I decided to write this article after a discussion with other colleagues.  We were wondering how and if the Percona Distribution for MySQL Operator(PDMO) could help a starting business or a new project inside an enterprise.  We were also wondering how it behaves in relation to already well established solutions, like Amazon RDS or Google SQL. 

In fact, we often see analysis and benchmarking covering huge datasets and instances with very high levels of resources . But we tend to forget that any application or solution started as something small, and it is important to identify a database platform that allows us to scale from zero to a decent amount of load, while keeping the cost as low as possible.  

This also happened while I was trying to get a bit more performance and stability out of Percona Operator for MySQL, and identify the possible minimal entrypoint and its usage. 

Given all the above I decided to perform an investigation comparing the Percona Distribution for MySQL Operatorin the two most used cloud services, AWS and Google cloud. I am working on the Azure AKS and will publish results when available.

The investigation was taking in consideration three possible scenarios:

• 1S - I have a new website and my editorial system is X, my content is mainly static (Mainly Read Only)
• 2S - I have a new social platform, what should I use (OLTP style)
• 3S - I have a small application dealing with orders/payments, highly transactional  (more TPC-C

And try to identify which is the best solution to use.

Of course, I do not pretend to cover all cases or even at 100% the one mentioned. But I think we can say a good 70% of cases nowaday match one of the three above

Tests

I have split the tests in three segment one for each major provider. This to be able to better compare the results, and following the assumption that if you are already registered on a platform you are more interested in comparing data locally than cross providers.

As usual for the tests I have used sysbench adapted version from here (https://github.com/Tusamarco/sysbench) and sysbench-tpcc style from here (https://github.com/Tusamarco/sysbench-tpcc). The customization for standard sysbench is in the table definition, and for tpcc like because I have improved the report layout. 

Given we are talking about an entry point for a new product, I was considering two different scenarios:

• When dataset fits in memory
• When dataset doesn’t fits in memory  

Data sets dimensions:

+--------------------+--------+-----------+----------+-----------+-----------+
| TABLE_SCHEMA       | TABLES | ROWS      | DATA (M) | INDEX (M) | TOTAL(M)  |
+--------------------+--------+-----------+----------+-----------+-----------+
| tpcc               |    180 | 293113897 | 44762.00 |   9033.07 |  53795.07 |
| tpcc_small         |     90 |  48899475 |  7611.15 |   1542.35 |   9153.51 |
| windmills_large    |     80 | 317752934 | 35360.00 |  38463.79 |  73823.79 |
| windmills_small    |     80 |  39731506 |  4447.50 |   6075.82 |  10523.32 |
+--------------------+--------+-----------+----------+-----------+-----------+

All tables are using InnoDB. Keep in mind that scenario 1S and scenario 2S use the same schemas (windmills_xx).

This is because normally when you start your new adventure, you do not have Terabytes of data in your database, but .. you plan to have them in the future. So most of the time you start with something very small and want to have a solution able to scale at least a bit.

I was also considering whether they need to be able to scale (a bit) so I was running tests using 64, 92,128, 256, 512 threads. 

Finally each test was executed for 100 minutes, multiple times in different moments of the day and of the week. When multiple crashes happen the result will be zero. 

Summarizing we have:

• 3 different scenarios
  • Read Only (when write are so scarce that do not represent a variable)
  • OLTP Read and write some are inside transaction some not (especially reads)
  • TPC-C like load almost all inside transactions
• 2 Dimensions of dataset  for each environment dimension 
  • Fits in memory (small) 
  • Large enough to cause more flushing and disk activity (large)
• 5 increasing number of threads 

Assumptions

Assumptions and clarification before starting the review:

Q: Why are you using only one RDS instance instead of also spinning a Replica, we can also have the PDMO using only one node and reducing the cost impact.

A:  We are testing minimal production solutions. Using only a node for PDMO is delivering a solution that has no High Availability. RDS with multi AZ is providing High Availability, with lower level of nines but still functional. See: https://aws.amazon.com/blogs/database/amazon-rds-under-the-hood-multi-az/   https://aws.amazon.com/blogs/database/amazon-rds-under-the-hood-single-az-instance-recovery/ 

Q: Did you tune any of the default settings coming with the service? 

A: No. The assumption was that the provider is already providing optimised setup in relation to the solution offered. 

Q: Did you modify the PDMO settings?

A: Yes. The basic config coming with the PDMO is set for testing on a very minimal/test setup. It is not production ready and should not be used in production as well. The settings for PDMO had been changed to match the resources available and the expected load. Nothing more than what we should expect from the service providers.

The layout

AWS

The following is a simplified layout of the tested solutions.

In general terms each solution spans across multiple subnets and AZ, each solution has a dedicated Application for testing and application connected directly to the closest endpoint provided. In case of EKS the nodes are distributed over 3 different subnets.

I am not reporting the whole internal details about how the Percona Distribution for MySQL Operator (PDMO) works, for that please refer to the documentation https://www.percona.com/doc/kubernetes-operator-for-pxc/index.html

The table below is reporting the high level description of the solutions tested:

*For PDMO we had to use machines with more resources to host additional services like HAProxy and the operator itself. But the resources allocated for the database services were reflecting the ones used AWS for Aurora and standard RDS. Said that, while for this exercise we have used machines with a bit more resources, it is common practice to have one powerful host node with multiple database clusters to optimise cost and resource utilization, I will cover this better later in the cost section.   

Down into the rabbit hole 

A lot of graphs can be found here: 

The name includes the test and the data set dimension.

In this section we will review the results of the tests by provider.

AWS

AWS is the most utilized Cloud provider and is also the one with more solutions when discussing Database services. I have tested and compared PDMO vs Aurora and standard RDS.

For clarity I am reporting two different sets of graphs for details. One is related to CPU, memory etc, the other is more MySQL specific. 

In this context we have the following:

• aurora-eks-test1-instance-1; Primary node of the Aurora service
• aurora-eks-test1-instance-1-eu-central-1a; Secondary node of the Aurora service
• Mt-rds-kube-test; RDS instance 
• pxc-cluster1-pxc-0 ; Pod serving the MySQL traffic for PDMO 

As you may notice from the graphs, the Aurora service automatically swap/failover from a node to another more than once.  

Checking Read Only traffic (scenario 1S)

Let us start with the easiest case, a simple new website with some editorial. As already mentioned this can be considered a Read Only case given the writes should be counted in few by minutes or even less. 

Small Dataset

As previously mentioned, this case is the first baby step, when your DB is really at an infant stage.

Y axis reports operations/sec

In case no matter if the dataset fits in memory or not, we can notice how the Percona MySQL operator is able to serve a higher amount of traffic in respect to the other AWS solutions. If we check what seems to be the limitation factor, we can see that:

 2 CPU

4 CPU

We can see how in both cases the AWs solutions reach 100% of CPU utilization, limiting the number of operations they can execute in memory.
Also notice how the load was automatically moved from one Aurora instance to another during the tests (CPUs graph).

Large Dataset

Now let's see what happens if I am successful and my website starts to increase in content, such that my dataset does not fit in memory, and as such I will have MySQL evicting old pages from memory

In case the dataset exceeds the dimension of the memory, PDMO is still performing significantly better than the other two solutions when using 4 CPUs. But it has a significant reduction of performance when using 2 CPUs.

2 CPUs

Disk Latency

The other relevant fact to notice is how the disk latency affected the reads for the PDMO when using a large dataset and 2 CPUs.
This negative effect doesn't show up in case of 4 CPUs 

4 CPUs 

Also in the case of 4 CPUs we see the same behavior, with AWS solutions not able to scale and failing to serve traffic. 

Disk Latency

Checking OLTP (scenario 2S)

This case is obviously more complicated given the writes and transactions. The level of reads is still high, around 93% but all operations are enclosed inside transactions. 

In this case we MUST review in parallel the SMALL and LARGE dataset to better understand what is happening.

Operation/sec

As you can see the solution using the Operator is overperforming in the case of small and large datasets. While AWS solutions are still suffering from the same CPU limitation we have seen in the read only load. 

2 CPUs small

2 CPUs Large

We can observe how CPUs resources in the case of AWS solutions are always at 100%, while PDMO can count on resource scaling.

2 CPUs large Disk Latency

4 CPUs small

4 CPUs Small Disk Latency

4 CPUs large

4 CPUs Large Disk Latency

With 4 CPUs available the Aurora solution can count in a bit more “space” to scale, but still not enough to result as efficient as PDMO.

As for the read only, we can see the disk latency significantly increase in the case of Large dataset, especially for the PDMO solution.  

It is interesting to see how Aurora is penalized not only by the CPU utilization but also by a very high disk latency. 

Latency 95pct (lower is better)

Latency metrics reflect the operation trend, highlighting how the low CPU-resources/Disk-latency affects all solutions, especially in case of large datasets. 

Checking TPC-c (inspired) load (scenario 3S)

Also in this case we have that all the queries are encapsulated in a transaction and in this kind of test the read and write operations are close to being split at the 50% mark, so we expect to see a significant load on disk write operation and for the data replication mechanism.  

Operation/sec

In this scenario PDMO results to be more penalized when the dataset is larger than the available memory. In some way this is expected, given the level of writing is such that the solution pays the cost of the virtual-synchronous replication mechanism used in PXC.

If we check the details we can see:

4 CPUs Large

4 CPUs Disck Latency

As we can see the PDMO node receiving the direct load is not crazy loaded, while the disk latency is higher than RDS with an average of 4ms write and 8ms for read operations.
But that is not enough to justify the delta we have in performance loss. What is preventing the solution from being as efficient as the other two?

PXC flow control

PDMO is using PXC as technology to offer service high availability and full data consistency. This means that the cluster must certify and apply all writes in any node at each commit, with the increase of the write operations this solution also increases the pressure on the replication layer and if any node has a delay in applying the write, that delay will affect the efficiency of the whole cluster. 

In our case we can see that nodes pxc-0 and pxc-2 are on hold for pxc-1 to perform the writes. 

We also have as side effect the increase of the sending queue on the primary node (pxc-0)

Latency

Latency is just confirming us the picture we had from the operations.  

What should I pick? 

Before indicating what should be a good solution for what, let us talk about costs.

Costs

To discuss the cost, we need to make some assumptions. 

The first assumption is that we need to calculate the cost by solutions and not by node. This is because we cannot consider a solution with a single node for PDMO or Aurora. To be effective both require their production minimum requirements, which for Aurora is two nodes and for PDMO given is based on PXC 3 nodes. Given that calculating by node is a fictitious artifact, while calculating by solution gives you the real cost.   

The second assumption is that we can have more PDMO solutions running on the same Node. Which means we will have a better relation cost/benefit in utilizing larger nodes and use them to run multiple PDMO solutions. In that case, the only cost that will need to be added is the cost of the attached EBS volume per delivered POD. 

Let me provide an example of PDMO calculation. 

If I consider the minimal Node resource requirements, as we did for these tests, I will need to have:

But if we use a larger node that can host multiple PDMO like a r5.4xlarge (16 CPU 128GB) that can host multiple PDMO :

As you can see also if a solution based on r5.4xlarge has higher absolute cost per node, the fact you can run multiple PDMO in a node allows you to reduce the cost in a significant way by solution.   

Said that let us summarize the cost of the solutions:

* For Aurora we had to include an average number of operations based on the test executed. Please refer to: https://calculator.aws/#/createCalculator/AuroraMySQL

As you can see PDMO is just a little bit more expensive than RDS when deployed in single instance mode. But it also provides better performance in most cases and always has a higher level of High Availability, given PXC is used in the background. 

There is no discussion if you can use multiple PDMO on larger nodes, in that case the cost goes lower than RDS.

So what to pick

Read-only

Choosing is not only based on performance, but also how efficient the solution we pick is to perform the maintenance of our platform. In this regard both operator and Aurora offer a good level of automation. But PDMO was always performing better than the other two solutions. On top of that Aurora based solutions are more expensive than the others, so in the end, for me, choosing PDMO is the logical thing to do, being ready to eventually do a small scale-up in case I need more resources. This is another advantage of PDMO, I can eventually decide to add fragments of the CPU cycle, without the need to move to a higher and more expensive system.  

OLTP

In this case there is no doubt that PDMO is the solution to go to. Aurora can be seen as a possible replacement given that the RDS solution has a lower level of HA in respect to the other two. But as we know the Aurora costs are higher than any other compared solutions here.

TPCC

RDS is the solution to go to serve the load and save the money, but you will have a lower level of HA. If you are willing to cover the cost Aurora is the solution offering you good performance and high level of High Availability. 

Finally we can say that PDMO was penalized by the replication/certification mechanism in use in PXC but, the significant saving in terms of cost can justify the adoption, especially considering that with a limited scaling up in resources we can gain more traffic served. Scaling will never be as expensive as if we adopt Aurora as a solution.     

Conclusions

Trying to identify an entry point solution for a starting activity is harder than choosing a platform for something already well defined. There is a lot of uncertainty, about the traffic, the kind of operations, the frequency of them, the growth (if any). 

In my opinion there are already so many variables and things that can possibly go wrong, that we should try to adopt, if not the safest solution, the one that gives us a good level of confidence in what we will be able to achieve. 

I was not surprised to see products such as Aurora shine when we have Tpcc like load, but I was pleased to see Percona solution with the operator performing in an admirable way. 

Of course I was not using the settings coming from the vanilla installation, given they are for testing. So if you install PDMO in production you MUST change them  (see here for my cr.yaml) and I was counting on version 1.9.0 at least, without which PDMO is not even in the game. 

But it is obvious that Percona is on the right path, with a long road ahead, but we it is moving in the right direction.

Percona Distribution for MySQL Operatoris still a baby, but is growing fast, and the more we have people using, testing and providing feedback, the fastest Percona will be able to provide software that not only early adopters implement, but also the more conservative ones can use for their production.

References

https://calculator.aws/

Disclaimer 

My blog, my opinions. 

I am not, in any way or aspects, presenting here the official line of the company I am working for.

Comments

If you have comments or suggestions please add them to the linkedin post

Preface

Percona is well known for its offer of several outstanding fully open source, free to download software packages. And while Percona started as a MySQL focus company, nowadays it covers different technologies such as MySQL, Postgres and Mongodb.

In its constant effort to make life easier for our utilizer Percona had moved from providing single software packages, to Percona Distributions for MySQL, MongoDB and Postgres. Percona Distributions are a set of software packages that Percona has tested and certifies working together. That it easier to deploy architectural solutions requiring the use of multiple components, such as proxy, topology manager, backup software and more. 

But we are going even further, and with the release of Percona Distribution for MySQL/MongoDB Operator we are providing a high level of automation to roll out and manage solutions based on Percona Distributions. 

One of my tasks, as MySQL technical leader, is to identify optimal architectures to serve several common cases. Such as Percona Distribution for MySQL: High Availability with Group Replication Solution. Or in the case of  Percona Distribution for MySQL Operator, identify the different dimensions (low/mid/high utilization) and suggest a Vanilla setup with the scope to get the most out of the solution deployed.  

This is a long exercise, which started with a lot of internal discussions to identify what can make sense as traffic, then testing, identifying the saturation points, testing again and so on. 

It is during this process that I found a small problem (Feature Request). This small issue is preventing us from easily and dynamically modifying some parameters in the checks the Operator  uses. Given that we had to put the testing on hold until the above FR is implemented. As you can see it is a small thing but it will give us better control over the Operator behavior and will help to have a well tuned platform. 

This article is to show the level of improvement you can have with small but targeted tuning.To do so I used the smallest solution we have identified. The solution is dimensioned to serve a small website or a simple application with low level of traffic. 

The environment

To help identify a balanced setup we were using sysbench and sysbench-tpcc. The whole stack on GCP was composed of Application nodes with sysbench, 2 ProxySQL nodes for R/W split only, 3 VMS 8 CPU 32GB RAM, with the Percona operator managing the MySQL service.

Tests

For this specific test we were running 68 - 96 -128 - 256 threads:

• Sysbench read only
• Sysbench read/write
• Sysbench Tpc-c like 

The tests were run multiple times and the data considered is the consolidation of the multiple runs. 

We always run first on basic environments for baseline. Meaning no tuning for the MySQL or Operator, just dimension correctly disk space and BufferPool (and related).

Then we apply some tuning and run the tests multiple times eventually refining when/where needed. 

The code can be found here and here

And now the results…

Sysbench r/w tests

I am not going to describe in detail the images that I think are clear enough. Just keep in mind on the left we have the results from our baseline, on the right the same tests on the same platform with the optimization applied.

Operations

Worth to mention that without tuning, the platform was not able to consistently scale up to 256 threads. While with a bit of adjustment not only it was able to serve 256 threads but we could have gone a bit further.

Reads

Writes

Comments

As you can see the sysbench tests clearly indicate that the platform with minor adjustment was acting better, and that it was able to serve more and with constant scaling. Let me add that almost all the tests runned on the “basic” platform had incidents, meaning as soon as the traffic was increasing, Sysbench was reporting connection interruptions or errors.

TPC-C

Operations

Reads

Writes

Comments

Also for Tpc-c like tests we have exactly the same trend. With our “optimised” solution able to serve up to 1516 qps while the “basic” one was able to reach only 322. In this case Also the “optimised” solution was not able to scale up to 256 threads, but that makes sense, given the more intense write workload present in this test and the small dimension of the platform.

Wooha what have you changed?

You may think we have done crazy things to get this difference, but we did not.

Let us jump back. As indicated at the beginning I had opened a FR (https://jira.percona.com/browse/K8SPXC-749) to be able to tune some/most of the timeouts existing in the operator.

Why? Think about this, when you install a cluster on iron, you do not set it to be able to work only when the load is low, and all the components of the server are able to answer in nanoseconds. What you do instead is tune the whole system to accommodate the increasing load, and you will give to some elements more space for “flexibility” eventually expecting to have delays in answer. When doing so you also need to correctly align all the parameters that will be affected on cascade. For instance if you know your data nodes will be very busy serving queries, they may also slow down in answering internal health checks, but if you relax the cluster health checks and not the checks used for testing the cluster from operator point of view, the platform will be unbalanced and will not work correctly.

At the same time, if you do not tune the solution at all, you may end up with a platform that is theoretically able to serve the load, but that is crashing for artificial limitations. 

The last one is exactly what was happening with our “basic” solution. As it is the operator comes with parameters that allow it to work well, but that are not designed to scale. Is like having a server where your CPUs are always at 20% and if the applications ask more, a controller will chop them in fear of having too much load. But the fact is that you want to have the CPUs at 80% or the server will be underutilized. 

Anyhow what we have changed was some Innodb parameters, to allow internal operations to work better. Then we force consistent reads in PXC, which actually SLOW down the operations, and finally we tune the PXC cluster to be more flexible in its internal checks, avoiding having it to expel nodes unless really needed to. 

All the above were done using the Operator configuration, but then we had to work manually changing all the timeouts parameters used by the operator checks to be aligned with what we had defined in the cluster. 

In particular what we have changed was:

script                      line    value
/usr/bin/clustercheckcron   33      TIMEOUT=10
liveness-check.sh           23      TIMEOUT=5
readiness-check.sh          21      TIMEOUT=10
/usr/local/bin/check_pxc.sh 15      TIMEOUT=${CUSTOM_TIMEOUT:-10}

Wait .. why ProxySQL?

Ok this is another long discussion and I will cover it better in another article. For now just consider that HAProxy does not allow r/w splitting or other nice functionalities like firewalling etcetera etcetera. So the idea is simple, let us use the operator with what fits it better, and then decouple the special needs, eventually adding proxysql in a separate deployment. 

If you are scared of the cost of adding an additional block to the architecture:

Where:
ProxySQL means: Application → ProxySQL → HAProxy.
HAProxy means: Application → HAProxy.

Hope this puts your worries at rest, of course this is using the “optimised” solution.

Conclusions

Percona Distribution for MySQL Operator, is a constantly growing/improving solution. It also has a lot of interesting features, like being able to manage your backup/restore, point in time recovery, and more. But its adoption is still limited and it is normal to have some drawbacks like this one. It is on us who play with real production environments, or as in this case playing to define certified solutions, to give feedback to improve how the operator works, in order to make it a stronger product able to serve you better day by day.

Now we are going to wait for the FR to be implemented, and then we will recover our dimensioning work.