A note about averages

Eagle-eyed observers may note that my chart’s “transactions per minute” averages are slightly different from HammerDB’s printed logs. This is because I truncated the results to fit neatly within 30 10 second intervals to create the chart.

Conclusions

PostgreSQL is faster than MariaDB at data management. PostgreSQL inserted 84% more orders than MariaDB while only needing 39% more transactions. This signals that the PostgreSQL global development group is doing a good job optimising available resources and minimizing the number of transactions needed for a unit of real work.

The transactions per minute figures are more consistent with MariaDB, but the metric we care about, orders per minute, isn’t broken out over time like transactions. I may try to break out orders with a time dimension in a future revision to this article. The MariaDB Foundation is handcuffed by its original purpose of offering a database with MySQL compatibility. However, this should only affect a client’s view of the database, not the actual transaction processing itself. It would be interesting to see how using the Spider storage engine would affect performance.

In a future article, we may rerun this test using multiple nodes comparing PostgreSQL to MariaDB with synchronous replication to see which one has better performance in large deployments.

contact

Ready to get started?

Get in touch with our support team who can further advise

Get in touch
By clicking “Accept All Cookies”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage, and assist in our marketing efforts. View our Privacy Policy for more information.